A multi-task joint high-voltage disconnector fault diagnosis method

Abstract

The application discloses a kind of high-voltage disconnecting switch fault diagnosis methods of multi-task combination, first utilize singular value decomposition (SVD) technique to the vibration signal of collection is carried out redundancy processing;Through short-time Fourier transform (STFT), one-dimensional vibration signal is converted into two-dimensional time-frequency feature map more conducive to feature convolutional neural network (CNN) extraction;Then utilize multi-branch convolutional neural network to build a multi-task joint fault diagnosis network with domain feature sharing mechanism, while detecting the fault degree and fault type of high-voltage disconnecting switch;According to signal sample characteristics, design a kind of attention module combining spatial attention and channel attention, realize fast and efficient fault feature extraction, obtain multi-task joint fault diagnosis model;With the high-voltage disconnecting switch state obtained by multi-task joint fault diagnosis, the training condition is visualized operation, and the current fault stage of current high-voltage disconnecting switch is diagnosed, timely decision and feedback are made.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of high-voltage disconnector health management, operation and maintenance, and in particular to a multi-task joint high-voltage disconnector fault diagnosis method. BACKGROUND

[0002] In recent years, the power grid construction has developed rapidly, and in-depth and comprehensive promotion of power system fault diagnosis research can better protect the safety of people's production and life. With the advent of the big data era, online monitoring and fault diagnosis technology can effectively warn of equipment failure, play a role in preventing trouble, and at the same time, the development of online fault diagnosis liberates labor and reduces the probability of workers being injured when detecting outside. With the advent of the big data era, online monitoring and fault detection technology can effectively warn of equipment failure, thereby improving the level of power grid equipment maintenance. Domestic and foreign actual operation shows that power equipment online monitoring technology has a wide application prospect.

[0003] However, at present, only the fault diagnosis technology of part of the power equipment develops rapidly, such as transformers, capacitors, etc., and the research on high-voltage disconnector fault detection is less, which becomes a short board of power system safety. High-voltage disconnector not only ensures normal power supply, but also can isolate the fault when the system fails, so fault monitoring of high-voltage disconnector has important practical value.

[0004] Since the opening and closing of the disconnector directly affects the power supply problem of residents and enterprises, sometimes it is necessary to continue to work the disconnector when the fault is detected, and then carry out power-off maintenance at a certain time node, so when detecting the fault, not only the type of the fault occurring needs to be detected, but also the fault degree of the circuit breaker needs to be determined, so as to determine whether it needs to be forced to stop and repair.

[0005] Aiming at the typical problems of high-voltage disconnectors, such as low voltage 85% UN, high voltage 110% UN, main knife linkage loose, inter-pole link loose, inter-phase link loose, base loose, lock loose, compound fault, unknown fault, etc., Chen Shigang et al. collect mechanical vibration signals through sensors installed on high-voltage disconnectors, and analyze the vibration signals by using wavelet denoising processing, empirical mode decomposition, clustering, support vector machine and BP neural network algorithm to diagnose the mechanical faults existing in high-voltage disconnectors. Liu Shibing et al. establish the functional relationship between the motor stator current fluctuation and the torque fluctuation, filter the current, and use the support vector machine to build a mechanical fault diagnosis model of high-voltage disconnector. However, the current researches mainly deal with the fault types and damage degrees of high-voltage disconnectors respectively. However, these researches in single-label system only diagnose one type of problem at a time. In the context of big data, this single-label system breaks the connection between different fault problems of high-voltage disconnectors, reduces the utilization rate of information, and has low fault diagnosis accuracy. In addition, the working environment of high-voltage disconnector is harsh, and there are many interferences. The existing researches cannot stably and accurately detect the multi-task fault. SUMMARY

[0006] In view of the shortcomings of the traditional high-voltage disconnector fault diagnosis model, the present application designs a multi-task joint high-voltage disconnector fault diagnosis method, which jointly diagnoses the fault type and fault degree of the disconnector. The multi-task joint network supervises each other, and the layers and modules share knowledge according to certain rules and strategies, and the attention mechanism is integrated.

