Method for generating classification model

Abstract

The embodiment of the invention relates to a method for generating a classification model. A method for creating a classification model includes obtaining at least one set of initial timing signals associated with at least one initial acquisition parameter, creating at least one set of simulated timing signals from the at least one set of initial timing signals, creating various test classification models from the set of initial or simulated timing signals, performance of each test classification model is evaluated, at least one set of final timing signals associated with at least one final acquisition parameter is obtained, and the classification model is created.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of French Patent Application No. 2306062, filed on June 14, 2023, which is incorporated herein by reference. Technical Field

[0003] The present invention generally relates to learning data and, in particular embodiments, to methods of generating classification models. Background Art

[0004] Computers can use classification models to classify data into categories or groups based on certain features. There are many types of classification models. Each type of classification model has its own advantages and disadvantages depending on the type of data to be classified.

[0005] Classification models include, among others, decision trees, neural networks, neighbor-based classification models such as k-nearest neighbor (KNN), and support vector machine (SVM) classification models.

[0006] Classification models can be used for time series signals. A time series signal is data that varies over time and contains measurements or observations at each acquisition time.

[0007] The timing signal may correspond to, for example, an audio signal, a video signal, a vibration signal or a physiological signal.

[0008] Specifically, the timing signal can be acquired according to various acquisition parameters, for example, the timing signal can be acquired according to the sampling frequency and the data volume set by the user.

[0009] The sampling frequency corresponds to the number of data samples of the time series signal per unit time. The higher the sampling frequency for the same amount of data, the higher the resolution of the time series signal in the time domain, which corresponds to the value of the sampling frequency divided by the amount of data. Conversely, a lower sampling frequency for the same amount of data may distort the time series signal.

[0010] The data amount corresponds to the data amount of the timing signal acquired during the acquisition period. The data amount is set by indicating the buffer memory size of the acquisition device used to acquire the timing signal. The larger the data amount at the same sampling frequency, the more likely it is to observe more accurate details of the timing signal in the frequency domain.

[0011] In addition, the data volume should be set according to the sampling frequency. In fact, if the sampling frequency is too high and the data volume is too small, the acquisition time may be too short to acquire sufficient information in the timing signal.

[0012] At a given sampling frequency, the larger the amount of data, the longer the acquisition time. This allows more behaviors to be observed in the acquired time series signal, but it also leads to higher energy consumption, longer waiting time for acquiring time series signals, and greater memory usage for storing the acquired data.

[0013] The classification model is trained from the acquired time series signal. The sampling frequency and data volume have an impact on the trained classification model.

[0014] Specifically, a classification model has performance that may vary depending on the sampling frequency and the amount of data used to acquire the time series signal provided to train the classification model.

[0015] The performance of the classification model is evaluated, in particular with respect to the accuracy of the classification model, the period used to acquire the time series signal to be used for the classification model (which affects the energy consumption of the acquisition and the waiting time for acquiring the time series signal) and the amount of the data set used to acquire the time series signal (which affects the occupancy of the memory used to store the acquired data).

[0016] Therefore, it is advantageous to set the sampling frequency and the amount of data so that a classification model with the best performance with respect to the user's needs can be obtained.

[0017] Conventionally, users perform multiple acquisitions at different sampling frequencies and different data amounts in order to test which combination of sampling frequency and data amount makes it possible to obtain a classification model with the best performance.

[0018] Therefore, the user performs a trial-and-error search to determine the combination of sampling frequency and data amount so that a classification model with optimal performance can be obtained.

[0019] This trial and error search requires multiple time series signal acquisitions. These time series signal acquisitions are time-consuming and expensive. In addition, trial and error search cannot ensure that a classification model with optimal performance is obtained.

