The invention concerns a method of treating, in particular fragmenting and/or weakening, a solid ma- terial, in particular rock or ore, by means of high vol- tage discharges as well as an arrangement for conducting the method according to the preambles of the independent claims .

BACKGROUND ART

It is known from prior art to treat material, like e.g. concrete or rock, by pulsed high voltage dis- charges in order to perform fragmentation and/or weaken- ing of the material, i.e. to reduce the particle size of the material and/or to generate cracks within the materi- al which facilitate fragmentation in a subsequent mech- anical fragmentation process.

However, in order to make it possible to em- ploy this technology in industrial scale production, it is crucial that a constant quality of the fragmented/wea- kened material can be ensured while at the same the effi- ciency can be optimized, which in particular is an unsol- ved problem in mineral processing applications, in which the material to be processed is a natural product which can vary in its physical properties to a large extend.

From WO 2013/053066 Al a method for fragmen- ting and/or weakening of material by means of high volta- ge discharges is known, in which during the fragmentation or weakening, respectively, of the material, process li- quid is discharged from the process zone and new process liquid, having a lower electric conductivity than the discharged process liquid, is fed into the process zone. By this, the energy efficiency and the ability to commi- nute hard and brittle materials can be substantially im- proved.

From WO 2015/058312 Al a method for fragmen- ting and/or weakening of material by means of high vol- tage discharges is known, in which the high voltage dis- charges are triggered subject to a continuously determi- ned process parameter which represents the situation with respect to the material located in the process zone. By doing so _f the process can be controlled such that the high-voltage discharges are only triggered in case there is a situation in the process zone in which a specific fragmentation work can be performed. Thus, the energy efficiency of the process can be considerably improved and an excessive fragmentation of the material can be prevented.

From DE 103 02 867 B3 a method for fragmen- ting and/or weakening of material by means of high vol- tage discharges is known, in which, depending on the dis- charge resistance and the ignition delay of the high- voltage discharges, the distance between the electrodes and the feeding of material is controlled. By this, the energy efficiency of the process can be improved.

From WO 2015/058311 Al methods for fragmen- ting and/or weakening of material by means of high vol- tage discharges are known, in which, depending on the turbidity of the process liquid, depending on the elec- trical resistance between the electrodes or depending on data representing an image of the material that has been processed or is to be processed, the generation of high voltage discharges, the distance between the electrodes, the feeding of material through the process zone and/or the feeding and discharging of process liquid is changed. With these methods it is possible to ensure a substan ^¬ tially constant quality of the processed material even when the feed material varies in quality, or to at least diminish the effect of variation of the feed material on the quality of the processed material. While the before mentioned method already re- present substantial improvements of the electrodynamic fragmentation/weakening process with regard to efficiency and control of the process, there is an ongoing need to further improve the process in this respect.

DISCLOSURE OF THE INVENTION

Hence, it is a general object of the inven- tion to provide a method of treating, in particular frag- menting and/or weakening, material by means of high-vol- tage discharges and an arrangement for conducting such method, which allow for further improvements of the pro- cess with respect to process efficiency and control of the quality of the processed material.

This object is achieved by the method and the arrangement according to the independent claims.

Accordingly, a first aspect of the invention concerns a method of treating, preferably fragmenting and/or weakening, solid material, like e.g. rock material or ore, by means of high-voltage discharges.

In this method, a process zone is provided between at least two electrodes, which are arranged at a distance relative to each other. This process zone is flooded with a process liquid and contains, arranged bet- ween the electrodes, the material that is to be treated, in particular is to be fragmented or weakened.

Between the electrodes, a high voltage dis- charge is generated by charging the electrodes with a high-voltage pulse provided by a tunable high voltage generator, which is set to predetermined pulse parame- ters.

Such pulse parameters can be, for example, the capacitance of the applied voltage (typically in the range between 1 nF and 500 nF) , the voltage of the pulse (typically in the range between 10 kV and 500 kV) , the voltage pulse front rise time (typically in the range between 1 nsec and 500 nsec) , the shape of the pulse (e.g. square, triangle or specific pattern) etc.

A signal representing at least one parameter of this high voltage discharge and/or representing an effect caused in the process zone by this high voltage discharge is determined and is compared, in its deter- mined form and/or in a processed form, with at least one reference.

Parameters of the high-voltage discharge are for example the discharge voltage, the discharge current, the discharge resistance and the ignition delay.

The term "reference" is to be understood here in its broadest possible interpretation. The at least one reference could for example be some earlier determined signal (s) in its/their determined form and/or processed form, or a threshold value or a range of values, which distinguishes between signals which represent an intended operational state/effect and signals which represent a not-intended operational state/effect.

It could for example be a set of reference signal curves, each curve representing a characteristic mode of operation or effect, by which reference curves, a determined signal curve is compared for evaluation of the reference curve it has the highest conformity with.

Or the at least one reference could for exam- pie be a representation, in particular in terms of para- meters of machine learning framework (support vectors and kernels for Support Vector Machines (SVM) , weights for neural networks, node weights for random forest algorithm etc.), into which the determined signal (s) in a processed form are integrated in order to decide if the high vol- tage discharge to which the determined signal (s) belong took place under intended operation conditions and/or generated an intended effect or not.

Depending on the result of the comparison of the signal with the reference, e.g. if according to the comparison, the high voltage discharge, the determined signal (s) belongs to, took place under intended operation conditions or has generated an intended effect, the pulse parameters of the high voltage generator are maintained or are changed.

Subsequently, a further high voltage dischar- ge is generated with the high voltage generator being set to the same (unchanged) or changed pulse parameters, the signal representing the at least one parameter of this high voltage discharge and/or of an effect caused in the process zone by this high voltage discharge is again de- termined and is again compared, in its determined form and/or in a processed form, with the at least one refer- ence and, depending on the result of the comparison of the signal with the reference, the pulse parameters of the high voltage generator are maintained or are changed.

By continuously repeating this sequence, it becomes possible to systematically adopt the pulse para- meters of the high voltage generator in order to achieve a constant and desired treatment result and a high effi- ciency even in applications, in which the material to be processed is a natural product and varies in its physical properties to a large extent.

To achieve this, it is preferred that the de- termined signal in its determined form and/or in a pro- cessed form is compared with the reference, and if a de- viation is determined or a deviation which exceeds a pre- determined tolerated deviation is determined, one or se- veral of the voltage pulse parameters of the high voltage generator are changed in such a manner that, when subse- quently the steps of generating a further high-voltage discharge, determining the signal and comparing the sig- nal with the reference are repeated, no deviation between the determined signal and the reference is detected or the detected deviation is smaller than before.

Alternatively, to achieve this, it is prefer- red that the determined signal in its determined form and/or in a processed form is compared with the referen- ce, and if no deviation is determined or a deviation which is smaller than a predetermined minimum deviation is determined, one or several of the voltage pulse para- meters of the high voltage generator are changed in such a manner that, when subsequently the steps of generating a further high-voltage discharge, determining the signal and comparing the signal with the reference are repeated, there is a deviation between the determined signal and the reference or the detected deviation is bigger than before.

In yet a further alternative, to achieve this, it is preferred that the determined signal in its determined form and/or in a processed form is compared with at least two references representing different high voltage discharge situations, one of which is the inten- ded high voltage discharge situation, and wherein in case the comparison reveals that the signal differs least from the reference representing the intended high voltage discharge situation, the voltage pulse parameters of the high voltage generator are kept unchanged, and in case the comparison reveals that the signal differs more from the reference representing the intended high voltage discharge situation than from another of the references, one or several of the voltage pulse parameters of the high voltage generator are changed.

These different alternatives represent dif- ferent strategies in deciding whether the pulse parame- ters of the high-voltage generator shall be maintained {i.e. not be changed) or shall be changed, and if so, how they shall be changed.

In a preferred embodiment of the method, a signal representing the acoustic emissions caused in the process zone by the high voltage discharge is determined and, in its determined form and/or in a processed form, is compared with the at least one reference. It has been found that by using these acoustic emissions, is becomes possible to very well distinguish between different operational modes and effects, respectively, e.g. between a) no discharge through the material;

b) surface discharge on material;

c) discharge through the material causing mainly

internal damages, in particular cracks; and d) discharge through the material causing

disintegration of the material.

