Cable fault edge diagnosis method fusing parallel network and attention mechanism

Abstract

The invention provides a cable fault edge diagnosis method fusing a parallel network and an attention mechanism, and the method comprises the steps: extracting features from a cable fault signal, and constructing a k-type fault sample set; constructing a parallel network fusing a deep convolutional neural network and a long-short term memory network, and adding a corresponding attention mechanism into the parallel network; training a cable fault edge diagnosis model by using the k-class fault sample set, and obtaining sample class features; the class center of each fault is calculated by using the sample class features, and a fault classifier is designed; screening feature samples from the trained old fault category samples to construct a combined feature set with fixed capacity; combining the t types of new fault samples with the combined feature set, and training a cable fault edge diagnosis model through knowledge distillation; updating the joint feature set by using t types of new fault samples; if a new type of fault continues to be generated, the previous two steps are repeated, and a final cable fault edge diagnosis model is obtained; and inputting a cable signal sample of which the fault needs to be diagnosed into the final cable fault edge diagnosis model to obtain class features of the sample, and determining a cable fault class of the sample by using a fault classifier.

Description

Technical Field

[0001] The present invention relates to the technical field of cable fault diagnosis, and in particular to a cable fault edge diagnosis method integrating a parallel network and an attention mechanism. Background Art

[0002] The traditional cable fault diagnosis idea is to centrally receive cable operation data on the cloud platform of the master station (or dispatching center), and then use the relevant algorithm model to perform fault diagnosis on the cables under its jurisdiction. It mainly faces three problems: first, the massive cable operation data is centralized to the cloud for processing, which has poor timeliness; second, the massive data is uploaded to the cloud, and the security of its data transmission is difficult to guarantee; third, there are differences in the cable operation environment and working conditions, and the cloud processing lacks pertinence. With the construction of new power systems, the development and introduction of new technologies such as edge computing, artificial intelligence, and 5G communications, the cable fault diagnosis idea on the edge side of the power grid, that is, through the combination of edge computing and artificial intelligence, can well solve these problems of cloud diagnosis.

[0003] However, the existing AI-based cable fault diagnosis methods currently face an important challenge: insufficient model generalization ability. The existing AI-based cable fault diagnosis model is static and can only diagnose fault conditions encountered during the training process. When faced with new faults with different characteristics caused by increased cable operation time and changes in the operating environment, it cannot make accurate diagnoses, that is, the existing AI-based cable fault diagnosis model cannot achieve scalable fault diagnosis. If the existing AI-based cable fault diagnosis model wants to diagnose faults in these new scenarios, it can only discard the currently trained model, then merge the new fault data with the original fault data, and retrain the cable fault diagnosis model to deploy it to the edge computing device. If the edge computing unit has infinite memory and does not consider data storage issues, the new fault can be diagnosed according to the above operations. However, in fact, the memory capacity of the edge computing device is limited, and if the previously resource-intensive training model is completely discarded, it will also cause a waste of resources.

[0004] Therefore, how to achieve scalable and accurate diagnosis of cable faults with limited memory in edge computing devices has become an urgent problem to be solved. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a cable fault edge diagnosis method integrating parallel network and attention mechanism, comprising the following steps:

[0006] S1, obtaining cable fault signal;

[0007] S2, extract features from cable fault signals and construct k-type fault sample sets;

[0008] S3. Construct a parallel network integrating deep convolutional neural network and long short-term memory network, and add corresponding attention mechanism to obtain the cable fault edge diagnosis model;

[0009] S4, using the k-type fault sample set to train the cable fault edge diagnosis model, and obtain the sample class features from the fully connected layer of the parallel network;

[0010] S5. Calculate the class center of each fault using the sample class features, and design a fault classifier based on the class features and class center of each fault sample;

[0011] S6. Using sample class features and fault class centers, feature samples are selected from old fault class samples that have completed training to construct a fixed-capacity joint feature set;