[0020] Therefore, there is a need to propose a solution that makes it possible to facilitate the search for acquisition parameters for acquiring time series signals for training classification models. Summary of the invention

[0021] According to one embodiment, a computer-implemented method for creating a classification model, referred to as a final classification model, is provided, the method comprising:

[0022] - obtaining at least one set of initial timing signals, the initial timing signals being associated with at least one initial acquisition parameter,

[0023] - creating at least one set of simulated timing signals from at least one set of initial timing signals, each set of simulated timing signals being associated with at least one simulated acquisition parameter different from at least one initial acquisition parameter,

[0024] - creating various test classification models, each test classification model being created from at least one set of initial timing signals or from at least one set of simulated timing signals, each test classification model being associated with at least one acquisition parameter corresponding to at least one initial or simulated acquisition parameter, the at least one initial or simulated acquisition parameter being associated with the initial or simulated timing signals used to create the test classification model,

[0025] - Evaluate the performance of each test classification model, the performance being associated with at least one acquisition parameter of the test classification model, and then

[0026] - obtaining at least one set of final time series signals associated with at least one final acquisition parameter selected from the analysis of the evaluated performance of each tested classification model, and then - creating a final classification model from the at least one set of final time series signals.

[0027] This approach makes it possible to simplify the determination of the acquisition parameters to be applied to acquire the final time series signal used to create the final classification model.

[0028] In practice, the acquisition of the initial time series signal is performed by an acquisition device such as a sensor according to at least one initial acquisition parameter. The performance of the test classification model is then evaluated based on the initial or simulated time series signal, not acquired by the acquisition device.

[0029] Thus, the acquisition parameters to be applied to acquire the final timing signal are determined based on an analysis of the performance of the placement of the acquisition parameters associated with the simulated timing signal.

[0030] Specifically, the acquisition parameters to be applied for acquiring the final timing signal may be selected by analyzing which acquisition parameters make it possible to obtain the best performance with respect to the performance criteria set by the user.

[0031] This makes it possible to avoid performing the acquisition of time series signals using various acquisition parameters in order to determine which acquisition parameters make it possible to create a classification model having the best performance with respect to the performance criteria set by the user.

[0032] By simplifying the search from at least one final acquisition parameter, such a method therefore makes it possible to reduce the overall cost of creating a final classification model, since the final classification model can be obtained more easily and quickly.

[0033] In an advantageous implementation, the at least one initial acquisition parameter comprises a combination of an initial sampling frequency and an initial amount of data from the initial timing signal.

[0034] Advantageously, the initial sampling frequency corresponds to a maximum sampling frequency allowed by an acquisition device used to acquire the initial timing signal.

[0035] In some embodiments, the initial data amount corresponds to a maximum data amount allowed by an acquisition device used to acquire the initial timing signal.

[0036] In an advantageous implementation, at least one analog acquisition parameter of the analog timing signal from each group includes an analog sampling frequency that is less than or equal to the initial sampling frequency and an analog data volume that is less than or equal to the initial data volume.

[0037] Advantageously, each set of initial timing signals is associated with a category indicated during acquisition of the initial timing signals.

[0038] In an advantageous implementation, each group of analog timing signals is associated with a category indicated for the group of initial timing signals from which the group of analog timing signals was created.

[0039] In some embodiments, the method further includes extracting feature values ​​of the initial and simulated timing signals, and each test classification model is created by analyzing the extracted feature values ​​and the categories associated with each set of initial timing signals or simulated timing signals used to create the test classification model.

[0040] In an advantageous implementation, the method further comprises extracting feature values ​​of the final time series signals, each final classification model being created based on an analysis of the extracted feature values ​​and the categories associated with each set of final time series signals.

[0041] In some embodiments, the method further includes indicating the performance of each test classification model relative to at least one acquisition parameter associated with the test classification model.

[0042] Advantageously, the indication of the performance of each classification model comprises displaying on the screen a performance graph comprising the performance of each classification model according to at least one associated acquisition parameter.

[0043] In an advantageous implementation, the evaluation performance of each test classification model includes accuracy, acquisition time of the time series signal, and the amount of acquired data of the time series signal.

[0044] Advantageously, the method further comprises creating a computer program product comprising instructions which, when executed by a computer, cause the computer to implement the final classification model.

