A method for processing measurements of a plurality of sensors installed at a technical system

The invention refers to a method for processing measurements of a plurality of sensors installed at a technical system.

Complex technical systems, e.g. dams, dikes, bridges, gas turbines and the like, are usually equipped with a network of a plurality of sensors measuring different parameters that can characterize the behavior of a technical system. This behavior can be assigned to an abnormal operation condition in case that measured signals of the sensors deviate from his- torical measurements related to a normal behavior. Such an abnormal behavior can be a sign of developing fault or failure, e.g. dike break or crack in a bridge.

In the prior art, several methods are known in order to de- tect an abnormal operation condition of a technical system based on sensor data of a plurality of sensors. E.g., a one- dimensional anomaly of the technical system, i.e. an anomaly where single measured signals of one or more sensors are out of range, can be detected and interpreted using uni-variate data analysis methods which are usually not able to detect and interpret multi—dimensional anomalies where several measured signals lead to the anomaly.

For multi-dimensional anomalies, multi-variate data analysis, e.g. multi-dimensional classifiers, are used. Those classifiers are trained by appropriate training data. After training, the classifier is applied to the measured signals of a plurality of sensors and gives an output indicating whether the underlying technical system shows a normal or an abnormal op- eration condition. The use of a multi-dimensional classifier has the disadvantage that it is not clear which sensors lead to the detected anomaly. Only the technical system as a whole is classified in the category of a normal or an abnormal operation condition.

It is an object of the invention to provide a method for pro- cessing measurements of a plurality of sensors based on multi-dimensional data analysis where core sensors responsible for an abnormal operation condition are identified.

This object is solved by the independent claims. Preferred embodiments of the invention are defined in the dependent claims .

According to the method of the invention, the measurements being processed are represented by one or more measurement vectors for one or more points in time, where a measurement vector at a respective point in time includes a measured signal for each sensor of the plurality of sensors. A first classifier trained by training data and designed for output- ting a normal or an abnormal operation condition of the tech- nical system is applied to the measurement vector for a corresponding point in time. I.e., the first classifier receives measured signals of all sensors as an input and provides as an output whether the technical system shows a normal or abnormal behavior. Here and in the following, the term "normal or abnormal operation condition" always refers to a normal or abnormal operation condition of the technical system.

In case of an abnormal operation condition outputted by the first classifier, second classifiers trained by the same training data as the first classifier and designed for out- putting a normal or an abnormal operation condition are applied to partial measurement vectors including a subset of at least two measured signals from the measurement vector until all so-called core partial measurement vectors out of the partial measurement vectors are identified. According to the invention, a core partial measurement vector comprises the following features:

i) The second classifier applied to the core partial measurement vector outputs an abnormal operation condition; ii) no smaller partial measurement vector exists which includes a subset of the measured signals from the core partial measurement vector, where the second classifier applied to the smaller partial measurement vector outputs an abnormal operation condition.

According to the above definition, the smallest subsets of measured signals leading to a classification in an abnormal operation condition are identified as core partial measurement vectors. As a consequence, all sensors which have meas- ured signals within the core partial measurement vectors are identified as core sensors providing measured signals leading to the abnormal operation condition.

The method of the invention has the advantage that - besides a classification in a normal and an abnormal operation condition - core sensors are identified which indicate the locations in the technical system where the cause of the abnormal behavior shall be present. Hence, those core sensors give an important information in order to further analyze the tech- nical system, e.g. by checking the section of the technical system where the core sensors are located. Thus, developing faults or failures in the technical system can be detected in an easy and efficient manner. In a preferred embodiment, the first and the second classifiers are so-called one-sided classifiers. A one-sided classifier is characterized by the feature that this classifier can only be trained by training data referring to the normal operation condition of the technical system. Hence, those clas- sifiers can also be used for complex technical systems where training data referring to an abnormal operation condition are rare. In a particularly preferred embodiment, the one-sided classifiers used in the method of the invention are based on Neural Clouds and/or Gaussian Mixture Models and/or hypercubes .

