Power cable fault sound recognition method and system based on multi-network fusion

Abstract

The present application relates to the technical field of power cable fault discharge sound identification, and proposes a power cable fault sound identification method and system based on multi-network fusion, which comprises: multi-level feature extraction and data envelope analysis on the sound signal sequence to generate multi-scale features and trend score graphs; using a self-attention mechanism to obtain similarity score graphs and value vectors; by splicing the trend and similarity graphs, constructing a gated feature and generating a hybrid weight graph; based on the hybrid weight graph, fusing the value vectors to complete feature fusion; finally, time series modeling and classification are performed to output the identification result. Through the deep learning model of multi-network fusion, the present application realizes the automatic extraction and fusion of multi-level and multi-type features of power cable discharge sound signals, and improves the intelligent level of fault identification and the generalization ability of the model. The present application overcomes the influence of the insufficient generalization ability of the classical sound signal description method and recognition method, and realizes the intelligent identification of power cable fault discharge sound.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power cable fault discharge sound recognition, in particular to a power cable fault sound recognition method and system based on multi-network fusion. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] As an important infrastructure of urban power distribution network, power cables play a key role in ensuring the safe and stable operation of power systems. Currently, urban power cables are usually laid underground in the form of ducts or direct burial. Although this method improves the safety of the line and the coordination of urban planning, it also makes the cable more susceptible to external environmental influences, such as moisture, corrosion, and mechanical extrusion, thereby causing various cable faults such as insulation aging, joint failure, and mechanical damage.

[0004] To timely detect and locate these faults, the power industry has developed various cable fault detection techniques, mainly including fault ranging and fault pinpointing. In the fault pinpointing stage, the current mainstream methods in engineering include acoustic measurement and acoustic-magnetic synchronization: acoustic measurement collects the sound signals generated by cable discharge through ground probes, which are processed by the host computer and played back by earphones for the test personnel to determine the fault point location; acoustic-magnetic synchronization method synchronously collects the sound signals and electromagnetic signals generated by discharge, and analyzes the waveform time difference to assist in determining the fault point.

[0005] The above methods are based on the premise of accurately obtaining and recognizing the discharge sound signals. However, in actual applications, the sound signals of the fault point are easily affected by various interferences, such as complex environmental noise (traffic noise, pedestrian noise, etc.) and the influence of different laying media (such as cement, sand, etc.) on the propagation characteristics of sound waves. These factors lead to a decrease in signal quality, making subsequent analysis difficult. To improve the accuracy of signal recognition, some current research uses traditional signal processing methods to describe and classify discharge sound, such as short-time Fourier transform, Mel frequency cepstral coefficient (MFCC), and wavelet transform. These methods usually rely on fixed feature extraction mechanisms and are difficult to adapt to complex scenarios with multiple media and multiple source noises. At the same time, existing deep learning models, such as convolutional neural networks, lack effective modeling capabilities for multi-level structural features in sound signals, such as local mutations and global trends; and have significant shortcomings in handling long-term dependencies of sound signals, which cannot fully capture the adaptability and generalization ability of discharge sound signals in complex scenarios. SUMMARY

[0006] To address the aforementioned problems, this invention proposes a method and system for power cable fault sound recognition based on multi-network fusion. Through a deep learning network model that integrates multiple networks, it enables the interactive fusion of multiple features, allowing for the automatic extraction of multi-level and multi-type features from cable discharge sound signals. This method highlights local detail features and global trend features while mitigating redundant features and models the long-term dependencies of sound signals. It overcomes the limitations of classical sound signal description and recognition methods in terms of generalization ability, thus achieving intelligent recognition of power cable fault discharge sounds.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] One or more embodiments provide a method for power cable fault sound recognition based on multi-network fusion, including the following steps:

[0009] The acquired audio signal sequence to be identified Multi-level feature extraction is performed to obtain the extracted multi-scale features. ;

[0010] The acquired audio signal sequence to be identified Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ;

[0011] Extracted multi-scale features Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors;

[0012] Trend Score Chart Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. ;

[0013] Based on hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain the fused features. ;

[0014] Targeting fusion features After performing time-series modeling and classification, the classification results of the sound signals are obtained.

