A circuit breaker hidden danger identification method and device based on feature space differentiation

Abstract

The present invention discloses a method and device for identifying circuit breaker hidden dangers based on feature space differentiation. The method comprises: collecting multi-dimensional time series signal data of the circuit breaker in the first time period before the current moment and performing signal preprocessing, wherein the multi-dimensional time series signal data includes a circuit breaker coil current signal, an acoustic signal, and a vibration signal; obtaining a circuit breaker state feature parameter group based on the multi-dimensional time series signal data of the circuit breaker based on a preset feature extraction method; and inputting a hidden danger identification model based on the circuit breaker state feature parameter group to obtain a circuit breaker hidden danger identification result. The present invention conducts a comprehensive analysis of the circuit breaker state based on the multi-dimensional time series signal data, fully utilizing the synchronization and differences in the performance of different dimensional time series signals when the circuit breaker is in different states, thereby improving the accuracy of circuit breaker hidden danger identification.

Description

Technical Field

[0001] The present invention relates to the field of power grid equipment safety technology, and in particular to a circuit breaker hidden danger identification method and device based on feature space differentiation. Background Art

[0002] Traditional methods for monitoring the condition of power transmission and transformation equipment rely on frequent manual inspections, periodic preventive maintenance, and testing. On-duty personnel regularly inspect equipment during operation, identifying potential anomalies based on visual appearance and indicator instruments to prevent accidents. In recent years, numerous latent problems have been uncovered through condition monitoring of various types of power transmission and transformation equipment put into operation. However, a single condition monitoring method and data alone makes it difficult to fully monitor the development of defects, and the ability to comprehensively analyze and monitor latent defects remains insufficient. Summary of the Invention

[0003] To address the problems of the prior art, the present invention provides a method and device for identifying circuit breaker hidden dangers based on feature space differentiation, which realizes circuit breaker hidden danger identification based on feature space differentiation at different time-frequency scales. The technical solution is as follows:

[0004] In a first aspect, a circuit breaker hidden danger identification method based on feature space differentiation is provided, comprising:

[0005] Collecting multi-dimensional time series signal data of the circuit breaker in a first time period before the current moment and performing signal preprocessing, wherein the multi-dimensional time series signal data includes a circuit breaker coil current signal, an acoustic wave signal, and a vibration signal;

[0006] Obtaining a circuit breaker state feature parameter group based on a preset feature extraction method for the multi-dimensional time series signal data of the circuit breaker;

[0007] Based on the state characteristic parameter group of the circuit breaker, a hidden danger identification model is input to obtain the circuit breaker hidden danger identification result.

[0008] In some embodiments, the training process of the hidden danger identification model includes:

[0009] Acquire multi-dimensional time series signal data of the circuit breaker in normal state and various abnormal states, and perform signal preprocessing;

[0010] Obtaining a state feature parameter group corresponding to the multi-dimensional time series signal data of the circuit breaker based on a preset feature extraction method;

[0011] Based on the state feature parameter group, a deep neural network is trained, and the trained deep neural network is obtained as a hidden danger identification model.

[0012] In some embodiments, the circuit breakers include circuit breakers with different operating mechanisms, and acceleration sensors are arranged at different positions of the circuit breakers with different operating mechanisms; and the circuit breaker abnormalities include various mechanical defects and faults.

[0013] In some embodiments, the characteristic parameter group includes statistical characteristics and information measurement characteristics in the time domain and frequency domain for each dimension of time series signal data;

[0014] The method of obtaining a circuit breaker state characteristic parameter group based on a preset feature extraction method includes:

[0015] For each dimension of time series signal data, multiple time series component signals are obtained based on signal decomposition;

[0016] Acquire A1 characteristic parameters and A2 characteristic parameters respectively based on multiple time series component signals;

[0017] For each time series component signal after signal decomposition, frequency domain conversion is performed to obtain different spectrum signals;

[0018] Based on the spectrum signal, B1 characteristic parameters, B2 characteristic parameters, and B3 characteristic parameters are obtained respectively.

[0019] In some embodiments, the A1 characteristic parameter is: a standard deviation of a plurality of time series component signals;

[0020] The A2 characteristic parameter is: the distance between the standard deviation of the multiple time series component signals of the current dimensional time series signal data and the center of the standard deviation of the multiple time series component signals of the dimensional time series signal data in a normal state;

[0021] The B1 characteristic parameter is: a first frequency proportion value, where the first frequency proportion value is: in the current spectrum signal, when there are k first frequencies among the accumulated n frequencies, the proportion of the first frequency energy in all frequency energies: The first frequency includes multiple frequency values ​​in a preset frequency range, and the first frequency represents the frequency in a normal state; A i is the amplitude of the ith first frequency, B j is the amplitude of the jth frequency in the current spectrum signal;

[0022] The B2 characteristic parameter is: a non-first frequency ratio, wherein the non-first frequency ratio is: a ratio of the frequency content of the non-first frequency to the frequency content of the first frequency in the current spectrum signal.