[0012] S7. When t types of new faults occur in the cable, the new fault samples are combined with the joint feature set, and the cable fault edge diagnosis model is trained through knowledge distillation to obtain a cable fault edge diagnosis model that can diagnose k+t types of faults;

[0013] S8, using the model trained in step S7 to obtain the class features and class centers of t types of new fault samples, and updating the joint feature set;

[0014] S9, repeating steps S7 and S8 until no new type of fault occurs, and obtaining a final cable fault edge diagnosis model;

[0015] S10, inputting the cable signal sample whose fault needs to be diagnosed into the final cable fault edge diagnosis model to obtain the class feature of the sample, and using the designed fault classifier to determine the cable fault category of the sample.

[0016] Optionally, in step S1, four observable electrical quantities of the cable are obtained from a number of observable electrical quantities of the cable, namely, ground wire current, zero-sequence current and AC two-phase current.

[0017] Optionally, in step S2, wavelet transform is used to extract waveform features and energy features in the grounding wire current to construct a statistical feature matrix; time series pair features are extracted from the four signals of grounding wire current, zero-sequence current and AC two-phase current to construct a shallow feature matrix, each fault sample is composed of a statistical feature matrix and a shallow feature matrix, and a number of fault samples constitute a fault sample set X = {X1, X2, ..., X k}, the fault type is k, X k ={x k 1 ,x k 2 ,...,x k n} represents the sample set of the k-th fault, and the number of samples of the k-th fault is n.

[0018] Optionally, the statistical feature matrix can be obtained by the following method: processing the grounding wire current signal through 4-layer wavelet decomposition to obtain 5 grounding wire current reconstruction signals; obtaining 9 waveform features including crest factor, pulse factor, clearance factor, shape factor, root mean square root, arithmetic mean, kurtosis, variance and skewness from each grounding wire current reconstruction signal, for a total of 45 waveform features; obtaining 9 energy features based on the energy relationship between the grounding wire current reconstructed signal and the original signal and the energy relationship between the reconstructed signals; combining the 45 waveform features with the 9 energy features to form a statistical feature matrix with a dimension of 6×9, wherein the first 5 rows of the matrix are waveform features, each row is the waveform feature of each reconstructed signal, arranged from low to high according to frequency, and the 6th row is the energy feature;

[0019] The shallow characteristic matrix can be obtained by the following method: taking the grounding wire current I s , zero sequence current I z , AC two-phase current I A with I C At the same time T, the amplitudes form a time series pair [I sT , I zT , I AT , I CT ], several time series pairs within a time window of length L constitute a shallow feature matrix in chronological order.

[0020] Optionally, in step S3, the convolutional neural network and the long short-term memory network are connected in parallel, and the two networks are connected by concat in the fully connected layer, wherein the statistical feature matrix is ​​used as the input of the convolutional neural network, and the shallow feature matrix is ​​used as the input of the long short-term memory network. A channel attention mechanism is added after the pooling layer 1 of the convolutional neural network to enhance the feature condensation effect of the convolutional neural network.

[0021] Optionally, in step S5, the class features of all samples of each fault are averaged, and the average value is the class center μ of this fault. k , the calculation formula is as follows:

[0022]

[0023] Where x is the sample in the input cable fault edge diagnosis model, It is expressed as a nonlinear mapping function of the cable fault edge diagnosis model;

[0024] Get the class center μ of k-type faults = {μ1, μ2, ..., μ k}, subtract the class features of all samples in each fault type from the k class centers, and compare these differences. The one with the smallest difference vector length is the fault category to which the sample belongs.

[0025] Optionally, in step S6, the joint feature set is represented by P = {P1, P2, ..., P k}, the capacity of the joint feature set is N U , from the sample set X of each fault category k Select m=N U / k samples, P k The characteristic sample set representing the k-th type of fault, the samples contained in it are called characteristic samples, and the selection of characteristic samples can be carried out according to the following rules:

[0026] The first feature sample selection: the sample set X k The class features and class center μ of each sample in k Do the difference, then compare the lengths of these difference vectors, and select the sample with the smallest difference vector length as the feature sample set P k The first feature sample is denoted as p k 1 ;

[0027] The second feature sample selection: p k 1 Remove it, and then replace X k The class features of the remaining samples are respectively k 1 The class features are averaged, and the obtained average vector is then compared with the class center μ k Do the difference and select the sample with the smallest difference vector length as the feature sample set P k The second feature sample is denoted as p k 2 ;

[0028] For the remaining feature sample selection: remove the feature samples that have been selected, and then select p k 2 The method starts from the sample set X k Select the remaining m-2 feature samples.