[0015] One or more embodiments provide a power cable fault sound recognition system based on multi-network fusion, including:

[0016] A multi-level feature extraction network is configured to extract the acquired sequence of sound signals to be identified. Multi-level feature extraction is performed to obtain the extracted multi-scale features. ;

[0017] The envelope feature extraction module is configured to extract the acquired sequence of sound signals to be recognized. Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ;

[0018] The fusion attention module is configured to extract multi-scale features. Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors;

[0019] The gating fusion module is configured to convert the trend score map Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. Based on hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain the fused features. ;

[0020] The timing capture and signal classification modules are configured to target fused features. After performing time-series modeling and classification, the classification results of the sound signals are obtained.

[0021] One or more embodiments provide a power cable fault sound recognition system based on multi-network fusion, including: a sound acquisition device and a processor;

[0022] The processor is configured to perform the steps of the above-described method for identifying power cable fault sounds based on multi-network fusion.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] The proposed power cable fault discharge sound recognition method can automatically extract multi-level and multi-type features from discharge sound signals, highlighting local detail features and global trend features, effectively weakening redundant information, and improving the pertinence and discriminativeness of feature representation. By fusing and modeling multi-scale structural information and trend features, it overcomes the limitations of traditional methods that rely on single features. The introduced gating fusion mechanism dynamically adjusts the fusion ratio based on the complementarity between features, enhancing the system's adaptability and robustness to different sound patterns. Furthermore, by utilizing a self-attention mechanism to model the long-term dependencies of sound signals, it achieves accurate characterization of complex dynamic evolution processes, significantly improving fault identification and location accuracy under multi-source noise interference and media differences.

[0025] The advantages of the present application and the advantages of additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0026] The accompanying drawings, which form a part of the specification, are included to provide further understanding of the application and are incorporated in and constitute a part of this specification. The illustrative embodiments of the present application and their description are used to explain the present application.

[0027] Figure 1 is a flow chart of the power cable fault sound recognition method of embodiment 1 of the present application;

[0028] Figure 2 is a model architecture diagram of the multi-network fusion deep learning network model of embodiment 1 of the present application;

[0029] Figure 3 is an example waveform diagram of the power cable fault discharge sound signal laid at the manhole cover during the construction of the training set of embodiment 1 of the present application;

[0030] Figure 4 is an example waveform diagram of the power cable fault discharge sound signal laid underground during the construction of the training set of embodiment 1 of the present application;

[0031] Figure 5 is an example waveform diagram of the non-discharge sound signal during the construction of the training set of embodiment 1 of the present application;

[0032] Figure 6 is a network structure diagram of the multi-level feature extraction network of embodiment 1 of the present application;

[0033] Figure 7 is a structure schematic diagram of the fusion attention module of embodiment 1 of the present application;

[0034] Figure 8 is a structure schematic diagram of the gate fusion module of embodiment 1 of the present application;

[0035] Figure 9 is a structure schematic diagram of the long short-term memory network of embodiment 1 of the present application;

[0036] Figure 10 is a structure schematic diagram of the time sequence capture and signal classification module of embodiment 1 of the present application. DETAILED DESCRIPTION

[0037] The present application will be further described below in conjunction with the drawings and embodiments.

[0038] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0039] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof. It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

[0040] Embodiment 1

[0041] In one or more embodiments disclosed in the technical solutions, as shown in Figures 1 to 10 A power cable fault sound recognition method based on multi-network fusion includes the following steps:

[0042] Step 1, multi-level feature extraction is performed on the obtained sound signal sequence to be recognized to obtain multi-scale features after extraction ;

[0043] Step 2, data envelope feature extraction is performed on the obtained sound signal sequence to be recognized , and a trend score graph is generated according to the envelope line obtained ;

[0044] Step 3, the extracted multi-scale features are subjected to self-attention mechanism operation to generate similarity score graph and value vector between query and key

[0045] Step 4, the trend score graph is spliced with the similarity score graph , convolution and activation are performed to generate gating features; the fusion ratio of the similarity score graph and the trend score graph is dynamically controlled according to the gating features, and a hybrid weight graph is generated ;

[0046] Step 5, based on the hybrid weight graph , the value vector generated using the multi-scale features is fused to obtain fusion features ;

[0047] Step 6, for the fusion features After the time series modeling and classification are performed, a classification result of the sound signal is obtained; based on the identified faulty discharge sound signal, further positioning can be performed to obtain a fault positioning result.