Those classifiers are well-known in the prior art. Particularly, the method of Neural Clouds is described in document [1] and the method of Gaussian Mixture Models is described in document [2] . In another embodiment of the invention, the second classifiers are applied to partial measurement vectors based on a scheme of several hierarchical levels of first and second classifiers, where the uppermost hierarchical level consists of the first classifier applied to the measurement vector and where a hierarchical level directly below the next higher hierarchical level includes all second classifiers applied to smaller partial measurement vectors comprising exactly one measured signal less than the partial measurement vectors to which the second classifiers of the next higher hierarchical level are applied. According to this hierarchical scheme, the second classifiers are applied subsequently from one hierarchical level to the next lower hierarchical level. In a preferred variant of this embodiment, the method of the invention terminates in case that all second classifiers in a hi- erarchical level output a normal operation condition. The above described hierarchical application of second classifiers enables a fast and efficient detection of core partial measurement vectors and core sensors. In another embodiment of the invention, one or more of the second classifiers and particularly all second classifiers have been trained before processing the measurements of the plurality of sensors. This enables a fast detection of core sensors. Nevertheless, in another embodiment, one or more of the second classifiers and particularly all second classifiers are trained when needed during processing the measurements of the plurality of sensors. Hence, the training of the second classifiers only takes place on demand, i.e. in case that a classification based on the respective second classifier is to be performed in the method of the invention. In another embodiment of the invention, additional (one- dimensional) third classifiers are applied to each single measured signal of at least a part of the measured signals for identifying single core sensors each providing a measured signal classified as an abnormal operation condition by the respective third classifier. Usually, no or only some core sensors found by the second classifiers are also identified by the third classifiers. The uni-variate analysis based on third classifiers may be used in parallel to the above described classification scheme based on second classifiers. Furthermore, in another embodiment, in case that one or more sensors are included in several second classifiers outputting an abnormal behavior, the measured signals of those sensors can be forwarded for further analysis to corresponding third classifiers .

In another embodiment of the invention, the plurality of sensors refers to a subset of sensors installed at the technical system, where the subset of sensors is installed at a predetermined part of the technical system. In this variant of the invention, the method for processing the measurements is preferably performed separately for several subsets of sensors. Hence, in case of complex technical systems, the method of the invention can be implemented efficiently by applying the steps of the method separately to subsets of sensors.

In another variant of the invention, one or more sensors of the plurality of sensors each are sensing means comprising several sub- sensors, where the measured signal of a sensing means is a combination of measured signals of the sub- sensors. Preferably, sensors having similar output form sub- sensors of a sensing means. As a measured signal of the sensing means, the average value of the measured signals of the sub-sensors may be used. In this embodiment, the number of sensors and thus the number of second classifiers is reduced and the method of the invention is less complex. The inventive method can be used for any technical system. In a preferred embodiment, the technical system is a structure, particularly a dam or building or bridge, where the sensors are preferably pore pressure sensors and/or inclination sensors and/or displacement sensors installed at the structure. However, the technical system may also be a turbine, particularly a gas turbine.

Besides the above method, the invention also refers to an apparatus for processing measurements of a plurality of sensors installed at a technical system, where the measurements are represented by one or more measurement vectors for one or more points in time, where a measurement vector at a respective point in time includes a measured signal for each sensor of the plurality of sensors. The apparatus includes a compu- ting means which is arranged to perform the steps of the method of the invention or one or more preferred embodiments of the method of the invention.

The invention also refers to a computer program product di- rectly loadable into the internal memory of a digital computer, comprising software code portions for performing the method of the invention or one or more preferred embodiments of the method of the invention when said product is run on a computer.

Furthermore, the invention refers to a computer program for controlling a computer to perform the method of the invention or one or more preferred embodiments of the method of the invention .

In the following, embodiments of the invention will be described with respect to the accompanying drawings, wherein Fig. 1 is a schematic view of a dam with sensors installed therein where the sensor data can be processed according to an embodiment of the invention;

Fig. 2 is a schematic diagram illustrating the processing of measurements according to an embodiment of the invention; Fig. 3 shows a schematic diagram of another embodiment of the invention.

The method according to the invention can be used for processing sensor data installed at any technical system. E.g., the measurements of sensors installed at a dam can be processed. Fig. 1 is the cross-sectional view of such a dam 1 with four pore pressure sensors SI, S2, S3 and S4 installed at the dam. Those four sensors SI to S4 are just shown as an example and the dam usually includes a much higher number of those sensors which are arranged at different positions along the dam. The dam 1 as shown in Fig. 1 prohibits water W on the left side of the dam having the water level WL1 to pass to the right side of the dam where the water W has the lower water level WL2.

