Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
METHOD AND APPARATUS FOR COMPRESSING DATA
Document Type and Number:
WIPO Patent Application WO/2019/004858
Kind Code:
A1
Abstract:
The invention relates to a method for compressing data D (nT) (65) that represent a time-dependent signal A(t) (45). The time-dependent signal A(t) (45) comprises a multiplicity of time-dependent signal elements Ai(t) (where i=1,..., NoS) (46-50). The method comprises the steps of: reception of a multiplicity of spectra Si(f) (where i=1,..., NoS) (106-110), wherein each spectrum Si(f) (106-110) corresponds to one of the time-dependent signal elements Ai(t) (46-50), wherein each spectrum Si(f) (106-110) comprises a multiplicity of frequencies fj (where j=1,, NoP) (111-116) and a multiplicity of amplitudes aj,i (where j=1,..., NoP and i=1,..., NoS) (121-126) of the multiplicity of frequencies fj (111-116); generation of a compressed data record (CDS); ascertainment of a respective number of coefficients (ARC) of an autoregressive model for the multiplicity of amplitudes aj,i (121-126) of each of the multiplicity of frequencies fj (111-116); and generation of the compressed data record (CDS), wherein the compressed data record (CDS) comprises at least the number of coefficients (ARC) and the frequencies fj(111-116) associated with the coefficients (ARC).

Inventors:
KOZIONOV ALEXEY PETROVICH (RU)
MANGUTOV OLEG VLADIMIROVICH (RU)
MOKHOV ILYA IGOREVICH (RU)
Application Number:
PCT/RU2017/000475
Publication Date:
January 03, 2019
Filing Date:
June 30, 2017
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
SIEMENS AG (DE)
International Classes:
H03M7/30
Domestic Patent References:
WO2013141741A12013-09-26
Foreign References:
US20110141845A12011-06-16
US20130262049A12013-10-03
CN103645052A2014-03-19
Download PDF:
Claims:
PATENT CLAIMS

1. A method for compressing data D(nT) (65) that represent a time-dependent signal A(t) (45) , wherein the time-dependent signal A(t) (45) comprises a multiplicity of time-dependent signal elements Ai(t) (where i=l, NoS) (46-50), having the steps of:

- reception of a multiplicity of spectra Sj.(f) (where i=l,

NoS) (106-110), wherein each spectrum S±(f) (106-110) corresponds to one of the time-dependent signal elements Ai(t) (46-50), wherein each spectrum S±(f) (106-110) comprises a multiplicity of frequencies f (where j=l, ... , NoP) (111-116) and a multiplicity of amplitudes a-j,i (where j=l, NoP and i=l, NoS) (121-126) of the multiplicity of frequencies fj (111-116) - generation of a compressed data record (CDS) , characterized by the steps of :

- ascertainment of a respective number of coefficients (ARC) of an autoregressive model for - the multiplicity of amplitudes ajri (121-126) of each of the multiplicity of frequencies fj (111-116) ;

- generation of the compressed data record (CDS) , wherein the compressed data record (CDS) comprises at least the number of coefficients (ARC) and the frequencies fj (111- 116) associated with the coefficients (ARC) . 2. The method as claimed in claim 1, in which a timestamp ti (where i=l, NoS) (56-60) is generated for each of the time-dependent signal elements Ai(t) (46-50), wherein the timestamp ti (56-60) represents a time at which the respective time-dependent signal element Ai(t) (46-50) was captured.

3. The method as claimed in claim 1 or 2 , in which the number of coefficients (ARC) is varied on the basis of a classification value, wherein the classification value represents a state of health of a technical system and is ascertained by processing the spectra Si(f) (106-110) .

4. The method as claimed in claim 3, in which a further step involves the classification value being ascertained by comparing the multiplicity of spectra Si(f) (106-110) with at least one threshold value . 5. The method as claimed in claim 3 or 4 , in which a further step involves a first classification value, which represents a first, in particular defective, state of health of the technical system, prompting a first, high number of coefficients (ARC) to be ascertained. 6. The method as claimed in one of claims 3 to 5, in which a further step involves a second classification value, which represents a second, in particular non-defective, state of health of the technical system, prompting a, in comparison with the first number, lower second number of coefficients (ARC) to be ascertained.