[0045] According to another embodiment, a computer program product is proposed, comprising instructions, which, when executed by a computer, cause the computer to perform a method for creating a classification model as described above.

[0046] According to yet another embodiment, a computer system is proposed, comprising: - a memory including a computer program product, the computer program product comprising instructions which, when executed by a computer, cause the computer to implement a method for creating a classification model such as previously described,

[0047] - a processing unit configured to execute the computer program product. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Other advantages and features of the present invention will become apparent upon examination of the detailed description of embodiments and implementations without limitation and the accompanying drawings, in which:

[0049] Figure 1 A computer system configured to implement a method for creating a classification model according to an embodiment is shown;

[0050] Figure 2 A computer-implemented method for creating a classification model for a time series signal according to an embodiment is shown; and

[0051] Figure 3 An example of a performance graph according to an embodiment is shown. DETAILED DESCRIPTION

[0052] Figure 1 An embodiment of a computer system SYS configured to implement a method for creating a classification model is shown.

[0053] The computer system SYS comprises a processing unit UT (eg a processor such as a microprocessor, a microcontroller, a central processing unit, etc.) and a memory MEM. Such a computer system SYS may be, for example, a personal computer or a server.

[0054] The memory MEM comprises a computer program PRG comprising instructions which, when executed by a processing unit UT of the computer system SYS, cause the computer system SYS to implement a method for creating a classification model MDL, as described below.

[0055] The memory MEM is also configured to store time series signals. These time series signals are used as learning data for creating the classification model MDL.

[0056] The timing signals may be provided by a user of the computer system SYS who wishes to generate a classification model from these timing signals.

[0057] The timing signal can be acquired by the sensor and then provided to the computer system SYS. The timing signal can correspond to an audio signal, a vibration signal or a physiological signal, for example. The acquired timing signal includes a set of data acquired over time. The timing signal can include multiple data lines.

[0058] Time series signals are associated with various categories. A category is a set of data (a data group) that has similar characteristics and is grouped together based on these characteristics.

[0059] Therefore, the timing signals are divided into a plurality of groups of timing signals, each group of timing signals being associated with a category indicated by the user.

[0060] For example, one set of timing signals may be associated with a class of signals corresponding to normal operation, while another set of timing signals may be associated with a class of signals corresponding to abnormal operation.

[0061] The processing unit UT is configured to create the classification model MDL by implementing the creation method. Specifically, the processing unit UT is configured to execute the computer program PRG by taking the timing signal as input to create the classification model MDL.

[0062] The timing signal is acquired by an acquisition device (not shown) configured to perform acquisition according to at least one acquisition parameter set by a user (eg, sampling frequency and data volume of the timing signal).

[0063] The sampling frequency corresponds to the number of data samples of the time series signal per unit time. The higher the sampling frequency for the same amount of data, the higher the resolution of the time series signal in the time domain, which corresponds to the value of the sampling frequency divided by the amount of data. Conversely, a lower sampling frequency for the same amount of data may distort the time series signal.

[0064] The data amount corresponds to the number of data of the timing signal acquired during the acquisition. The data amount is set by indicating the buffer memory size of the acquisition device used to acquire the timing signal. The larger the data amount at the same sampling frequency, the more likely it is to observe more accurate details of the timing signal in the frequency domain.

[0065] In addition, the data volume should be set according to the sampling frequency. In fact, if the sampling frequency is too high and the data volume is too small, the acquisition time may be too short to acquire sufficient information in the timing signal.

[0066] At a given sampling frequency, the larger the amount of data, the longer the acquisition time. This allows more behaviors to be observed in the acquired time series signal, but it also leads to higher energy consumption, longer waiting time for acquiring time series signals, and greater memory usage for storing the acquired data.

[0067] The acquisition device enables a user to select a sampling frequency that is less than or equal to a maximum sampling frequency allowed by the acquisition device. The acquisition device enables a user to select a data amount that is less than or equal to a maximum data amount allowed by the acquisition device.

[0068] Figure 2An implementation of a method for creating a classification model for a time series signal implemented by a computer is shown.