[0048] The embodiment is aimed at the problem that a large amount of environmental noise is mixed in the collected original sound data, causing insufficient modeling capability, and proposes multi-level feature extraction, which can significantly improve the extraction capability of the sound signal feature and enhance the feature expression capability. In view of the problem that the convolutional neural network is difficult to depict the overall change trend of the signal, the data envelopment analysis method is used to model the time domain profile of the sound signal, to generate a trend score map, thereby effectively modeling the overall change trend of the sound signal. In view of the problem that the model has poor long-time dependence relationship capturing capability for the sound signal, a fusion attention module is proposed, and the extracted multi-scale features are used to calculate the relationship between the query, key and value, to generate a similarity score map and a corresponding value vector. The fusion proportion of the trend score map and the similarity score map is dynamically controlled through the gating fusion, which significantly enhances the representation capability of the features extracted by the model for the global trend and local mutation. Finally, the above method realizes effective modeling and classification of the fault sound, and then the classification result can be combined to perform fault signal positioning.

[0049] The power cable fault discharge sound recognition method proposed in the embodiment can realize automatic extraction of multi-level and multi-type features in the discharge sound signal, highlight local detail features and global trend features, effectively weaken redundant information, and improve the pertinence and discrimination of feature expression. Through the fusion modeling of multi-scale structural information and trend features, the limitations of traditional methods on single feature dependence are overcome; the introduced gating fusion mechanism dynamically adjusts the fusion proportion according to the complementarity between features, enhancing the adaptability and robustness of the system to different sound patterns. Further, the long-time dependence relationship of the sound signal is modeled by using the self-attention mechanism, realizing accurate depiction of the complex dynamic evolution process, and significantly improving the fault recognition and positioning accuracy in the background of multi-source noise interference and medium difference.

[0050] Step 1 can be implemented through the constructed multi-level feature extraction network, as shown in Figure 6 The structure of the multi-level feature extraction network includes a first CR module, a splitting module, a lightweight processing branch and a deep extraction branch, a splicing module and a second CR module; the two CR modules include a convolution layer and an activation layer, which perform convolution and activation operations on the input features; in the embodiment, the activation layer adopts a ReLU activation function; the multi-level feature extraction network can significantly improve the extraction capability of the sound signal feature and enhance the feature expression capability.

[0051] Based on the constructed multi-level feature extraction network, the obtained sound signal sequence to be recognized is subjected to multi-level feature extraction to obtain multi-scale features after extraction , multi-scale features , the extraction method comprising the following steps:

[0052] Step 11, the acquired sound signal sequence to be identified is convolved and activated, and the features are split along the channel dimension and input to the lightweight processing branch and the deep extraction branch;

[0053] Specifically, the input sound signal sequence is , which is input to the multi-level feature extraction network of the double branch to capture multi-level local detail information in the sound signal. The CR module combined with convolution and ReLU activation function is used to process the input data, and the features are split along the channel dimension and input to the lightweight processing branch and the deep extraction branch; the formula for feature processing is:

[0054] (1) ;

[0055] (2) ;

[0056] Wherein, represents the input sound signal sequence data, represents the convolution operation, represents the ReLU activation function, represents the combination of convolution operation and ReLU activation function, represents the splitting operation; represents the features input to the lightweight processing branch after CR module operation, represents the features input to the deep extraction branch after CR module operation.

[0057] Step 12, in the deep extraction branch, the input features , multi-scale information is obtained by using deep progressive multi-layer coding based on convolutional neural network, and multi-scale information is connected by using a jump connection block to obtain multi-level features ; the formula for the deep extraction branch is:

[0058] (3) ;

[0059] (4) ;

[0060] (5) ;

[0061] Wherein, represents the input features of the deep extraction branch, represents the multi-level features obtained by step-by-step extraction through the jump connection block, represents the convolution operation, denotes a ReLU activation function, denotes a combination of a convolution operation and a ReLU activation function, denotes a skip connection block.

[0062] Figure 6 wherein n denotes an input signal dimension; c denotes a feature dimension of the signal; m denotes a number of skip connection blocks in deep feature extraction; n c denotes a feature dimension after a split operation; n m + 2) c denotes a feature dimension after splicing of the deep extraction branch and the lightweight processing branch.