[0023] The B3 characteristic parameters are: Among them H j is the energy proportion of the jth frequency in the current spectrum signal,

[0024] In some embodiments, the step of inputting a hidden danger identification model based on a state characteristic parameter group of the circuit breaker to obtain a circuit breaker hidden danger identification result includes:

[0025] Input the time series signal data of each dimension into the time series feature extraction module of the hidden danger identification model to obtain the time series features of the time series signal of each dimension. The time series feature extraction module is implemented based on a recurrent neural network;

[0026] Based on the time series characteristics and state characteristic parameter group of each dimension time series signal, the feature fusion analysis module of the hidden danger identification model is input to obtain the fusion characteristics of the multi-dimensional time series signal data of the circuit breaker;

[0027] Based on the fusion characteristics of the multi-dimensional time series signal data of the circuit breaker, the hidden danger classification module of the hidden danger identification model is input to obtain the circuit breaker hidden danger status represented by the multi-dimensional time series signal data of the circuit breaker.

[0028] In some embodiments, the feature fusion analysis module includes a feature weight assignment module and a feature fusion module. The feature fusion module is implemented based on a convolutional neural network. The feature weight assignment module is used to filter according to the importance of features of different dimensions. The weight acquisition method of the feature weight assignment module includes:

[0029] Based on the time series features and state feature parameter groups of each dimension time series signal, an M-dimensional feature vector is formed. For the p-th type of circuit breaker abnormality, Lp samples are sampled, where p = 1, 2, ..., P, and P is the total number of circuit breaker abnormality types. The M-dimensional feature vectors of the Lp samples of the p-th type of circuit breaker abnormality form an Lp*M feature matrix, and P types of circuit breaker abnormalities form P Lp*M feature matrices.

[0030] Based on P Lp*M feature matrices, the first correlation coefficient R1 and the second distinction coefficient R2 of the m-th dimension feature in the M-dimensional feature vector are respectively obtained, and the weight of the m-th dimension feature in the M-dimensional feature vector is determined based on (R2-R1) / (R2+R1), where the first correlation coefficient R1 represents the difference of the m-th dimension feature in the same type of circuit breaker abnormal samples, and the second distinction coefficient R2 represents the difference of the m-th dimension feature in different types of circuit breaker abnormal samples, where m=1, 2, ..., M.

[0031] In some embodiments, the method for obtaining the first correlation coefficient R1 includes:

[0032] The average value of the distribution distance between the elements in the mth column of a single feature matrix: d1, d1 is the average value of the distance between any two elements in the mth column of a single feature matrix;

[0033] The average value of the distribution distance of the elements in the mth column of P feature matrices: d1';

[0034] The center of the m-th column element of a single feature matrix: d2, d2 is the average value of all elements in the m-th column of a single feature matrix;

[0035] The average value d2' of the distance between the elements of the mth column of a single feature matrix and its center d2;

[0036] The average value d2″ of the distance d2′ between the elements of the mth column of P feature matrices and their center d2;

[0037] Determine the first correlation coefficient R1 of the m-th dimension feature of all samples of P-type circuit breaker abnormality based on d1′ and d2″;

[0038] The method for obtaining the second discrimination coefficient R2 includes:

[0039] For the Lp*M feature matrix composed of Lp samples collected for the p-th type of circuit breaker anomaly, convert it into a 1*M feature matrix. The conversion method is to use d2 corresponding to all elements of the m-th column vector of the Lp*M feature matrix to represent the elements of the m-th column of the 1*M feature matrix;

[0040] The average distribution distance between the elements in the mth column of the P transformed feature matrices: D1, D1 is the average distance between the elements d2 in the mth column of any two 1*M feature matrices;

[0041] The center of the m-th column element of the P transformed feature matrices: D2, D2 is the average value of the m-th column element d2 of the P 1*M feature matrices;

[0042] The average value D2' of the m-th column elements of the P transformed feature matrices to their center D2;

[0043] The second discrimination coefficient R2 of the m-th dimension feature of all samples of the P-type circuit breaker abnormality is determined based on D1 and D2'.

[0044] In a second aspect, a circuit breaker hidden danger identification device based on feature space differentiation is provided, comprising:

[0045] A monitoring signal acquisition unit, configured to acquire multi-dimensional time series signal data of the circuit breaker in a first time period before the current moment and perform signal preprocessing, wherein the multi-dimensional time series signal data includes a circuit breaker coil current signal, an acoustic wave signal, and a vibration signal;

[0046] A signal feature extraction unit, configured to obtain a circuit breaker state feature parameter group based on a preset feature extraction method for the multi-dimensional time series signal data of the circuit breaker;

[0047] The circuit breaker hidden danger identification unit is used to input the hidden danger identification model based on the state characteristic parameter group of the circuit breaker and obtain the circuit breaker hidden danger identification result.

[0048] In a third aspect, a computer-readable storage medium is provided, on which computer instructions are stored. When the instructions are executed by a processor, the steps of the circuit breaker hidden danger identification method based on feature space differentiation are implemented.