[0069] The method aims to help select at least one acquisition parameter for acquiring a time series signal to be used for creating a classification model MDL.

[0070] For example, the method is intended to help select the combination of sampling frequency and data amount to acquire these time series signals.

[0071] The method includes obtaining 20 at least one set of initial timing signals.

[0072] The initial timing signal is acquired at the maximum sampling frequency allowed by the acquisition device. The initial timing signal is also acquired at the maximum data volume allowed by the acquisition device.

[0073] In this way, the initial timing signals are accurate enough to distinguish the various classes associated with them.

[0074] The method then comprises creating 21 at least one set of simulated time series signals from the initial time series signals. Simulated time series signals are thus obtained for the various classes under investigation. The simulated time series signals make it possible to simulate the acquisition of time series signals with acquisition parameters different from the acquisition parameters used to acquire the initial time series signals.

[0075] Specifically, analog timing signals are generated at various sampling frequencies that are less than or equal to the maximum sampling frequency.

[0076] Simulated timing signals are also created using various data amounts that are less than or equal to the maximum data amount.

[0077] More specifically, each analog timing signal uses the data of the initial timing signal. Using a sampling frequency less than the maximum sampling frequency makes it possible to retain only a portion of the data of the timing signal. Using a data amount less than the maximum data amount makes it possible to reduce the amount of data in the analog timing signal relative to the initial timing signal. Each analog timing signal can have the data of the initial timing signal repeated in the analog timing signal. Each analog timing signal can also have multiple lines different from the initial timing signal. Each line includes the same amount of data. In order to evaluate the test classification model described below, it may be important to have a sufficient number of lines for each category. For example, it may be advantageous to use a minimum of 100 lines per category.

[0078] Simulated timing signals are used to evaluate the effects of sampling frequency and data volume on classification accuracy.

[0079] Specifically, the method comprises extracting 22 feature values ​​that are considered for classifying the initial time series signal and the simulated time series signal.

[0080] For example, the characteristics of a timing signal may include amplitude, frequency, duration, minimum value, maximum value, mean value, standard deviation, etc.

[0081] The method then comprises creating 23 a test classification model for each set of time series signals. Each test classification model is created from extracted feature values ​​of the time series signals used to create the classification model.

[0082] In some embodiments, a test classification model is created by using a relatively fast method. For example, a test classification model can be created by using a decision tree, a random forest classifier, a support vector machine (SVM), or a k-nearest neighbor (KNN) method. Therefore, a test classification model can be created without performing cross validation or dividing the initial time series signal into a set of learning signals and a set of test signals.

[0083] Each test classification model is associated with a sampling frequency and a data amount corresponding to the sampling frequency and data amount of the timing signal used to create the test classification model.

[0084] Subsequently, the method includes evaluating 24 the performance of each test classification model. Specifically, evaluating the performance of each test classification model includes determining the accuracy of the test classification model.

[0085] The accuracy of the test classification model is determined by comparing the categories identified by the test classification model with the categories entered by the user during acquisition of the initial timing signal.

[0086] Evaluating the performance of each tested classification model also involves determining when the timing signal is acquired.

[0087] Evaluating the performance of each test classification model may also include calculating a score. The score is then calculated based on the various performance criteria to be evaluated. For example, the performance score of each test classification model may be calculated based on the accuracy of the test classification model, the acquisition time, and the amount of data associated with the test classification model. The performance score may be set, in particular, by weighting the various performance criteria evaluated. The user may also set test constraints, such as a maximum acquisition cycle and / or a maximum amount of data per row.

[0088] The method then includes indicating 25 the performance of each test classification model created with respect to the sampling frequency and the amount of data associated with the test classification model.

[0089] For example, the performance may be displayed on a screen of a computer system SYS. The performance graph may, for example, be created by the computer system and then displayed on the screen.

[0090] An example of such a performance graph is Figure 3The performance graphs show the performance in relation to the accuracy PRC and the acquisition time ACQ_T of the time series signal according to the sampling frequency SPL_R (in Hz) and the amount of data DAT_L (in bits) associated with these models.