Reference is made in this regard also to the "Study conducted regarding the acoustic characterization of solid materials pre-weakening using electrical discharges", which in greater detail is described herein later on.

Preferably, the signal is determined by one or several acoustic sensors placed in the process zone and/or outside the process zone, in the latter case e.g. at the outer surface of a process vessel containing the process zone.

In case the acoustic sensors are placed in or near to the process zone, it is preferred that they are placed near to the at least two electrodes, preferably with a distance of less than 50 mm from a line defined by a shortest distance between those two electrodes between which the high voltage discharges are generated. By doing so, the signal disturbances through reflections and noise created through movement of the material in the process zone can be kept low.

Preferably, one or several optical acoustic- sensitive fiber sensors are used, preferably sensors with fibers with Bragg grating or with Fabry Perot fibers and/or fiber interferometers. Such kind of optical sen ^¬ sors have the advantage that they are neutral to the strong electromagnetic fields generated by the high-vol ^¬ tage discharges and can be placed close to the discharge area. In that case, it is further preferred that the fiber of the sensor is pre-strained. By doing so, it is possible to tune the sensitivity of the sensor in the frequency domain.

Furthermore, when using optical acoustic-sen- sitive fiber sensors, it is of advantage that the fiber axis together with a line defined by a shortest distance between those two electrodes, between which the high-vol- tage discharges are generated, forms an angle in the ran- ge between 15° and 75°. By this, the sensitivity of the determination of the acoustic emissions can be optimized, wherein it is advisable to experimentally determine the angle for which a maximum sensitivity is achieved.

By advantage, the adjustment of the angle may be done manually or automatically to a predetermined va- lue.

In a further preferred embodiment of the method, in which a signal representing the acoustic emis- sions caused in the process zone by the high voltage dis- charge is determined, several acoustic sensors are placed at different pre-defined positions, preferably in or near to the process zone, and from the signals determined by these sensors and their known positions the location of the high voltage discharge is determined. Suitable algo- rithms how to calculate the location of the high-voltage discharge from these data are known to the person skilled in the art and therefore are not further described here.

If this determined location of the high-vol- tage discharge, which is a processed form of the determi- ned signals of the sensors, is compared with the at least one reference, which is preferred, the possibilities to distinguish between several different operational modes and/or effect of the discharge on the material are sub- stantially improved.

In still a further preferred embodiment of the method, a signal representing one or several electri- cal parameters of the high voltage discharge is determi- ned and, in its determined form and/or in a processed form, is compared with the at least one reference.

Preferably, a signal representing the elec- trical current and/or the voltage of the high-voltage discharge is determined and, in its determined form and/- or in a processed form, is compared with the at least one reference.

Determining such signals has the advantage that they are relative safe and easy to determine and that typically they show little disturbance only.

In still a further preferred embodiment of the method, a signal representing the electromagnetic fields caused in the process zone by the high-voltage discharge is determined and, in its determined form and/or in a processed form, is compared with the at least one reference.

Preferably, this signal is determined by one or several Pockels-Kerr cells placed in or near to the process zone and/or outside the process zone.

In still another preferred embodiment of the method, a signal representing the light emissions caused in the process zone by the high voltage discharge is de- termined and is compared with the at least one reference.

In particular, if such signals representing the electromagnetic fields and/or the light emissions caused in the process zone by the high voltage discharge are used in combination with other signals, e.g. signals representing the acoustic emissions caused in the process zone by the high voltage discharge and/or signals repre ^¬ senting one or several electrical parameters of the high voltage discharge, the possibilities to distinguish bet- ween several different operational modes and/or effects of the discharge on the material are substantially im- proved.

In still a further preferred embodiment of the method, the determined signal is processed by decom- position before it is compared with the at least one re- 5 ference, preferably by wavelet decomposition, polynomial decomposition/ functional decomposition or Fourier decora- position. These and other types of decomposition are known to the skilled person and therefore are not des- cribed here more into detail. By advantage, it is the result of the decomposition of the signal that is com- pared with the at least one reference.

In still a further preferred embodiment of the method, a support vector machine (SVM) , a random forest algorithm and/or an artificial neural network is used for comparing the determined signal in its deter- mined form and/or in a processed form with the at least one reference. These mathematical methods are known to the skilled person and therefore are not described here more into detail.

In still a further preferred embodiment of the method, online machine learning is applied in com- paring the determined signal in its determined form and/- or in a processed form with the at least one reference. Different methods of machine learning are known to the skilled person and therefore are not described here more into detail.

The above methods have proven to be especial- ly well suited if signals representing the acoustic emis- sions caused in the process zone by the high-voltage dis- charge are determined and compared with a reference. For a practical example how these methods can be applied in this regard, reference is made to the "Study conducted regarding the acoustic characterization of solid mate ^¬ rials pre-weakening using electrical discharges", which in greater detail is described herein later on.

In still a further preferred embodiment of the method, a pre-determined reference is used. For pre ^¬ determining this reference, sample material is treated with high voltage discharges between the electrodes with different pulse parameter settings of the high-voltage generator, and the physical effect of each high-voltage pulse or high-voltage discharge, respectively, on the sample material is investigated, for example by visual inspection of the material, and is correlated to the de- termined signal.

Preferably, the physical effect of the high- voltage pulse or the high-voltage discharge, respective- ly, on the sample material is expressed in a particle size distribution, in a bond index and/or in an axb-value and is correlated to the determined signal.

Also, it is preferred to use a determined signal or a number of determined signals, which are representing high-voltage discharges that caused a desired effect on the sample material, as reference, in the determined form and/or in a processed form.

By doing so, the reference can better be tai- lored to represent a specific effect on the material that is treated.

In some cases, it is of advantage that different references are pre-determined for different effects on the sample material.

For example, if the sample material is a brittle material, in particular concrete, rock material or ore, it is preferred that references are predetermined for at least two of the following effects:

a) no discharge through the material;

b) surface discharge on material; c) discharge through the material causing mainly internal damages, in particular cracks;

d) discharge through the material causing disintegration of the material.

For a practical example how sample material can be used to establish pre-determined references, also here reference is made to the "Study conducted regarding the acoustic characterization of solid materials pre- weakening using electrical discharges", which in greater detail is described herein later on. By doing so, it can be distinguished between effects that are relevant for the result of the treatment .

In still a further preferred embodiment of the method/ during the generation of high-voltage dis- charges between the electrodes, the material that is to be treated is fed through the process zone, preferably by means of a conveyor carrying said material. This allows the treatment of a continuous stream of material accor- ding to the method.

In still a further preferred embodiment of the method, during the generation of high-voltage dis- charges between the at least two electrodes, process li- quid is fed to the process zone and process liquid is discharged from the process zone. By doing so, the opera- ting conditions in the process zone can be kept quite stable.

Furthermore, it is preferred that in the method according to the invention, at least the comparing of the signal with the reference and the maintaining or changing of the pulse parameters of the high-voltage generator is performed automatically by an electronic control system. Preferably, also the determining of the signal is performed automatically by the electronic control system. By doing so, a fully automatic process can be establish that adopts the pulse parameters to changing conditions in order to achieve stable treatment results. However, it is also envisaged to perform these steps manually.

A second aspect of the invention concerns an arrangement for conducting the method according to the first aspect of the invention.

This arrangement comprises a process zone formed between at least two electrodes which are arranged at a distance relative to each other, which process zone in the intended operation is flooded with a process li- quid, for example with water.

The arrangement further comprises a tunable high-voltage generator for generating, at predetermined pulse parameters of the generator, high-voltage dischar- ges between the at least two electrodes in the intended operation.

Such pulse parameters can be, for example, the capacitance of the applied voltage (typically in the range between 1 nF and 500 nF) , the voltage of the pulse (typically in the range between 10 kV and 500 kV) , the voltage pulse front rise time (typically in the range between 1 nsec and 500 nsec) , the shape of the pulse (e.g. square, triangle or specific pattern) etc.

It also comprises means for determining a signal representing at least one parameter of the high- voltage discharge and/or an effect caused in the process zone by the high voltage discharge, as well as a control system.

Parameters of the high-voltage discharge are for example the discharge voltage, the discharge current, the discharge resistance and the ignition delay.

The control system is adapted for comparing the determined signal, in its determined form and/or in a processed form, with at least one reference and for kee- ping the pulse parameters of the high voltage generator unchanged or changing one or several of the pulse para- meters of the high voltage generator, depending on the result of the comparison. The term "reference" has alrea- dy been explained earlier under the first aspect of the invention.

