An intelligent detection method for robustness of anomaly recognition software

Through variational modal decomposition and baseline exchange processing, combined with the recognition results of the abnormal identification software, the robustness of the abnormal identification software is evaluated, and the problem of insufficient accuracy of robustness detection in the existing technology is solved, and more accurate Scenario robustness evaluation.

CN118193395BActive Publication Date: 2024-11-15ZHUNJIAN HEBEI TESTING TECH SERVICE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410384238.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-01
Publication Date
2024-11-15
Estimated Expiration
2044-04-01

AI Technical Summary

Technical Problem

In the prior art, the robustness detection accuracy of abnormal recognition software is poor, and it cannot fully reflect the impact of different scenarios, resulting in the inability to reliable detection results.

Method used

By obtaining the detection timing signals of different scenarios, performing variational modal decomposition, extracting the baseline timing signals and performing baseline exchange processing, combining the software to be detected to perform abnormal identification of the baseline exchange timing signals, and calculating the abnormal detection indicators to evaluate the robustness of the software .

Benefits of technology

It improves the robustness detection accuracy of abnormal identification software in different scenarios, ensuring the effectiveness and reliability of the software in multiple environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118193395B_ABST
    Figure CN118193395B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of software testing, and in particular to an intelligent detection method for robustness of abnormal recognition software, which method obtains detection timing signals in different scenes, uses abnormal recognition software to be detected to perform abnormal recognition on the detection timing signals, and determines each abnormal value; performs variational mode decomposition on the detection timing signals, and determines the baseline timing signals of the detection timing signals in different scenes according to the decomposition results; the baseline timing signals, performs baseline exchange processing on the detection timing signals of different scenes, and obtains the baseline exchange timing signals of the detection timing signals of different scenes; according to the position difference of the abnormal values ​​in the detection timing signals and the baseline exchange timing signals in the same scene, the abnormal detection index of the detection timing signals of different scenes is determined, thereby determining the robustness detection result of the abnormal recognition software to be detected. The present invention effectively improves the accuracy of robustness detection of abnormal recognition software.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an intelligent detection method for robustness of abnormality recognition software. Background Art

[0002] Anomaly Detection software is a system that automatically detects and identifies unexpected or abnormal behaviors by analyzing data patterns and behaviors. When using anomaly detection systems to identify abnormal data, there may be uncertainty in the data in actual applications and may face various complex environments and scenarios, so there will be higher requirements for the robustness of the anomaly detection system. Therefore, in order to ensure the accuracy of abnormal data identification, it is necessary to test the robustness of the anomaly detection software to evaluate the reliability of software detection and ensure the quality and performance of the software.

[0003] In the prior art, when testing the robustness of anomaly recognition systems, usually only the data changes in one type of scenario are targeted, that is, the robustness of the anomaly recognition system is tested based on the anomaly recognition results of the anomaly recognition system in a single scenario. However, since the scenarios cannot be exhaustive, the impact of the scenarios cannot be fully reflected, and different scenarios may be interrelated, the accuracy of the robustness test of the anomaly recognition software cannot be guaranteed. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent detection method for robustness of abnormality recognition software, which is used to solve the problem of poor accuracy of robustness detection of abnormality recognition software in the prior art.

[0005] To solve the above technical problems, the present invention provides an intelligent detection method for robustness of abnormal recognition software, comprising the following steps:

[0006] Acquire detection timing signals in different scenarios, use the abnormality recognition software to be detected to perform abnormality recognition on the detection timing signals of different scenarios respectively, and determine each abnormal value in the detection timing signals of different scenarios;

[0007] Performing variational mode decomposition on the detection time series signals in different scenes, and determining the baseline time series signals of the detection time series signals in different scenes according to the decomposition results;

[0008] Using the baseline timing signals of the detection timing signals of different scenes, performing baseline exchange processing on the detection timing signals of different scenes to obtain the baseline exchange timing signals of the detection timing signals of different scenes;

[0009] Using the abnormality recognition software to be detected to perform abnormality recognition on the baseline exchange timing signals of the detection timing signals of different scenes respectively, and determine each abnormal value in the baseline exchange timing signals of the detection timing signals of different scenes;

[0010] According to the position difference of the outliers in the detection timing signal and the baseline exchange timing signal in the same scenario, the anomaly detection index of the detection timing signal in different scenarios is determined, and according to the anomaly detection index, the robustness detection result of the anomaly recognition software to be detected is determined.

[0011] Furthermore, variational mode decomposition is performed on the detection time series signals in different scenarios, including:

[0012] The detection time series signal in each scene is used as the target time series signal, and at the current decomposition level, the target time series signal is subjected to variational mode decomposition to obtain various component signals;

[0013] Performing time-frequency conversion on each of the component signals, determining a frequency spectrum curve corresponding to each of the component signals, and determining each of the extreme value points on each of the component signals and its frequency spectrum curve, and determining an abnormal probability value corresponding to each of the extreme value points in each of the component signals according to whether the signal corresponding to each of the extreme value points in the target time series signal is an abnormal value;

[0014] Determine the data feature value corresponding to each component signal according to the difference in signal amplitude between the target time series signal and each component signal at each time point, the abnormal probability value corresponding to each extreme point in each component signal, the time point interval between each two adjacent extreme points of the same type in each component signal, and the total number of extreme points in the frequency spectrum curve corresponding to each component signal;

[0015] Determine the instantaneous frequency of each component signal at each time point, and determine the similarity tolerance threshold corresponding to each component signal according to the data feature value corresponding to each component signal, the signal amplitude and instantaneous frequency of each component signal at each time point, and the time point corresponding to each abnormal value in the target time series signal; determine the sample entropy corresponding to each component signal according to each component signal and its corresponding similarity tolerance threshold;

[0016] Obtaining the frequency of occurrence of the minimum sample entropy among the sample entropies corresponding to all component signals obtained by performing variational mode decomposition on the target time series signal at each of the current decomposition number and each of the previous decomposition number that is less than the current decomposition number, and determining the objective function according to the frequency of occurrence of the minimum sample entropy corresponding to each of the current decomposition number and each of the previous decomposition number that is less than the current decomposition number, and the distribution discreteness of the data eigenvalues ​​corresponding to each component signal obtained by performing variational mode decomposition on the target time series signal at the current decomposition number;

[0017] The objective function is solved to determine the current decomposition layer number corresponding to the maximum value of the objective function, so as to obtain the optimal decomposition layer number, and the optimal component signals obtained by performing variational modal decomposition on the target time series signal under the optimal decomposition layer number are determined as decomposition results.

