Cable joint early fault diagnosis model training method and application

Abstract

The invention provides a cable joint early fault diagnosis model training method and application, and belongs to the technical field of cable diagnosis. Acquiring cable joint gas concentration data related to the early fault types of the plurality of cable joints; based on a mean shift algorithm, performing data enhancement processing on the cable joint gas concentration data related to the early fault type of each cable joint; training a fault diagnosis model according to the cable joint gas concentration data after data enhancement processing; and inputting the concentration data of the to-be-detected cable joint into the trained fault diagnosis model, and determining the early fault type of the to-be-detected cable joint. According to the method, the fault identification accuracy is improved, and the method has huge practical application potential in early fault detection of the high-voltage cable joint. The present invention is based on less data, and it is expected to acquire analog data through data augmentation. According to the method, a part of process points in the mean shift training process are regarded as simulation data, so that a simple network is trained, and a good effect is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of cable diagnosis, and in particular relates to a cable joint early fault diagnosis model training method and application based on data enhancement and a neural network model. Background Art

[0002] Cable joints are common key components in high-voltage cables. Factors such as aging of joint materials, improper installation, and changes in ambient temperature and humidity may cause damage to cable joints, which in turn may cause partial discharge, heating, and other phenomena. This not only affects the normal operation of urban power lines, but also easily causes major fire hazards. Therefore, it is very important to reasonably and effectively evaluate the early failure state of pipe-cable joints. Under the action of electrical stress, thermal stress, mechanical stress, and chemical reaction, the solid insulating materials in the cable joints may react, deteriorate, and degrade, and release various gases. These gases mainly include carbon dioxide (CO2), carbon monoxide (CO), methane (CH4), ethylene (C2H4), ethane (C2H6), and hydrogen (H2). By analyzing and comparing the gas concentrations when the solid material cracks and escapes, the early failure of the cable joint can be judged.

[0003] Transformer dissolved gas analysis (DGA) can provide reference for cable joint fault diagnosis to a certain extent, such as Duval triangle method, IEC standard specifications, etc. These methods compare the concentration of certain characteristic gases with the specified level to evaluate the condition of the transformer. However, these methods are still difficult to be directly applied to cable joint fault diagnosis. This is because, on the one hand, these methods mainly rely on long-term accumulated measurements and laboratory transformer dissolved gas data to analyze and establish the association between faults and characteristic gas levels, and cannot be directly applied to cable joint fault diagnosis; on the other hand, even in transformer fault diagnosis, these methods have varying degrees of deficiencies in diagnostic accuracy, anti-interference ability, and applicability. With the advancement of intelligent algorithms, the use of pre-trained network models in the field of fault detection has become a major trend. However, due to the long data collection cycle of cable joint gas concentration, it is difficult to obtain a large amount of data. In the case of insufficient data volume, how to train a simple and effective network model is a direction that urgently needs to be explored. Summary of the invention

[0004] The technical problem to be solved by the present invention is how to train a simple and effective model using a limited data set.

[0005] The present invention solves the above technical problems through the following technical means:

[0006] A cable joint early fault diagnosis model training method comprises the following steps:

[0007] S1 obtains cable joint gas concentration data related to early failure types of multiple cable joints;

[0008] S2 performs data enhancement processing on the cable joint gas concentration data related to the early fault type of each cable joint based on the mean shift algorithm; the data enhancement processing is performed in the following steps:

[0009] S21 initializes the position of each sample in the cable joint gas concentration dataset;

[0010] S22 takes any position as the cluster center position of the data set and performs kernel density estimation;

[0011] S23 calculates the average displacement vector;

[0012] S24 updates the cluster center position according to the average displacement vector;

[0013] S25 determines whether the modulus of the average displacement vector is less than the set convergence threshold. If so, the iteration ends and several process point positions are obtained. If not, return to step S22 and recalculate the kernel density based on the updated cluster center position.

[0014] S26 sets the cluster center position to the position at the end of the iteration;

[0015] S27 calculates the distance between the iterative process point position and the final cluster center position, determines whether it is less than the set sample threshold, and uses the process point position less than the sample threshold as the generated data sample to complete the data enhancement process;

[0016] S3 trains the fault diagnosis model based on the cable joint gas concentration data after data enhancement processing.