7. The method as claimed in claim 5 or 6 , in which the first and/or the second number of coefficients (ARC) of the autoregressive model is determined by the Akaike information criterion (AIC) . 8. The method as claimed in claim 5 or 6, in which the first and/or the second number of coefficients (ARC) of the autoregressive model is determined by the Bayesian information criterion (BIC) .

9. The method as claimed in claim 5 or 6, in which the first and/or the second number of coefficients (ARC) of the autoregressive model is determined such that the first and/or second number of coefficients (ARC) corresponds to the number of frequencies fj (111-116) .

10. The method as claimed in one of the preceding claims, in which, as a further step, the ascertainment of reconstructed data D' (nT) from the compressed data record (CDS) is performed.

11. The method as claimed in claim 10, in which a projected spectrum Si' (f) is ascertained from the number of coefficients (ARC) in the compressed data record (CDS) per timestamp ti (56-60) .

12. The method as claimed in claim 11, in which the projected spectrum Si' (f) is ascertained for at least one of the timestamps ti (56-60) .

13. A computer program product that can be loaded directly into the internal memory of a digital signal processing unit and comprises software code sections that are used to carry out the steps in accordance with one of the preceding claims when the product runs on the signal processing unit. 14. A signal processing unit that is set up to compress data D(nT) (65) in accordance with one of claims 1 to 12, comprising :

- a reception unit (130) for receiving the multiplicity of spectra Si(f) (where i=l, Nos) (106-110); - an autoregressive coefficient determination unit (140) for determining the respective number of coefficients (ARC) of an autoregressive model for the multiplicity of amplitudes a-j,i (121-126) of each of the multiplicity of frequencies fj (111-116); - a memory unit (100) for storing the compressed data record (CDS) .

15. The signal processing unit as claimed in claim 14, further comprising:

- a correlation unit (150) for determining a correlation coefficient between the data D(nT) (65) and the reconstructed data D1 (nT) .

16. The signal processing unit as claimed in claim 14, further comprising:

- a spectrum unit (105) for each of the time-dependent signal elements Ai(t) (46-50) for computing the respective spectrum Si (f) (106-110).

Description:
METHOD AND APPARATUS FOR COMPRESSING DATA

The invention relates to a method for compressing data that represent a time-dependent signal, wherein the time-dependent signal comprises a multiplicity of time-dependent signal elements. The method comprises the steps of reception of a multiplicity of spectra, wherein each spectrum corresponds to one of the time-dependent signal elements and wherein each spectrum comprises a multiplicity of frequencies and a multiplicity of amplitudes of the multiplicity of frequencies. The method further comprises the steps of generation of a compressed data record. The invention further relates to a computer program product and to a signal processing unit that is set up to compress data in accordance with the method according to the invention. The compression of data plays an important part in the management of large volumes of data. For example, large volumes of data accumulate in modern technical systems for monitoring the technical system. Management of the data, part of which takes place at the location of the technical system and another part of which takes place remotely therefrom, comprises storage and transmission of the data. Efficient data management is achievable by intelligent data compression technologies, as a result of which firstly the storage space required for storing the data can be reduced and the bandwidth required for transmitting the data can be reduced.

In the case of a continual data stream that represents a time-dependent signal that has been collected during uninterrupted monitoring of a technical system, e.g. the vibration monitoring of a machine, during operation thereof, the processing and storage of Fourier transformations, i.e. spectra of the time-dependent signal, requires a large amount of memory space. Standard methods of compressing data spectra comprise e.g. universal lossless compression tools (such as e.g. ZIP) or the use of memory-optimized coding. The compression rates, i.e. the ratio of the memory space required for storing the compressed data relative to the memory space for the measurement data, is comparatively low, however. There is therefore a need for improved methods for compressing data .

WO 2013/141741 Al discloses a method for compressing data that represent a time-dependent signal, wherein the time- dependent signal comprises a multiplicity of time-dependent signal elements. This method involves a multiplicity of spectra being received, each spectrum corresponding to one of the time-dependent signal elements. Each spectrum comprises a multiplicity of frequencies and a multiplicity of amplitudes of the multiplicity of frequencies. The multiplicity of amplitudes is normalized for each spectrum in order to obtain normalized amplitudes. The normalized amplitudes are processed, based on a distribution of the normalized amplitudes in view of different threshold values. On that basis, a compressed data record is generated.