[0007] To achieve the above purpose, the present application provides the following technical scheme:

[0008] A multi-task joint high-voltage disconnector fault diagnosis method, comprising the following steps:

[0009] Step 1: Use singular value decomposition (SVD) technology to process the vibration signals collected by the sensor to reduce the redundant information in the signals.

[0010] Use singular value decomposition (SVD) technology to analyze the vibration signals collected by the sensor, use phase space reconstruction theory to construct an m*n order Hankel matrix, and assume that the vibration signals collected by the sensor are x(i)(i=1, 2,..., N), then the Hankel matrix is as follows:

[0011]

[0012] In formula (1), m+n-1=N, N is the total length of the vibration signal, and A is the Hankel matrix. Singular value decomposition is performed on A to obtain:

[0013]

[0014] where: U is an m-order orthogonal matrix, V is an n-order orthogonal matrix, ∑ = diag(σ1, σ2,..., σ r )∑ = diag(σ1, σ2,..., σ r ) is an r-order diagonal matrix, and σ r is a non-zero singular value of the matrix A, O is a zero matrix, and when the SVD method is used for signal processing, the singular value decomposition of A is written in a simplified vector form:

[0015]

[0016] where S = [diag(σ1, σ2,..., σ r ), O] or its transpose, depending on the size of m and n, S ∈ R m×n , O is a zero matrix, R m×n is an m × n order real number, A i ∈ R m×n , U i ∈ R m×m , V i ∈ R n×n , i = 1, 2,..., r, r = min(m, n);

[0017] x(i) is composed of useful signals containing fault information and redundant useless information, and the singular value obtained by decomposing the matrix A is λ, which reflects the concentration of the energy of the useful signals and the redundant useless information; the singular value size represents the contribution degree of the signal, and a smaller singular value represents redundant useless information, which is set to zero, and the inverse process of formula 2 is used to reconstruct the vibration signal to remove the redundant signal;

[0018] Step 2: convert the one-dimensional vibration signal into a two-dimensional time-frequency feature map more suitable for feature convolutional neural network (CNN) extraction by short-time Fourier transform (STFT);

[0019] Since the working environment of the disconnecting switch is complex and variable, the collected vibration signal is not smooth, so after obtaining the one-dimensional time-domain vibration signal, it is converted into a two-dimensional time-frequency map by using the STFT technology, effectively extracting the fault features in the non-stationary signal, and the subsequent fault detection system uses a two-dimensional convolutional neural network. Under the same conditions, the convolutional neural network processes two-dimensional data, which is generally better than one-dimensional data. STFT uses a time window to slide on the time-domain signal of the vibration signal, cuts the time-domain signal into multiple segments, respectively performs Fourier transform on each segment, and finally obtains the frequency domain characteristics at each time. Let the window function be r(t), and the short-time Fourier transform of the signal x(t) is defined as:

[0020]

[0021] wherein f represents frequency (unit: Hz), and is an arbitrary real number; t represents time (unit: s), and r*(τ-t) is a time slice near time t;

[0022] Step 3: a multi-task joint fault diagnosis network with a domain feature sharing mechanism is built by using a multi-branch convolutional neural network, and the fault degree and fault type of the high-voltage disconnecting switch are detected; and high-voltage disconnecting switch fault signal sample features are obtained;

[0023] The multi-task joint fault diagnosis network connects two networks through a domain sharing module to form a layer-to-layer and module-to-module knowledge sharing according to certain rules and strategies, automatically determines the shared layer for different tasks, and classifies the collected high-voltage disconnecting switch fault signals and judges the fault degree through end-to-end learning; the calculation process of the domain feature sharing unit is shown in formula 5; let the feature maps of two tasks be x1 and x2, and the feature maps after feature sharing of the domain feature sharing unit for different tasks be and α AB and α BA represent the sharing weights between different tasks, and α AA and α BB represent the sharing weights between the same tasks; the weight matrix encodes the relationship between the two tasks, and the sharing degree is set by training α AB , α BA , α AA and α BB ; the greater the allocated weight value, the higher the sharing degree of the feature plane of the two tasks of the layer;