[0029] Optionally, in step S7, g y (x) represents the distribution probability of the old fault samples in the joint feature set obtained in step S6 in the old fault diagnosis model, g y (x) is considered to be the knowledge distilled from the old fault diagnosis model, called soft label. When the old fault samples in the joint feature set are input into the cable fault edge diagnosis model for training, the distillation loss is as follows:

[0030]

[0031] in, a y (x i ) is the output value of the fault sample input to the cable fault edge diagnosis model, θ is the parameter of the cable fault edge diagnosis model, i = 1, 2, ..., m;

[0032] When the samples in the new fault sample set are input into the cable fault edge diagnosis model for training, the classification loss is as follows:

[0033]

[0034] Among them, δ is the label of the sample, called the hard label;

[0035] In order to prevent the cable fault edge diagnosis model from overfitting, the L2 regularization term is used to limit it. The formula is as follows:

[0036]

[0037] Among them, w i is the weight parameter of the cable fault edge diagnosis model, n w is the number of weight parameters, λ w is the weight penalty factor, which is 1×10 -5 ;

[0038] The total loss function L of the cable fault edge diagnosis model is constructed by the distillation loss of the old fault samples, the classification loss of the new fault samples and the L2 regularization term. T (θ), as shown in the following formula:

[0039] L T (θ) = ∑(L1+L2)+L2.

[0040] Optionally, in step S8, the update rule of the joint feature set is as follows:

[0041] After adding t new fault categories, the total number of fault categories becomes k+t, and the number of samples of each fault category in the new joint feature set is m′=N U / (k+t);

[0042] For k old fault categories, the last m-m' feature samples in their respective feature sample sets are removed. For t new fault categories, m' feature samples are selected from their respective feature sample sets according to the method in claim 7. The updated joint feature set is expressed as P'={P1', P2', ..., P k ′,P k+1 ′,…,P k+t ′}.

[0043] After adopting the above technical solution, the present invention has at least the following beneficial effects:

[0044] 1. The present invention takes into account the limited memory of edge computing devices and uses a fixed-capacity set to characterize the practical problem of limited memory of edge computing devices. This set is called a joint feature set. By performing deep sample selection on fault samples that have completed training, more representative samples are selected to construct a fixed-capacity joint feature set.

[0045] 2. The cable fault edge diagnosis model of the present invention integrates a parallel network of a deep convolutional neural network and a long short-term memory network, and adds a corresponding attention mechanism thereto. By combining new fault samples with the old joint feature set to form a new joint feature set, and then continuously updating the cable fault edge diagnosis model through knowledge distillation, it is possible to achieve scalable and accurate diagnosis of cable faults based on artificial intelligence without abandoning the previous diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be 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 labor.

[0047] Figure 1 This is a simulation diagram of a 10kV three-outlet cable network;

[0048] Figure 2 A flow chart of a cable fault edge diagnosis method integrating a parallel network and an attention mechanism provided in an embodiment of the present disclosure;

[0049] Figure 3 It is a schematic diagram of the structure and working principle of the cable fault edge diagnosis model;

[0050] Figure 4 Schematic diagram of the channel attention mechanism;

[0051] Figure 5 Confusion matrix diagram of the cable fault edge diagnosis method results that integrates parallel networks and attention mechanisms.