It is an object of the invention to specify a method for compressing data and a signal processing unit that further increase the compression of the data.

These objects are achieved by a method in accordance with the features of claim 1 and a signal processing unit in accordance with the features of claim 14. Advantageous configurations are obtained from the dependent patent claims.

The invention proposes a method for compressing data D(nT) that represent a time-dependent signal A(t) . The time- dependent signal A(t) comprises a multiplicity of time- dependent signal elements Ai(t) . First of all, a multiplicity of spectra Si(f) are received, wherein each spectrum Si(f) corresponds to one of the time -dependent signal elements Ai(t) . Each spectrum Si(f) comprises a multiplicity of frequencies f j and a multiplicity of amplitudes a- j ,i of the multiplicity of frequencies f j . Further, a compressed data record is generated. The method is distinguished in that a respective number of coefficients of an autoregressive model for the multiplicity of amplitudes a-^i of each of the multiplicity of frequencies f j is ascertained. Thereafter, the compressed data record is generated, wherein the compressed data record comprises at least the number of coefficients and the frequencies f j associated with the coefficients .

The compressed data record can be used to severely reduce large volumes of data. This involves the memory space required for storing the data being severely reduced. Further, the bandwidth required for transmitting the captured data in the form of the compressed data record is low, which means that further processing of the data contained in the compressed data record can take place at a different location than the capture.

In accordance with one expedient configuration, a timestamp ti is generated for each of the time -dependent signal elements Ai(t) . The timestamp ti represents a time at which the respective time-dependent signal element Ai(t) was captured. The timestamp ti can be used to ascertain the order of occurrence of the time -dependent signal elements Ai(t) . This allows indexing of the multiplicity of amplitudes a j ,i for each spectrum and subsequent processing of said amplitudes for the purpose of ascertaining a respective number of coefficients.

In accordance with a further expedient configuration, the number of coefficients is varied on the basis of a classification value. The classification value represents a state of health of a technical system. The classification value is ascertained by processing the spectra Si(f) . The technical system may be, by way of example, a motor, a turbine and the like. The state of health of the technical system indicates whether the technical system behaves normally or is in an unusual state. An unusual state of this kind could be, by way of example, the occurrence of unusual or disproportionately high vibrations and the like.

In accordance with a further expedient configuration, a further step involves the classification value being ascertained by comparing the multiplicity of spectra Si(f) with at least one threshold value. The at least one threshold value may have been stipulated in advance by experiments or numerical ascertainment. By way of example, one or more threshold values can be used to stipulate whether vibrations in a technical system are in a standard range, in a still tolerable range or in a no longer tolerable range .

In accordance with a further expedient configuration, a further step involves a first classification value, which represents a first, in particular defective, state of health of the technical system, prompting a first, in particular high, number of coefficients to be ascertained. This ensures that the number of coefficients can be used for the most accurate possible reproduction of the spectra ascertained from the time-dependent signal.

In accordance with a further configuration, a further step involves a second classification value, which represents a second, in particular non-defective, state of health of the technical system, prompting a, in comparison with the first number, lower second number of coefficients to be ascertained. As a result, when the state of health of the technical system is "good", it is possible for comparatively coarse reproduction of the multiplicity of spectra using a comparatively small number of classification values to take place, which means that the extent of the data available following performance of the determination of the number of coefficients of a regressive model can be reduced to a maximum degree .

In other words, this means that the compression ratio is adjusted based on the state of health of the technical system. When the technical system is in a defective state, a low compression ratio is chosen, whereas when the technical system behaves normally, the compression ratio is chosen to be high. This ultimately allows - when the compression of the data is considered over a longer period - compression of the data to the maximum possible degree.

In accordance with a further expedient configuration, the first and/or the second number of coefficients of the autoregressive model is determined by the Akaike information criterion (AIC) .

In accordance with one alternative configuration, the first and/or the second number of coefficients of the autoregressive model is determined by the Bayesian information criterion (BIC) .

In accordance with a further alternative configuration, the first and/or the second number of coefficients of the autoregressive model is determined such that the first and/or second number of coefficients corresponds to the number of frequencies fj . The latter can be effected manually by an operator.