[0024]

[0025] The main body of the two networks is a convolutional neural network, which includes a convolutional layer, a pooling layer, a fully connected layer and a softmax classification layer; the convolutional layer is a kind of feedforward neural network with deep structure, which includes convolutional calculation, and the main function of the convolutional layer is to extract fault features by sliding the convolutional kernel on the feature plane; the convolutional layer includes multiple convolutional kernels, each of which corresponds to a weight coefficient and a bias amount, and the feature image of the next layer is extracted after the feature of the previous layer image is extracted by moving the convolutional kernel, and the calculation method is as follows:

[0026]

[0027] The summation part in the formula is equivalent to solving a cross-correlation, b is a bias amount, y is an output result of a previous neuron, ω is a weight, and Z l (i,j) and Z l+1(i,j) represents the convolutional input and output of the (l+1)th layer, where L l+1 For Z l+1 Given the size of the feature map, assuming the feature map has the same length and width, Z(i,j) corresponds to the pixel in the i-th row and j-th column of the feature map, and K... l Let f be the number of channels in the l-th feature map, and s0 and p be the convolutional layer parameters, corresponding to the kernel size, stride, and number of padding layers, respectively. The convolutional layer uses the ReLU function, which is expressed as:

[0028] σ r (x)=max(0,x) (7)

[0029] Pooling layers compute the nth feature map of the l-th pooling layer. It manifests as:

[0030]

[0031] It is the nth input mapping, which is the nth input mapping of the previous convolutional layer. pool() is the max pooling equation, p is the pooling size, and s is the stride. The fully connected layer transforms the two-dimensional image into one-dimensional data, flattens the data, and achieves the final classification through the final Softmax layer.

[0032] Step 4: After obtaining the fault signal sample features of the high-voltage disconnector through the convolutional neural network, an attention module combining spatial attention and channel attention is designed to achieve fast and efficient fault feature extraction; the final multi-task joint fault diagnosis network model is obtained.

[0033] The attention module that combines spatial attention and channel attention specifically includes a spatial attention module and a channel attention module.

[0034] For a feature map F in a multi-task network, the channel attention mechanism focuses on what is meaningful in the input feature map. It uses max pooling and average pooling to compress the feature map in spatial dimension, obtaining two feature descriptors. These two feature descriptors are then fed into the elements of a multilayer perceptron (MLP) and accumulated one by one. Through activation operations, the output channel attention feature M is obtained. C (F), the calculation process is shown in equation (9), and it is multiplied with the original feature map to perform adaptive feature refinement, and finally a new feature map F' is obtained;

[0035] The spatial attention channel primarily focuses on location information. First, it uses max pooling and average pooling to obtain two distinct feature maps. These two feature maps are then merged, followed by convolution and activation operations to obtain feature map M. C(F'), the calculation process is shown as formula (10), the obtained result is multiplied by the input feature map, and finally the required feature F'' is obtained; the whole process is shown as formula (11), wherein is multiplied with the corresponding element;

[0036] M c (F) = sigma (MLA (Pvgpool (F)) + MLA (Maxpool (F))) (9)

[0037] M s (F') = sigma (f 7*7 ([AvgPool (F) ; MaxPool (F)])) (10)

[0038]

[0039] Step 5: The multi-task joint fault diagnosis model established above is trained by using the vibration signal, the training condition is visualized, and then the state of the detected high-voltage disconnecting switch is analyzed, and the running stage of the current high-voltage disconnecting switch in the whole life cycle is diagnosed, so that decisions and feedbacks can be made in time.

[0040] The trained multi-task joint high-voltage disconnecting switch fault diagnosis is used to detect the fault type and fault degree of the high-voltage disconnecting switch at the same time, when there is no fault, the model outputs the current state, and no alarm is issued, when a fault occurs, but the degree is lower than the threshold, the system issues an alarm, prompts the current state and the fault type, when the fault degree is higher than the threshold, that is, the fault degree reaches the critical point, the model alarm is prominent, and the manual participation is informed, so that the problem can be solved in time and major losses can be avoided.