[0018] Further, the data characteristic value corresponding to each of the component signals is determined, and the corresponding calculation formula is:

[0019]

[0020] Among them, γ j represents the data characteristic value corresponding to the jth component signal; h i represents the signal amplitude of the jth component signal at the i-th time point; F i represents the signal amplitude of the target time series signal at the i-th time point; n represents the total number of time points corresponding to the target time series signal; G j represents the total number of extreme value points in the frequency spectrum curve corresponding to the jth component signal; u represents the average abnormal probability value corresponding to the uth pair of adjacent extreme value points of the same type in the jth component signal; t u represents the time interval between the uth pair of adjacent extreme value points of the same type in the jth component signal; m represents the total number of adjacent extreme value points of the same type in the jth component signal; norm[] represents the normalization function.

[0021] Further, the similarity tolerance threshold corresponding to each of the component signals is determined, and the corresponding calculation formula is:

[0022]

[0023] Among them, r j represents the similarity tolerance threshold corresponding to the j-th component signal; γ j represents the data characteristic value corresponding to the jth component signal; i represents the signal amplitude of the jth component signal at the i-th time point; f i represents the instantaneous frequency of the jth component signal at the i-th time point; n represents the total number of corresponding time points of the target time series signal; represents the average signal amplitude of the jth component signal at all time points; || represents the absolute value sign; represents an abnormal point judgment function. When the signal corresponding to the target time series signal at the i-th time point is an abnormal value, otherwise

[0024] Further, determine the objective function, and the corresponding calculation formula is:

[0025]

[0026] Where Q represents the objective function; γ j represents the data eigenvalue corresponding to the jth component signal obtained by performing variational mode decomposition on the target time series signal at the current decomposition level K; represents the average data eigenvalues ​​corresponding to all the component signals obtained by performing variational modal decomposition on the target time series signal at the current decomposition level K; K represents the current decomposition level; num K represents the total number of all component signals obtained by performing variational modal decomposition on the target time series signal at the current decomposition level K; p(S e,min ) represents the minimum sample entropy S among the sample entropies corresponding to all component signals obtained by performing variational mode decomposition on the target time series signal at the decomposition level e. e,min The frequency of occurrence of; K0 represents the minimum number of decomposition levels.

[0027] Furthermore, the method further includes: before solving the objective function, determining the maximum value of the current decomposition level, and the corresponding calculation formula is:

[0028]

[0029] Among them, K max represents the maximum value of the current decomposition layer; K0 represents the set minimum decomposition layer; Y represents the number of abnormal values ​​in the target time series signal; A represents the approximate entropy corresponding to the target time series signal; represents the rounding up symbol; tanh() represents the hyperbolic tangent function.

[0030] Further, determining a baseline timing signal of the detection timing signal includes:

[0031] A decomposed signal corresponding to a minimum frequency value in the decomposition result of the detection timing signal is determined, and the decomposed signal corresponding to the minimum frequency value is used as a baseline timing signal of the detection timing signal.

[0032] Furthermore, the baseline exchange timing signals of the detection timing signals of different scenes are obtained, including:

[0033] Sorting the detection timing signals of all scenes to obtain a detection timing signal sequence;

[0034] The baseline timing signal corresponding to each detection timing signal in the detection timing signal sequence is superimposed on the subsequent detection timing signal, and the baseline timing signal corresponding to the subsequent detection timing signal itself is removed; and the baseline timing signal corresponding to the last detection timing signal in the detection timing signal sequence is superimposed on the first detection timing signal, and the baseline timing signal corresponding to the first detection timing signal itself is removed, thereby obtaining the baseline exchange timing signals of the detection timing signals of different scenes.

[0035] Furthermore, the abnormal detection indicators of the detection time series signals in different scenarios are determined, including:

[0036] According to the position of the abnormal value in the detection timing signal in each scenario, the signal amplitude at the same position in the corresponding detection timing signal is set to a first value, and the signal amplitude at other positions in the corresponding detection timing signal is set to a second value, so as to obtain the abnormal flag sequence corresponding to the detection timing signal in each scenario;

[0037] According to the position of the abnormal value in the baseline exchange timing signal of the detection timing signal in each scenario, the signal amplitude at the same position in the corresponding baseline exchange timing signal is set to a first value, and the signal amplitude at other positions in the corresponding baseline exchange timing signal is set to a second value, so as to obtain the abnormal flag sequence corresponding to the baseline exchange timing signal of the detection timing signal in each scenario;

[0038] The sequence distance between the abnormal flag sequences corresponding to the detection timing signal and the baseline exchange timing signal in the same scenario is calculated, and the normalized value of the sequence distance is determined as the abnormality detection index of the detection timing signal of the corresponding scenario.

[0039] Furthermore, the robustness test results of the anomaly recognition software to be tested are determined, including:

[0040] According to the anomaly detection index of the detection time series signals of different scenarios, determine the average anomaly detection index of the detection time series signals of all scenarios;

[0041] If the average anomaly detection index is less than the set detection index threshold, it is determined that the anomaly recognition software to be detected is robust; otherwise, it is determined that the anomaly recognition software to be detected is not robust.