[0063] Step 13, splicing the multi-level features obtained by the deep extraction branch and the bottom layer information features obtained by the lightweight processing branch in the channel dimension, and reducing the dimension of the spliced features to obtain fused multi-scale features ; the fusion processing process is as follows:

[0064] (6)

[0065] (7)

[0066] wherein, denotes a bottom layer feature of the lightweight processing branch, denotes a feature input into the deep extraction branch, denotes a multi-level feature obtained by step-by-step extraction through the skip connection block, denotes a convolution operation, denotes a ReLU activation function, denotes a combination of a convolution operation and a ReLU activation function, denotes a splicing operation along the channel dimension.

[0067] In the above feature extraction scheme in the embodiment, the spliced features are reduced in dimension, the parameter amount of the network is reduced, and the network performance and efficiency are balanced. The multi-level feature network with double branches realizes efficient and comprehensive feature extraction through the synergistic effect of the deep extraction branch and the lightweight processing branch. The deep extraction branch extracts multi-scale information and deep-level features through multiple convolution operations and skip connection blocks, while the lightweight processing branch quickly extracts basic feature information. The fusion of the two enables the model to more comprehensively understand the input data, improves the performance and robustness of the model, and is suitable for various complex tasks.

[0068] In step 2, the obtained to-be-recognized sound signal sequence is subjected to data envelope feature extraction, and a trend score graph is generated according to the obtained envelope line ​​, trend score map The generation method is as follows:

[0069] Step 21, envelope feature extraction is performed on the obtained to-be-recognized sound signal sequence x, that is, Hilbert transform is performed to obtain an envelope line, and the formula is as follows:

[0070] (8);

[0071] Wherein, The Hilbert transform of the input signal . .

[0072] Step 22, a trend score map is generated using the envelope line, and the formula is as follows:

[0073] (9);

[0074] Wherein, .

[0075] In step 3, the obtained multi-scale feature is subjected to a self-attention mechanism operation to generate a correlation feature, including the following steps:

[0076] Step 31, the multi-scale feature is mapped into a query vector , a key vector and a value vector ;

[0077] Step 32, the mapped matrix is subjected to linear transformation through a learnable weight matrix , , ; the formula is as follows:

[0078] (10);

[0079] (11);

[0080] (12);

[0081] Step 33, the similarity between each query vector and all key vectors is calculated using dot product to obtain a similarity score map between the query and the key, which is calculated by the following formula:

[0082] (13);

[0083] (14);

[0084] Wherein, is the dimension of the key vector, denotes the transpose of .

[0085] Step 4 is implemented by a gating fusion module, as shown in the structural diagram Figure 8 , the trend score map obtained in step 2 is spliced with the similarity score map obtained in step 3, and then a convolution and a Sigmoid activation function are used to generate a gating feature gate. The gating feature is used to dynamically control the proportion of the similarity score map and the trend score map, and a convolution is used to generate a hybrid weight map , the formula is as follows:

[0086] (15);

[0087] (16);

[0088] wherein, , denotes the weight matrix of convolution, and denotes the bias of convolution, represents the dot product operation, , gate denotes the gating feature;

[0089] In step 5, based on the hybrid weight map , a multi-scale feature is used to generate a value vector for fusion, and a final self-attention output feature is obtained as follows:

[0090] (17);

[0091] Step 6, the method for time series modeling and classification of the fused feature , includes the following steps:

[0092] Step 61, for the fused feature after dynamic weight allocation of the self-attention mechanism, a long short-term memory network is used to model the time series relationship, capture the time series features, and obtain an output feature to better understand the dynamic changes of the sound signal in the fault occurrence process.

[0093] The LSTM is used to model the time series relationship of the fused feature after dynamic weight allocation of the self-attention mechanism, and the final output feature is a feature representation after time series modeling. These features can better reflect the dynamic changes of the sound signal in the fault occurrence process, including long-term dependencies and short-term dynamic changes, thereby improving the accuracy of fault detection and the robustness of the model.