[0041] The present application has the beneficial technical effects:

[0042] In view of the prior art, various faults are diagnosed respectively, and in the big data background, the connection between different fault problems of mechanical equipment is broken, the model can analyze the fault type and the fault degree of the high-voltage disconnecting switch at the same time, the various fault features are exchanged through the domain feature sharing layer, the features are mutually referred, the connection between various problems is strengthened, the feature extraction capability is improved, the prediction accuracy and stability are better than those of the prior art, and the feature attention mechanism is added, so that the fault diagnosis efficiency is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 It is a flow chart of the multi-task joint high-voltage disconnecting switch fault diagnosis method of the present application;

[0044] Figure 2 It is a technical roadmap of the multi-task joint high-voltage disconnecting switch fault diagnosis method of the present application;

[0045] Figure 3 A domain feature sharing module diagram proposed for the present application;

[0046] Figure 4 An attention module diagram proposed for the present application;

[0047] Figure 5 A high-voltage disconnecting switch training situation result visualization diagram proposed for the present application. DETAILED DESCRIPTION

[0048] The establishment process and training result visualization of the present application, a multi-task joint high-voltage disconnecting switch fault diagnosis method, are described below in combination with the drawings and examples.

[0049] In traditional high-voltage disconnecting switch fault diagnosis research, much attention is paid to high-voltage disconnecting switch fault types, and few to both high-voltage disconnecting switch fault type judgment and high-voltage disconnecting switch damage degree judgment. In actual production, most enterprises pursue maximum benefit, and most high-voltage disconnecting switches are replaced under the condition that they cannot be replaced. Therefore, an improved multi-task joint high-voltage disconnecting switch fault diagnosis method is proposed, which can diagnose the fault type and give the damage degree of the high-voltage disconnecting switch at the same time. Users can decide when to replace the high-voltage disconnecting switch according to the fault type and the corresponding damage degree to achieve maximum benefit.

[0050] A multi-task joint high-voltage disconnecting switch fault diagnosis method, the overall flow chart is as shown in Figure 1 Figure 2 The technical roadmap of the multi-task joint high-voltage disconnecting switch fault diagnosis method in the present application is as shown in Figure 2 It is a multi-task joint high-voltage disconnecting switch fault diagnosis method based on convolutional neural network, which can diagnose the fault type and the damage degree of the disconnecting switch at the same time. The two tasks are performed simultaneously, and the fault feature sharing is performed in the middle layer of the neural network, which strengthens the connection between various fault diagnosis tasks and improves the generalization ability of the model and the accuracy of fault diagnosis. The method includes the following steps:

[0051] Step 1: Use singular value decomposition (SVD) technology to process the vibration signals of the disconnecting switch collected by the sensor to reduce the amount of redundant information in the signals.

[0052] Use singular value decomposition (SVD) technology to analyze the vibration signals collected by the sensor, use phase space reconstruction theory to construct an m×n order Hankel matrix, and assume that the vibration signals collected by the sensor are x(i)(i=1,2,...,N). The Hankel matrix is as follows: ​

[0053]

[0054] In formula (1), m+n-1=N, N is the total length of the vibration signal, A is a Hankel matrix, and singular value decomposition is performed on A to obtain:

[0055]

[0056] wherein U is an m-order orthogonal matrix, V is an n-order orthogonal matrix, ∑=diag(σ1,σ2,...,σ r )∑=diag(σ1,σ2,...,σ r ) is an r-order diagonal matrix, and σ r is a non-zero singular value of the matrix A, and O is a zero matrix; when the SVD method is used for signal processing, the singular value decomposition of A is written in a simplified vector form:

[0057]

[0058] wherein S=[diag(σ1,σ2,...,σ r ),O] or its transpose, depending on the size of m and n, S∈R m×n , O is a zero matrix, R m×n is an m*n-order real number, A i ∈R m×n , U i ∈R m×m , V i ∈R n×n , i=1,2,...,r, and r=min(m,n);