In a preferred embodiment of the arrangement, the means for determining a signal are adapted for deter- mining a signal representing the acoustic emissions cau ^¬ sed in the process zone by the high-voltage discharge. In this embodiment, the control system is adapted for compa- ring this signal, in its determined form and/or in a pro- cessed form, with the at least one reference.

Preferably, in that case the means for deter- mining signals representing the acoustic emissions com- prise one or several optical acoustic sensitive fiber sensors, by advantage sensors with fibers with Bragg gra- ting or with Fabry Perot fibers and/or fiber interfero- meters, wherein preferably the fibers of the sensors are in each case arranged pre-strained in a fiber holder.

Such kind of optical sensors have the advan- tage that they are neutral to electromagnetic fields and can be placed close to the discharge area. Furthermore, they can be tuned by the pre-straining with regard to the sensitivity of the fiber in the frequency domain.

By advantage, the fiber axis of the sensor or the sensors can be aligned with respect to a line defined by a shortest distance between two electrodes, between which in the intended operation the high voltage dischar- ges are generated, such that it forms, together with said line, an angle in the range between 15° and 75°. By this, the sensitivity of the determination of the acoustic emissions can be optimized.

Preferably, the arrangement comprises several acoustic sensors which are arranged at different pre-de- fined positions. In that case, the control system prefer- ably is adapted to determine the location of the high- voltage discharge from the signals determined by these acoustic sensors and their known positions and to compare the determined location of the high voltage discharge with the at least one reference.

In a further preferred embodiment of the ar- rangement, the means for determining the signal are adap ^¬ ted for determining a signal representing one or several electrical parameters of the high voltage discharge, like e.g. the discharge voltage, the discharge current, the discharge resistance and/or the ignition delay time, and the control system is adapted for comparing this signal, in its determined form and/or in a processed form, with the at least one reference. Such signals can relatively safe and easy be determined and typically show little disturbance.

In still a further preferred embodiment of the arrangement, the means for determining a signal are adapted for determining a signal representing the elec- tromagnetic fields caused in the process zone by the high-voltage discharge. In that case, the control system is adapted for comparing this signal with the at least one reference. Preferably, the means for determining a signal comprise one or several Pockels-Kerr cells for doing so.

In still a further preferred embodiment of the arrangement, the means for determining the signal are adapted for determining a signal representing the light emissions caused in the process zone by the high-voltage discharge and the control system is adapted for comparing this signal with the at least one reference.

In still a further preferred embodiment, the control system of the arrangement is adapted to process the determined signal by decomposition before it is com- pared with the at least one reference, in particular by wavelet decomposition, polynomial decomposition, func- tional decomposition or Fourier decomposition. These and other types of decomposition are known to the person skilled in the art and therefore are not described here more into detail.

In still a further preferred embodiment, the control system of the arrangement is adapted to use a support vector machine (SVM) , a random forest algorithm and/or an artificial neural network for comparing the determined signal in its determined form and/or in a processed form with the at least one reference. These mathematical methods are known to the skilled person and therefore are not described here more into detail. In still a further preferred embodiment, the control system of the arrangement is adapted to apply online machine learning in comparing the determined sig- nal in its determined form and/or in a processed form with the at least one reference. Different methods of machine learning are known to the skilled person and therefore are not described here more into detail.

BRIEF DESCRIPTION OP THE DRAWINGS

The invention will be better understood and objects other than those set forth above will become ap- parent when consideration is given to the following de- tailed description thereof. Such description makes refer- ence to the annexed drawings, wherein: Fig. 1 is a schematic representation of a first arrangement according to the invention;

Fig. 2 is a schematic representation of a second arrangement according to the invention;

Fig. 3 is a schematic representation of a third arrangement according to the invention;

Fig. 4 is a schematic representation of the process zone of the arrangement of Fig. 1;

Fig. 5 is a schematic representation of a holder for pre-strained optical fiber sensors;

Fig. 6 is a schematic representation of a preferred variant of the process zone of the arrangement of Fig. 1;

Fig. 7a shows the μν/μβ-signal representing the acoustic emissions generated in the process zone by a high-voltage discharge as determined with the fiber axis of the sensor oriented parallel with a line defined by a shortest distance between the two electrodes;

Fig. 7b shows the uV/ps-signal representing the acoustic emissions generated in the process zone by a high-voltage discharge as determined with the fiber axis of the sensor oriented perpendicular to a line defined by a shortest distance between the two electrodes;

Fig. 8 shows examples of transparent artifi- cial material samples before the treatment with electric discharges;

Fig. 9 shows examples of transparent artifi- cial material samples after the treatment with electric discharges together with the corresponding data of the acoustic emissions caused by the discharges;

Fig. 10 shows two different discrete wavelet trans-formation (DWT) schemes;

Fig. 11 is a table (Table I) with classifica- tion categories and training and test acoustic emission datasets;

Fig. 12 shows examples of transparent artifi- cial material samples after the treatment with electric discharges resulting in different effects together with the corresponding data of the acoustic emissions caused by the discharges;

Fig. 13 shows the characteristics of a 4-band data adaptive wavelet;

Fig. 14 shows principal component analysis (PCA) weight maps for frequency bands extracted with a 4- band discrete wavelet transformation (DWT); and

Fig. 15 shows a table (Table II) with classi- fication test accuracy results.

MODES FOR CARRYING OUT THE INVENTION

Fig. 1 shows a schematic representation of a first arrangement for pre-weakening or fragmenting of rock material 4 by means of high-voltage discharges 5 according to the invention.

As can be seen, the arrangement comprises a tunable high-voltage generator 1, a process chamber 2 with a process zone formed between two electrodes 20, 20' arranged at a distance relative to each other, which pro ^¬ cess zone is flooded with water 21 as process liquid. The rock material 4 that is to be weakened or fragmented is arranged between the two electrodes 20, 20' immersed in the water 21.

The arrangement further comprises a control system 3 (see dotted line) comprising an acoustic emis- sion sensor 31, which is immersed in the process chamber 2 in the water 21 close to the electrodes 20, 20' and is connected via a data-line B with a signal processing unit 32. By means of the control system 3, the pulse parame- ters of the tunable high-voltage generator 1 can be set via a control line 33.

In the intended fragmentation operation, the high-voltage generator 1 is set to certain pulse parame- ters by the control system 3, and a high-voltage dischar- ge 5 is generated between the two electrodes 20, 20' . The line defined by a shortest distance between the two elec- trodes 20, 20' between which the high voltage discharge 5 is generated, is denominated with 22.

The acoustic emission sensor 31 determines a signal representing the acoustic emissions 30 in the pro- cess zone caused by this high voltage discharge 5 and sends it via the data-line B to the signal processing unit 32, which processes it and compares it with a refer- ence that has been determined before in trial runs with sample material and which represents the effect "dischar- ge through the material causing disintegration of the material (fragmentation)".

In case the comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has resulted in a fragmentation of the rock material 4 , for the next high voltage discharge 5 the pulse parameters of the high-voltage generator 1 are kept unchanged or the control system 3 sets them identical to the parameters of the pulse that generated the previous high-voltage dis- charge 5 , respectively. in case the comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has not resulted in a fragmentation of the rock material 4, for the next high voltage discharge 5 the pulse parameters of the high-voltage generator 1 are changed by the control system 3 in such a manner that the likelihood of a frag- mentation of the material 4 is increased, e.g. by increa- sing the voltage of the next pulse.

In the intended pre-weakening operation, the high-voltage generator 1 is set to certain pulse parame- ters by the control system 3, and a high-voltage dischar- ge 5 is generated between the two electrodes 20, 20' . The acoustic emission sensor 31 determines a signal represen- ting the acoustic emissions 30 in the process zone caused by this high voltage discharge 5 and sends it via the data-line B to the signal processing unit 32, which pro- cesses it and compares it with a reference that has been determined before in trial runs with sample material and which represents the effect "discharge through the mate- rial causing mainly internal damages, in particular cracks (pre-weakening)".

In case the comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has resulted in a pre-weakening of the rock material 4, for the next high-voltage discharge 5 the pulse parameters of the high-voltage generator 1 are kept unchanged or the control system 3 sets them identical to the parameters of the pulse that generated the previous high-voltage discharge 5, respectively.