[0049] The present invention provides a method and device for identifying hidden dangers of circuit breaker based on feature space differentiation, which has the following beneficial effects: the present invention comprehensively analyzes the circuit breaker status based on multi-dimensional time series signal data, fully utilizes the synchronization and differences in the performance of time series signals of different dimensions when the circuit breaker is in different states, and cross-utilizes the outstanding advantages of time series signals of different dimensions in expressing the normal and various abnormal states of the circuit breaker, thereby achieving accurate analysis and judgment of the normal and various abnormal states of the circuit breaker. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a circuit breaker hidden danger identification method based on feature space differentiation provided by an embodiment of the present application;

[0051] Figure 2 Schematic diagram of a method for obtaining a circuit breaker state characteristic parameter group according to an embodiment of the present application;

[0052] Figure 3 This is a schematic diagram of the data processing flow of the hidden danger identification model of the embodiment of the present application on input data;

[0053] Figure 4 This is a structural schematic diagram of a circuit breaker hidden danger identification device based on feature space differentiation provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0055] See also Figure 1 The embodiment of the present application provides a circuit breaker hidden danger identification method based on feature space differentiation, including:

[0056] Step 1: Collect multi-dimensional time series signal data of the circuit breaker in the first time period before the current moment and perform signal preprocessing, wherein the multi-dimensional time series signal data includes a circuit breaker coil current signal, an acoustic wave signal, and a vibration signal;

[0057] Step 2: obtaining a circuit breaker state feature parameter group based on a preset feature extraction method for the multi-dimensional time series signal data of the circuit breaker;

[0058] Step 3: Based on the state characteristic parameter group of the circuit breaker, a hidden danger identification model is input to obtain the circuit breaker hidden danger identification result.

[0059] In the embodiment of the present application, a comprehensive analysis of the circuit breaker status is performed based on multi-dimensional time series signal data, fully utilizing the synchronization and differences in the performance of time series signals of different dimensions when the circuit breaker is in different states, and cross-utilizing the outstanding advantages of time series signals of different dimensions in expressing the normal and various abnormal states of the circuit breaker, thereby achieving accurate analysis and judgment of the normal and various abnormal states of the circuit breaker.

[0060] In one embodiment, in step 3 above, the training process of the hidden danger identification model includes:

[0061] Step 41: Acquire multi-dimensional time series signal data of the circuit breaker in normal state and various abnormal states, and perform signal preprocessing;

[0062] Step 42: obtaining a state feature parameter group corresponding to the multi-dimensional time series signal data of the circuit breaker based on a preset feature extraction method;

[0063] Step 43: Based on the state feature parameter group, a deep neural network is trained, and the trained deep neural network is obtained as a hidden danger identification model.

[0064] In an embodiment of the present application, for training the hidden danger identification model, historical data is first obtained, including multi-dimensional time series signal data of the circuit breaker in normal state and various abnormal states. Based on the historical time series data, the corresponding normal and various abnormal labels of the circuit breaker are obtained, and features are extracted based on the historical time series data. Feature fusion analysis is performed, and finally the hidden danger type is identified through the classification layer of the hidden danger identification model.

[0065] In one embodiment, the circuit breakers include circuit breakers with different operating mechanisms, and acceleration sensors are arranged at different positions of the circuit breakers with different operating mechanisms; and the circuit breaker abnormalities include various mechanical defects and faults.

[0066] In one embodiment, in the above step 2, the feature parameter group includes statistical features and information measurement features in the time domain and frequency domain for each dimension of time series signal data;

[0067] In the above step 2, the circuit breaker state characteristic parameter group is obtained based on the preset feature extraction method, including:

[0068] Step 21: for each dimension of time series signal data, obtain multiple time series component signals based on signal decomposition;

[0069] Step 22, respectively obtaining A1 characteristic parameters and A2 characteristic parameters based on multiple time series component signals;

[0070] Step 23: performing frequency domain conversion on each time series component signal after the signal decomposition to obtain different spectrum signals;

[0071] Step 24: Obtain B1 characteristic parameters, B2 characteristic parameters, and B3 characteristic parameters based on the spectrum signal.

[0072] In the embodiments of the present application, signal decomposition can be implemented based on various time-domain signal decomposition algorithms such as empirical mode decomposition and variational mode decomposition, and frequency domain conversion can be based on existing time-domain-frequency domain conversion algorithms, such as Fourier transform, etc. This embodiment of the present application does not limit this.

[0073] In an embodiment of the present application, for various types of time series signals collected by the circuit breaker, various types of feature parameters are extracted based on various methods including time domain analysis, frequency domain analysis, statistical analysis, information metric analysis, etc., to achieve feature space differentiation of the circuit breaker collected signals, thereby providing a basis for improving the accuracy of subsequent circuit breaker hidden danger identification.