[0091] Based on the indicated performance, the user can select acquisition parameters for acquiring the final timing signal.

[0092] Specifically, the user can select the sampling frequency and the amount of data associated with the classification model that has obtained the best performance with respect to the user's needs. In some embodiments, the selection is automatically performed based on the criteria input by the user. Alternatively, some combination of sampling frequency and data amount that obtains the best performance can be suggested to the user.

[0093] The user can then use the acquisition device to acquire the final timing signal groups associated with the various class sets. These final timing signals are acquired by using a combination of sampling frequency and selected data volume.

[0094] The method then comprises obtaining 26 these final timing signals. In particular, these final timing signals are provided as input to a computer program PRG.

[0095] The method then comprises extracting 27 characteristic values ​​of the final time series signal.

[0096] The method then comprises creating 28 a final classification model from these final time series signals.

[0097] Specifically, a final classification model is created from the extracted feature values ​​of the final time series signal used to create the classification model. The classification model can be obtained by a method different from that used to create the test classification model. For example, the final classification model can be created according to a more advanced method than the method used to create the test classification model. This makes it possible to obtain a more accurate final classification model. Specifically, the final classification model can be an artificial neural network.

[0098] The method then comprises creating 29 a computer program comprising instructions which, when executed by a computer, cause the computer to implement the final classification model.

[0099] Such a computer program may then be used by a computer, for example by a microcontroller, to perform classification of time series signals of the same type as those used to train the final classification model, by using the final classification model.

[0100] More specifically, the performance criteria set by the user for analyzing the performance of the test classification model may be selected according to the computer that will execute the final classification model.

[0101] This approach makes it possible to simplify the search for a combination of sampling frequency and amount of data for obtaining the final time series signal for creating the final classification model.

[0102] In practice, the acquisition of the initial time series signal is performed according to a single combination of sampling frequency and data volume. The performance of the test classification model is then evaluated based on simulated time series signals, rather than being acquired by the acquisition device.

[0103] By simplifying the search for a combination of sampling frequency and amount of data for obtaining the final time series signal, this approach may also reduce the overall cost of creating the final classification model, since the final classification model may be obtained more quickly.

[0104] This approach also has the advantage of being able to be fully automated, in which case the final choice of a combination of sampling frequency and data volume is made automatically based on performance criteria entered by the user.

Claims

1. A method for creating a classification model, the method comprising: Acquire at least one set of initial timing signals, wherein the initial timing signals are associated with at least one initial acquisition parameter; creating at least one set of simulated timing signals from the at least one set of initial timing signals, the analog timing signals in each set of analog timing signals being associated with at least one simulated acquisition parameter different from the at least one initial acquisition parameter; creating various test classification models, each test classification model being created from at least one set of initial timing signals or from the at least one set of simulated timing signals, each test classification model being associated with at least one acquisition parameter corresponding to the at least one initial acquisition parameter or the at least one simulated acquisition parameter associated with the initial timing signals or the simulated timing signals used to create the test classification model; evaluating a performance of each test classification model, the performance being associated with the at least one acquired parameter of the test classification model; obtaining at least one set of final timing signals associated with at least one final acquisition parameter selected from the analysis of the evaluated performance of each tested classification model; as well as The classification model is created from the at least one set of final time series signals. 2 . The method of claim 1 , wherein the at least one initial acquisition parameter comprises a combination of an initial sampling frequency and an initial amount of data from an initial timing signal. 3 . The method according to claim 2 , wherein the initial sampling frequency corresponds to a maximum sampling frequency allowed by an acquisition device used to acquire the initial timing signal. 4 . The method according to claim 2 , wherein the initial data amount corresponds to a maximum data amount allowed by an acquisition device for acquiring the initial timing signal.

5. The method according to claim 2, wherein the at least one analog acquisition parameter of the analog timing signal from each group of analog timing signals includes an analog sampling frequency less than or equal to the initial sampling frequency and an analog data amount less than or equal to the initial data amount. The method of claim 1 , wherein each set of initial timing signals is associated with a category indicated during acquisition of the initial timing signals.