[0094] AsFigure 9 The fusion feature is input into a long short-term memory network to perform time series relationship modeling, and the process is as follows:

[0095] Three gates that constitute a long short-term memory network (LSTM) are constructed, including a forgetting gate, an input gate and an output gate, to control the flow of information and capture long-time dependency. The specific formula is as follows:

[0096] (18) ;

[0097] wherein, is the output of the forgetting gate, is a sigmoid activation function, is a weight matrix of the forgetting gate, is a bias vector of the forgetting gate, is a hidden state of the previous time step, is an input feature of the current time step;

[0098] (19) ;

[0099] (20) ;

[0100] wherein, is the output of the input gate, is a candidate cell state, is a hyperbolic tangent activation function, and are weight matrices of the input gate and the candidate cell state respectively, and are bias vectors of the input gate and the candidate cell state respectively;

[0101] (21) ;

[0102] wherein, is a cell state of the current time step, is a cell state of the previous time step.

[0103] (22) ;

[0104] (23) ;

[0105] wherein, is the output of the output gate, is a hidden state of the current time step, is a weight matrix of the output gate, is a bias vector of the output gate.

[0106] ​Through the gating mechanism of the LSTM, the LSTM can process the input features time step by time , update the hidden state , and update the cell state . Finally, the output features of the LSTM contain dynamic change information on the sequence.

[0107] Step 62, the long short-term memory network (LSTM) output feature is processed through an activation operation, and then the output is converted into a probability distribution through a fully connected layer and a softmax function, to obtain a determination result of the type of the sound signal, which can include a fault or no fault;

[0108] Optionally, the activation operation can be a ReLU activation operation.

[0109] As shown in Figure 10 , a fully connected classification structure diagram is given. After the long short-term memory network output feature is processed through a ReLU activation function, the feature is mapped to a classification space through a fully connected layer, and the output is converted into a probability distribution through a softmax function, so as to determine the type of the sound signal.

[0110] Further technical solutions also include constructing a multi-network fusion deep learning network for implementing the processes of steps 1 to 6, including a feature extraction module, a fusion attention module, and a time sequence capturing and signal classification module.

[0111] The feature extraction module is configured to perform steps 1 and 2, including a multi-level feature extraction network and an envelope feature extraction module:

[0112] The multi-level feature extraction network is configured to perform multi-level feature extraction on the obtained sound signal sequence to be recognized , to obtain multi-scale features after extraction .

[0113] The envelope feature extraction module is configured to perform data envelope feature extraction on the obtained sound signal sequence to be recognized , to generate a trend score map according to the envelope line .

[0114] The fusion attention module is configured to perform a self-attention mechanism operation on the extracted multi-scale features , to generate a similarity score map between queries and keys and a value vector.

[0115] The gating fusion module is configured to fuse the trend score map and the similarity score map The spliced, convolutional and activated generated gating features control the fusion ratio of the similarity score map and the trend score map dynamically according to the gating features, and generate a hybrid weight map ; based on the hybrid weight map using multi-scale features The generated value vector is fused to obtain a fusion feature ;

[0116] The time sequence capture and signal classification module is configured to perform time sequence modeling and classification on the fusion feature to obtain a classification result of the sound signal; including a long short-term memory network, a full connection layer and an activation module; the activation module can use a Softmax function to activate and convert the output into a probability distribution;

[0117] Further, the method for training the multi-network fusion deep learning network comprises the following steps:

[0118] Step S1, obtaining cable fault discharge sound signals and non-discharge sound signals, and constructing a training data set after marking;

[0119] Optionally, the cable fault point instrument is used to collect sound signals, and cable fault discharge sound signals and non-discharge sound signals at different positions are collected;

[0120] In this embodiment, signals at two positions of the well cover and the ground underground are selected; the well cover is the maintenance installation or maintenance opening of the cable, and the sound signal is transmitted through the cavity under the well cover; the sound signal at the ground underground position is the sound signal transmitted through the soil; the data set constructed by the signals at different positions is used to verify the generalization performance of the model.

[0121] Specifically, the cable fault discharge sound signals and non-discharge sound signals at the well cover can be collected to construct a data set A; the cable fault discharge sound signals and non-discharge sound signals at the ground underground can be collected to construct a data set B. For the data set A and B, the cable fault discharge sound signals are marked as 1, and the non-discharge sound signals are marked as 0.

[0122] Figure 3 The waveform diagram of the cable fault discharge sound signal at the well cover is shown. Figure 4 The waveform diagram of the cable fault discharge sound signal at the ground underground is shown. Figure 5 The waveform diagram of the non-discharge sound signal is shown.

[0123] Step S2, dividing the data set into a training data set and a test data set according to a set ratio;

[0124] Specifically, the data set A and the data set B are divided into a training data set and a test data set according to a ratio of 7:3.