[0074] In one embodiment, the characteristic parameters in steps 22 and 24 are:

[0075] The A1 characteristic parameter is: the standard deviation of multiple time series component signals;

[0076] The A2 characteristic parameter is: the distance between the standard deviation of the multiple time series component signals of the current dimensional time series signal data and the center of the standard deviation of the multiple time series component signals of the dimensional time series signal data in a normal state;

[0077] The B1 characteristic parameter is: a first frequency proportion value, where the first frequency proportion value is: in the current spectrum signal, when there are k first frequencies among the accumulated n frequencies, the proportion of the first frequency energy in all frequency energies: The first frequency includes multiple frequency values ​​in a preset frequency range, and the first frequency represents the frequency in a normal state; A i is the amplitude of the ith first frequency, B j is the amplitude of the jth frequency in the current spectrum signal;

[0078] The B2 characteristic parameter is: a non-first frequency ratio, wherein the non-first frequency ratio is: a ratio of the frequency content of the non-first frequency to the frequency content of the first frequency in the current spectrum signal.

[0079] The B3 characteristic parameters are: Among them H j is the energy proportion of the jth frequency in the current spectrum signal,

[0080] In one embodiment, in step 3 above, based on the state characteristic parameter group of the circuit breaker, inputting the hidden danger identification model to obtain the circuit breaker hidden danger identification result includes:

[0081] Step 31: Input the time series signal data of each dimension into the time series feature extraction module of the hidden danger identification model to obtain the time series features of the time series signal of each dimension. The time series feature extraction module is implemented based on a recurrent neural network.

[0082] Step 32: Based on the time series features and state feature parameter group of each dimensional time series signal, the data are input into the feature fusion analysis module of the hidden danger identification model to obtain the fusion features of the multi-dimensional time series signal data of the circuit breaker;

[0083] Step 33: Based on the fusion features of the multi-dimensional time series signal data of the circuit breaker, the fusion features are input into the hidden danger classification module of the hidden danger identification model to obtain the hidden danger status of the circuit breaker represented by the multi-dimensional time series signal data of the circuit breaker.

[0084] In the embodiment of the present application, the time series feature extraction module can be divided into three paths, which respectively perform time series feature extraction on the three signals of the circuit breaker coil current signal, the acoustic wave signal, and the vibration signal. The time series feature extraction module is implemented based on a recurrent neural network. In one embodiment, it can be implemented based on the GRU network, a variant of the LSTM network. The time series feature extraction module model has few parameters, which effectively improves the training speed of the hidden danger identification model. In addition, in the embodiment of the present application, the hidden danger identification model combines the manually designed feature parameters A1, A2, B1, B2, B3 and the features extracted by the neural network (implemented by the time series feature extraction module) to achieve the effectiveness of feature extraction.

[0085] In one embodiment, in the above step 32, the feature fusion analysis module includes a feature weight assignment module and a feature fusion module. The feature fusion module is implemented based on a convolutional neural network. The feature weight assignment module is used to filter according to the importance of features of different dimensions. The weight acquisition method of the feature weight assignment module includes:

[0086] Step 321: Based on the time series features and state feature parameter groups of each dimensional time series signal, an M-dimensional feature vector is formed. Lp samples are sampled for the p-th type of circuit breaker abnormality, where p = 1, 2, ..., P, and P is the total number of circuit breaker abnormality types. The M-dimensional feature vectors of the Lp samples of the p-th type of circuit breaker abnormality form an Lp*M feature matrix, and P types of circuit breaker abnormalities form P Lp*M feature matrices.

[0087] Step 322: Based on P Lp*M feature matrices, a first correlation coefficient R1 and a second discrimination coefficient R2 of the m-th dimension feature in the M-dimensional feature vector are respectively obtained, and the weight of the m-th dimension feature in the M-dimensional feature vector is determined based on (R2-R1) / (R2+R1), where the first correlation coefficient R1 represents the difference of the m-th dimension feature in the same type of circuit breaker abnormality samples, and the second discrimination coefficient R2 represents the difference of the m-th dimension feature in different types of circuit breaker abnormality samples, where m=1, 2, ..., M.

[0088] In an embodiment of the present application, the temporal feature extraction module of the hidden danger identification model is implemented based on a recurrent neural network, and the feature fusion module is implemented based on a convolutional neural network. The recurrent neural network is combined to mine the temporal features of the data, and the convolutional neural network is combined to mine the spatial features, thereby realizing deep feature extraction of multi-dimensional features, effectively improving the richness of the extracted features. At the same time, in this application, the feature weight assignment module is added to the hidden danger identification model for the various feature parameters A1, A2, B1, B2, and B3 designed manually in the early stage, and the feature parameters are screened and filtered according to the importance of features of different dimensions, thereby effectively reducing redundant features, avoiding interference from irrelevant features, and avoiding the invalidity and interference of manually designed features, thereby improving the hidden danger identification accuracy of the hidden danger identification model.

[0089] At the same time, in an embodiment of the present application, the feature weight assignment module analyzes the importance of each dimensional feature based on all collected samples, and analyzes the differences of the m-th dimensional feature in the same type of circuit breaker abnormal samples and the differences of the m-th dimensional feature in different types of circuit breaker abnormal samples through the first correlation coefficient R1 and the second distinction coefficient R2, thereby improving the accuracy of the feature weight assignment module in analyzing the importance of features of different dimensions.