[0052] Attached Figure 1Note: C1 is the cable line, C2 is the cable line, C3 is the cable line, C4 is the cable line, C5 is the cable line, C6 is the cable line, Bus1 is the bus, Bus2 is the bus, Bus3 is the bus, Bus4 is the bus, Bus5 is the bus, Bus6 is the bus, T1 is the distribution substation, T2 is the distribution substation, T3 is the distribution substation, T4 is the distribution substation, T5 is the distribution substation. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] In PSCAD / EMTDC software, if Figure 1 As shown in the figure, a 10kV three-outlet cable network simulation diagram is built, and 8 cable states are simulated, including 3 disturbances and 5 faults. The disturbances are static load mutation, dynamic load mutation and capacitor switching; the faults are intermittent arc fault, permanent fault and 3 types of weak faults. The 10kV squirrel cage asynchronous motor model is used to simulate the dynamic load mutation, the line end parallel capacitor group is used to simulate the capacitor switching, and the Mayr arc model is used to simulate the intermittent arc fault. For permanent faults, two models, metallic grounding and transition resistance grounding, are used for simulation. For weak faults, three models are used for simulation: (1) the outer insulation of the cable is damaged due to mechanical stress or chemical corrosion, resulting in the metal sheath being grounded at the damaged point; (2) the insulation between the cable core and the metal sheath is degraded, resulting in a high resistance fault; (3) the insulation degradation fault at a certain point of the cable caused by corrosion or moisture intrusion after the metal sheath is grounded. To simplify the description of the weak fault model, these three models are referred to as weak fault 1, weak fault 2 and weak fault 3.

[0055] The present disclosure provides a cable fault edge diagnosis method integrating parallel network and attention mechanism. The specific process is as follows: Figure 2 As shown, the following steps are included:

[0056] S1, obtaining cable fault signal;

[0057] Eight different cable state models are applied at different locations of the 10kV three-outlet cable network simulation diagram, and four data signals, namely ground wire current, zero-sequence current and AC two-phase current, are measured at the line head end.

[0058] S2, extract features from cable fault signals and construct a k-type fault sample set;

[0059] Wavelet transform is used to extract the waveform characteristics and energy characteristics of the grounding wire current and construct a statistical feature matrix. Time series pair characteristics are extracted from the four signals of grounding wire current, zero-sequence current and AC two-phase current to construct a shallow feature matrix. Each fault sample consists of a statistical feature matrix and a shallow feature matrix.

[0060] The statistical feature matrix can be obtained by the following method: the grounding wire current signal is processed by 4-layer wavelet decomposition to obtain 5 grounding wire current reconstruction signals; 9 waveform features including crest factor, pulse factor, clearance factor, shape factor, root mean square root, arithmetic mean, kurtosis, variance and skewness are obtained from each grounding wire current reconstruction signal, totaling 45 waveform features; 9 energy features are obtained according to the energy relationship between the grounding wire current reconstruction signal and the original signal and the energy relationship between the reconstructed signals; the 45 waveform features and the 9 energy features are combined to form a statistical feature matrix with a dimension of 6×9, in which the first 5 rows of the matrix are waveform features, each row is the waveform feature of each reconstructed signal, arranged from low to high according to frequency, and the sixth row is the energy feature;

[0061] The shallow characteristic matrix can be obtained by the following method: take the grounding current I s , zero sequence current I z , AC two-phase current I A with I C At the same time T, the amplitudes form a time series pair [I sT , I zT , I AT , I CT ], several time series pairs within the time window of length L constitute a shallow feature matrix in chronological order;

[0062] In this embodiment, the time window length is set to 61, taking the moment when the grounding wire current suddenly changes as the standard, including 10 points before the moment when the grounding wire current suddenly changes and 50 points after the moment when the grounding wire current suddenly changes.

[0063] The above feature extraction method is used to process the simulation data and part of the measured data, and a total of 11943 samples of 8 cable states are obtained. These samples are randomly shuffled, 85% of them are used as training samples and 15% as test samples. The sample distribution is shown in Table 1, where the brackets are the number of measured data.

[0064] Table 1 Sample distribution

[0065]

[0066] The initial cable fault types are set to 4, that is, k=4, and one cable fault is gradually added later to simulate the process of increasing cable fault types.