[0125] For data set A and data set B, each data set corresponds to 1260 groups of training samples, of which 630 groups are discharge sound and 630 groups are non-discharge sound; the test sample is 540 groups, of which 270 groups are discharge sound and 270 groups are non-discharge sound. The training data sets of data set A and data set B are used for training Figure 2 The parameters of the multi-network fusion deep learning network shown in the figure are used to test the recognition performance of the network model obtained by training.

[0126] Step S3, input the discharge sound and non-discharge sound signals in the training data set into the multi-feature interaction fusion deep network model, obtain the optimal network model through multiple iteration learning and model parameter optimization, including the following steps:

[0127] Step S31, multi-level feature extraction is performed on the discharge sound and non-discharge sound signals in the training data set to obtain the extracted multi-scale features;

[0128] Step S32, data envelope feature extraction is performed on the discharge sound and non-discharge sound signals in the training data set, and a trend score graph is generated according to the envelope line obtained;

[0129] Step S33, the extracted multi-scale features are subjected to self-attention mechanism operation to generate similarity score graph between query and key and value vector;

[0130] Step S34, the trend score graph and the similarity score graph are spliced, and convolution and activation are performed to generate gating features; the fusion ratio of the similarity score graph and the trend score graph is dynamically controlled according to the gating features to generate a mixed weight graph;

[0131] Step S35, based on the mixed weight graph, the value vector generated using the multi-scale features is fused to obtain the fusion features;

[0132] Step S36, after the fusion features are subjected to time series modeling and classification, the classification results of the sound signals are obtained;

[0133] Step S37, the loss value of the overall loss function is calculated according to the fault recognition result, and the model parameters of the multi-network fusion deep learning network are adjusted, and through multiple iteration training and back propagation, until the convergence condition is met to obtain the trained multi-network fusion deep learning network model;

[0134] Further, the loss function constructed includes a classification loss and a regularization loss:

[0135] (24);

[0136] The classification loss is:

[0137] (25);

[0138] This loss term is used to optimize the binary classification performance of the model. Wherein, is the true label, is the probability of positive class predicted by the model. By minimizing the difference between the predicted probability and the true label, this loss term can improve the classification accuracy of the model in the fault detection task.

[0139] The regularization loss is:

[0140] (26);

[0141] This loss term prevents the model from overfitting by L2 regularization. Wherein, is the parameter of the model, is the regularization coefficient. By adding a regularization term to the loss function, the complexity of the model can be effectively reduced, and the generalization ability of the model can be improved.

[0142] Further, using the test sound data set, the multi-network fusion deep learning network model is tested, and the model performance evaluation index is calculated, and the calculation formula is as follows:

[0143] (27);

[0144] Wherein, TP represents the number of samples correctly judged as fault discharge sound signal, TN represents the number of samples correctly judged as non-discharge sound signal, FP represents the number of samples incorrectly judged as non-discharge sound signal, and FN represents the number of samples incorrectly judged as non-discharge sound signal.

[0145] Table 1 and Table 2 respectively show the recognition effect of the optimal training model on the cable fault discharge sound of a single laying medium. Among them, Table 1 is to use the training samples in data set A to construct the model, and use the test samples in data set A to verify the recognition effect. Table 2 is to use the training samples in data set B to construct the model, and use the test samples in data set B to verify the recognition effect.

[0146] Table 1 uses test samples in data set A to verify the recognition effect;

[0147]

[0148] Table 1 gives the confusion matrix of the multi-feature interaction fusion deep network model for recognizing the cable fault discharge sound and non-discharge sound signal at the well cover. Finally, after ten cross experiments, the average recognition accuracy of the sound is 97.03%.

[0149] Table 2 uses test samples in data set B to verify the recognition effect;

[0150]

[0151] Table 2 shows the confusion matrix of one experiment when the multi-feature interaction fusion deep network model is used for the identification of cable fault discharge sound and non-discharge sound signals at the manhole cover. Finally, through ten cross-validation experiments, the average accuracy of sound recognition is 98.11%.

[0152] Table 3 shows the effect of power cable fault discharge sound identification across laying media, i.e. using training samples in dataset A to build the model and using test samples in dataset B to verify the identification effect, verifying the generalization ability of the multi-network fusion deep learning network architecture of the embodiment in sound recognition.