[0090] Furthermore, the training process of the hidden danger identification model in the embodiment of the present application includes:

[0091] Step 51: Preprocess the collected multi-dimensional time series signal data for the circuit breaker in normal state and various abnormal states and obtain the characteristic parameters A1, A2, B1, B2, and B3 according to the above.

[0092] Step 52: The feature weight assignment module assigns weights to the feature parameters A1, A2, B1, B2, and B3 based on all sample data of the multi-dimensional time series signal data under normal and various abnormal states of the circuit breaker using the above steps 321-322 to obtain feature weight assignment results for different feature parameters;

[0093] Step 53: The multi-dimensional time series signal data collected in step 1 is input into a time series feature extraction module to obtain the time series features of each dimensional time series signal;

[0094] Step 54: input the weight-adjusted state feature parameter group obtained in step 52 and the time series feature obtained in step 53 into a feature fusion module for convolution processing;

[0095] Step 55: input the convolution output result of step 54 into the hidden danger classification module of the hidden danger identification model to obtain the hidden danger identification result;

[0096] Step 56: Based on the hidden danger identification result of step 55 and the hidden danger type labeling data corresponding to the preset collected sample data, the model parameters of the temporal feature extraction module, the feature fusion module, and the hidden danger classification module in the hidden danger identification model are calculated and corrected through the loss function;

[0097] Step 57 : Based on the preset number of iterations of steps 51 to 56 , a trained hidden danger identification model is obtained.

[0098] In one embodiment, the method for obtaining the first correlation coefficient R1 in step 322 includes:

[0099] The average value of the distribution distance between the elements in the mth column of a single feature matrix: d1, d1 is the average value of the distance between any two elements in the mth column of a single feature matrix;

[0100] The average value of the distribution distance of the elements in the mth column of P feature matrices: d1';

[0101] The center of the m-th column element of a single feature matrix: d2, d2 is the average value of all elements in the m-th column of a single feature matrix;

[0102] The average value d2' of the distance between the elements of the mth column of a single feature matrix and its center d2;

[0103] The average value d2″ of the distance d2′ between the elements of the mth column of P feature matrices and their center d2;

[0104] Determine the first correlation coefficient R1 of the m-th dimension feature of all samples of P-type circuit breaker abnormality based on d1′ and d2″;

[0105] In one embodiment, the method for obtaining the second discrimination coefficient R2 in step 322 includes:

[0106] For the Lp*M feature matrix composed of Lp samples collected for the p-th type of circuit breaker anomaly, convert it into a 1*M feature matrix. The conversion method is to use d2 corresponding to all elements of the m-th column vector of the Lp*M feature matrix to represent the elements of the m-th column of the 1*M feature matrix;

[0107] The average distribution distance between the elements in the mth column of the P transformed feature matrices: D1, D1 is the average distance between the elements d2 in the mth column of any two 1*M feature matrices;

[0108] The center of the m-th column element of the P transformed feature matrices: D2, D2 is the average value of the m-th column element d2 of the P 1*M feature matrices;

[0109] The average value D2' of the m-th column elements of the P transformed feature matrices to their center D2;

[0110] The second discrimination coefficient R2 of the m-th dimension feature of all samples of the P-type circuit breaker abnormality is determined based on D1 and D2'.

[0111] It can be understood that for the m-th dimension feature in the M-dimensional feature vector, it is necessary to measure the similarity and difference of the dimension feature in the P-type circuit breaker abnormality samples. If the difference of the dimension feature in the same type of circuit breaker abnormality samples is small and the difference in different types of circuit breaker abnormalities is large, it means that the dimension feature has an important contribution to identifying different circuit breaker abnormalities. If the difference of the dimension feature in the same type of circuit breaker abnormality samples is large, then consider assigning a smaller weight to the dimension feature. Even if the difference of the dimension feature in the same type of circuit breaker abnormality samples is too large, greater than the difference of the dimension feature in different types of circuit breaker abnormalities, then consider assigning 0 to the dimension feature, that is, deleting the dimension feature.

[0112] The difference of the m-th dimension feature (m=1, 2, ..., M) in the M-dimensional feature vector in the same type of circuit breaker abnormal samples is referred to as the "first correlation coefficient R1" in this application. The method for obtaining the first correlation coefficient R1 includes:

[0113] The average value of the distribution spacing of the elements in the mth column of a single feature matrix: d1, d1 is the average value of the distance between any two elements in the mth column of a single feature matrix, that is, the average value of the distance between any two elements in the mth column of a single feature matrix;

[0114] The average value of the distribution distance of the elements in the mth column of P feature matrices is: d1',

[0115] The center of the m-th column element of a single feature matrix: d2, d2 is the average value of all elements in the m-th column of a single feature matrix;

[0116] The average value d2' of the distance between each element in the mth column of a single feature matrix and its center d2;

[0117] The average value d2″ of d2′ corresponding to the m-th column element of P feature matrices;

[0118] Determine the first correlation coefficient R1 of the m-th dimension feature of all samples of P-type circuit breaker abnormality based on d1′ and d2″;

[0119] It can be understood that d1 represents the difference in the performance of the mth dimension feature of the same type of circuit breaker anomaly in different samples, which is a representation of intra-class difference. d1' is the average of P d1s, which represents an average of intra-class difference.