In accordance with a further expedient configuration, as a further step, the ascertainment of reconstructed data D 1 (nT) from the compressed data record is performed. In order to be able to assess the Q factor of the compression, there may furthermore be provision for a correlation coefficient between the data D(nT) and the reconstructed D' (nT) to be ascertained. The correlation coefficient can be ascertained during the monitoring of the technical system, this then taking place particularly before the transmission of the compressed data to a remote computation unit. If a reconstruction of the data D(nT) with insufficient consistency is obtained from the reconstructed data D' (nT) , then the compression can be repeated using a larger number of coefficients. Fundamentally, the correlation coefficient can also be used to optimize the compression method in advance, so that an already suitable number of correlation coefficients is determined on the basis of the state of health.

In accordance with a further expedient configuration, a projected spectrum Si 1 (f) is ascertained from the number of coefficients in the compressed data record per timestamp ti. The projected spectrum Si' (f) can then be used to analyze the technical system. On condition that there is a sufficiently good correlation with the spectra ascertained from the data D(nT), it is possible for a qualified appraisal of the state of the technical system to be made.

In accordance with a further expedient configuration, the projected spectrum Si 1 (f) is ascertained for at least one of the timestamps ti. In particular, there may be provision for the projected spectrum Si' (f) to be ascertained for each of the timestamps ti.

The invention further proposes a computer program product that can be loaded directly into the internal memory of a digital signal processing unit and comprises software code sections that are used to carry out the steps of the method described herein when the product runs on the signal processing unit. The computer program product may be realized in the form of a DVD, a CD-ROM, a USB memory stick and the like. However, the computer program product may also be available as a stored program that can be loaded wirelessly or via a wired connection.

A signal processing unit is further proposed that is set up to compress data D(nT) in accordance with the method described herein. The signal processing unit comprises a reception unit, an autoregressive coefficient determination unit and a memory unit. The reception unit is designed to receive the multiplicity of spectra Si(f) . The autoregressive coefficient determination unit is designed to determine a respective number of coefficients of an autoregressive model for the multiplicity of amplitudes a-j,i of each of the multiplicity of frequencies fj . The memory unit is designed to store the compressed data record.

In accordance with one expedient configuration, the signal processing unit further comprises a correlation unit for determining a correlation coefficient between the data D(nT) and the reconstructed data D 1 (nT) .

In accordance with a further expedient configuration, the signal processing unit comprises a spectrum unit for each of the time -dependent signals Ai(t) for computing the respective spectrum Si ( f ) .

The invention for compressing data that represent a time- dependent signal Ai(t) will now be described with reference to the attached drawings . The drawings and the embodiments described below are used to describe an exemplary embodiment, but are not restricted thereto. In the figures below, like elements are provided with like reference symbols. In the drawings :

Fig. 1 shows a schematic depiction of a state monitoring system that comprises a data capture unit and a signal conditioning unit; Fig. 2 shows data that represent a time-dependent signal;

Fig. 3 shows various units of a signal conditioning unit;

Fig. 4 shows a multiplicity of spectra;

Fig. 5 shows an amplitude array; Fig. 6 shows a flowchart that illustrates the various method steps of the method according to the invention;

Fig. 7 shows a flowchart that illustrates a further sub- aspect of the method according to the invention as a flowchart;

Fig. 8 shows the time profile of a time-dependent FFT amplitude of measurement data and compressed data;

Fig. 9 shows a compression ratio that illustrates the compression on the basis of the number of autoregressive coefficients; and

Fig. 10 shows the profile of the R 2 coefficient on the basis of the number of correlation coefficients.

Fig. 1 shows a state monitoring system 10 for monitoring the health of a technical system 20, e.g. a motor. The state monitoring system 10 comprises a data capture unit 30 and a signal processing unit 40. The data capture unit 30 captures a time-dependent signal A(t) 45, which is a continual signal (characterized by the independent time variable "t") that relates, by way of example, to vibrations of the motor 20. The data capture unit 30 provides the signal processing unit 40 with data D(nT) 65. The data D(nT) 65 are a discrete-time (characterized by the independent variable "nT") representation of the time-dependent signal A(t) 45, which is progressively processed in order to compress the data D(nT) 65.