[0067] S3. Construct a parallel network integrating deep convolutional neural network and long short-term memory network, and add corresponding attention mechanism to obtain cable fault edge diagnosis model;

[0068] The cable fault edge diagnosis model constructed is as follows Figure 3 As shown in the figure, the convolutional neural network and the long short-term memory network are connected in parallel, and the two networks are connected by concat in the fully connected layer, where the statistical feature matrix is ​​used as the input of the convolutional neural network, and the shallow feature matrix is ​​used as the input of the long short-term memory network;

[0069] In this embodiment, the relevant parameters of the convolutional neural network are shown in Table 2, and the long short-term memory network parameters are shown in Table 3.

[0070] Table 2 Convolutional neural network parameters

[0071]

[0072]

[0073] Table 3 Parameters of long short-term memory network

[0074]

[0075] A channel attention mechanism is added after the pooling layer 1 of the convolutional neural network to enhance the feature condensation effect of the convolutional neural network;

[0076] Channel attention mechanism such as Figure 4 As shown in Figure 1, suppose the feature after pooling 1 layer is Fp1, and its dimension is H×W×C, where H is the height of the two-dimensional matrix, W is the width of the two-dimensional matrix, and C is the number of channels, that is, the number of two-dimensional matrices; perform global maximum pooling and global average pooling on Fp1 respectively, and obtain two features Gpm(Fp1) and Gpa(Fp1) with dimensions of 1×1×C, where Gpm and Gpa are global maximum pooling and global average pooling operations respectively; divide Gpm(Fp1) and Gpa(Fp1) into The two features are input into a shared two-layer fully connected layer to obtain two new features with dimensions of 1×1×C, where the number of neurons in the first layer of the full connection is K, and the number of neurons in the second layer is C; the two 1×1×C features are added together, and then the channel attention weight matrix Wc is obtained by the Sigmoid function; the feature Fp1 is multiplied by Wc to obtain a new feature with dimensions of H×W×C, which can be used as the input of the convolution 2 layer, where the values ​​of H, W, C and K are 4, 3, 9, and 3 respectively.

[0077] S4, using the k-type fault sample set to train the cable fault edge diagnosis model, and obtain the sample class features from the fully connected layer of the parallel network;

[0078] In this embodiment, the cable fault edge diagnosis model constructed in step S3 is trained using four types of fault sample sets. After the training is completed, data features are obtained in the fully connected layer of the parallel network, which are called sample class features.

[0079] S5. Calculate the class center of each fault using the sample class features, and design a fault classifier based on the class features and class center of each fault sample;

[0080] Several fault samples constitute a fault sample set X = {X1, X2, ..., X k}, the fault type is k, X k ={x k 1 ,x k 2 ,...,x k n} represents the sample set of the k-th fault, and the number of samples of the k-th fault is n;

[0081] Each fault X in the fault sample set k The class features of all samples in are averaged, and this average is X k The class center μ k , the calculation formula is as follows:

[0082]

[0083] Where x is the sample in the input cable fault edge diagnosis model, It is expressed as a nonlinear mapping function of the cable fault edge diagnosis model;

[0084] Get the class center μ of k-type faults = {μ1, μ2, ..., μ k}, subtract the class features of all samples in each fault from the k class centers respectively, and compare these differences. The one with the smallest difference vector length is the fault category to which the sample belongs, which is the fault classifier.

[0085] S6, using sample class features and fault class centers to select a fixed-capacity joint feature set from old fault class samples that have completed training;

[0086] The memory of edge computing devices is limited. Here, a joint feature set with a fixed capacity is constructed to characterize this practical problem. The joint feature set is a screening of old fault category samples that have completed training. The samples it contains can better reflect the characteristics of the fault category to which it belongs.