[0153] Table 3 is the test sample in dataset B to verify the identification effect of the model built by the training sample in dataset A;

[0154]

[0155] Table 3 shows the multi-feature interaction fusion deep network model using cable fault discharge sound and non-discharge sound at the manhole cover for model training and optimization; then, the trained model is used for generalization experiments of ground and underground cable fault discharge sound and non-discharge sound recognition. The confusion matrix of one experiment is shown. Finally, through ten cross-validation experiments, the average accuracy of sound recognition is 87.07%.

[0156] Ablation experiments were conducted to compare the individual effects of the proposed "multi-level feature extraction network" and "signal envelope trend score" in dataset A and dataset B, and the multi-scale feature and trend score chart The accuracy after fusion is shown in Table 4.

[0157] Table 4 shows the ablation test results.

[0158]

[0159] From the check mark in Table 4, it can be seen that when only the basic model (i.e. the convolutional neural network CNN and LSTM of the embodiment) is used, the accuracy of the model on dataset A and dataset B is 90.56% and 89.72% respectively, and the performance is relatively low.

[0160] After introducing the multi-level feature extraction network, the accuracy is significantly improved to 95.43% (dataset A) and 96.25% (dataset B), indicating that this module can effectively mine deeper feature information.

[0161] When the signal envelope trend score is extracted by using the envelope feature extraction module alone, the accuracy reaches 94.32% and 93.35%, which is also higher than the basic model, indicating that the feature can provide important discriminant information.

[0162] On this basis, the fusion attention mechanism is introduced, and the multi-scale features and the trend score are comprehensively utilized, and the accuracy of the model on dataset A and dataset B reaches 97.03% and 98.11% respectively, achieving the best performance.

[0163] In summary, the ablation experiment fully shows that the multi-level feature extraction network and the signal envelope trend score play an important role in improving the performance of the model, and the fusion attention mechanism further enhances the comprehensive utilization ability of the model to different features, significantly improving the accuracy and robustness of fault detection.

[0164] The power cable fault discharge sound intelligent recognition method proposed in this embodiment can realize automatic extraction of multi-level and multi-type features of cable discharge sound signals, can highlight local detail features and global trend features, weaken redundant features, and can model long-time dependence of sound signals, overcome the influence of insufficient generalization ability of classical sound signal description methods and recognition methods, realize the intelligentization of power cable fault discharge sound recognition, avoid the subjective influence brought by relying on the experience of test personnel, and provide a powerful guarantee for the safe and stable operation of the power system.

[0165] Embodiment 2

[0166] Based on embodiment 1, the power cable fault sound recognition system based on multi-network fusion is provided in this embodiment, which includes:

[0167] The multi-level feature extraction network is configured to perform multi-level feature extraction on the acquired sound signal sequence to be recognized to obtain multi-scale features after extraction .

[0168] The envelope feature extraction module is configured to perform data envelope feature extraction on the acquired sound signal sequence to be recognized to generate a trend score graph according to the envelope line obtained .

[0169] The fusion attention module is configured to perform self-attention mechanism operation on the extracted multi-scale features to generate a similarity score graph between the query and the key and a value vector

[0170] The gating fusion module is configured to fuse the trend score graph and the similarity score graph The spliced images are subjected to convolution and activation to generate gating features; a hybrid weight map is generated based on the gating features and a dynamic control of the fusion ratio of the similarity score map and the trend score map ; and the hybrid weight map is used to fuse the value vectors generated using the multi-scale features to obtain fused features

[0171] The time series capture and signal classification module is configured to perform time series modeling and classification on the fused features to obtain a classification result of the sound signal.

[0172] It should be noted that each module in this embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0173] Embodiment 3

[0174] Based on Embodiment 1, this embodiment provides a power cable fault sound recognition system based on multi-network fusion, which comprises a sound collecting device and a processor.

[0175] The sound collecting device can be a cable fault point instrument, which is used to collect fault discharge sound signals and non-discharge sound signals.

[0176] The processor is configured to perform the steps of the power cable fault sound recognition method based on multi-network fusion described in Embodiment 1.