[0017] Furthermore, the early failure types of the cable joints are several types of polyurethane thermal failure, silicone rubber thermal failure, XLPE thermal failure, EVA thermal failure, EPDM thermal failure, polyurethane discharge failure, silicone rubber discharge failure, XLPE discharge failure, EVA discharge failure, and EPDM discharge failure.

[0018] Furthermore, the cable joint gas concentration data are concentrations of several gases including carbon dioxide (CO2), carbon monoxide (CO), methane (CH4), ethylene (C2H4), ethane (C2H6), and hydrogen (H2).

[0019] Furthermore, in step S21, a training data set is constructed based on the cable joint gas concentration data; the data set includes a number of sample data D = {x1, x2, x3, ..., x N}, each sample x i Represents a d-dimensional vector; for the i-th sample, its position is Ci =x i ,i=1,2,…,N; at the same time, the convergence threshold ε, number of iterations T and bandwidth parameter h must be initialized.

[0020] Further, in step S22, initially, any sample position among the N samples can be used as the cluster center position. The kernel density estimate g(C t ):

[0021]

[0022] In the formula, C t Indicates the location where the density value is calculated in the current iteration process, C i represents the location of each sample in the data set, and h represents the bandwidth parameter of the kernel density estimation;

[0023] In step S23, the average displacement vector is calculated according to the following formula:

[0024]

[0025] Where, ΔC t represents the average drift vector of the t-th iteration;

[0026] In step S24, the current cluster center position is updated according to the average displacement vector according to the following formula:

[0027] C t+1 =C t +ΔC t ;

[0028] In the formula, C t+1 Indicates the updated position.

[0029] Further, in step S25, according to |ΔC t |<ε to judge, where |ΔC t | represents the displacement vector ΔC t The modulus of , ε represents the convergence threshold.

[0030] Furthermore, according to ∥C t -C C ∥<λ, where ∥C t -C C ∥ represents the location of the cluster center process point C of each category t and the final cluster center position C C The L2 distance between them, λ represents the set sample threshold.

[0031] Furthermore, in step S3, a fully connected neural network is used as a fault diagnosis model.

[0032] The present invention also provides a method for using the above-mentioned diagnostic model to diagnose early-stage faults of cable joints, inputting the concentration data of the cable joint to be tested into the trained fault diagnosis model to determine the type of early-stage fault of the cable joint to be tested.

[0033] The present invention also provides a processing device, characterized in that it includes at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method as described above.

[0034] The advantages of the present invention are:

[0035] Due to the long data collection cycle of cable joint gas concentration, it is difficult to obtain a large amount of data. In order to improve the accuracy of the analysis, the present invention generates simulated data based on the existing data based on the data enhancement method of the mean shift algorithm for further fault diagnosis analysis. The mean shift algorithm is a non-parametric clustering algorithm that adaptively finds the clustering structure of the data by maximizing the information entropy of the data. In this algorithm, at each iteration, the cluster center drifts and updates in the direction of the maximum density and is attracted to the area with higher data density, that is, the cluster center of the cluster. This means that the cluster center will gradually tend to the dense area of ​​similar data points. Therefore, during the iteration process, the cluster center will gradually approach the final cluster center along a certain path. The regional elements on the path close to the cluster center have similar properties to the final cluster center. Taking this into account, the present invention generates a training data set for fault type diagnosis by restoring the intermediate iteration point and using it as the generated simulation element. Therefore, the mean shift algorithm finds the cluster center by drifting the mean of the data point in the direction of the maximum density until it reaches the local maximum density area.

[0036] The present invention firstly enhances the data by a method based on mean shift, and then uses the enhanced data to train a fault diagnosis model, which is then used for early fault diagnosis of cable joints, thereby improving the accuracy of fault identification and having great practical application potential in early fault detection of high-voltage cable joints. The implementation of this diagnostic tool in the distribution network can improve reliability, reduce maintenance costs, and prevent catastrophic failures through timely intervention.