In case the comparison reveals that, accor ^¬ ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has not resulted in a pre-weakening of the rock material 4, the signal is compared with the reference that represents the effect "discharge through the material causing disinte- gration of the material (fragmentation)".

In case this comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has re- suited in a fragmentation of the rock material 4, for the next high-voltage discharge 5 the pulse parameters of the high-voltage generator 1 are changed by the control sys- tem 3 in such a manner that the likelihood of a fragmen- tation of the material is decreased, e.g. by decreasing the voltage of the next pulse.

In case this comparison reveals that, accor- ding to this template, the high-voltage discharge 5 to which the determined and processed signal belongs has not resulted in a fragmentation of the rock material 4, for the next high-voltage discharge 5 the pulse parameters of the high-voltage generator 1 are changed by the control system 3 in such a manner that the likelihood of a pre- weakening or fragmentation of the material is increased, e.g. by increasing the voltage of the next pulse.

The above steps are repeated for each high- voltage discharge until the desired fragmentation or weakening result is achieved.

Fig. 2 shows a schematic representation of a second arrangement for pre-weakening or fragmenting of rocks 4 by means of high-voltage discharges according to the invention.

As can be seen, this arrangement as well com- prises a tunable high-voltage generator 1, a process chamber 2 with a process zone formed between two electro- des 20, 20' arranged at a distance relative to each other, which process zone is flooded with water 21 as process liquid. Also here, the rock material 4 that is to be pre-weakened or fragmented is arranged between the two electrodes 20, 20' immersed in water 21.

The control system 3 of this arrangement comprises a signal processing unit 32, by means of which also here the pulse parameters of the tunable high-vol- tage generator 1 can be set via a control line 33. This signal processing unit 32 is furthermore connected with the high-voltage generator 1 via a data-line A. Via this data-line A, signals representing parameters of the high- voltage discharges caused by the high-voltage pulse of the generator 1 can be determined and send to the signal processing unit 32. Such signals are, for example, the discharge voltage, the discharge current, the discharge resistance and the ignition delay.

In the intended fragmentation operation, the high-voltage generator 1 is set to certain pulse parame- ters by the control system 3 and a high-voltage discharge 5 is generated between the two electrodes 20, 20' . The line defined by a shortest distance between the two elec- trodes 20, 20' between which the high voltage discharge 5 is generated, is denominated with 22.

From the high-voltage generator 1, signals representing a parameter of the high-voltage discharge are determined and send to the signal processing unit 32, for example the discharge voltage and the ignition delay, which processes them and compares them with a reference range or with a reference that has been determined before in trial runs with sample material, and which represents the effect ^discharge through the material causing disin- tegration of the material (fragmentation)".

In case the comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has re- sulted in a fragmentation of the rock material 4, for the next high-voltage discharge 5 the pulse parameters of the high-voltage generator 1 are kept unchanged or the con- trol system 3 sets them identical to the parameters of the pulse that generated the previous high-voltage dis- charge 5, respectively.

In case the comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has not resulted in a fragmentation of the rock material 4, for the next high-voltage discharge 5 the pulse parameters of the high-voltage generator 1 are changed by the control system 3 in such a manner that the likelihood of a fragmentation of the material is increased, e.g. by increasing the voltage of the next pulse.

In the intended pre-weakening operation, the high-voltage generator 1 is set to certain pulse parame- ters by the control system 3 and a high-voltage discharge 5 is generated between the two electrodes 20, 20' .

From the high-voltage generator 1, signals representing a parameter of the high-voltage discharge 5 are determined and are send to the signal processing unit 32, for example the discharge voltage and the ignition delay, which processes them and compares them with a re- ference range or with a reference that has been determi- ned before in trial runs with sample material and which represents the effect "discharge through the material causing mainly internal damages, in particular cracks (pre-weakening) ".

In case the comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has resulted in a pre-weakening of the rock material 4, for the next high-voltage discharge 5 the pulse parameters of the high-voltage generator 1 are kept unchanged or the control system 3 sets them identical to the parameters of the pulse that generated the previous high-voltage discharge 5, respectively.

In case the comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has not resulted in a pre-weakening of the rock material 4, the signal is compared with the reference that represents the effect "discharge through the material causing disinte ^¬ gration of the material (fragmentation)". 5 in case this comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has resulted in a fragmentation of the rock material 4, for the next high-voltage discharge 5 the pulse parameters of0 the high-voltage generator 1 are changed by the control system 3 in such a manner that the likelihood of a frag- mentation of the material is decreased, e.g. by decrea- sing the voltage of the next pulse.

In case this comparison reveals that, accor- ding to this reference, the high-voltage discharge 5 to which the determined and processed signal belongs has not resulted in a fragmentation of the rock material 4, for the next high voltage discharge 5 the pulse parameters of the high-voltage generator 1 are changed by the control system 3 in such a manner that the likelihood of a pre- weakening or fragmentation of the material 4 is increa- sed, e.g. by increasing the voltage of the next pulse.

The above steps are repeated for each high- voltage discharge until the desired fragmentation or pre- weakening result is achieved.

Fig. 3 shows a schematic representation of a third arrangement for pre-weakening or fragmenting of rock material 4 by means of high-voltage discharges 5 according to the invention. This arrangement is a combi- nation of the two before described arrangements. Its con- trol system 3 processes the signals representing the acoustic emissions 30 as well as the signals representing the parameters of the high voltage discharge and compares them with a combined reference or with separate referen- ces.

Fig. 4 shows a schematic representation of the process zone of the arrangement of Fig. 1, with the acoustic sensor 31, which is an acoustic-sensitive fiber sensor, alternatively positioned with the axis L of the fiber parallel to the line 22 defined by a shortest dis ^¬ tance between the two electrodes 20, 20' between which the high voltage discharge 5 is generated (designated with reference numeral 311) or at an angle Θ of 90° thereto (designated with reference numeral 312) . The dis- tance between the line 22 that is defined by a shortest distance between the two electrodes 20, 20' and the acoustic fiber sensor 311 or 312, respectively, is desig- nated with d.

As can be taken from the Figures 7a and 7b, which are showing the μV/μs-signals representing the acoustic emissions 30 generated in the process zone by a high voltage discharge 5 through the material 4 when the fiber axis L is parallel to the line 22 (Fig. 7a) and when the fiber axis L is at an angle Θ of 90° to the line 22 (Fig. 7b), the sensitivity of the sensor 311 or 312, respectively, can be significantly increased by tilting the fiber axis L with regard to the line 22.

Fig. 5 shows a schematic representation of a holder 6 for keeping optical fiber of the sensor 31 tightened. The strength of the strain of the fiber of the optical sensor 31 affects its sensitivity. The strain of a fused silica fiber can for example be in the range bet- ween 10 - 5000 um/m. In case of a polymer optical fiber, it can largely be extended up to 10000 um/m. By varying the strain of the fiber and the length between its fixed parts, it is possible to tune the sensitivity of the sen- sor 31 in frequency domain, while together with the orientation of its fiber axis L with regard to the line 22 it is possible to tune it in amplitude frequency do- main. The holder 6 comprises a frame 60 and clamps 61 for tightening the fiber from both sides of the sensor area C with a pre-strain onto the holder 6. In Fig. 5, a light source 62 is shown transmitting light in direction of the fiber axis L through the fiber. Preferably, the pre- strain of the fiber can be changed manually or automati- cally during operation in order to tune the sensitivity of the sensor 31. Fig. 6 shows a schematic representation of a preferred variant of the process zone of the arrangement of Fig. 1 having eight optical fiber sensors 31 being equally arranged at an equal distance d from the line 22 defined by a shortest distance between the two electrodes 20, 20' around the process zone. All optical fiber sen- sors 31 or the fiber axis L thereof, respectively, are tilted at the same angle Θ with respect to the line 22. With this arrangement, from the signals determined by these sensors 31 and their known positions, the location of the high-voltage discharge 5 can be determined.