[0059] x(i) is composed of useful signals containing fault information and redundant useless information, the singular value obtained by decomposing the matrix A is λ, and then the concentration of the energy of the useful signals and the redundant useless information is reflected; the singular value size represents the contribution degree of the signal, and a smaller singular value represents redundant useless information, the value of which is set to zero, the vibration signal is reconstructed by using the inverse process of formula 2, and the redundant signal is removed;

[0060] Step 2: converting the one-dimensional vibration signal into a two-dimensional time-frequency feature map more conducive to feature convolutional neural network (CNN) extraction by short-time Fourier transform (STFT);

[0061] Because the working environment of disconnectors is complex and changeable, the collected vibration signals are unstable, so the one-dimensional time-domain vibration signals are obtained, and the STFT technology is used to convert them into two-dimensional time-frequency diagrams, effectively extracting the fault features in the non-stationary signals, and the subsequent fault detection system uses a two-dimensional convolutional neural network, which is generally better than one-dimensional data in the same situation; STFT uses a time window to slide on the time-domain signal of the vibration signal, cuts the time-domain signal into multiple segments, respectively performs Fourier transform, and calculates the frequency domain characteristics of each segment, and finally obtains the frequency domain characteristics at each time; Let the window function be r(t), the short-time Fourier transform of the signal x(t) is defined as:

[0062]

[0063] Where f represents the frequency (unit: Hz), which is any real number; t represents time (unit: s), r*(τ-t) is the time slice near time t;

[0064] Step 3: A multi-task joint fault diagnosis network with domain feature sharing mechanism is built by using a multi-branch convolutional neural network, which simultaneously detects the fault degree and fault type of the high-voltage disconnector;

[0065] Since disconnectors directly affect the power supply problem of residents and enterprises, not only the fault type needs to be detected, but also the fault degree of the circuit breaker needs to be determined to determine whether forced shutdown repair is needed. Therefore, an improved multi-task deep neural network is proposed, which performs two tasks simultaneously and shares fault features through a shared module to strengthen the connection between various fault diagnosis tasks and improve the generalization ability of the model and the accuracy of fault diagnosis.

[0066] The multi-task joint fault diagnosis network connects the two networks through a domain sharing module, forms a layer-to-layer and module-to-module knowledge sharing according to certain rules and strategies, and automatically learns the features that need to be shared by adding a domain feature sharing module between the feature layers of the two networks. The domain feature sharing module is a coefficient matrix; different tasks automatically determine the shared layers, and through end-to-end learning, the collected high-voltage disconnector fault signals are classified and the fault degree is judged; the calculation process of the domain feature sharing unit is shown in formula 5; let the feature maps of the two tasks be x1 and x2, and the feature maps after feature sharing of the feature maps of different tasks by the domain feature sharing unit are and α AB and α BA represent the shared weights between different tasks, and α AA and α BBrepresent the same task; the weight matrix encodes the relationship between two tasks, and the degree of sharing is set by training AB , α BA , α AA and α BB , the greater the assigned weight indicates that the two tasks of the layer have a higher degree of sharing of feature planes;

[0067]

[0068] The specific formula of the learning strategy of the module is as follows:

[0069]

[0070]

[0071] The main body of the two networks is a convolutional neural network, which includes convolutional layers, pooling layers, fully connected layers and softmax classification layers; the convolutional layer is a kind of feedforward neural network with deep structure, which includes convolutional calculation, and the convolutional layer can extract fault features by sliding the convolutional kernel on the feature plane;

[0072] 1) Convolutional layer. The convolutional layer is a kind of feedforward neural network with deep structure, which includes convolutional calculation, and the convolutional layer can extract fault features. The convolutional layer includes multiple convolutional kernels, each of which corresponds to a weight coefficient and a bias, and the feature image of the next layer is extracted by moving the convolutional kernel to the feature image of the previous layer. The calculation method is as follows:

[0073]