[0042] The present invention has the following beneficial effects: the present invention performs variational mode decomposition on detection timing signals in different scenes, extracts baseline timing signals affected by scenes from the detection timing signals, and uses the extracted baseline timing signals to perform baseline exchange processing on the detection timing signals of different scenes, thereby associating the detection timing signals affected by different scenes, and obtaining baseline exchange timing signals of the detection timing signals of different scenes. The detection timing signals and baseline exchange timing signals of different scenes are both subjected to abnormal identification using the abnormal identification software to be detected, and each abnormal value is identified. When the abnormal identification software to be detected has good robustness, the positions of the abnormal values ​​identified in the detection timing signals and the baseline exchange timing signals in the same scene should be the same, so according to the position difference of the abnormal values ​​identified in the detection timing signals and the baseline exchange timing signals in the same scene, the abnormal detection index is determined, and finally the robustness detection result of the abnormal identification software to be detected is determined. The present invention extracts the baseline timing signals affected by the scenes from the detection timing signals in different scenes, thereby correlating the detection timing signals in different scenes, and performs combined detection on the detection timing signals in different scenes, thereby effectively improving the accuracy of robustness detection of abnormality recognition software. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 Flow chart of an intelligent detection method for robustness of abnormal recognition software according to an embodiment of the present invention;

[0045] Figure 2 The present invention is a flowchart of performing variational mode decomposition on detection time series signals in different scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. In addition, all parameters or indicators in the formulas involved in this article are normalized values ​​that eliminate the influence of dimensions.

[0048] In order to solve the problem of poor accuracy in robustness detection of abnormality recognition software in the prior art, this embodiment provides an intelligent robustness detection method for abnormality recognition software. The corresponding flowchart of the method is as follows: Figure 1 As shown, the following steps are included:

[0049] Step S1: Acquire detection timing signals in different scenes, use the abnormality recognition software to be detected to perform abnormality recognition on the detection timing signals of different scenes respectively, and determine each abnormal value in the detection timing signals of different scenes.

[0050] For a certain type of data for abnormal identification detection, the abnormal characteristics of the data obtained in different scenarios will be affected by the environment and will be different. To test the robustness of the abnormal identification software, it is necessary to first obtain the detection data in different scenarios. Therefore, for a certain type of detection data, obtain its data in different scenarios, where different scenarios refer to different environmental conditions, and arrange the data in time sequence to obtain the detection timing signals in different scenarios. For example, when it is necessary to perform abnormal detection on humidity data in different scenarios, the different scenarios can be living environments, bathroom environments, etc. At this time, in order to obtain humidity data in different scenarios, the humidity sensor is placed in different environments, and the humidity sensor collects the humidity in the corresponding environment according to the set sampling frequency. The humidity data collected by the humidity sensor at each time point in each environment is arranged in chronological order, thereby obtaining a humidity data sequence, and the humidity data sequence is used as the detection timing signal in the corresponding scene of each environment. In this embodiment, the length of the detection timing signals in all different scenarios is the same.

[0051] After obtaining the detection timing signals in different scenes, the detection timing signals of different scenes are respectively identified by the abnormal recognition software to be detected, and each abnormal value in the detection timing signals of different scenes is determined. In this embodiment, the abnormal recognition software to be detected refers to software that uses the analysis of input data, screens and processes abnormal values ​​in the input data, for example, it can be a Python-based Pyod abnormal detection tool library, which contains traditional algorithms and cutting-edge algorithms, such as integrated abnormal detection and deep learning abnormal detection, and the embodiment of the present invention does not limit it. Since the focus of the embodiment of the present invention is to detect the recognition robustness of the abnormal recognition software, and the recognition process of the detection timing signals of different scenes by the abnormal recognition software belongs to the prior art and is not the focus of the embodiment of the present invention, the recognition process of the detection timing signals of different scenes by the abnormal recognition software is not repeated here. Since the change of the detection data will be limited by the scene, the abnormal recognition software will be affected by different scenes in the process of abnormal detection of the detection data in different scenes, and the abnormal recognition software with good robustness should be able to cope with the changes brought to the data by different scenes.

[0052] Step S2: performing variational mode decomposition on the detection time series signals in different scenes, and determining the baseline time series signals of the detection time series signals in different scenes according to the decomposition results.

[0053] Considering that anomaly recognition software with better robustness means that more ideal detection results can be obtained under the influence of different scenarios, and the impact of scene changes on sensor data signals is mainly reflected through changes in signal baselines, therefore, in order to detect sensor data under different environments and judge the robustness of anomaly detection, it is necessary to decompose the data signal, and then change the scene influence on the detection data under each scenario, and then perform anomaly recognition on the detection data after the scene change, and judge the degree of difference in anomaly recognition under the influence of different scenarios. If the difference is small, it means that the anomaly recognition software has a certain degree of robustness.

[0054] Based on the above analysis, in order to detect the robustness of the anomaly recognition software, it is necessary to separate the baseline data affected by environmental factors in the detection data in different scenarios. Therefore, for the detection time series signals in different scenarios, the signal decomposition algorithm is used to perform variational mode decomposition, and the baseline data is selected. In order to obtain an accurate baseline and avoid problems such as modal aliasing as much as possible, the embodiment of the present invention selects to use the variational mode decomposition algorithm (Variational Mode Decomposition, VMD) to perform variational mode decomposition on the detection time series signal of each scene, thereby decomposing the detection time series signal of each scene into a series of modal components with different frequencies and amplitudes. Considering that when the VMD algorithm is used to perform variational mode decomposition on the detection time series signal of each scene, it is necessary to manually select the decomposition level of the signal, and different decomposition levels have a greater impact on the decomposition result, thereby affecting the accuracy of subsequent baseline selection, so it is necessary to adaptively select the appropriate number of decomposition layers.

[0055] In order to adaptively select the appropriate number of decomposition layers, so as to facilitate the subsequent decomposition of the baseline signal of environmental influence as completely as possible, it is necessary to perform feature analysis on the decomposition results at each decomposition layer in an iterative manner to select the optimal number of decomposition layers. The optimal number of decomposition layers should be able to minimize the mutual interference between the component signals, so as to separate the baseline components representing environmental factors as much as possible.