[0037] The data enhancement method based on the mean shift algorithm provided by the present invention effectively solves the problems of long data collection period and small sample size of gas concentration in cable joints; and is suitable for popularization and use in this field;

[0038] The data enhancement method based on the mean shift algorithm provided by the present invention can quickly obtain effective simulation data by maximizing the information entropy of the data and adaptively finding the clustering structure of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the fully connected neural network structure;

[0040] Figure 2 Schematic diagram of neuron structure;

[0041] Figure 3 A schematic diagram of a flow chart of a method for early fault diagnosis of a cable connector provided by an embodiment of the present invention;

[0042] Figure 4 A schematic diagram of a cable joint gas concentration data enhancement process based on a mean shift algorithm provided in an embodiment of the present invention;

[0043] Figure 5 Schematic diagram of t-SNE visualization results obtained by the cable joint early fault diagnosis method of the present invention; wherein (a) corresponds to the simulation test set before classification, and (b) corresponds to the simulation test set after classification. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are 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.

[0045] Example 1

[0046] This embodiment describes a method for early fault diagnosis of a cable joint, using a fully connected neural network as a fault diagnosis model. Specifically, Figure 1 As shown in the figure, the fully connected neural network consists of an input layer, a hidden layer, and an output layer. The input layer has 6 inputs, which are the concentrations of 6 characteristic gases: carbon dioxide (CO2), carbon monoxide (CO), methane (CH4), ethylene (C2H4), ethane (C2H6), and hydrogen (H2). The hidden layer has 3 layers of networks, containing 64, 128, and 64 neurons respectively. The number of neurons in the output layer is 10, corresponding to the probabilities of 10 types of faults.

[0047] Each layer consists of multiple neurons. In each layer, the output of the previous layer is used as input, and data is propagated unidirectionally from the input layer to the output layer. Each connection has a weight, which is learned through training data. Each neuron calculates the weighted sum of all its inputs and generates an output through an activation function such as ReLU, sigmoid, or tanh. The ReLU activation function is selected in this embodiment because it can alleviate the gradient vanishing problem, which is particularly beneficial when training deep neural networks. Unlike Sigmoid and Tanh functions, ReLU provides sparse activation, which improves the efficiency of the learning process.

[0048] Neuronal structure Figure 2 As shown. Assume that the input of a single neuron is {x1,x2,…,x i}, add the weighted sum of the input and bias, activate the activation function f, and get the corresponding output.

[0049] Therefore, the calculation of the i-th neuron in the j-th hidden layer can be expressed as:

[0050]

[0051] The ReLU function is used as the activation function, and the activation function is as follows:

[0052]

[0053] In order to enhance the nonlinear fitting ability of the model, the hidden layer adopts the Dropout strategy to randomly discard some neurons. The output of the output layer is normalized by the sigmoid function to obtain the probability of each fault category, and the category with the highest probability is selected as the predicted fault type. The sigmoid function is as follows:

[0054]

[0055] In the formula, y l Represents the output result of the lth neuron in the output layer.

[0056] The above fully connected neural network is trained using back propagation, and the loss function is the cross entropy loss function:

[0057]

[0058] In the formula, y m represents the prediction result of the mth fault type, represents the corresponding label, and M represents the total number of fault types.

[0059] Based on the above fault diagnosis model, this embodiment provides a cable connector early fault diagnosis model training method. Figure 3 As shown, it includes the following steps:

[0060] S1. Obtain cable joint gas concentration data related to multiple cable joint early failure types.

[0061] In this embodiment, the cable joint material types mainly include polyurethane, silicone rubber, XLPE, EVA, and EPDM (ethylene propylene diene monomer rubber); therefore, the cable joint early failure types here are polyurethane thermal failure, silicone rubber thermal failure, XLPE thermal failure, EVA thermal failure, EPDM thermal failure, polyurethane discharge failure, silicone rubber discharge failure, XLPE discharge failure, EVA discharge failure, and EPDM discharge failure. The cable joint gas concentration data include carbon dioxide (CO2), carbon monoxide (CO), methane (CH4), ethylene (C2H4), ethane (C2H6), and hydrogen (H2) gas concentrations.