According to the invention as described in the following, the processing of the sensor data can determine an abnormal behavior or operation condition of the dam 1 in combination with an identification of core sensors. The positions of those core sensors correspond to the locations at the dam where the abnormal behavior occurs so that those locations can be checked by service staff and appropriate counter measures can be initiated. In general, the time series of several sensors installed at the corresponding technical system is analyzed by the method of the invention. For each point in time of the time series, a measurement vector including measured values for all sensors or a part of the sensors of the technical system is processed by the method of the invention. In the embodiment shown in Fig. 2, the measurement vector for a corresponding point in time of a technical system with sensors SI, S2, Sn is designated as MV and includes the measured values or signals pi, p2 , pn of the respective sensors SI, S2 , Sn. In the method described herein, the measurement vector is classified by a one-sided classifier CL(pl, p2, pn) which is also designated as CL1 and corresponds to a first classifier in the terminology of the claims. This multi -dimensional classifier has been trained beforehand by appropriate training data of historical measurements. Due to the use of a onesided classifier, only training data referring to the normal behavior of the technical system can be used. As examples of one-sided classifiers, Neural Clouds, Gaussian Mixture Models or hypercubes can be applied to the measurement vector MV. The classifier CL1 and also the classifiers CL2 described in the following usually output a confidence value, e.g. a nor- malized confidence value between zero and one. In case that the confidence value exceeds a certain threshold, the operation state or condition of the technical system is classified as normal. Contrary to that, if the confidence value does not exceed the threshold, the operation state or condition of the technical system is classified as abnormal.

The embodiment of Fig. 2 includes a scheme of several hierarchical levels HO, H(m-l), Hm where the uppermost hierarchical level HO is represented by the classifier CL1. The lower levels succeeding this level HO include classifiers CL2 corresponding to second classifiers in the terminology of the claims. A corresponding hierarchical level includes all classifiers which use as an input a number of measured signals including one measured signal less than the input of the classifiers of the next higher hierarchical level. The lowest hierarchical level Hm includes the second classifiers having two measured values as an input. Besides the hierarchical level HO, Fig. 2 shows the structure of the hierarchical level H(ra-l) and the lowest hierarchical level Hra. For the level H(m-l) three classifiers CL(pl, p2 , p3 ) , CL(pl, p2 , p4) and CL(p(n-2), p(n-l), pn) are illustrated. For the level Hm, three classifiers CL(pl, p2) , CL(pl, p3) and CL(p(n-l), pn) are shown. Both hierarchical levels H(m-l) and Hm include more than those classifiers because there are more combinations of two or three measured signals out of the measured signals of all sensors. As can be seen from Fig. 2, the meas- ured values forming an input of classifiers CL2 are designated as partial measurement vectors MV' . For sake of clarity, only some of the classifiers CL2 and partial measurement vectors MV' are indicated by those reference numerals. In the embodiment of Fig. 2, all second classifiers CL2 have been learned beforehand by using the same training data as for the first classifier CL1. This offline training enables a fast processing of sensor data, e.g. during monitoring of the corresponding technical system. When performing the method of the invention, the second classifiers CL2 in the lower hierarchical levels are applied to the corresponding partial measurement vectors MV' in case that the first classifier CL1 outputs an abnormal operation condition based on the measurement vector MV. By applying those second classifiers, those sensors which are responsible for the abnormal operation condition or behavior can be isolated out of the plurality of sensors as will be described in the following. As indicated by the arrow P in Fig. 2, the lower diagram shows the step of applying the classifiers to the measured signals for a particular example. In this diagram, all classifiers which output an abnormal operation condition are surrounded by a frame F. According to Fig. 2, the second classifiers are applied for one hierarchical level to the next in the direction to lower hierarchical levels, i.e. all classifiers CL2 in one hierarchical level are applied to the partial measurement vectors MV' before the classifiers of the next lower hierarchical level are used. The hierarchical application of the classifiers has the advantage that the method can be stopped in case that all second classifiers in one hierarchical level output a normal operation condition. According to Fig. 2, the classifiers CL(pl, p2 , p3), CL(pl, p2, p4) and CL(p(n-2), p(n-l), pn) in the hierarchical level H(m-l) and the classifier CL(pl, p2) in the hierarchical level Hm output an abnormal operation condition. Contrary to that, the other classifiers in the levels Hm and H(m-l) output a normal operation condition. Furthermore, those classi- fiers in the hierarchical levels between HO and H(m-l) (not shown in Fig. 2) which include all measured signals of one or more classifiers outputting an abnormal operation condition in the hierarchical level H(m-l) also output an abnormal operation condition.

After having applied the second classifiers as shown in the lower diagram of Fig. 2, so-called core partial measurement vectors and so-called core sensors are identified. This is done by determining those partial measurement vectors having the following features:

i) The second classifier applied to the partial measurement vector outputs an abnormal operation condition;

ii) no smaller partial measurement vector exists which includes a subset of the measured signals from the partial measurement vector, where the second classifier applied to the smaller partial measurement vector outputs an abnormal operation condition.