7. The method of claim 6, wherein the analog timing signals of each set of analog timing signals are associated with the category indicated for the initial set of timing signals from which the set of analog timing signals was created.

8. The method according to claim 7 further includes extracting characteristic values ​​of the initial timing signal and characteristic values ​​of the simulated timing signal, and creating each test classification model from analysis of the extracted characteristic values ​​and the categories associated with each group of initial timing signals or each group of simulated timing signals used to create the test classification model. 9 . The method according to claim 8 , further comprising extracting feature values ​​of the final time series signals, the classification model being created based on analysis of the extracted feature values ​​and the categories associated with each set of final time series signals.

10. The method of claim 1, further comprising creating a computer program product comprising instructions that, when executed by a computer, cause the computer to implement the classification model.

11. A method for creating a classification model, the method comprising: Acquire at least one set of initial timing signals, wherein the initial timing signals are associated with at least one initial acquisition parameter; creating at least one set of simulated timing signals from the at least one set of initial timing signals, the analog timing signals in each set of analog timing signals being associated with at least one simulated acquisition parameter different from the at least one initial acquisition parameter; creating various test classification models, each test classification model being created from at least one set of initial timing signals or from at least one set of simulated timing signals, each test classification model being associated with at least one acquisition parameter corresponding to at least one initial acquisition parameter or simulated acquisition parameter associated with the initial timing signals or the simulated timing signals used to create the test classification model; evaluating a performance of each test classification model, the performance being associated with at least one acquired parameter of the test classification model; as well as The performance of each test classification model relative to the at least one acquisition parameter associated with the test classification model is indicated.

12. The method of claim 11, wherein the indication of the performance of each classification model comprises displaying a performance graph on a screen, the performance graph comprising the performance of each classification model according to at least one associated acquisition parameter. 13 . The method of claim 11 , wherein the evaluated performance of each test classification model includes accuracy, acquisition time of a time series signal, and an amount of data acquired for the time series signal.

14. A computer system comprising: A memory comprising a computer program, the computer program comprising instructions to: Obtaining at least one set of initial timing signals, wherein the initial timing signals are associated with at least one initial acquisition parameter; creating at least one set of simulated timing signals from the at least one set of initial timing signals, each set of the simulated timing signals being associated with at least one simulated acquisition parameter different from the at least one initial acquisition parameter; creating various test classification models, each test classification model being created from at least one set of initial timing signals or from the at least one set of simulated timing signals, each test classification model being associated with at least one acquisition parameter corresponding to at least one initial acquisition parameter or simulated acquisition parameter associated with the initial timing signals or the simulated timing signals used to create the test classification model; evaluating a performance of each test classification model, the performance being associated with the at least one acquired parameter of the test classification model; obtaining at least one set of final timing signals associated with at least one final acquisition parameter selected from an analysis of the evaluated performance of each tested classification model; as well as creating the classification model from the at least one set of final time series signals; and a processor configured to execute the computer program.

15. The computer system of claim 14, wherein the at least one initial acquisition parameter comprises a combination of an initial sampling frequency and an initial amount of data from an initial timing signal. 16 . The computer system of claim 15 , wherein the initial sampling frequency corresponds to a maximum sampling frequency allowed by an acquisition device for acquiring the initial timing signal. 17 . The computer system according to claim 15 , wherein the initial data amount corresponds to a maximum data amount allowed by an acquisition device for acquiring the initial timing signal.

18. The computer system of claim 15, wherein the at least one analog acquisition parameter of the analog timing signal from each group includes an analog sampling frequency less than or equal to the initial sampling frequency and an analog data amount less than or equal to the initial data amount.

19. The computer system of claim 14, wherein each set of initial timing signals is associated with a category indicated during acquisition of the initial timing signals.

20. The computer system of claim 14, further comprising indicating the performance of each tested classification model by displaying a performance graph on a screen, the performance graph comprising the performance of each classification model according to the at least one associated acquisition parameter, the screen being coupled to the computer system.

Images