The data capture unit 30 comprises a sensor 31, a signal conditioning unit 32 and an analog-to-digital converter 33 (ADC) . The sensor 31 captures the time-dependent signal A(t) 45 and makes it available to the signal conditioning unit 32 for processing. The signal conditioning unit 32 conditions the time-dependent signal A(t) 45 and makes a conditioned, time-dependent signal A' (t) available to the analog-to- digital converter 33. The analog-to-digital converter 33 digitizes the conditioned time-dependent signal A' (t) , as a result of which the data D(nT) 65 are generated. These are provided to the signal processing unit 40, wherein the data D(nT) 65 represent the time-dependent signal A(t) . The time-dependent signal A(t) 45 can be captured persistently (continually) by the data capture unit 30. The latter can be regarded as a multiplicity of time-dependent signal elements Ai(t) (where i=l, NoS) 46-50 (cf. fig. 2) . The time-dependent signal elements Ai(t) can be regarded as the time-dependent signal A(t) 45 captured during a multiplicity of different time intervals Ati (where i=l, NoS) 51-55 or as a time-dependent signal A(t) 45 provided with windows in a time domain, wherein the window intervals correspond to the time intervals Ati 51-55. Alternatively, the data capture unit 30 can capture the time- dependent signal A(t) 45 during the time intervals Ati 51-55 using a time domain provided with windows, as a result of which the time-dependent signal A(t) 45 can be regarded as a successive capture of the signal elements Ai(t) 46-50. The periods of the time intervals Ati 51-55 may have been defined by a user or be defined in a fixed or variable manner, depending on the type of the time-dependent signal A(t) 45, signal processing requirements, properties of the data capture unit 30 and/or of the signal processing unit 40, etc. The periods may be microseconds, milliseconds or seconds. The time intervals 51-55 are always contiguous. They could, however, also overlap or be separated from one another by particular time spans.

In the present description, the term "signal element" is defined as a section of the time -dependent signal A(t) 45 that has been captured by the data capture unit 30 during respective time intervals Ati 51-55.

In this description, "NoS" is a dimensionless unit and relates to a cardinal number of the signal elements Ai(t) 46- 50 that have been captured during the time intervals t± 51- 55 and in total form the time -dependent signal A(t) 45. In other words, NoS represents the number of signal elements Ai(t) .

A multiplicity of timestamps ti (where i=l, NoS) 56-60 from the time -dependent signal A(t) 45 are determined, wherein each timestamp ti 56-60 represents a starting time for a respective signal element Ai(t) 46-50 that has been captured during the time intervals Ati 51-55. For example, the timestamp ti 56 denotes the starting time for the signal element Ai(t) 46.

The data D(nT) 56 are a discrete-time representation of the time-dependent signal A(t) 45. For this reason, the signal elements Ai(t) 46-50 result in a multiplicity of discrete- time signal elements Dj . (nT) (i=l, NoS; n=l, NoP) 66-70 (see Fig. 2, lower graph) , wherein the data elements Di(nT) 66-70 are consistent with the data D (nT) 65 that are available in the time intervals Ati 51-55. The data elements Di(nT) 66-70 can likewise be obtained by means of discrete- time "domain windowing" of the time-dependent signal A(t) 45. For this reason, the data D(nT) 65 comprise discrete-time representations of the multiplicity of signal elements Ai(t) 51-55, i.e. of the data elements Di (nT) 66-70. Furthermore, the data D(nT) 65 may also be a digital representation of the discrete-time equivalent of the time-dependent signal A(t) 45.

In this description, "NoP" is a dimensionless unit and relates to a cardinal number of the samples that are contained in each discrete-time signal element Di(nT) 66-70. "NoP" is likewise consistent with a cardinal number of the frequencies that is determinable by an NoP-point discrete Fourier transformation (DFT) that has been applied to the data elements Di (nT) 66-70.

"NoP" can be varied for the data elements Di (nT) 66-70 by varying the sampling rate of the analog-to-digital converter 33. "NoP" can also be modified by zero padding, as well as the duration of the time intervals Ati 51-55 being able to be changed on the basis of the effect of a change of "NoP" . A variation of "NoP" accordingly varies the cardinal number of the frequencies that is determinable by an NoP-point DFT.

The timestamps ti 56-60 likewise represent the time of the start of the data elements Di(nT) 66-70. Therefore, the timestamps ti 56-60 comprise information regarding the time of capture of the data elements Di(nT) 66-70, i.e. the starting time for the time intervals t ± 51-55, and are therefore advantageous for the compression of the data D(nT) 65, as well as for the reconstruction of the data D(nT) 65. The timestamps ti 56-60 can also represent periods for which respective signal elements Ai(t) and data elements Di(nT) have been captured.