[0120] d2' represents the difference between the performance of the mth dimension feature of the same type of circuit breaker anomaly in different samples and its mean, which is another representation of intra-class variability. d2″ is the average of P d2's, which represents another average of intra-class variability.

[0121] The first correlation coefficient R1 combines two different intra-class differences to represent the difference of the m-th dimension feature in the same type of circuit breaker abnormality samples. It can be understood that the smaller the first correlation coefficient R1, the greater the contribution to identifying different circuit breaker abnormalities.

[0122] The difference of the m-th dimension feature (m=1, 2, ..., M) in the M-dimensional feature vector in different types of circuit breaker abnormality samples is referred to as the "second discrimination coefficient R2" in this application. The method for obtaining the second discrimination coefficient R2 includes:

[0123] For the Lp*M feature matrix composed of Lp samples collected for the p-th type of circuit breaker anomaly, convert it into a 1*M feature matrix. The conversion method is to use d2 corresponding to all elements of the m-th column vector of the Lp*M feature matrix to represent the elements of the m-th column of the 1*M feature matrix. All samples collected for the P-type circuit breaker anomaly form P 1*M feature matrices.

[0124] The average distribution distance between the elements in the mth column of the P transformed feature matrices (i.e., P 1*M feature matrices): D1, where D1 is the average distance between the elements in the mth column d2 of any two 1*M feature matrices, that is, the average distance between the elements in the mth column of each of the P 1*M feature matrices;

[0125] The center of the m-th column element of the P transformed feature matrices: D2, D2 is the average value of the m-th column element d2 of the P 1*M feature matrices;

[0126] The average value D2' of each element of the mth column of the P transformed feature matrices to their center D2;

[0127] Determine the second discrimination coefficient R2 of the m-th dimension feature of all samples of P-type circuit breaker abnormality based on D1 and D2';

[0128] It can be understood that each 1*M feature matrix represents the M-dimensional features corresponding to one type of circuit breaker anomaly. D1 represents the performance difference of the m-th dimension features in the features of different types of circuit breaker anomalies, which is a representation of inter-class differences.

[0129] D2' represents the difference between the performance of the m-th dimension feature and its mean in the features of different types of circuit breaker anomalies, which is another representation of inter-class difference.

[0130] The second discrimination coefficient R2 combines two different inter-class differences to represent the differences of the m-th dimension feature in different types of circuit breaker abnormality samples. It can be understood that the larger the second discrimination coefficient R2, the greater the contribution to identifying different circuit breaker abnormalities.

[0131] In one embodiment, the method for obtaining the first correlation coefficient R1 and the second discrimination coefficient R2 includes:

[0132] R1 = λ1(d1' and d2"); R2 = λ2(D1 and D2'), where λ1 and λ2 are weight coefficients;

[0133] The mth column element data of a feature matrix is ​​represented by a circle with d2 as the center and d1 as the radius, represented by |d2 p1 -d2 p2 |-d1 p1 -d1 p2 Representing the distance between circles p1 and p2, obtain the sum D3 of the distances between the two circles corresponding to the mth column of any two feature matrices in the P feature matrices, that is, the sum of the distances between the two circles corresponding to the mth column of each pair of feature matrices in the P feature matrices. When the sum of the distances is greater than 0, assign a number greater than λ1 to λ2 according to the size of the sum of the distances. In a preferred embodiment, λ1 = 1, λ2 = e D3 .

[0134] An embodiment of the present application provides a circuit breaker hidden danger identification device based on feature space differentiation, comprising:

[0135] A monitoring signal acquisition unit, configured to acquire multi-dimensional time series signal data of the circuit breaker in a first time period before the current moment and perform signal preprocessing, wherein the multi-dimensional time series signal data includes a circuit breaker coil current signal, an acoustic wave signal, and a vibration signal;

[0136] A signal feature extraction unit, configured to obtain a circuit breaker state feature parameter group based on a preset feature extraction method for the multi-dimensional time series signal data of the circuit breaker;

[0137] The circuit breaker hidden danger identification unit is used to input the hidden danger identification model based on the state characteristic parameter group of the circuit breaker and obtain the circuit breaker hidden danger identification result.

[0138] In some embodiments, the circuit breaker hidden danger identification device based on feature space differentiation provided by the embodiments of the present invention can be implemented using a combination of software and hardware. As an example, the circuit breaker hidden danger identification device based on feature space differentiation provided by the embodiments of the present invention can be directly embodied as a combination of software modules executed by a processor. The software module can be located in a storage medium, and the storage medium is located in a memory. The processor reads the executable instructions included in the software module in the memory and, in combination with the necessary hardware, completes the circuit breaker hidden danger identification method based on feature space differentiation provided by the embodiments of the present invention. The specific definition of the circuit breaker hidden danger identification device can be found in the definition of the circuit breaker hidden danger identification method above, and will not be repeated here.