STUDY CONDUCTED REGARDING THE ACOUSTIC CHARACTERIZATION OF SOLID MATERIALS PRE-WEAKENING USING ELECTRICAL

DISCHARGES

In the following, a study is described that was carried out in order to improve the monitoring of the pre-weakening quality of material treated with high-vol- tage discharges. This study presents a preferred way of processing and comparing the determined signals represen- ting the acoustic emissions (AE) in the process zone cau- sed by the high voltage discharges, which has practically been verified with artificial polymer based materials and can, without any problem for the person skilled in the art, be adapted for other materials, like e.g. rock mate- rial or ore, and be implemented as step d) into the meth- od according to claim 1 for providing the result used in step e) of said method. Such a method implementing this way of processing and comparing the determined signals represents a preferred embodiment of the method according to the first aspect of the invention.

I. EXPERIMENTAL SETUP AND ACOUSTIC DATA ACQUISITION A. Artificial samples The study was carried out using transparent artificial samples (TAS) that allowed visual monitoring of the cracks formation induced by the electric dischar- ge.

Fig. 8 shows some of the transparent artifi- cial samples (TAS) before electric discharge: Fig. 8 (a) shows a side view of poly methyl methacrylate (PMMA) sample without inclusions, Fig. 8 (b) shows top and side views of pressed PMMA samples with mineral inclusions, its diameter and height are 50 mm and 20 mm respectively; Fig. 8 (c) shows top and side views of pressed epoxy sample with cylindrical glass inclusions and size 50x20 mm; Fig. 8 (d) shows a schematic view of the sample location in the discharge chamber, filled with water.

According to the visual control, all AE sig- nals were labeled according to the observed cracks . The labeling criteria as well as the corresponding pre-weak- ening categories are postponed until Section III. Here, the types and properties of TAS are discussed.

Two types of TAS were manufactured from two polymers based materials; poly methyl methacrylate (PMMA) and epoxy resin. The dielectric constants of the selected materials is around 3 for PMMA and around 4 for epoxy which fit into the range of dielectric constants of most natural solid materials (natural rocks) that are between 3 and 20. The dielectric properties of the TAS can be compared to the ones of quartz, which is a component of a broad range of natural solid materials. In contrast, the dielectric strength of the PMMA and epoxy varies in the range of 15-20 MV/m as compared to the one of natural solid materials of 1.9 - 7 MV/m. Hence, to facilitate the dielectric breakdown of the TAS, a metallic pin was introduced inside the sample to provide an electric field enhancement and thus provoke the electric discharge through the TAS medium. The pin is shown in Fig. 8a, and its position with respect to electrode is schematically demonstrated in Fig. 8d.

TAS with and without inclusions were produced for collecting AE data. The TAS with inclusions were used to simulate the disordered grain structure of real solid materials in a simplified two-component model. The vari- ations in TAS inner structure provided the stochastic de- velopment of discharges inducing pre-weakening, in accor- dance with the mechanisms described above.

Several types of inclusions were incorporated into the TAS at various concentrations and different po- sitions. As inclusions, glass pieces of various shape (ball, cylinder, cube) and sizes (from 5 to 40 mm) as well as table salt or mineral particles (from 2 to 8 mm) from magnetite (Fe304) and hematite (or-Fe203) were em- ployed and some examples are given in Fig. 8b and c.

The TAS made of PMMA without inclusions pos- sess a homogeneous medium and were manufactured by cut- ting of 50 mm square rod of pure PMMA into slices of 20 mm thickness as shown in Fig.8a. In contrast, the TAS made of PMMA (or PMMA TAS) with inclusions were produced by hot pressing under vacuum of acrylic hot mounting re- sin powder (Clarofast from Struers) at 170 °C and 25MPa for 15 min and then cooled down. The inclusions were in- corporated inside the powder before hot pressing.

Finally, the TAS from epoxy were produced by chemical reaction of the two components, of a clear epoxy resin, inside a 50 mm square mold. To add inclusions, the epoxy samples were composed of two layers (as shown in Fig.8c). After the solidification of the first layer, the inclusions were placed inside the second layer at defined positions.

B. High voltage generator The discharge events inside the TAS were ini- tiated using a big scale voltage generator from Selfrag AG (Kerzers, Switzerland) . It allowed tuning of the ope- rating voltage and storage capacitance in the range of 90 - 200 kV, negative polarity, and 2.5 - 75 nF, respective- ly. The voltage exposure of the TAS was carried out in a chamber filled with water. The setup is a standard indus- trial environment to provoke discharge preferentially in- side the solid materials in a given voltage range and it is schematically represented in Fig. 8d. The gap between the electrodes was 50 mm with the cathode (as the polari- ty is negative) touching the pin electrode of the sample.

C. Acoustic emission of electric discharge pre-weakening

The detection of the acoustic signals was made directly inside the water filled chamber using an acoustic hydrophone sensor R30UC (Physical acoustics cor- poration, USA) . It was grounded and placed at a distance of 20 cm from the electrode gap (see Fig. 8d) . The AE signals were recorded with a 10 MHz sampling rate and an electrical signal amplification of 20 dB. The recording time was 16 ms and it was initiated via a record trigger synchronized with the discharge initiation.

All TAS were exposed to electrical pulses and the corresponding AE signals were recorded. Two examples of TAS with inclusions and their corresponding AE data are shown in Fig. 9. Fig. 9 (al) and (bl) are top views of TAS 1 and TAS 2, respectively, after an electric dis- charge, where with the white markers point the glass inclusions. In Fig. 9 (al) the inclusions are glass balls placed inside the sample and the insert shows the

delamination of a ball. In Fig. 9 (a2) the broken glass inclusions can be observed around the center of TAS 2 and a detail is visible in the insert. Fig. 9 (a2) and (b2) are AE signals of the electric discharge in TAS 1 and TAS 2, respectively. In Fig. 9 (b2) , the top right insert is a zoom of the crack acoustic echo showing the abrupt changes in frequency content. Fig. 9 (a3) and (b3) are sonograms of AE from both TAS. Both samples were exposed to the same electrical pulse with voltage and capacitance of 120 kV and 2.5 nF, respectively. The metal pin electrode is placed at the center of the TAS.

Both presented TAS are epoxy made with glass inclusions of diameter of 5 mm (Fig. 9 (al) ) and 6-8 mm (Fig. 9 (bl)). The inclusions were placed at the periphe- ry in the first TAS, whereas in the second one, they were located in the center. Both TAS were exposed to identical electrical pulses with a voltage and capacitance of 120 kV and 5 nF, respectively. The process resulted in the occurrence of multiple cracks without samples disintegra- tion (pre-weakening) . Under such circumstances, the stress wave propagation can be described as a cylindri- cally symmetrical process that starts from the discharge area in the middle of the TAS and propagates to its periphery. This is confirmed when observing the cracks induced by discharges in Fig. 9 (al) (TAS 1) and Fig. 9 (bl) (TAS 2), which radially propagate in different direc- tions starting from the pin electrode in the middle of the TAS. The stochastic nature in cracks formation is ob- vious when comparing Fig. 9 (al) and Fig. 9 (bl) . Their number and positions are very different between both sam- ples. Additionally, damage of some inclusions can be ob- served in TAS 2 against TAS 1, where only one glass ball inclusion was deliminated. The corresponding recorded AE is shown in Fig. 9 (a2) and Fig. 9 (b2) . Different inclu- sion distributions and their sizes affect the AE and lead to different distributions of frequencies in the time do- main. As an example, the flux in AE of TAS 2 in Fig. 9 (b2) can be seen 10ms after electric discharge onset. It corresponds to crack occurrences, which are absent in TAS 1 in Fig. 9 (bl) . Besides, different contents are visible when comparing both signals from 2 to 7 ms after the process initiation (see Figs 9 (a2) and 9 (b2)). The non- stationary behavior of the AE is shown in the insert in Fig. 9 (b2), where the individual crack echo is depicted. As can be observed it is characterized by abrupt changes in the frequency content.

The frequency energy distribution for both signals is presented in Fig. 9 (a3) and Fig. 9 (b3) . The crack formation frequencies are in the range of 0-160

KHz, although the distribution of those within this range is different for each sample. This diversity is provided by the presence of inclusions and different distribution of accumulated mechanical energy within each sample and evidence of this is visible in Fig. 9 (a3) and Fig. 9

(b3) on separate fluxes marked with the arrows. Each sig- nal includes a random number of fluxes that occurs at random time. The challenge in classification of such sig- nals is in the extraction of a fixed combination of fre- quencies that uniquely characterize each pre-weakening state. The methodology for this will be described in section III.