[0056] Based on this, Figure 2 As shown, the detection time series signal in each scene is used as the target time series signal. Under the current decomposition layer number K, the VMD algorithm is used to perform variational mode decomposition on the target time series signal to obtain various component signals. In the process of using the VMD algorithm to decompose the target time series signal to obtain various component signals, the various component signals are arranged in order of increasing frequency. As the component number increases, the corresponding frequency component also increases, and the high-frequency component usually has lower energy, while the low-frequency component usually has higher energy.

[0057] Use Fourier transform to perform time-frequency conversion on each component signal to obtain the spectrum diagram or spectrum curve corresponding to each component signal, and use the peak detection algorithm to determine the peak value, that is, the extreme value point, in the spectrum curve, and count the total number of extreme value points. At the same time, use the peak detection algorithm to determine the peak value, that is, the extreme value point, in each component signal. Determine whether the signal corresponding to each extreme value point in each component signal in the target time series signal is an abnormal value. If it is an abnormal value, the abnormal probability value corresponding to the corresponding extreme value point in the component signal is set to the first value, otherwise, the abnormal probability value corresponding to the corresponding extreme value point in the component signal is set to the second value. The value ranges of the first value and the second value are both 0-1, and the first value is greater than the second value. In this embodiment, the value of the first value is set to 1 and the value of the second value is set to 0. It should be understood that when using anomaly identification software to identify anomalies of the target timing signal, if the probability value of each signal in the target timing signal being an anomaly can be identified, then the probability value of the signal corresponding to each extreme point in each component signal in the target timing signal being an anomaly can be directly used as the abnormal probability value corresponding to each extreme point in each component signal.

[0058] Furthermore, based on the difference in signal amplitude between the target time series signal and each component signal at each time point, the abnormal probability value corresponding to each extreme point in each component signal, the time point interval between each two adjacent extreme points of the same type in each component signal, and the total number of extreme points in the spectrum curve corresponding to each component signal, a feature analysis is performed on each component signal to determine the data feature value corresponding to each component signal.

[0059] In this embodiment, the data characteristic value corresponding to each component signal is determined, and the corresponding calculation formula is:

[0060]

[0061] Among them, γ j represents the data characteristic value corresponding to the jth component signal; h i represents the signal amplitude of the jth component signal at the i-th time point; F i represents the signal amplitude of the target time series signal at the i-th time point; n represents the total number of time points corresponding to the target time series signal; G j represents the total number of extreme value points in the frequency spectrum curve corresponding to the jth component signal; u represents the average abnormal probability value corresponding to the uth pair of adjacent extreme value points of the same type in the jth component signal; t u represents the time interval between the uth pair of adjacent extreme value points of the same type in the jth component signal; m represents the total number of adjacent extreme value points of the same type in the jth component signal; norm[] represents the normalization function.

[0062] In the above data feature value calculation formula, the possibility that the current component signal is a baseline signal is used to reflect its corresponding signal feature. On the one hand, reflects the energy in the current component signal, Reflects the energy in the initial target timing signal, through To reflect the energy proportion of the current component signal in the initial target time series signal, the larger the proportion, the higher the contribution of the current component signal, the greater the probability that it is the baseline signal, and the smaller the value of the corresponding data characteristic value, that is, the data characteristic value can reflect the possibility that the current component signal is the baseline signal. On the other hand, using Calculate the average time interval between two adjacent extreme points of the same type in the current component signal. At the same time, in order to reduce the interference of abnormal points on the actual characteristics of the signal, use the average abnormal probability of two adjacent extreme points of the same type in the initial target time series signal to reduce the contribution of the time interval corresponding to the abnormal extreme point, and finally accurately obtain the average time series difference between two adjacent extreme points of the same type in the current component signal. j It reflects the number of frequency components contained in the current component signal, and the inverse of the average timing difference is used to calculate the G j Multiplying, on the other hand, reflects the probability that the current component signal is the baseline signal. When the average timing difference is larger and G j When the value of is small, it means that the timing difference of the same type of extreme points in the current component signal is larger, and the number of frequency components in the current component signal is smaller. At this time, it means that the component signal changes relatively evenly and smoothly, which conforms to the characteristics of the baseline signal. The current component signal is more likely to be the baseline signal, and the value of the data characteristic value corresponding to the current component signal is smaller.

[0063] After analyzing the data change characteristics of each component signal and obtaining the data characteristic values ​​of each component signal used to reflect the possibility of the baseline signal of each component signal, in order to facilitate the subsequent adaptive selection of the optimal decomposition level, it is also necessary to determine the sample entropy (SE) of each component signal. The component signal corresponding to the minimum sample entropy is the trend item of the target time series signal. When the value of the decomposition level is appropriate, the sample entropy of the trend item will become smaller, and then as the number of decomposition levels increases, the sample entropy will gradually stabilize. However, when determining the sample entropy of each component signal, it is necessary to select appropriate parameters according to the specific application scenario and signal characteristics.

[0064] In order to determine the sample entropy of each component signal, in an embodiment of the present invention, a Hilbert transform is performed on each component signal, the signal obtained after the transform is used as an analytical signal, and then the phase of the analytical signal is differentiated to obtain the instantaneous frequency of the corresponding component signal at each time point. Since the specific implementation process of performing a Hilbert transform on each component signal to obtain the instantaneous frequency of each component signal at each time point belongs to the prior art, it will not be repeated here.

[0065] When calculating the sample entropy of each component signal, the larger the number of subsequences whose maximum difference between adjacent elements in the component signal does not exceed the similarity tolerance threshold, the more repetitive patterns or self-similarity there are in the data, that is, the stronger the regularity of the data, the lower the complexity. Therefore, for different component signals, in order to express the complexity of the signal sequence, the similarity tolerance threshold should be calculated in combination with the actual performance characteristics of the component signal.