[0139] An embodiment of the present application provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the above-mentioned circuit breaker hidden danger identification method based on feature space differentiation. The computer-readable storage medium can be phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), non-volatile random access memory (NVRAM), electrically erasable programmable read-only memory (EEPROM), etc.

[0140] The present invention is not limited to the above-mentioned specific implementation methods. Various changes made by ordinary technicians in this field based on the above-mentioned concept without creative work are all within the scope of protection of the present invention.

Claims

1. A circuit breaker hidden danger identification method based on feature space differentiation, characterized in that: include: Collecting multi-dimensional time series signal data of the circuit breaker in a first time period before the current moment and performing signal preprocessing, wherein the multi-dimensional time series signal data includes a circuit breaker coil current signal, an acoustic wave signal, and a vibration signal; Obtaining a circuit breaker state feature parameter group based on a preset feature extraction method for the multi-dimensional time series signal data of the circuit breaker; Based on the state characteristic parameter group of the circuit breaker, a hidden danger identification model is input to obtain the circuit breaker hidden danger identification result; The characteristic parameter group includes statistical features and information measurement features in the time domain and frequency domain for each dimension of time series signal data; the circuit breaker state characteristic parameter group is obtained based on the preset feature extraction method, including: for each dimension of time series signal data, obtaining multiple time series component signals based on signal decomposition; obtaining A1 characteristic parameters and A2 characteristic parameters based on the multiple time series component signals; performing frequency domain conversion on each time series component signal after signal decomposition to obtain different spectrum signals; and obtaining B1 characteristic parameters, B2 characteristic parameters, and B3 characteristic parameters based on the spectrum signals; The A1 characteristic parameter is: the standard deviation of multiple time series component signals; the A2 characteristic parameter is: the distance between the standard deviation of multiple time series component signals of the current dimensional time series signal data and the center of the standard deviation of multiple time series component signals of the dimensional time series signal data under normal conditions; the B1 characteristic parameter is: the first frequency proportion value, and the first frequency proportion value is: in the current spectrum signal, when there are k first frequencies in the accumulated n frequencies, the proportion of the first frequency energy in all frequency energies: The first frequency includes multiple frequency values ​​in a preset frequency range, and the first frequency represents the frequency in a normal state; A i is the amplitude of the ith first frequency, B j is the amplitude of the jth frequency in the current spectrum signal; the B2 characteristic parameter is: the non-first frequency ratio, and the non-first frequency ratio is: the ratio of the frequency content of the non-first frequency to the frequency content of the first frequency in the current spectrum signal, The B3 characteristic parameters are: Among them H j is the energy proportion of the jth frequency in the current spectrum signal, 2. The circuit breaker hidden danger identification method based on feature space differentiation according to claim 1 is characterized in that: The training process of the hidden danger identification model includes: Acquire multi-dimensional time series signal data of the circuit breaker in normal state and various abnormal states, and perform signal preprocessing; Obtaining a state feature parameter group corresponding to the multi-dimensional time series signal data of the circuit breaker based on a preset feature extraction method; Based on the state feature parameter group, a deep neural network is trained, and the trained deep neural network is obtained as a hidden danger identification model.

3. The circuit breaker hidden danger identification method based on feature space differentiation according to claim 2 is characterized in that: The circuit breakers include circuit breakers with different operating mechanisms, and acceleration sensors are arranged at different positions of the circuit breakers with different operating mechanisms; the circuit breaker abnormalities include various mechanical defects and faults.

4. The circuit breaker hidden danger identification method based on feature space differentiation according to claim 1, characterized in that: The circuit breaker-based state characteristic parameter group is input into the hidden danger identification model to obtain the circuit breaker hidden danger identification result, including: Input the time series signal data of each dimension into the time series feature extraction module of the hidden danger identification model to obtain the time series features of the time series signal of each dimension. The time series feature extraction module is implemented based on a recurrent neural network; Based on the time series characteristics and state characteristic parameter group of each dimension time series signal, the feature fusion analysis module of the hidden danger identification model is input to obtain the fusion characteristics of the multi-dimensional time series signal data of the circuit breaker; Based on the fusion characteristics of the multi-dimensional time series signal data of the circuit breaker, the hidden danger classification module of the hidden danger identification model is input to obtain the circuit breaker hidden danger status represented by the multi-dimensional time series signal data of the circuit breaker.