As shown in Fig. 2, for the classifiers CL(pl, p2, p3) and CL(pl, p2 , p4) , a smaller partial measurement vector (pi, p2) exists for which the corresponding classifier CL(pl, p2) leads to an abnormal behavior. Hence, the partial measurement vectors (pi, p2, p3 ) and (pi, p2 , p4 ) do not form core partial measurement vectors. Contrary to that, the partial meas- urement vectors (pi, p2) included in the classifier CL(pl, p2) in the hierarchical level Hm and the partial measurement vector (p(n-2), p(n-l), pn) included in the classifier CL(p(n-2), p(n-l), pn) in the hierarchical level H(m-l) cannot be associated with a smaller partial measurement vector according to the above definition. Hence, those partial measurement vectors form core measurement vectors which are indi- cated by reference numerals MV' ' in Fig. 2. As a consequence, all measured signals included in the core partial measurement vectors MV' ' refer to core sensors. I.e., in the embodiment shown in Fig. 2, sensors SI, S2, S(n-2), S(n-l) and Sn are core sensors which provide signals leading to an abnormal op- eration condition of the technical system. In a next step, those core sensors can be investigated or analyzed further in order to find the causes for the abnormal behavior and initiate corresponding counter measurements, e.g. the repairment of the dam as shown in Fig. 1 or positions of abnormal behav- ior.

In the embodiment as described in Fig. 2, all second classifiers have been trained before performing the method of the invention. However, in an alternative embodiment, it is also possible that the corresponding second classifiers are not trained beforehand but only on demand, i.e. in case that the respective classifier CL2 shall be applied to the partial measurement vector. This embodiment of the invention may be used in case that the response time for identifying core sen- sors is not a critical target.

In case of a large number of sensors, the number of second classifiers used in the method of the invention may be very large. A straightforward calculation leads to the result that (2 ⁿ -l-n) first and second classifiers shall be constructed for n sensors. E.g., for ten sensors, 1013 classifiers need to be considered. To avoid problems with such a high number of classifiers, the technical system may be divided into several sections, each section only including a part of the classifiers where the method of the invention is applied separately to each section. E.g., in case of a dam, sensors in different cross-sections can be identified and processed sep- arately based on the method of the invention. Also in other technical systems, e.g. in a gas turbine, different sections with corresponding sensors can be defined. Furthermore, in another embodiment, the sensor data of several sensors may be combined to one (virtual) sensing device. This will also reduce the number of classifiers. E.g., adjacent sensors usually having similar measured values can be combined to one sensing device providing a measured value which is the average value of the measured signals of the combined sensors.

Fig. 3 shows another embodiment of the method of the invention. In this embodiment, one-dimensional classifiers CL3 are used in addition to the above described multi-dimensional classifiers CLl and CL2. As shown in Fig. 3, the method pro- cesses the corresponding measured signals pi, p2, pn of a plurality of sensors at a corresponding point in time . In block Bl, the method as described with respect to Fig. 2 based on the classifiers CLl and CL2 is performed leading to an output 01 indicating the core sensors. Additionally, the measured signal of each sensors is processed by a separate one-dimensional classifier indicated by CL3. This processing is performed in block B2 of Fig. 3. As an output 02, sensors are identified having a measured value classified as an abnormal operation condition by a classifier CL3. According to block B3, the outputs 01 and 02 are further processed in order to combine the results of the multi-dimensional classifiers with the results of the one-dimensional classifiers.

E.g., the one-dimensional classifiers may be classifiers giving an abnormal operation condition in case that the measured signal lies outside certain limits, thus indicating a malfunction of the corresponding sensor. In such an instance, those core sensors of the output 01 can be identified which do not work properly. Hence, the result 03 of block B3 also indicates which of the core sensors are defect.

The invention as described in the foregoing has several advantages. Particularly, a multi-dimensional classification method not only determines whether a complex technical system shows a normal or an abnormal behavior but also gives information with respect to core sensors. At the locations of those core sensor, the causes of the abnormal operation con- dition of the technical system shall be present.

List of References:

[1] Lang, B, Poppe, T, , Minin, A., Mokhov, I., Kuperin, Y. , Mekler, A., Liapakina, I.: Neural Clouds for Monitoring Complex Systems. Optical Memory and Neural Networks (Information Optics), 2008, Vol. 17. NO. 3, pp. 183-192

[2] D. Reynolds, Gaussian Mixture Models,

http: //www.11.mit . edu/mission/communications/ist/publica tions/0802_Reynolds_Biometrics-GMM . pdf