[0066] Based on the above analysis, according to the data characteristic value corresponding to each component signal, the signal amplitude and instantaneous frequency of each component signal at each time point, and each abnormal value in the target time series signal, the actual performance characteristics of each component signal are analyzed to determine the similarity tolerance threshold corresponding to each component signal. The corresponding calculation formula is:

[0067]

[0068] Among them, r j represents the similarity tolerance threshold corresponding to the j-th component signal; γ j represents the data characteristic value corresponding to the jth component signal; i represents the signal amplitude of the jth component signal at the i-th time point; f i represents the instantaneous frequency of the jth component signal at the i-th time point; n represents the total number of corresponding time points of the target time series signal; represents the average signal amplitude of the jth component signal at all time points; || represents the absolute value sign; represents an abnormal point judgment function. When the signal corresponding to the target time series signal at the i-th time point is an abnormal value, otherwise

[0069] In the above calculation formula of similarity tolerance threshold, Represents the standard deviation of the signal amplitude at each time point in the current component signal, which is used as the adjustment base of the similarity tolerance threshold. In order to ensure that the abnormal signal component can be excluded from the baseline component as much as possible during the decomposition process, the proportion of the abnormal value time series points in the target time series signal in the process of calculating the similarity tolerance threshold should be reduced. Therefore, the abnormal point judgment function is constructed As a parameter for adjusting the standard deviation calculation, for a normal signal in the target time series signal, this value is 1, and for an abnormal value in the target time series signal, the instantaneous power x i ×f i The bigger, The smaller the value of , the adjusted standard deviation is obtained, which reduces the possibility that the abnormal baseline signal is affected by abnormal data when calculating the sample entropy. At the same time, the data eigenvalue γ corresponding to the current component signal j The larger the value of is, the smaller the probability of being a baseline signal is, the higher the signal frequency is, and the greater the complexity of the signal sequence is. In order to more completely reflect the repetitive pattern in the signal data, when comparing similar samples in the signal data, the similarity tolerance threshold should be smaller. Therefore, using the unified logical relationship (1-γ j ) is multiplied by the adjusted standard deviation to obtain the similarity tolerance threshold for calculating sample entropy.

[0070] Furthermore, according to each component signal and its corresponding similarity tolerance threshold, the sample entropy corresponding to each component signal is determined.

[0071] Among them, the sample entropy corresponding to each component signal is determined, and the corresponding calculation formula is:

[0072]

[0073] Among them, S j represents the sample entropy corresponding to the jth component signal; B m (r j ) indicates that the maximum difference between adjacent elements in the subsequence of length m in the jth component signal does not exceed r j The number of subsequences of m+1 (r j ) indicates that the maximum difference between adjacent elements in the subsequence of length m+1 in the jth component signal does not exceed r j The number of subsequences; m represents the parameter of sample entropy, and in this embodiment, m=2 is set.

[0074] In the above sample entropy calculation formula, the lower the sample entropy value, the higher the self-similarity of the corresponding component signal, and the larger the sample entropy value, the more complex the corresponding component signal. Since the specific implementation process of calculating the sample entropy corresponding to each component signal according to each component signal and its corresponding similarity tolerance threshold is a prior art, it will not be repeated here.

[0075] Through the above method, it is possible to determine the data eigenvalues ​​and sample entropies corresponding to each component signal obtained by performing variational mode decomposition on the target time series signal under the current decomposition layer number K. Since this process is an iterative signal decomposition process, it is also possible to obtain the sample entropies corresponding to each component signal under each decomposition layer number in each previous decomposition layer number that is less than the current decomposition layer number K. That is, when the minimum decomposition layer number K0 is set, the minimum decomposition layer number K0 can be reasonably set according to the scene corresponding to the corresponding target time series signal. In this embodiment, the minimum decomposition layer number K0=2 is set, and the variational mode decomposition is performed on the target time series signal, so that each decomposition signal can be obtained. According to the above method of determining the sample entropy corresponding to each decomposition signal under the current decomposition layer number K, the sample entropy corresponding to each decomposition signal corresponding to the minimum decomposition layer number K0=2 can be determined. Then, at the next decomposition layer number K0+1=3 of the minimum decomposition layer number K0, the target time series signal is again subjected to variational mode decomposition, so as to obtain the sample entropy corresponding to each decomposition signal corresponding to the next decomposition layer number K0+1=3. By analogy, at the decomposition level K-1, the target time series signal is subjected to variational mode decomposition, thereby obtaining the sample entropy corresponding to each decomposition signal corresponding to the decomposition level K-1; finally, at the current decomposition level K, the target time series signal is subjected to variational mode decomposition, thereby obtaining the data eigenvalues ​​and sample entropy corresponding to each component signal corresponding to the current decomposition level K.

[0076] In order to screen out the optimal current decomposition level K, each decomposition level K0, K0+1, K0+2, ..., K-1 less than the current decomposition level K is called the previous decomposition level of the current decomposition level K. For any one of the current decomposition level and the previous decomposition level less than the current decomposition level, determine the occurrence frequency of the minimum sample entropy in the sample entropies corresponding to all component signals obtained by performing variational mode decomposition on the target time series signal at the decomposition level. Then, according to the occurrence frequency of the minimum sample entropy corresponding to each decomposition level in the current decomposition level K and the previous decomposition level less than the current decomposition level, and the distribution discreteness of the data eigenvalues ​​corresponding to each component signal obtained by performing variational mode decomposition on the target time series signal at the current decomposition level, the objective function is determined, and the corresponding calculation formula is:

[0077]

[0078] Where Q represents the objective function; γ j represents the data eigenvalue corresponding to the jth component signal obtained by performing variational mode decomposition on the target time series signal at the current decomposition level K; represents the average data eigenvalues ​​corresponding to all the component signals obtained by performing variational modal decomposition on the target time series signal at the current decomposition level K; K represents the current decomposition level; num K represents the total number of all component signals obtained by performing variational modal decomposition on the target time series signal at the current decomposition level K; p(S e,min ) represents the minimum sample entropy S among the sample entropies corresponding to all component signals obtained by performing variational mode decomposition on the target time series signal at the decomposition level e. e,min The frequency of occurrence of; K0 represents the minimum number of decomposition levels.