II. TIME FREQUENCY ANALYSIS USING M-BAND WAVELETS

A. M-band wavelets

Wavelet transform (WT) was introduced Daube- chies (Daubechies I., Ten Lectures on Wavelets; CBMS-NSF Lecture Notes nr. 61, SIAM (1992)) as an alternative to Fourier technique, expanding the analysis from the fre ^¬ quency to the time-frequency domain. This possibility is given by employing short oscillating functions as a sig- nal decomposition basis. They are known as wavelets and are localized in both time and frequency domains. The basis wavelets are prototyped from a single wavelet func ^¬ tion called mother wavelet and the WT of continuous sig ^¬ nal f(t) is defined as: where ψ* is the wavelet function scaled by j and trans- lated by k _f and t is the time stamp. In practical compu- tations, the continuous wavelet transform is replaced by its discrete counterpart called discrete wavelet trans- form (DWT) , which operates with discrete sampled signals f [i] and has a limited number of computational steps. DWT is defined as: where the wavelet function is defined as

is the total number of samples i in the signal and j is the resolution level. The DWT can be also defined in ter- ms of filter banks (Daubechies I., Ten Lectures on Wave- lets; CBMS-NSF Lecture Notes nr. 61, SIAM (1992)), in which both the scaling and wavelet functions act as fil- ter channels, extracting the low and the high frequency content of the signal correspondently. The definition of DWT is as follows:

where for simple DWT, M is equal to 2, <p(n) is the cor- responding scaling function at the decomposition level j, ψ() is the wavelet, ho, h«-i are the low pass and high pass filters, respectively. The wavelet function is character- ized by vanishing moments N, satisfying the condition: where k-1 ,.,,Ν. The greater number of vanish- ing moments allows representing the complex signals with smaller number of wavelet coefficients.

Fig. 10 shows schemes of two different DWTs.

Fig. 10 (a) shows the scheme of a Standard DWT, where hO and hi are the low pass and high pass filters respecti- vely; Fig. 10 (b) of a M-band DWT with M-channels, where hO is the low pass, hl-h.M-2 are narrow band and hM-1 is the high pass filters.

The full DWT for several scales is carried out using a pyramidal scheme presented in the Fig. 10 (a), which includes several stages and the sequential split of the low frequency content is carried out at each stage. This results the division of the signal frequency content into narrow frequency bands, which are localized in time. The two-channel DWT can be extended to multiple channels providing more detailed partitioning of the time-frequency space. This is achieved by involving M-l wavelets instead of one that are associated with the scale function. Each wavelet is applied to the individual subspace of the signal thus providing with more precise decomposition (Gupta A., Joshi S.D., Prasad S., A new approach for estimation of statistically matched wavelet; IEEE transactions on signal processing, vol. 563, 5, p.1778-1793 (2005)). The M-band decomposition scheme is presented in Fig. 10 (b) and its application results in extraction of low, several narrow and high frequency bands. In the present contribution, the information from the channels, marked as outputs in Fig. 10 were taken for further analysis. The relative energies were computed for each output and considered as features.

The energy of the individual frequency band is defined as:

where d are the wavelet or scale function coefficients extracted from Eqs (3) and (4). Following Eq. (5), the relative energies are the normalized version of the sub band energies and defined as:

where is the total energy accumulated in all frequency bands at resolution level j.

fhe decomposition of the signal into separate narrow frequency bands makes wavelets suitable for noise reduction. In the present work, the intrinsic machine noises were measured and the corresponding frequency bands were excluded from further analysis.

B. Design of data adaptive M-band wavelets. The AE signals are characterized by changes in frequencies and amplitudes as shown in the insert of Fig. 9 (b2) . This results in the presence of a wide va- riety of differently shaped patterns within the acoustic signal due to complex interferences of acoustic echoes coming from multiple cracks. To achieve a better wavelet approximation operating with signals of such statistical diversity, the analysis was carried out using data adap- ted wavelets.

Several techniques for wavelet construc- tion were developed in the last decade. However, in this investigation, the computational complexity, and the de- sign constraints were taken into consideration. Mallat (Mallat P., Zhang P., Matching pursuits with time-fre- quency dictionaries; IEEE, Trans. Sign. Proc; vol 41,12, p.3397-3415 (1993)) and Krim (Krim P., Tucker P., Mallat P., Donoho P., On denoising and best signal represent- tation; IEEE, Trans. Inf. Theory, vol.45, 7, pp.2225-2238 (1999)) proposed methods where the wavelet is constructed as an optimal combination from several pre-defined wave- lets. The main limit of the methods is that the result depends on the initial wavelet collection that originally may not be optimal for a specific application. Tewik (A. H. Tewfik, D. Sinha and P. Jorgensen: On the optimal choice of a wavelet for signal representation. IEEE

Trans, on In-formation Theory 38(2) (March 1992) 747-765) and Gopinath (R.A. Gopinath, J.E. Odergard, C.S. Burrus, Optimal wavelet representation of signals and the wavelet sampling theorem; IEEE transactions on Circuits and sys- tems-II: Analog and digital processing, vol.41, 4, 1994) used the method of error approximation minimization in both the frequency and time domains. Their results, pre- sented for the two-channel case, leads to an exponential raise of computational complexity when extended to M- bands. In Unser (Aldroubi, M.Unser, Families of multire- solution and wavelet spaces with optimal properties, Numer .Func.Anal. , vol.14, no.5/6, pp.417-446, 1993) , another method was proposed for biorthogonal wavelets construction, limiting the design to existing wavelet basis. In our study, the method proposed by Gupta et al. (Gupta A., Joshi S.D., Prasad S., A new approach for es- timation of statistically matched wavelet; IEEE transac- tions on signal processing, vol. 563, 5, p.1778-1793

(2005) ) was used which adapts the M-band wavelets to the signal in a statistical manner taking into account the diversity of the signals content. The method exploits the self-similarity as a global likelihood criterion between wavelet approximation and the original signal. It sup- ports the construction of both, orthogonal and bio-ortho- gonal bases. In present contribution both bases were in- vestigated applicably to the AE decomposition and the one with the minimum approximation error was chosen. Its cha- racteristics are described in section III, B.

C. Feature space reduction

Before classification, only the most informa- tive features were selected using principal component analysis (PCA) (I.T. Jolliffe, Principal Component Analy- sis, second edition (Springer), 2002). The disposal of non-informative features decreases the noise in classi- fication and additionally reduces the computational com- plexity (I.T. Jolliffe, Principal Component Analysis, second edition (Springer), 2002). The features selection with PCA is carried out by projecting the features from their original featu- re space Uo into the linear subspace Ui that has a reduced dimensionality and is a linear approximation of Uo . The new features coordinates in the subspace Ui are defined through projection, which is defined as:

where Xuo is the matrix of dimensions [n, p] , p is the number of features, n is the number of measurements, w is the principal components matrix that includes the weights are the new features coordinates in the re- mapped space. The projection w is constructed to maximize the variance of Eq. (7) :

where S is the covariance matrix of Xoo (I.T. Jolliffe, Principal Component Analysis, second edition (Springer) , 2002) . The solution can be obtained by a diagonalization of S using singular value decomposition, selecting eigen- vectors with the highest eigenvalues. The elements ob- tained from w are known as principal components and the selection of informative features is achieved by employ- ing only w with the greatest variance values.

In case when the number of observations exceeds the feature space dimensions, a linear projection into low dimensions is possible and this is presently the case. In the opposite situation, this is addressed to non-linear PCA and more details for both can be found in (I.T. Jolliffe, Principal Component Analysis, second edition (Springer), 2002).

D. Support vector machine classifier

Support vector machine (SVM) is a statistical machine learning technique proposed by Cortes and Vapnik (Cortes, C, Vapnik, V., Support-vector networks, Machine Learning 20 (3): 273, (1995), doi:10.1007/BF00994018) . For a simple binary case, the objective of the classifier is to separate high dimensional feature sets that belong to two categories into two distinct groups, which are labeled This is performed by constructing a decision hyperplane that separates the sets in the fea- ture space providing the maximum margin between them. The construction of hyperplane is a quadric programming pro- blem which is resolved during an SVM training procedure using pre-labelled feature sets, called training sets. The data points from the two features sets that are neighboring the decision hyperplane are known as support vectors and are the essential margin characteristic. The constructed hyperplane allows defining the decision func- tion for classification of unlabeled data, defined as: where f(y) is the label for the current feature vector and is equal to either 1 or -1, xi are the values

of the feature vector X, s _sV is the support vector, αi is the Lagrangian multiplier and Jb is a bias, computed du- ring the training procedure (Cortes, C, Vapnik, V., Support-vector networks, Machine Learning 20 (3): 273, (1995), doi:10.1007/BF00994018; Hofraann, Thomas; Schol- kopf, Bernhard; Smola, Alexander J. Kernel Methods in Machine Learning, (2008)).