[0087] The joint feature set is represented as P = {P1, P2, ..., P k}, the capacity of the joint feature set is N U , from the sample set X of each fault category kSelect m=N U / k samples, P k The characteristic sample set representing the k-th type of fault, the samples contained in it are called characteristic samples, and the selection of characteristic samples can be carried out according to the following rules:

[0088] The first feature sample selection: the sample set X k The class features and class center μ of each sample in k Do the difference, then compare the lengths of these difference vectors, and select the sample with the smallest difference vector length as the feature sample set P k The first feature sample is denoted as p k 1 ;

[0089] The second feature sample selection: p k 1 Remove it, and then replace X k The class features of the remaining samples are respectively k 1 The class features are averaged, and the obtained average vector is then compared with the class center μ k Do the difference and select the sample with the smallest difference vector length as the feature sample set P k The second feature sample is denoted as p k 2 ;

[0090] For the remaining feature sample selection: remove the feature samples that have been selected, and then select p k 2 The method starts from the sample set X k Select the remaining m-2 feature samples.

[0091] S7. When t types of new faults occur in the cable, the new fault samples are combined with the joint feature set, and the cable fault edge diagnosis model is trained through knowledge distillation to obtain a cable fault edge diagnosis model that can diagnose k+t types of faults;

[0092] In this embodiment, t=1 is set, that is, one cable fault is added on the basis of the four cable faults. The feature samples in the joint feature set obtained in the previous step are the trained old fault samples. When the old fault samples in the joint feature set are input into the cable fault edge diagnosis model for training, the distillation loss is as follows:

[0093]

[0094] in, is the distribution probability of the old fault samples in the old fault diagnosis model, which is the knowledge distilled from the old fault diagnosis model, called soft label, a y (xi ) is the output value of the fault sample input to the cable fault edge diagnosis model, θ is the parameter of the cable fault edge diagnosis model, i = 1, 2, ..., m;

[0095] When the samples in the new fault sample set are input into the cable fault edge diagnosis model for training, the classification loss is as follows:

[0096]

[0097] Among them, δ is the label of the sample, called the hard label;

[0098] In order to prevent the cable fault edge diagnosis model from overfitting, the L2 regularization term is used to limit it. The formula is as follows:

[0099]

[0100] Among them, w i is the weight parameter of the cable fault edge diagnosis model, n w is the number of weight parameters, λ w is the weight penalty factor, which is 1×10 -5 ;

[0101] The total loss function L of the cable fault edge diagnosis model is constructed by the distillation loss of the old fault samples, the classification loss of the new fault samples and the L2 regularization term. T (θ), as shown in the following formula:

[0102] L T (θ) = ∑(L1+L2)+L2.

[0103] S8, using the model trained in step S7 to obtain the class features and class centers of t types of new fault samples, and updating the joint feature set;

[0104] After adding t new fault categories, the total number of fault categories becomes k+t, and the number of samples of each fault category in the new joint feature set is m′=N U / (k+t);

[0105] The selection of feature samples in the joint feature set is achieved by the distance between the class feature of the sample and the class center, which means that the feature sample is k The earlier the order in the sequence, the more representative the feature sample is. Therefore, for the k old fault categories, the last m-m' feature samples in their respective feature sample sets are removed. For the t new fault categories, m' feature samples are selected from their respective feature sample sets according to the method in step S6. The updated joint feature set is expressed as P'={P1', P2', ..., P k ′,P k+1 ′,…,Pk+t ′}.

[0106] S9, repeating steps S7 and S8 until no new type of fault occurs, and obtaining a final cable fault edge diagnosis model;

[0107] In this embodiment, steps S7 and S8 are repeated 4 times in total, so that the cable fault types are increased to 8 types.

[0108] S10, inputting the cable signal sample to be diagnosed into the final cable fault edge diagnosis model to obtain the class feature of the sample, and using the designed fault classifier to determine the cable fault category of the sample;

[0109] Based on experience and related calculations, the parameters used in the scalable diagnostic training of the model are set as follows: the capacity of the joint feature set N U The value is 4000, the learning rate is 0.1, the size of the training batch is 64, the number of training rounds is 200, and the optimizer uses the Adam (Adaptive Moment Estimation) algorithm with adaptive moment estimation.