5. The circuit breaker hidden danger identification method based on feature space differentiation according to claim 4 is characterized in that: The feature fusion analysis module includes a feature weight assignment module and a feature fusion module. The feature fusion module is implemented based on a convolutional neural network. The feature weight assignment module is used to filter according to the importance of features of different dimensions. The weight acquisition method of the feature weight assignment module includes: Based on the time series features and state feature parameter groups of each dimension time series signal, an M-dimensional feature vector is formed. For the p-th type of circuit breaker abnormality, Lp samples are sampled, where p = 1, 2, ..., P, and P is the total number of circuit breaker abnormality types. The M-dimensional feature vectors of the Lp samples of the p-th type of circuit breaker abnormality form an Lp*M feature matrix, and P types of circuit breaker abnormalities form P Lp*M feature matrices. Based on P Lp*M feature matrices, the first correlation coefficient R1 and the second distinction coefficient R2 of the m-th dimension feature in the M-dimensional feature vector are respectively obtained, and the weight of the m-th dimension feature in the M-dimensional feature vector is determined based on (R2-R1) / (R2+R1), where the first correlation coefficient R1 represents the difference of the m-th dimension feature in the same type of circuit breaker abnormal samples, and the second distinction coefficient R2 represents the difference of the m-th dimension feature in different types of circuit breaker abnormal samples, where m=1, 2, ..., M.

6. The circuit breaker hidden danger identification method based on feature space differentiation according to claim 5, characterized in that: The method for obtaining the first correlation coefficient R1 includes: The average value of the distribution distance between the elements in the mth column of a single feature matrix: d1, d1 is the average value of the distance between any two elements in the mth column of a single feature matrix; The average value of the distribution distance of the elements in the mth column of P feature matrices: d1'; The center of the m-th column element of a single feature matrix: d2, d2 is the average value of all elements in the m-th column of a single feature matrix; The average value d2' of the distance between the elements of the mth column of a single feature matrix and its center d2; The average value d2″ of the distance d2′ between the elements of the mth column of P feature matrices and their center d2; Determine the first correlation coefficient R1 of the m-th dimension feature of all samples of P-type circuit breaker abnormality based on d1′ and d2″; The method for obtaining the second discrimination coefficient R2 includes: For the Lp*M feature matrix composed of Lp samples collected for the p-th type of circuit breaker anomaly, convert it into a 1*M feature matrix. The conversion method is to use d2 corresponding to all elements of the m-th column vector of the Lp*M feature matrix to represent the elements of the m-th column of the 1*M feature matrix; The average distribution distance between the elements in the mth column of the P transformed feature matrices: D1, D1 is the average distance between the elements d2 in the mth column of any two 1*M feature matrices; The center of the m-th column element of the P transformed feature matrices: D2, D2 is the average value of the m-th column element d2 of the P 1*M feature matrices; The average value D2' of the m-th column elements of the P transformed feature matrices to their center D2; The second discrimination coefficient R2 of the m-th dimension feature of all samples of the P-type circuit breaker abnormality is determined based on D1 and D2'.

7. A circuit breaker hidden danger identification device based on feature space differentiation, characterized in that: include: A monitoring signal acquisition unit, configured to acquire multi-dimensional time series signal data of the circuit breaker in a first time period before the current moment and perform signal preprocessing, wherein the multi-dimensional time series signal data includes a circuit breaker coil current signal, an acoustic wave signal, and a vibration signal; A signal feature extraction unit, configured to obtain a circuit breaker state feature parameter group based on a preset feature extraction method for the multi-dimensional time series signal data of the circuit breaker; A circuit breaker hidden danger identification unit is used to input a hidden danger identification model based on a state characteristic parameter group of the circuit breaker and obtain a circuit breaker hidden danger identification result; In the signal feature extraction unit, the feature parameter group includes statistical features and information measurement features in the time domain and frequency domain for each dimensional time series signal data; the circuit breaker state feature parameter group is obtained based on the preset feature extraction method, including: for each dimensional time series signal data, obtaining multiple time series component signals based on signal decomposition; obtaining A1 feature parameters and A2 feature parameters based on the multiple time series component signals; performing frequency domain conversion on each time series component signal after signal decomposition to obtain different spectrum signals; and obtaining B1 feature parameters, B2 feature parameters, and B3 feature parameters based on the spectrum signals; The A1 characteristic parameter is: the standard deviation of multiple time series component signals; the A2 characteristic parameter is: the distance between the standard deviation of multiple time series component signals of the current dimensional time series signal data and the center of the standard deviation of multiple time series component signals of the dimensional time series signal data under normal conditions; the B1 characteristic parameter is: the first frequency proportion value, and the first frequency proportion value is: in the current spectrum signal, when there are k first frequencies in the accumulated n frequencies, the proportion of the first frequency energy in all frequency energies: The first frequency includes multiple frequency values ​​in a preset frequency range, and the first frequency represents the frequency in a normal state; A i is the amplitude of the ith first frequency, B j is the amplitude of the jth frequency in the current spectrum signal; the B2 characteristic parameter is: the non-first frequency ratio, and the non-first frequency ratio is: the ratio of the frequency content of the non-first frequency to the frequency content of the first frequency in the current spectrum signal, The B3 characteristic parameters are: Among them H j is the energy proportion of the jth frequency in the current spectrum signal, 8. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images