The signal processing unit 40, which is depicted schematically in fig. 3, comprises a processor 90 and a memory unit 100 for storing compressed data. For the further explanations, NoP = 6 and NoS = 5 are chosen in fig. 4.

The processor 90 comprises a spectrum unit 105, a spectral data reception unit 130, an autoregressive coefficient determination unit 140 and an optional correlation unit 150. The aforementioned units 105, 130, 140 and 150 are designed and set up to compress the data D(nT) .

The spectrum unit 105 receives the data elements Di(nT) 66-70 and computes a multiplicity of spectra Si(f) (where i=l, NoS) 106-110. The spectrum unit 105 computes the spectra Si(f) 106-110 as a multiplicity of NoP-point (discrete Fourier transformations) DFTs . The multiplicity of spectra Si(f) 106-110 comprises a multiplicity of frequencies fj (where j=l, ..·, NoP) 111-116 and a multiplicity of amplitudes a- j ,i (where j=l, NoP; i=l, NoS) 121-126 of the frequencies f j 111-116. The computed amplitudes a-j,i 121-126 are associated with the frequencies j 111-116 and scales of the proportions of the multiplicity of the frequencies f j 111-116 in the signal elements Ai(t) 46-50. The association of the amplitudes aj,i 121-126 with frequencies f j 111-116 is a definition that a person skilled in the art knows for a spectrum of a time-dependent signal, i.e. the amplitudes are a measurement of the proportion of the associated frequency with respect to the subordinate time-dependent signal.

Schematic operation of the spectrum unit 105 is identified by "105" in fig. 4. For each spectrum S ± (f) 106-110, the horizontal axis "f " represents a "frequency" (f j 111-116) . The vertical axis "A" represents an "amplitude", i.e. the amplitudes a j ,i 121-126 of the frequencies f j 111-116 of the underlying time-dependent signal A(t) . The operation of the spectrum unit 105 is explained below. The data elements Di (nT) 66-70 are received by the signal processing unit 40 and buffer- stored as data blocks of a particular length in the memory unit 100. Each data block corresponds to each of the data elements Di(nT) 66-70 that has been captured during each of the time intervals At± 51- 55. Each data block is provided with a respective timestamp ti 56-60 for identifying the starting time of the capture of the data elements Di(nT) 66-70. The data blocks that correspond to the data elements Di (nT) 66-70 are then received and processed for the purpose of computing the spectra Si(f) 106-110.

In the exemplary embodiment described here, the amplitudes a- j,! 121-126 and the frequencies f j 111-116 of the spectra Si(f) 106-110 are depicted as two dimensional amplitude arrays 128 (cf. fig. 5) that comprise rows and columns. By way of example, the array shown in fig. 5 comprises NoP rows, which represent the NoP f j 111-116 frequencies, and NoS columns, which represent the NoS data elements Di(nT) 66-70. Each of the spectra Si(f) 106-110 is computed by respective data elements Di (nT) 66-70. For this reason, "NoS" also represents the number of spectra.

The spectral data reception unit 130 receives the spectra Si(f) 106-110 from the spectrum unit 105. The spectral data reception unit 130 generates a vector of the amplitudes a-^i 121-126 and one of the frequencies f j 111-116 contained in the spectra S±(f) 106-110. This associates a vector a ,i 121- 126 with a respective timestamp ti 65-60.

In the amplitude array 128, each column "i" of the NoS columns is denoted by a respective timestamp ti 65-60 that corresponds to the respective data elements Di(nT) 66-70. Correspondingly, each row "j" of the NoP rows is denoted by a respective frequency fj 111-116. For example, the first column is identified by the timestamp ti 56 and contains all amplitudes a- j, i 121-126 of the frequencies f j 111-116 that have been computed from the spectra S^f) 106. Correspondingly, the first row is identified by fi 111 and contains all amplitudes a j , ± 121-126 of the frequency fi 111 that have been computed from the number of spectra Si(f) 106- 110.