[0110] The first four types of fault sample sets are respectively input into the cable fault edge diagnosis model for training. Then, according to the frequency of adding one new type of fault sample each time, 5-8 types of fault samples are added to the model for scalable diagnosis training. After each scalable diagnosis is completed, the corresponding test samples are used to evaluate the scalable diagnosis trained model, and the scalable diagnosis confusion matrix of the cable fault edge diagnosis model in 5-type state, 6-type state, 7-type state and 8-type state is obtained, as shown in Fig. Figure 5 The scalable diagnosis results of the sample set are shown in Table 4.

[0111] Table 4 Scalable diagnosis results of sample sets

[0112]

[0113]

[0114] Although the present invention has been disclosed as above by way of embodiments, it is not intended to limit the present invention. Any person skilled in the art may make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the claims.

Claims

1. A cable fault edge diagnosis method integrating parallel network and attention mechanism, characterized in that: The following steps are involved: S1, obtaining cable fault signal; S2, extract features from cable fault signals and construct k-type fault sample sets; S3. Construct a parallel network integrating deep convolutional neural network and long short-term memory network, and add corresponding attention mechanism to obtain the cable fault edge diagnosis model; S4, using the k-type fault sample set to train the cable fault edge diagnosis model, and obtain the sample class features from the fully connected layer of the parallel network; S5. Calculate the class center of each fault using the sample class features, and design a fault classifier based on the class features and class center of each fault sample; S6. Using sample class features and fault class centers, feature samples are selected from old fault class samples that have completed training to construct a fixed-capacity joint feature set; S7. When t types of new faults occur in the cable, the new fault samples are combined with the joint feature set, and the cable fault edge diagnosis model is trained through knowledge distillation to obtain a cable fault edge diagnosis model that can diagnose k+t types of faults; S8, using the model trained in step S7 to obtain the class features and class centers of t types of new fault samples, and updating the joint feature set; S9, repeating steps S7 and S8 until no new type of fault occurs, and obtaining a final cable fault edge diagnosis model; S10, inputting the cable signal sample whose fault needs to be diagnosed into the final cable fault edge diagnosis model to obtain the class feature of the sample, and using the designed fault classifier to determine the cable fault category of the sample.

2. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 1 is characterized in that: In the step S1, four observable electrical quantities of the cable are obtained from a plurality of observable electrical quantities of the cable, namely, ground wire current, zero sequence current and AC two-phase current.

3. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 2 is characterized in that: In the step S2, wavelet transform is used to extract the waveform characteristics and energy characteristics of the ground wire current, and a statistical characteristic matrix is ​​constructed; The time series features are extracted from the four signals of ground line current, zero sequence current and AC two-phase current to construct a shallow feature matrix. Each fault sample consists of a statistical feature matrix and a shallow feature matrix. Several fault samples constitute a fault sample set X = {X1, X2, ..., X k }, the fault type is k, represents the sample set of the k-th fault, and the number of samples of the k-th fault is n.

4. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 3 is characterized in that: The statistical feature matrix can be obtained by the following method: processing the grounding wire current signal through 4-layer wavelet decomposition to obtain 5 grounding wire current reconstruction signals; respectively obtaining 9 waveform features including crest factor, pulse factor, clearance factor, shape factor, root mean square root, arithmetic mean, kurtosis, variance and skewness from each grounding wire current reconstruction signal, for a total of 45 waveform features; According to the energy relationship between the reconstructed signal and the original signal of the ground wire current and the energy relationship between the reconstructed signals, 9 energy features are obtained; The 45 waveform features and 9 energy features are combined to form a statistical feature matrix with a dimension of 6×9, where the first 5 rows of the matrix are waveform features, each row is the waveform feature of each reconstructed signal, arranged from low to high frequency, and the sixth row is the energy feature; The shallow characteristic matrix can be obtained by the following method: taking the grounding wire current I s , zero sequence current I z , AC two-phase current I A with I C At the same time T, the amplitudes form a time series pair [I sT , I zT , I AT , I CT ], several time series pairs within a time window of length L constitute a shallow feature matrix in chronological order.

5. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 1 is characterized in that: In the step S3, the convolutional neural network and the long short-term memory network are connected in parallel, and the two networks are connected by concat in the fully connected layer, wherein the statistical feature matrix is ​​used as the input of the convolutional neural network, and the shallow feature matrix is ​​used as the input of the long short-term memory network. A channel attention mechanism is added after the pooling layer 1 of the convolutional neural network to enhance the feature condensation effect of the convolutional neural network.

6. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 1 is characterized in that: In step S5, the class features of all samples in each fault are averaged, and this average is the class center μ of this fault. k , the calculation formula is as follows: Where x is the sample in the input cable fault edge diagnosis model, It is expressed as a nonlinear mapping function of the cable fault edge diagnosis model; Get the class center μ of k-type faults = {μ1, μ2, ..., μ k }, subtract the class features of all samples in each fault from the k class centers respectively, and compare these differences. The one with the smallest difference vector length is the fault category to which the sample belongs. This is the constructed fault classifier.

7. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 1 is characterized in that: In step S6, the joint feature set is represented as P = {P1, P2, ..., P k }, the capacity of the joint feature set is N U , from the sample set X of each fault category k Select m=N U / k samples, P k The characteristic sample set representing the k-th type of fault, the samples contained in it are called characteristic samples, and the selection of characteristic samples can be carried out according to the following rules: The first feature sample selection: the sample set X k The class features and class center μ of each sample in k Do the difference, then compare the lengths of these difference vectors, and select the sample with the smallest difference vector length as the feature sample set P k The first feature sample is denoted as p k 1 ; The second feature sample selection: p k 1 Remove it, and then replace X k The class features of the remaining samples are respectively k 1 The class features are averaged, and the obtained average vector is then compared with the class center μ k Do the difference and select the sample with the smallest difference vector length as the feature sample set P k The second feature sample is denoted as p k 2 ; For the remaining feature sample selection: remove the feature samples that have been selected, and then select p k 2 The method starts from the sample set X k Select the remaining m-2 feature samples.

8. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 1 is characterized in that: In step S7, g y (x) represents the distribution probability of the old fault samples in the joint feature set obtained in step S6 in the old fault diagnosis model, g y (x) is considered to be the knowledge distilled from the old fault diagnosis model, called soft label. When the old fault samples in the joint feature set are input into the cable fault edge diagnosis model for training, the distillation loss is as follows: in, a y (x i ) is the output value of the fault sample input to the cable fault edge diagnosis model, θ is the parameter of the cable fault edge diagnosis model, i = 1, 2, ..., m; When the samples in the new fault sample set are input into the cable fault edge diagnosis model for training, the classification loss is as follows: Among them, δ is the label of the sample, called the hard label; In order to prevent the cable fault edge diagnosis model from overfitting, the L2 regularization term is used to limit it. The formula is as follows: Among them, w i is the weight parameter of the cable fault edge diagnosis model, n w is the number of weight parameters, λ w is the weight penalty factor, which is 1×10 -5 ; The total loss function L of the cable fault edge diagnosis model is constructed by the distillation loss of the old fault samples, the classification loss of the new fault samples and the L2 regularization term. T (θ), as shown in the following formula: L T (θ)=∑(L1+L2)+L2。 9. The cable fault edge diagnosis method integrating parallel network and attention mechanism according to claim 7 is characterized in that: In step S8, the update rule of the joint feature set is as follows: After adding t new fault categories, the total number of fault categories becomes k+t, and the number of samples of each fault category in the new joint feature set is m′=N U / (k+t); For k old fault categories, the last m-m' feature samples in their respective feature sample sets are removed. For t new fault categories, m' feature samples are selected from their respective feature sample sets according to the method in claim 7. The updated joint feature set is expressed as P'={P1', P2', ..., P k ′,P k+1 ′,…,P k+t ′}.

Images