In case of non-linearly, separable features sets SVM employs a so called kernel trick

(Hofmann, Thomas; Scholkopf, Bernhard; Smola, Alexander J. Kernel Methods in Machine Learning, (2008)), where the features sets are remapped into a higher dimensional space wherein they are linearly separated. For this contribution, the decision function is defined as: where k( _f ) is a kernel function. In the present work, the radial based function (RBF) was used as it showed the best classification results. It is defined as

The classification scheme, described abo- ve, can be applied for cases with more than two catego- ries using a cascade from binary classifiers which was used in the present work. The basic implementation of SVM was taken from work (C.-C. Chang and C.-J. Lin. LIBSVM : a library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2:27:1—27:27, 2011) . III. RESULTS AND DISCUSSION

A. Categorization and labeling of AE dataset

This investigation was carried out on 205 ma- nufactured TAS, from which 86 were made using PMMA and the other 119 from epoxy. In the latter, glass inclusions were encapsulated with various concentrations; simulating the non-uniformity in natural materials and providing di- versity in the AE signals, as described in section I, C. The samples were exposed to pulses with voltage and capa- citance in the range of 90 - 180 kV and 2.5-10 nF, res- pectively. Most samples were treated with multiple pulses until the desired pre-weakening was reached. The AE sig- nal from each electrical pulse was recorded leading to a total of 500 AE signals. Each TAS was visually controlled after each pulse treatment. According to this, the AE signals were divided into categories and subcategories. The categories scheme with their corresponding number of included signals and TAS are presented in Fig. 11 (Table I) . The first classification level includes three catego- ries that describe electric discharge propagation, while the second classification level contains several subcate- gories that describe the electric discharge efficiency in terms of pre-weakening. The classification of categories precedes the one of subcategories.

Fig. 12 shows optical microscope view of TAS subjected to different electric discharges resulting in different effects and the respective AE signals recorded. Fig. 12 (a) is a top view of TAS made of epoxy from the category Discharge in TAS, subcategory Pre-weakening, The applied electrical pulse is 90 kV and 2.5 nF. The arrow indicates the exit of the discharge. Cracks propagated from the discharge path but are contained within the sample. Most of the glass inclusions are also cracked. Fig. 12 (b) is a top view of TAS made of epoxy from the category Discharge in TAS, subcategory Break-down. The applied electrical pulse is 130 kV and 8 nF. The arrow indicates the material loss in the central area of the pin electrode. Fig. 12 (c) is a side view of the pin electrode area of a TAS made of PMMA for the category Discharge fail, subcategory Discharge tree. The applied electrical pulse is 120 kV and 2.5 nF. Fig. 12 (d) is a side view of a TAS made of PMMA from the category Surface discharge, subcategory Without pre-weakening. Fig. 12 (e) is a top view of a TAS made of epoxy from the category Surface discharge, sub category With pre-weakening. The applied electrical pulse is 120 kV and 2.5 nF. The glass inclusion (large square of 40mm x 40mm x 15mm) has many cracks induced by the surface discharge. The epoxy matrix is intact around the inclusion. Fig. 12 (f) is a side view of a TAS made of PMMA after a discharge exposure of 120 kV and a capacitance of 8 nF, where left are cracks around the pin electrode zone and right is a zoom in the area of discharge path.

The category Dlechnrge in TAS includes the AE that describes the pre-weakening due to the electric dis- charge propagation inside the TAS medium. The TAS from the corresponding subcategories pre-weakening and break- down are presented in Fig. 12 (a) and Fig. 12 (b) , res- pectively. In both TAS, the discharge area is visible at the center and multiple cracks are propagating from the TAS center where the metallic pin is placed. In the case of pre-weakening, the shorter cracks are confined inside the TAS without propagation to its edges and thus it pre- vents the disintegration of the TAS in pieces (as shown in Fig. 12 (a). The crack formation area around the elec- tric discharge propagation path for this subcategory is additionally depicted in Fig. 12 (f) and can be observed in all TAS after the discharge treatment. Multiple cracks compactly surround the pin electrode area as shown in Fig. 12 {f) (left) . The cross-section of the same area in Fig. 12 (f) (right) shows the presence of the discharge channel responsible for the surrounding cracks. The TAS belongings to the subcategory break-down are easily reco- gnized as the cracks are propagating to the edges leading to the breakage of the TAS in several separate parts as seen in Fig. 12 (b) . In several cases the break-down pro- cess was accompanied by the loss of material. This can be observed in Fig. 12 (b) , where the corresponding area is where the pin electrode placement is marked with an arrow. The signals from both subcategories {pre-weakening and break-down) showed similarities and signal attenu- ations to the noise level at about 15 ms. It is important to mention that the category discharge in TAS is the most important one as it provides pre-weakening at a minimum energy consumption. Its differentiation using AE from other categories will be discussed in the next subsec- tion.

The category Discharge fall includes par- tial or no discharge and includes two subcategories: no discharge and discharge tree (see Fig. 11 (Table I)). The subcategory no discharge occurs when the voltage of the electrical pulse is lower than the dielectric strength of the TAS. Consequently, all TAS preserved their integrity with no visible damage. In the case of no discharge, the TAS were treated iteratively by increasing the voltage until the dielectric break-down level is exceeded so that the electric discharge happened. The signals from this subcategory were used for estimating the intrinsic noise of the generator. The corresponding frequency sub-bands were removed from further analysis according to descrip- tion given in Section II-A. The subcategory discharge tree includes AE from events when the applied voltage was enough to provide the growth of the streamer towards the counter electrode but the energy of the pulse was still low so it was absorbed by the AS medium before reaching the counter electrode. In such cases, the branching chan- nels of the damping discharge propagation inside the TAS medium were observed and depicted in Fig. 12 (c) . The nature of this structure is a surface interface inside the TAS medium created by plasma formation. The AE sig- nals from both subcategories (no discharge and discharge tree) are characterized by low acoustic energy and short duration and evidence of this is seen from the corres- ponding AE signals in Fig. 12 (c) .

The category Surface discharge contains the

AE signals when the discharge occurred in the surrounding water environment or along the sample surface. This hap- pens when break-down voltage of the interface TAS/water is lower than the one of TAS. The surface discharge leads to two events that are described by the corresponding subcategories in Fig. 11 (Table I) . The subcategory of surface discharge without pre-weakening is recognized by scratches on the TAS surface as shown in Fig. 12 (d) . Actually, in this case, no cracks inside the TAS medium are observed but only surface scratches indicating the discharge propagation path. Noteworthy, despite having a discharge occurring outside the TAS, crack formations are still possible due to propagation of pressure waves indu- ced by the outer electric discharge through the sample. However, to achieve this state, the energy of pressure waves entering the sample has to be equal or higher than the energy required for cracks initiation. Thus, this phenomenon is observed almost exclusively in TAS with a brittle inclusion such as glass. This subcategory surface discharge with pre-weakening is illustrated in Fig. 12 (e), where the cracks formation in TAS induced by the surface discharge is visible. As can be seen, the cracks occur mainly close to the TAS surface and gradually dis- appeared while propagating into the TAS center. This is in contrast with discharge inside where the cracks propa- gate from the center of the TAS as presented in Fig. 12 (a) and Fig. 12 (b) .

Observing the AE signals from all TAS presen- ted in Fig. 12, it can be perceived that ones from dis- charge in TAS and the surface discharges have close iden- tity in both time duration and amplitude levels. Taking into account the different number of cracks caused by each individual electric discharge, the AE differentia- tion is challenging and its feasibility will be discussed in the next subsection. In practice, the events 1.1 and 3.1 from Table I (see Fig. 11) are of the highest impor- tance as they provide the desired solid materials treat- ment.