[0074] The summation part in the formula is equivalent to solving a cross-correlation, B is the bias, and Z l and Z l+1 represent the convolutional input and output of the l+1 layer, L l+1 is the size of Z l+1 , it is assumed that the feature map has the same length and width, Z(i,j) corresponds to the pixels of the feature map, K is the number of channels of the feature map, f, s0 and p are the parameters of the convolutional layer, corresponding to the size of the convolutional kernel, the convolutional step and the padding layer. The convolutional layer uses the Relu function, which is expressed as:

[0075] σ r (x)=max(0,x) (9)

[0076] 2) Pooling layer. The role of the pooling layer is to reduce the dimension of the data and reduce the amount of calculation, which is equivalent to a filter. The pooling region of the pooling layer is selected in a similar way to the convolutional kernel scanning the feature map, which is controlled by the pooling size, step and padding. Mathematically, the nth feature map of the lth pooling layer can be calculated as follows: may be expressed as:

[0077]

[0078] is the n-th input map, that is, the n-th input map of the previous convolutional layer, pool() is the maximum pooling equation, p is the pooling size, and s is the step size.

[0079] 3) Full connection layer. It is responsible for converting the two-dimensional feature map output by convolution into a one-dimensional vector, integrating the feature representation into a value, thereby realizing an end-to-end learning process, which has the advantages of reducing the influence of feature position on classification results and improving the robustness of the entire network.

[0080] 4) Softmax classification layer. Softmax is used in the multi-classification process, which maps the outputs of multiple neurons of the full connection layer to the (0, 1) interval to form a probability value output, thereby realizing the classification of the fault form and damage degree of the high-voltage disconnecting switch.

[0081] Step 4: After obtaining the signal sample features by the convolutional neural network, an attention module combining spatial attention and channel attention is designed to realize fast and efficient fault feature extraction; Figure 3 is a structural diagram of the domain feature sharing module in the present application. A final multi-task joint fault diagnosis network model is obtained;

[0082] Figure 4 is a structural diagram of the attention mechanism model in the present application. When a multi-task neural network is used to analyze the time-frequency image, the information in the figure does not have an effect on all regions, such as the fault type of the high-voltage disconnecting switch, which should pay attention to the position of the characteristic frequency line, and the fault degree needs to be observed. The color of the characteristic frequency position in the time-frequency image, the deeper the color, the greater the amplitude.

[0083] Therefore, the present application sets multiple CBAM (Convolutional Block Attention Module) attention mechanisms in the multi-task deep learning network, as shown in Figure 4 , so that the network pays more attention to useful information and pays less attention to useless features and interference information, which not only enables the multi-task neural network to more effectively extract more expressive features, thereby improving the accuracy of fault diagnosis, but also reduces the calculation amount and improves the efficiency of fault diagnosis.

[0084] For one feature map F of the multi-task network, the channel attention mechanism of CBAM mainly focuses on what is meaningful in the input feature map. It uses max pooling and average pooling to compress the feature map in the spatial dimension, obtains two feature descriptors, and then adds the elements of the two feature descriptors one by one to accumulate through a multi-layer perception (MLP), and outputs the channel attention feature M through an activation operation C (F), the calculation process is shown in formula (11), and is multiplied by the original feature map to perform adaptive feature refinement, and finally a new feature map F' is obtained. The spatial attention channel mainly focuses on position information. First, two different feature maps are obtained by using max pooling and average pooling, the two feature maps are combined, and then convolution operation and activation operation are performed to obtain a feature map M C (F′), the calculation process is shown in formula (12), and the result is multiplied by the input feature map to finally obtain the required feature F". The whole process is shown in formula (13), wherein is multiplied by the corresponding elements.

[0085] M c (F)=σ(MLP(Avgpool(F))+MLP(Maxpool(F))) (11)

[0086] M s (F′)=σ(f 7*7 ([AvgPool(F);MaxPool(F)])) (12)

[0087]

[0088] Step 5: The multi-task joint fault diagnosis model established above is trained by using the vibration signal, the training condition is visualized, and then the state of the high-voltage disconnecting switch being detected is analyzed, and the running stage of the high-voltage disconnecting switch in the whole life cycle is diagnosed, so that decisions and feedbacks can be made in time.