[0079] In the above objective function calculation formula, It represents the variance of the data eigenvalues ​​corresponding to each component signal obtained by performing variational mode decomposition on the target time series signal at the current decomposition level K. Since the component signals obtained by ideal signal decomposition should have relatively independent signal characteristics, the larger the value of the variance, the more discrete the data eigenvalues ​​corresponding to each component signal at the current decomposition level K, which means that the characteristics of each component signal at the current decomposition level are more independent. Considering that the performance characteristics of the component signal alone cannot fully represent the pros and cons of the decomposition level, which may lead to over-decomposition, the change law of sample entropy is used as a feature to assist in judging the pros and cons of the decomposition level. When the current decomposition level K is appropriately selected, the sample entropy of the trend item becomes smaller, and as the decomposition level increases, the sample entropy gradually stabilizes. Therefore, in order to avoid over-decomposition, the information entropy of the sample entropy at the current decomposition level K and each previous decomposition level that is less than the current decomposition level is calculated. When the sample entropy tends to be stable, its information entropy begins to decrease. At this time, over-decomposition may occur, so the information entropy of the sample entropy should also be maximized.

[0080] The above objective function is solved to determine the current decomposition level corresponding to the maximum value of the objective function, and the current decomposition level determined is used as the optimal decomposition level. Under the optimal decomposition level, the VMD algorithm is used to perform variational mode decomposition on the target time series signal to obtain each optimal component signal, and each optimal component signal is determined as the decomposition result.

[0081] Furthermore, in order to reduce the amount of calculation in the process of solving the above objective function while ensuring that a better number of decomposition layers is obtained, it is also necessary to analyze the target time series signal to determine the maximum number of decomposition layers K for signal decomposition of the target time series signal. max At this time, the value range of the decomposition layer number corresponding to the target time series signal is the minimum decomposition layer number K0 ~ the maximum decomposition layer number K maxIn the process of solving the above objective function, it is only necessary to solve it within the value range of the decomposition layer number, thereby effectively reducing the amount of calculation.

[0082] Considering that when decomposing the signal data, in order to ensure the integrity of the segmentation of the baseline signal, it is necessary to consider the outliers existing in the current signal and the complexity of the signal itself. When there are more outliers and the complexity of the signal itself is higher, the corresponding number of decomposition layers should be larger. Therefore, according to each outlier in the target time series signal, the proportion of the number of outliers in the target time series signal is determined, and at the same time, according to the signal amplitude at each time point in the target time series signal, the approximate entropy corresponding to the target time series signal is determined. In the process of determining the approximate entropy corresponding to the target time series signal, this embodiment sets the parameter m=2, the effective threshold r=0.15×σ, and σ is the standard deviation of the signal amplitude at each time point in the target time series signal. Since the process of calculating the approximate entropy corresponding to the target time series signal belongs to the prior art, it will not be repeated here.

[0083] Combined with the number of outliers in the target time series signal and the approximate entropy corresponding to the target time series signal, the maximum number of decomposition layers K for signal decomposition of the target time series signal is determined. max , the corresponding calculation formula is:

[0084]

[0085] Among them, K max represents the maximum value of the current decomposition layer; K0 represents the set minimum decomposition layer; Y represents the number of abnormal values ​​in the target time series signal; A represents the approximate entropy corresponding to the target time series signal; represents the rounding up symbol; tanh() represents the hyperbolic tangent function; μ represents the adjustment factor, and in this embodiment, μ=1 is set.

[0086] In the above calculation formula for the maximum value of the current decomposition layer, the approximate entropy A is used to measure the probability of generating a new pattern in the target time series signal. The more complex the time series signal, the larger the corresponding approximate entropy. The quantity proportion Y reflects the abnormality of the target time series signal. The quantity proportion Y and the approximate entropy A together reflect the complexity and abnormality of the current target time series signal. Therefore, according to the product of the quantity proportion Y and the approximate entropy A, the parameter μ+tanh(Y×A) for adjusting the maximum number of iterations is determined. When the target time series signal contains more abnormal values ​​and the higher the complexity, the current target time series signal should have more layers to be decomposed. At this time, the value of the parameter μ+tanh(Y×A) for adjusting the maximum number of iterations should be larger. Multiply the parameter μ+tanh(Y×A) by the set minimum number of decomposition layers K0 and round it up to get the maximum number of decomposition layers K. max .

[0087] According to the above method, the decomposition result of the detection timing signal of each scene at its corresponding optimal decomposition layer number can be determined, and then the decomposition signal corresponding to the minimum frequency value in the decomposition result can be determined, and the decomposition signal corresponding to the minimum frequency value can be used as the baseline timing signal of the detection timing signal of the corresponding scene. Since the detection timing signal of each scene is decomposed by obtaining a more accurate decomposition layer number, the baseline timing signal obtained by decomposition will be less affected by abnormal values, the components will be more independent, and it will better meet the requirements of baseline decomposition, which is conducive to improving the accuracy of subsequent robustness detection.

[0088] Step S3: using the baseline timing signals of the detection timing signals of different scenes, performing baseline exchange processing on the detection timing signals of different scenes to obtain the baseline exchange timing signals of the detection timing signals of different scenes.

[0089] After determining the baseline timing signals of the detection timing signals of different scenes through the above steps, the detection timing signals of different scenes are subjected to baseline exchange processing, that is: the detection timing signals of all scenes are sorted, and the sorting method is not limited to obtain a detection timing signal sequence; the baseline timing signal corresponding to each detection timing signal in the detection timing signal sequence is superimposed on the subsequent detection timing signal, and the baseline timing signal corresponding to the subsequent detection timing signal itself is removed, that is, the baseline timing signal corresponding to the subsequent detection timing signal itself is replaced with the baseline timing signal corresponding to the previous detection timing signal, to obtain the baseline exchange timing signal corresponding to the subsequent detection timing signal, and the baseline timing signal corresponding to the last detection timing signal in the detection timing signal sequence is superimposed on the first detection timing signal, and the baseline timing signal corresponding to the first detection timing signal itself is removed, that is, the baseline timing signal corresponding to the first detection timing signal itself is replaced with the baseline timing signal corresponding to the last detection timing signal, to obtain the baseline exchange timing signal corresponding to the first detection timing signal, thereby obtaining the baseline exchange timing signals of the detection timing signals of different scenes.