[0177] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0178] The above describes the specific embodiments of the present application in conjunction with the accompanying drawings, but is not intended to limit the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.​​

Claims

1. A power cable fault sound recognition method based on multi-network fusion, characterized in that, Comprising the steps of: To the acquired to be identified sound signal sequence Multi-level feature extraction is performed to obtain the extracted multi-scale features ; To the acquired sound signal sequence to be recognized Carrying out data envelope feature extraction, and generating a trend score graph according to the obtained envelope line ; extracted multi-scale features performing a self-attention mechanism operation to generate a similarity score map between the query and the key and a value vector; a trend score map a similarity score map splicing, convolution and activation to generate a gating feature; dynamically control the fusion ratio of the similarity score map and the trend score map according to the gating feature to generate a hybrid weight map ; Hybrid weight map based To use multi-scale features The generated value vectors are fused to obtain fused features ; For fused features After timing modeling and classification, a classification result of the sound signal is obtained. Extracting multi-scale features The multi-level feature extraction network is implemented by constructing a multi-level feature extraction network, which includes a first CR module, a splitting module, a light processing branch and a deep extraction branch, a splicing module and a second CR module connected in turn; two CR modules include convolution layers and activation layers, which perform convolution and activation operations on input features; Multi-scale features are extracted by a method comprising the steps of: The acquired to-be-recognized sound signal sequence Convolution and activation processing are performed, and the features are split along the channel dimension and input to the lightweight processing branch and the deep extraction branch. In the deep extraction branch, the input features are encoded by a deep progressive multi-layer coding based on a convolutional neural network to obtain multi-scale information, and a skip connection block is used to connect the multi-scale information to obtain multi-level features; The multi-level features obtained by the deep extraction branch and the bottom information features obtained by the lightweight processing branch are spliced in the channel dimension, and the spliced features are reduced in dimension to obtain fused multi-scale features ; Dynamically control the proportion of the similarity score map and the trend score map using a gating feature, and generate a hybrid weight map using convolution , as follows: ； ； wherein, , represents a weight matrix of convolution, and represents a bias of convolution, denotes a dot product operation, , gate represents a gating feature. 2.The multi-network convergence based power cable fault sound recognition method of claim 1, wherein, Trend score graph The generation method of the trend score graph is as follows: The acquired to-be-recognized sound signal sequence A Hilbert transform is performed, and then envelope feature extraction is performed to obtain an envelope line; A trend score map is generated using the envelope line. 3.The multi-network convergence based power cable fault sound recognition method of claim 1, wherein, The obtained multi-scale features performing a self-attention mechanism operation, including the following steps: mapping the multi-scale features into matrices, including a query matrix , a key matrix and a value matrix ; The mapped matrix is linearly transformed by a learnable weight matrix; The similarity scores between the query and the keys are computed using dot product to calculate the similarity of each query vector with all key vectors . 4.The power cable fault sound recognition method based on multi-network fusion according to claim 1, characterized in that: For fusing features A method of temporal modeling and classification, comprising the steps of: The long short-term memory network is used for dynamically allocating weights to the fusion features of the self-attention mechanism Time series relationship modeling is performed to capture time sequence characteristics and obtain output features ; Long short-term memory network output features After the activation operation processing, the output is converted into a probability distribution through a fully connected layer and a softmax function to obtain a determination result of the type of the sound signal.

5. The multi-network fusion-based power cable fault sound recognition system according to claim 1, characterized in that, Comprising: The multi-level feature extraction network is configured to perform multi-level feature extraction on an acquired sound signal sequence to be recognized to obtain multi-scale features after extraction ; an envelope feature extraction module configured to extract envelope features from the obtained sound signal sequence performing data envelope feature extraction, generating a trend score map from the obtained envelope ; a fusion attention module configured to fuse the extracted multi-scale features performing a self-attention mechanism operation to generate a similarity score map between the query and the keys and a value vector; The gating fusion module is configured to fuse the trend score map with the similarity score map to generate a gating feature; dynamically control the fusion ratio of the similarity score map and the trend score map according to the gating feature, and generate a hybrid weight map ; fuse the value vector generated based on the hybrid weight map using the multi-scale feature to obtain a fused feature ; The time series capture and signal classification module is configured to perform time series modeling and classification on the fused features After the time series modeling and classification, a classification result of the sound signal is obtained.

6. A power cable fault sound recognition system based on multi-network fusion, characterized by, The sound acquisition device and the processor; The processor is configured to perform the steps of the power cable fault sound recognition method based on multi-network fusion according to any one of claims 1-4. ​

Images