[0089] The trained multi-task joint high-voltage disconnecting switch fault diagnosis is used to detect the fault type and fault degree of the high-voltage disconnecting switch at the same time. When there is no fault, the model outputs the current state and does not issue an alarm. When a fault occurs, but the degree is lower than the threshold, the system issues an alarm to prompt the current state and the fault type. When the fault degree is higher than the threshold, that is, the fault degree reaches the critical point, the model alarm is highlighted, and manual participation is notified to solve the problem in time and avoid causing significant losses.

[0090] Figure 5The present application adopts the T-SNE (T-Distributed Stochastic Neighbor Embedding) method to reduce the dimension of the high-dimensional classification result data and perform visualization, so that the training condition can be better identified, and the training method can be properly adjusted to obtain better training effect.

Claims

1. A multi-task collaborative fault diagnosis method for high-voltage disconnect switches, characterized in that, Includes the following steps: Step 1: Use Singular Value Decomposition (SVD) technology to perform redundancy removal processing on the vibration signal of the disconnector acquired by the sensor, reducing a large amount of redundant information in the signal; Step 2: The one-dimensional vibration signal is converted into a two-dimensional time-frequency feature map that is more conducive to feature extraction by the convolutional neural network (CNN) through short-time Fourier transform (STFT); Step 3: Build a multi-task joint fault diagnosis network with a domain feature sharing mechanism using a multi-branch convolutional neural network, and detect the fault degree and fault type of the high-voltage disconnector switch; obtain the fault signal sample features of the high-voltage disconnector switch. Step 4: After obtaining the fault signal sample features of the high-voltage disconnector through the convolutional neural network, an attention module combining spatial attention and channel attention is designed to achieve fast and efficient fault feature extraction. The final multi-task joint fault diagnosis network model is obtained; The attention module that combines spatial attention and channel attention specifically includes a spatial attention module and a channel attention module. For a feature map F in a multi-task network, the channel attention mechanism focuses on what is meaningful in the input feature map. It uses max pooling and average pooling to compress the feature map in spatial dimension, obtaining two feature descriptors. These two feature descriptors are then fed into the elements of a multilayer perceptron (MLP) and accumulated one by one. The output channel attention feature is then generated through activation operations. The calculation process is shown in equation (9), and the result is multiplied with the original feature map to perform adaptive feature refinement, finally yielding a new feature map. ; The spatial attention channel primarily focuses on location information. First, it uses max pooling and average pooling to obtain two distinct feature maps. These two feature maps are then merged, followed by convolution and activation operations to obtain the final feature map. The calculation process is shown in equation (10). The result is multiplied by the input feature map to finally obtain the required features. The entire process is shown in equation (11), where Multiply corresponding elements in the same position; （9）； （10）； （11）； Step 5: Use vibration signals to train the multi-task joint fault diagnosis model established above, visualize the training results, analyze the status of the detected high-voltage disconnect switch, diagnose the current operating stage of the high-voltage disconnect switch in its entire life cycle, and make timely decisions and feedback. The trained multi-task joint high-voltage disconnector fault diagnosis system simultaneously detects the fault type and severity of the high-voltage disconnector. When there is no fault, the model outputs the current status without issuing an alarm. When a fault occurs but the severity is below the threshold, the system issues an alarm, indicating the current status and fault type. When the fault severity exceeds the threshold, i.e., the fault severity reaches the critical point and shutdown for maintenance is necessary, the model alarm is highlighted, notifying human intervention to resolve the problem promptly and avoid significant losses.