Fig. 5 shows an example of the amplitude array 128 for five spectra Si(f) 106-111 (NoS = 5) and six frequencies fj 111- 116 (NoP = 6) , as is generated by the spectral data reception unit 130. The column values a ltl 121 to a 6 ,i 126 represent, by way of example, the values of the six frequencies f x 111 to f 6 116 that have been determined by the spectrum Si(f) 106 and computed for the data elements Di(nT) 66 with the timestamp tx 56. The column values a 3 ,i 123 to a 3 , 5 123 represent the values of the third frequency f 3 113 that have been determined by the spectrum Si(f) 106 to S 5 (f) 110 and respectively computed for the data elements Ό (nT) 66 to D 5 (nT) 70 and each bear the timestamp ti 56 to t 5 60.

The volume of data accumulating in the process can assume a very large extent . In order to be able to reduce the memory space required for storage, but also the bandwidth required for transmission, there is provision for compression. After the amplitude array 128, as described above, has been ascertained, the method described in fig. 6 is now used to perform the further processing.

In a step 300, the spectra Si(f) 106-110 are ascertained, as described above, so that the multiplicity of amplitudes a j( i 121-126 of each of the multiplicity of frequencies f j 111-116 is available for each timestamp ti 56-60. A step 310 involves a respective number of coefficients of an autoregressive model for the multiplicity of amplitudes a-j,i 121-126 of each of the multiplicity of frequencies fj 111-116 being ascertained for each spectrum Si(f) 106-110. This means that the number of coefficients of the chosen autoregressive model is computed for the column values a. l t l 121 to a 6 ,i 126, that the values of the six frequencies fi 111 to f 6 116 that have been determined for the spectrum Si(f) 106, for the data elements Di(nT) 66 and for the timestamp ti 56. Correspondingly, a corresponding procedure is used for the further spectra S 2 (f) 107 to S 5 (f) 110.

In a step 320, the entirety of the coefficients of the autoregressive model is stored in the memory unit 100 as a compressed data record CDS. Alternatively or additionally, the ascertained coefficients of the autoregressive model for the multiplicity of amplitudes a- j ,i 121-126 of each of the multiplicity of frequencies f j 111-116 of the spectra Si(f) 106-110 (i.e. the compressed data record CDS) can also be transmitted via a communication link, which is not shown in the figures, to a receiver, which is likewise not shown, for storage and further processing.

Since further analysis requires all of the accumulating data (i.e. the autoregressive coefficients ascertained using the method above), it is possible for this, sometimes still substantial, volume of data to be further reduced using the method described in fig. 7. This further reduction is based on a classification value that represents a state of health of the technical system. This is ascertained by processing the spectra Si(f) 106-110 ascertained in step 300. Adaptation of the compression rate, which can be influenced by means of the chosen number of coefficients of the autoregressive model for the multiplicity of amplitudes of each of the multiplicity of frequencies, is then performed on the basis of the classification value. The adaptation method starts with step 400. In step 410, the number of spectra Si(f) 106-110 and the resultant amplitude array 128 are determined, as has been explained above. In step 420, the classification value is determined by comparing the multiplicity of spectra Si(f) 106-110 with a threshold value . The threshold value may have been determined beforehand by experiments or numerical simulation. The threshold value can be used to stipulate whether the technical system has a fault, e.g. excessively strong vibrations of the motor 20, or whether the technical system is in a fault-free state. To this end, it is possible to stipulate a first classification value that represents a first, defective state of health of the technical system, i.e. the motor 20. A second classification value may be associated with a second, fault-free state of health of the technical system, i.e. the motor 20.

In step 430, a check is performed to determine which of the classification values is available. If the first classification value is available, then a defective state of health of the technical system is inferred and the method continues with step 440. If the second classification value has been ascertained, which represents the second, non- defective (i.e. fault-free) state of health of the technical system, then the method continues with step 450. In step 440, which is performed when there is a first, defective state of health of the technical system, a first, high number of coefficients of the autoregressive model for the multiplicity of amplitudes a- j; i 121-126 of each of the multiplicity of frequencies f j 111-116 is ascertained. This attains a comparatively lower compression rate for the data

D(nT) that are to be compressed.

In step 450, which is carried out when there is the second, non-defective state of health of the technical system, a second number of coefficients of the autoregressive model for the multiplicity of amplitudes a- j ^ 121-126 of each of the multiplicity of frequencies f j 111-116 is ascertained, wherein the second number of coefficients is (substantially) lower in comparison with the first number of coefficients. This attains a comparatively higher compression rate.