[0090] Step S4: using the abnormality recognition software to be detected to perform abnormality recognition on the baseline exchange timing signals of the detection timing signals of different scenes respectively, and determining each abnormal value in the baseline exchange timing signals of the detection timing signals of different scenes.

[0091] The abnormality recognition software to be detected is used to perform abnormality recognition on the baseline exchange timing signal of the detection timing signal of each scene, so that each abnormal value in the baseline exchange timing signal of the detection timing signal of each scene can be obtained.

[0092] Step S5: Determine the anomaly detection index of the detection timing signal of different scenarios according to the position difference of the anomaly values ​​in the detection timing signal and the baseline exchange timing signal under the same scenario, and determine the robustness detection result of the anomaly recognition software to be detected according to the anomaly detection index.

[0093] For the detection timing signal in each scenario, according to the position of the abnormal value in the detection timing signal, the signal amplitude at the same position in the corresponding detection timing signal is set to the first value, and the signal amplitude at other positions in the corresponding detection timing signal is set to the second value, so as to obtain the abnormal flag sequence corresponding to the detection timing signal in each scenario. Similarly, for the baseline exchange timing signal of the detection timing signal in each scenario, according to the position of the abnormal value in the baseline exchange timing signal, the signal amplitude at the same position in the corresponding baseline exchange timing signal is set to the first value, and the signal amplitude at other positions in the corresponding baseline exchange timing signal is set to the second value, so as to obtain the abnormal flag sequence corresponding to the baseline exchange timing signal of the detection timing signal in each scenario. In this embodiment, the value of the first value is 1, and the value of the second value is 0. It should be understood that the purpose of setting the first value and the second value is to distinguish the types of abnormal values ​​and normal values ​​in the detection timing signal and the baseline exchange timing signal in each scenario. Therefore, under the condition that the values ​​of the first value and the second value are not equal, the specific values ​​of the first value and the second value are not limited.

[0094] Calculate the sequence distance between the abnormal flag sequence corresponding to the detection timing signal and the baseline exchange timing signal in the same scene, and determine the normalized value of the sequence distance as the abnormal detection index of the detection timing signal of the corresponding scene. In this embodiment, the Hamming distance between the abnormal flag sequence corresponding to the detection timing signal and the baseline exchange timing signal in the same scene is calculated, and the Hamming distance is used as the sequence distance between the two abnormal flag sequences, and the sequence distance is normalized by the hyperbolic tangent function, and the normalized value of the sequence distance is determined as the abnormal detection index of the detection timing signal of the corresponding scene. When the value of the abnormal detection index of the detection timing signal of a certain scene is smaller, it means that when the abnormal recognition software to be detected performs abnormal data detection on the detection timing signal and the baseline exchange timing signal in the scene, the detected abnormal results are more similar, and the less affected by the scene change.

[0095] In order to determine the robustness of anomaly detection of the anomaly recognition software to be detected in different scenarios, a detection index threshold is pre-set, and the detection index threshold can be reasonably set as needed. In this embodiment, the detection index threshold is set to be 0.3. The average value of the anomaly detection index of the detection timing signal of all scenarios is determined, so as to obtain the average anomaly detection index of the detection timing signal of all scenarios, and the average anomaly detection index is compared with the set detection index threshold. If the average anomaly detection index is less than the set detection index threshold, it is determined that the anomaly recognition software to be detected is robust, otherwise it is determined that the anomaly recognition software to be detected is not robust.

[0096] It should be noted that the above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An intelligent detection method for robustness of abnormal recognition software, characterized in that: The following steps are involved: Acquire detection timing signals in different scenarios, use the abnormality recognition software to be detected to perform abnormality recognition on the detection timing signals of different scenarios respectively, and determine each abnormal value in the detection timing signals of different scenarios; Performing variational mode decomposition on the detection time series signals in different scenes, and determining the baseline time series signals of the detection time series signals in different scenes according to the decomposition results; Using the baseline timing signals of the detection timing signals of different scenes, performing baseline exchange processing on the detection timing signals of different scenes to obtain the baseline exchange timing signals of the detection timing signals of different scenes; Using the abnormality recognition software to be detected to perform abnormality recognition on the baseline exchange timing signals of the detection timing signals of different scenes respectively, and determine each abnormal value in the baseline exchange timing signals of the detection timing signals of different scenes; Determine anomaly detection indicators of detection timing signals of different scenarios according to the position difference of anomaly values ​​in detection timing signals and baseline exchange timing signals under the same scenario, and determine robustness detection results of anomaly recognition software to be detected according to the anomaly detection indicators; Perform variational mode decomposition on the detection time series signals in different scenarios, including: The detection time series signal in each scene is used as the target time series signal, and at the current decomposition level, the target time series signal is subjected to variational mode decomposition to obtain various component signals; Performing time-frequency conversion on each of the component signals, determining a frequency spectrum curve corresponding to each of the component signals, and determining each of the extreme value points on each of the component signals and its frequency spectrum curve, and determining an abnormal probability value corresponding to each of the extreme value points in each of the component signals according to whether the signal corresponding to each of the extreme value points in the target time series signal is an abnormal value; Determine the data feature value corresponding to each component signal according to the difference in signal amplitude between the target time series signal and each component signal at each time point, the abnormal probability value corresponding to each extreme point in each component signal, the time point interval between each two adjacent extreme points of the same type in each component signal, and the total number of extreme points in the frequency spectrum curve corresponding to each component signal; Determine the instantaneous frequency of each component signal at each time point, and determine the similarity tolerance threshold corresponding to each component signal according to the data feature value corresponding to each component signal, the signal amplitude and instantaneous frequency of each component signal at each time point, and the time point corresponding to each abnormal value in the target time series signal; determine the sample entropy corresponding to each component signal according to each component signal and its corresponding similarity tolerance threshold; Obtaining the frequency of occurrence of the minimum sample entropy among the sample entropies corresponding to all component signals obtained by performing variational mode decomposition on the target time series signal at each of the current decomposition number and each of the previous decomposition number that is less than the current decomposition number, and determining the objective function according to the frequency of occurrence of the minimum sample entropy corresponding to each of the current decomposition number and each of the previous decomposition number that is less than the current decomposition number, and the distribution discreteness of the data eigenvalues ​​corresponding to each component signal obtained by performing variational mode decomposition on the target time series signal at the current decomposition number; The objective function is solved to determine the current decomposition layer number corresponding to the maximum value of the objective function, so as to obtain the optimal decomposition layer number, and the optimal component signals obtained by performing variational modal decomposition on the target time series signal under the optimal decomposition layer number are determined as decomposition results.