2. The multi-task collaborative high-voltage disconnector fault diagnosis method according to claim 1, characterized in that, Step 1 is as follows: The vibration signals acquired by the sensor are analyzed using singular value decomposition (SVD) technology, and phase space reconstruction theory is used to construct... The order Hankel matrix, assuming the vibration signal collected by the sensor is... The Hankel matrix is ​​as follows: （1）； In equation (1), , The total length of the vibration signal. For the Hankel matrix, Singular value decomposition yields: （2）； Where: U is an m-order orthogonal matrix, and V is an n-order orthogonal matrix. It is an r-order diagonal matrix, and Let O be the non-zero singular values ​​of matrix A, and let O be the zero matrix. When using the SVD method for signal processing, the singular values ​​of A are decomposed into a simplified vector form: （3）； in, Or its transpose, depending on the sizes of m and n. O is the zero matrix. For a real number of order m×n, , , , i=1,2,…,r, r=min(m, n); Composed of useful signals containing fault information and redundant useless information, the singular values ​​obtained from the decomposition of matrix A are This reflects the concentration of useful signal and redundant useless information energy; the magnitude of the singular value represents the degree of contribution of the signal, and a smaller singular value represents redundant useless information. By setting its value to zero, the vibration signal is reconstructed using the inverse process of Equation 2 to remove the redundant signal.

3. The multi-task collaborative high-voltage disconnector fault diagnosis method according to claim 1, characterized in that, Step 2 is as follows: Because the working environment of disconnect switches is complex and variable, the collected vibration signals are not stable. Therefore, after obtaining the one-dimensional time-domain vibration signal, STFT technology is used to convert it into a two-dimensional time-frequency graph, effectively extracting fault features from the non-stationary signal. Furthermore, the subsequent fault detection system uses a two-dimensional convolutional neural network. Under the same conditions, convolutional neural networks generally perform better than one-dimensional data processing. STFT uses a time window to slide across the time-domain signal of the vibration signal, truncating the time-domain signal into multiple segments, performing Fourier transforms on each segment, and obtaining the frequency domain characteristics of each segment, ultimately deriving the frequency domain characteristics at each moment. Let the window function be r(t), then the short-time Fourier transform of the signal x(t) is defined as: （4）； Where f represents frequency (in Hz) and is any real number; t represents time (in seconds). This is a time slice near time t.

4. The multi-task collaborative high-voltage disconnector fault diagnosis method according to claim 1, characterized in that, The multi-task joint fault diagnosis network described in step 3 connects two networks through a domain sharing module, forming a layer-to-layer and module-to-module knowledge sharing system based on certain rules and strategies. The sharing layer is automatically determined for different tasks, and the collected high-voltage disconnector fault signals are classified and their severity is judged through end-to-end learning. The calculation process of the domain feature sharing unit is shown in Equation 5; let the feature maps of the two tasks be x1 and x2 respectively. The image after feature sharing of the feature maps of different tasks by the domain feature sharing unit is as follows. and , and This represents the shared weights between different tasks. and This represents the shared weights between tasks; the weight matrix encodes the relationship between two tasks, and is trained... , , and To set the degree of sharing, the larger the weight assigned, the higher the degree of sharing of the feature surfaces of the two tasks in that layer; （5）； The core of both networks is a convolutional neural network, which includes convolutional layers, pooling layers, fully connected layers, and softmax classification layers. A convolutional layer is a deep feedforward neural network that performs convolutional calculations. Its main function is to extract fault features by sliding convolutional kernels across the feature plane. Each convolutional layer contains multiple kernels, each corresponding to a weight coefficient and a bias. It extracts the feature image of the next layer by moving the kernels through the previous layer's image. The calculation method is as follows: （6）； The summation part in the formula is equivalent to solving for the first-order cross-correlation, where b is the bias and y is the output of the previous neuron. As weight, and This represents the convolutional input and output of the (l+1)th layer. for Given the dimensions of the feature map, assume the length and width are the same. The corresponding pixel in the i-th row and j-th column of the feature map, Let f be the number of channels in the l-th feature map, and s0 and p be the convolutional layer parameters, corresponding to the kernel size, stride, and number of padding layers, respectively. The convolutional layer uses the ReLU function, which is expressed as: （7）； Pooling layers compute the nth feature map of the l-th pooling layer. It manifests as: （8）； It is the nth input mapping, which is the nth input mapping of the previous convolutional layer. It is the max pooling equation, where p is the pooling size and s is the stride; while the fully connected layer transforms the two-dimensional image into one-dimensional data, flattens the data, and finally achieves classification through the Softmax layer.

Images