Although the present exemplary embodiment has been provided merely to distinguish between a defective and a non-defective state of health of the technical system, the definition of a plurality of threshold values allows finer grading of different fault states of the technical system to be performed. The scale of the state of health is then taken as a basis for choosing the number of coefficients of the autoregressive model for the multiplicity of amplitudes a- j ,i 121-126 of each of the multiplicity of frequencies fj 111- 116.

Following the determination of the coefficients of the autoregressive model for the multiplicity of amplitudes a-j (i 121-126 of each of the multiplicity of frequencies fj 111- 116, the method ends in step 460.

The first and/or the second number of coefficients of the autoregressive model can be determined in different ways. To this end, it is possible, by way of example, to use the known Akaike information criterion (AIC) or the known Bayesian information criterion (BIC) . Similarly, manual judgement by an operator is possible, the operator being able to choose the arrangement, i.e. the number of coefficients of the autoregressive model, such that the number corresponds to the number of frequencies f j . In the exemplary embodiment described previously, this would correspond to the number " 6 11 . In order to be able to perform a data analysis and visualization of the compressed data from the coefficients of the autoregressive model (i.e. the compressed data record CDS) , the spectra can be reconstructed from the coefficients for a respective timestamp ti 56-60. To this end, a projected spectrum Si' (f) is ascertained from the number of coefficients in the compressed data record CDS per timestamp ti 56-60. This reconstruction can be used particularly also for denoising the original spectrum Si(f) 106-110. An exemplary embodiment is shown in fig. 8, which illustrates the amplitude A of the ascertained spectrum Si(f) in comparison with the reconstructed spectrum Si' (f) for a timestamp tj . . It can clearly be seen that the reconstructed spectrum Si' (f) corresponds to a projection of the originally ascertained spectrum Si(f) . The projection is all the more accurate the larger the number of coefficients of the autoregressive model that has been chosen for compression.

This approach described here can be used for any spectrum Si(f) for any timestamp ti. Fig. 9 and 10 show the efficiency of the proposed method by way of example. The efficiency of the method is judged by a compression coefficient K and what is known as the "coefficient of determination" R 2 .

The compression coefficient K is the ratio of the initial number of input values to the number of values that need to be stored after compression. It can therefore be computed from the ratio of the size of the input data to the size of the compressed data. This results, on the basis of the number of coefficients ARC of the regressive model, in the profile shown in fig. 9, wherein the compression coefficient K is depicted over the number of coefficients ARC used for the compression. R 2 indicates how well data points fit into a statistical model, such as a line or curve. R 2 is determined as follows:

Here, σ denotes the variance in the model error of a model and Oy denotes the variance in the real spectrum Si(f). Fig. 10 shows the coefficient R 2 over the number of coefficients ARC used for the compression. From this, it can clearly be seen that a significant compression of the spectrum is obtained with just a small number of coefficients ARC of the autoregressive model. The proposed method allows adaptive compression that is based on the actual state of the technical system to be monitored. The compression of the spectra ascertained in a known manner by representation using a number of coefficients of an autoregressive model allows high compression rates. At the same time, the reconstruction of the originally determined spectra from the number of coefficients of the autoregressive model can involve filtering being performed, with the filter strength resulting from the number of coefficients previously used for the compression. The data compressed using the method described can be stored as a compressed data record CDS and/or transmitted to another computation unit for further processing.

List of reference symbols

10 State monitoring system

20 Motor

30 Data capture unit

31 Sensor

32 Signal conditioning unit

33 Analog- to-digital converter

40 Signal processing unit

45 Time -dependent signal A(t)

46-50 Multiplicity of time -dependent signal elements A j (t)

51-55 Multiplicity of time intervals Ati

56-60 Multiplicity of times t j (timestamps)

65 Datum D (nT)

66-70 Multiplicity of discrete time signal elements D j (nT)

90 Processor

100 Memory unit

105 Spectrum unit

106-110 Multiplicity of spectra Si(f) 111-116 Multiplicity of frequencies f j 121-126 Multiplicity of amplitudes a-j,i

128 Amplitude array

130 Spectral data reception unit

140 Autoregressive coefficient determination unit

150 Correlation unit

300, 310, 320, 400, 410, 420, 430, 440, 450, 460 Method step