2. According to claim 1, a method for intelligent detection of robustness of abnormal recognition software is characterized in that: Determine the data characteristic value corresponding to each of the component signals, and the corresponding calculation formula is: ; in, Indicates data characteristic values ​​corresponding to the component signals; Indicates The component signals are at The signal amplitude at each time point; Indicates that the target timing signal is The signal amplitude at each time point; n Indicates the total number of time points corresponding to the target timing signal; Indicates The total number of extreme value points in the frequency spectrum curve corresponding to each of the component signals; Indicates Of the component signals The average abnormal probability value corresponding to adjacent extreme value points of the same type; Indicates Of the component signals The time interval between adjacent extreme points of the same type; Indicates The total number of pairs of adjacent extreme value points of the same type in the component signals; Represents the normalization function.

3. The method for detecting the robustness of abnormal recognition software according to claim 1, characterized in that: Determine the similarity tolerance threshold corresponding to each of the component signals, and the corresponding calculation formula is: ; in, Indicates A similarity tolerance threshold corresponding to each of the component signals; Indicates data characteristic values ​​corresponding to the component signals; Indicates The component signals are at The signal amplitude at each time point; Indicates The component signals are at The instantaneous frequency at a time point; n Indicates the total number of time points corresponding to the target timing signal; Indicates The average signal amplitude of each of the component signals at all time points; Indicates the absolute value sign; represents an abnormal point judgment function. When the target timing signal is at When the signal corresponding to a time point is an abnormal value, ,otherwise .

4. The method for detecting the robustness of abnormal recognition software according to claim 1, characterized in that: Determine the objective function, and the corresponding calculation formula is: ; in, represents the objective function; Indicates the number of decomposition levels at the current time Next, the first data characteristic values ​​corresponding to the component signals; Indicates the number of decomposition levels at the current time , performing variational modal decomposition on the target time series signal to obtain average data eigenvalues ​​corresponding to all the component signals; Indicates the current decomposition level; Indicates the number of decomposition levels at the current time , the total number of all the component signals obtained by performing variational modal decomposition on the target time series signal; Indicates the number of decomposition levels The minimum sample entropy among the sample entropies corresponding to all component signals obtained by performing variational mode decomposition on the target time series signal is The frequency of occurrence; Indicates setting the minimum number of decomposition levels.

5. The method for intelligent detection of robustness of abnormal identification software according to claim 1, characterized in that: The method further includes: before solving the objective function, determining the maximum value of the current decomposition level, the corresponding calculation formula is: ; in, Indicates the maximum value of the current decomposition level; Indicates setting the minimum number of decomposition levels; Indicates the proportion of abnormal values ​​in the target time series signal; Represents the approximate entropy corresponding to the target time series signal; Indicates the rounding up symbol; represents the hyperbolic tangent function.

6. The method for intelligent detection of robustness of abnormal identification software according to claim 1, characterized in that: Determining a baseline timing signal of the detection timing signal includes: A decomposed signal corresponding to a minimum frequency value in the decomposition result of the detection timing signal is determined, and the decomposed signal corresponding to the minimum frequency value is used as a baseline timing signal of the detection timing signal.

7. The method for intelligent detection of robustness of abnormal identification software according to claim 1, characterized in that: The baseline exchange timing signals of the detection timing signals of different scenes are obtained, including: Sorting the detection timing signals of all scenes to obtain a detection timing signal sequence; The baseline timing signal corresponding to each detection timing signal in the detection timing signal sequence is superimposed on the subsequent detection timing signal, and the baseline timing signal corresponding to the subsequent detection timing signal itself is removed; and the baseline timing signal corresponding to the last detection timing signal in the detection timing signal sequence is superimposed on the first detection timing signal, and the baseline timing signal corresponding to the first detection timing signal itself is removed, thereby obtaining the baseline exchange timing signals of the detection timing signals of different scenes.

8. The method for intelligent detection of robustness of abnormal identification software according to claim 1, characterized in that: Determine the anomaly detection indicators of the detection time series signals in different scenarios, including: According to the position of the abnormal value in the detection timing signal in each scenario, the signal amplitude at the same position in the corresponding detection timing signal is set to a first value, and the signal amplitude at other positions in the corresponding detection timing signal is set to a second value, so as to obtain the abnormal flag sequence corresponding to the detection timing signal in each scenario; According to the position of the abnormal value in the baseline exchange timing signal of the detection timing signal in each scenario, the signal amplitude at the same position in the corresponding baseline exchange timing signal is set to a first value, and the signal amplitude at other positions in the corresponding baseline exchange timing signal is set to a second value, so as to obtain the abnormal flag sequence corresponding to the baseline exchange timing signal of the detection timing signal in each scenario; The sequence distance between the abnormal flag sequences corresponding to the detection timing signal and the baseline exchange timing signal in the same scenario is calculated, and the normalized value of the sequence distance is determined as the abnormality detection index of the detection timing signal of the corresponding scenario.

9. The method for intelligent detection of robustness of abnormal identification software according to claim 1, characterized in that: Determine the robustness test results of the anomaly recognition software to be tested, including: According to the anomaly detection index of the detection time series signals of different scenarios, determine the average anomaly detection index of the detection time series signals of all scenarios; If the average anomaly detection index is less than the set detection index threshold, it is determined that the anomaly recognition software to be detected is robust; otherwise, it is determined that the anomaly recognition software to be detected is not robust.