The two-level classification presented in Table I (see Fig. 11) was carried out with four cascade SVM classifiers using "one against all" classification scheme. The classifiers were trained and tested on sepa- rate datasets for which the total number of included AE signals is also given in Table I. During training, the first classification level (categories level) incorpo- rates all signals from the lower level (subcategories level) .

B. Data adaptive wavelet

The choice of the most suitable wavelet for collected AE signals was done once before analysis of en ^¬ tire dataset involving Gupta et al. (Gupta A., Joshi S.D. , Prasad S., A new approach for estimation of statis- tically matched wavelet; IEEE transactions on signal pro- cessing, vol. 563, 5, p.1778-1793 (2005)). The optimal wavelet was generated using several signals, selected from different categories (see Fig. 11 (Table I)). When designing the wavelet the number of vanishing moments and the number of the channels was adjusted to minimize the approximation error of selected AE signals. The optimal combination of parameters that provided the minimum ap- proximation error was at 5-band wavelet and six scales (see Eqs (3) and (4) and the scheme in Fig. 10 (b) .

The characteristics of the 4-band data adap- tive wavelet computed with the method described in I.T. Jolliffe, Principal Component Analysis, second edition (Springer) , 2002] using signals collected from the TAS are presented in Fig. 13.

The wavelet basis was taken as a biorthogonal and was constructed using several signals taken from the different categories (see Fig. 11 (Table I)). The appli- cation of the 4-band wavelet allowed to tile the time - frequency space into separate frequency bands of width in the range of approx. 9.7 kHz to 2.5 MHz. For further ana- lysis, the energies of those frequency bands were consi- dered as features and computed according to Eq. (6) . In this work, the sequence of those is the input information for the SVM classifier.

C. Feature space reduction

The selection of the most informative featu- res was carried out using principal component analysis (PCA) . As already mentioned, four classifiers were used to provide a two-level classification according to the structure presented in Table I of Fig. 11. Hence, the first classifier is used to separate the three categories and the other three classifiers are used to separate the subcategories within each category. To do so, four trai- ning datasets were formed and the PCA was performed for each dataset separately (the description of those are presented in Fig. 11 (Table I)). The solution of features space reduction in PCA is achieved by searching the ele- ments that provide the maximum cumulative variance (see description in section II-C) .

Fig. 14 shows PCA weight maps for frequency bands extracted with the 4-band DWT, wherein: Fig. 14 (a) shows the weight map for classification of the categories Discharge in TAS, Discharge fail and Surface discharge (categories 1 , 2, 3 in Fig. 11 (Table I)), Fig. 14 (b) shows the weight map for classification of the subcatego- ries from the category Discharge in TAS (categories 1.1 and 1.2 in Fig. 11 (Table I)), Fig. 14 (c) shows the weight map for classification of the subcategories from the category Discharge fail (categories 2 . 1 and 2.2 in Fig. 11 (Table I)) and Fig. 14 (d) shows the weight map for classification of the subcategories from the category Surface discharge (categories 3. 1 and 3.2 in Fig. 11 (Table I)). The horizontal marker on the scale bar defi- nes the threshold for the features selection. As outputs, the narrow and high frequency bands are numbered accor- ding to the scheme presented in Fig. 10.

In Fig. 14 , the computed variances in the low dimensional feature space are presented for all four da- tasets. The different shades of grey encode the relative variance, which was computed as the variance normalized by the cumulative variance. As can be seen, most of the variance is provided by the frequency bands ranging from 9. 765 to 156. 25 kHz. They are concentrated mostly in the time span of 0-3 ms after the signal start (see Fig. 14 (a-c) ) . In contrast, for the surface discharge subcatego- ries (Fig. 14 (d) ) , the features with the maximum varian ^¬ ce are in the time span of 0-9 ms after signal start. As the AE signals have a longer duration in comparison to the ones from other categories (see signal duration in

Fig. 12 ) , the variance for surface discharge is distribu- ted among a greater number of features than the other cases. Consequently, the most informative features, taken for further analysis, were selected according to a fixed threshold, which is marked by a red horizontal marker on the scale bars in Fig. 14. This threshold was selected after an exhaustive search to determine the optimal com- promise between classification accuracy and the features number.

D. Classification tests results The training of the classifiers and their tests were carried out using the different datasets des- cribed in Fig. 11 (Table I) . Signals from different sub- categories were decomposed and the extracted features we- re then fed top the cascade of classifiers. The results are presented in Fig. 15 (Table II), where in raws the classification accuracy for each category and subcategory is presented and can be compared to the ground truth reference. Here the accuracy for each subcategory (marked as dark grey in Fig. 15 (Table II)) is computed as:

the error rates (marked as light grey in Fig. 15 (Table II)) are computed as:

while the accuracy and error rate of each category is defined as:

5 where for (14) and (15) parameter n=l,2 corresponds to the number of subcategories, included into the current category.

Based on the data from Fig. 15 (Table II) for the first classification level (classification of catego-0 ries) , the accuracy of the category Discharge in ICRS, Discharge fail and Surface discharge (see 1. in Fig. 15 (Table II)) are as high as 87%, 93% and 84%, respecti- vely.

The category Discharge in TAS includes the5 pre-weakening events which is of upmost importance for practical applications, in particular for the mining in- dustry. 10 % of the misclassification are coming from the category Surface discharge (see 3. in Fig. 15 (Table II)). The signals from the latter have, in some cases, the same duration and intensity as the ones from the category Discharge in TAS which is the cause of errors. Few errors (3%) are coming from the categories Discharge rail (see 2. in Fig. 15 (Table II)). This can be explain- ed by the shorter duration of the AE signals as compared to the ones from the other categories and the lower in- tensity levels of generated AE (see Fig. 12 (c) ) . The second level classification for the subcategories Pre- weaking (see 1.1 in Fig. 15 (Table II)) and Break-down (see 1.2. in Fig. 15 (Table II)) have an accuracy of 85% and 89%, respectively. The main sources of misclassifi- cation for those subcategories are in their mutual over- lapping. This is certainly due to the closeness of the two categories. The only difference is that, in the sub- category Break-down, the cracks propagate until the sur- face of the sample thus breaking the latter in several pieces.

The category Discgarge full embraces the sub- categories Discharge tree (see 2.1 in Fig. 15 (Table II)) and No discharge (see 2.2 in Fig. 15 (Table II)). Despite an error of 7% for Discharge in TAS due to the closeness in duration and the same level of the AE intensity in some cases, the classification of this category and sub- categories showed the highest accuracy. This is attribu- ted to the short duration of the signals as compared to the other categories.

The category Surface discharge incorporates the subcategories Surface discharge with pre-weakening (see 3.1 in Fig. 15 (Table II)) and Surface discharge without pxe-weakening (see 3.2 in Fig. 15 (Table II)). With a classification rate of 84%, this category is the least accurate. It is seen that 15% of the misclassifica- tion is made with Discharge in TAS. Although, this mis- classification is not significant, this is most critical error since the overlapping of the categories Surface discharge with the Discharge in HAS decreases slightly the solids materials processing efficiency. Therefore, additional investigations are pursued to enhance this classification rate. Similarly, the subcategories Surface discharge with pre-ireakening and Surface discharge without pre-weakening remain the least accurate with 73% and 81%, respectively. However, most of the errors come from cross-classification between the two subcategories. This can be explained by the high variability of signals fea- tures of these subcategories. The AE from cracks forma- tions during the surface discharge is very weak and is hardly detectable compared to the strong AE created by the Shockwave in water. That brings to the mutual over- lapping of both subcategories as shown in Fig. 15 (Table II) . Another source of errors is in the classification of Pre-weakening (3% in 1.1, Fig. 15 (Table II)) or Break- down (7+5% in 1.2, Fig. 15 (Table II)) signals in these subcategories. Finally, the low classification accuracy for the surface discharge subcategories does not have a significant impact in practical applications. The reason is that the solid materials subjected to a surface dis- charge are not sufficiently damaged and/or pre-weakened, and they have to be additionally processed until the desired pre-weakening quality is achieved. Under such circumstances, it does not matter whether there is a misclassification of between the subcategories Surface discharge with pre-weakening and Surface discharge without pre-weakening. Actually, only the classification errors with the subcategories Pre-weakening and Break- down are problematic but as already mentioned, there are very limited.

Although it is presently shown and described preferred embodiments of the invention, it has to be distinctly understood that the invention is not limited thereto but it may be instead variously embodied and practiced within the scope of the following claims.