[0062] S2. Based on the mean shift algorithm, the cable joint gas concentration data related to the early fault type of each cable joint is enhanced.

[0063] like Figure 4 As shown, in this embodiment, the cable joint gas concentration data related to each cable joint early fault type is subjected to data enhancement processing according to the following steps:

[0064] S21 initializes the position of each sample in the cable joint gas concentration dataset.

[0065] According to the cable joint gas concentration data related to each cable joint early fault type, a training data set for the corresponding fault type is constructed. The training data set includes a number of sample data D = {x1, x2, x3, ..., x N}, each sample x i Represents a d-dimensional vector (x i ∈R d ). In this embodiment, d=6, indicating that 6 gas concentrations are included.

[0066] For the i-th sample, its position is C i =x i ,i=1,2,…,N; at the same time, the convergence threshold ε, number of iterations T and bandwidth parameter h must be initialized.

[0067] S22 takes any position as the cluster center position of the data set and performs kernel density estimation.

[0068] Initially, any sample position among the N samples can be used as the cluster center position.

[0069] This step calculates the kernel density estimate g(C t ):

[0070]

[0071] In the formula, C t Indicates the location of the density value calculated in the current iteration process, that is, the center position in the clustering process, C i Represents the location of each sample in the data set, and h represents the bandwidth parameter of the kernel density estimation.

[0072] S23 calculates the average displacement vector.

[0073] This step calculates the average displacement vector according to the following formula:

[0074]

[0075] Where, ΔC t represents the average drift vector at the t-th iteration.

[0076] S24 updates the cluster center position according to the average displacement vector.

[0077] In this step, the position of the current sample is updated according to the average displacement vector according to the following formula:

[0078] C t+1 =C t +ΔC t ;

[0079] In the formula, C t+1 Indicates the updated position.

[0080] S25 determines whether the modulus of the average displacement vector is less than the set convergence threshold. If so, the current iteration ends and several process point positions are obtained; if not, return to step S22 and recalculate the kernel density based on the updated cluster center position.

[0081] In this step, according to |ΔC t |<ε to judge, where |ΔC t | represents the displacement vector ΔC t The modulus of , ε represents the convergence threshold.

[0082] S26 sets the cluster center position to the position at the end of each iteration, denoted as C C .

[0083] S27 calculates the distance between the iterative process point position and the final cluster center position, determines whether it is less than the set sample threshold, and uses the process point position less than the sample threshold as the generated data sample to complete the data enhancement processing.

[0084] In this step, according to ∥C t -C C∥<λ, where ∥C t -C C ∥ represents the location of the cluster center process point C of each category t and the final wrestling center position C C The L2 distance between them is λ, and λ represents the set sample threshold. The process points near the final cluster center are selected and used as the generated data samples to achieve data augmentation.

[0085] When the above convergence condition is still not met after reaching the number of iterations T, it is necessary to terminate the iteration process, adjust the bandwidth parameter h of the kernel density estimation, and repeat the above steps S22-S27. For each fault type data set, repeat the above steps S21-S27 to achieve data enhancement for each fault type data set.

[0086] S3 trains the fault diagnosis model based on the cable joint gas concentration data after data enhancement processing.

[0087] In this embodiment, the fault diagnosis model can be trained by conventional means.

[0088] Example 2

[0089] Based on the fault detection model trained in Example 1, the concentration data of the cable joint to be tested is input into the trained fault diagnosis model to determine the early fault type of the cable joint to be tested.

[0090] Experimental example

[0091] Based on 40 sets of experimental basic data, this experimental example uses these 40 sets of experimental basic data as the initialization position and adopts the base mean shift data enhancement algorithm (previous step S2) to obtain 400 sets of simulation data. Each set of data includes the concentrations of six characteristic gases: carbon dioxide (CO2), carbon monoxide (CO), methane (CH4), ethylene (C2H4), ethane (C2H6) and hydrogen (H2). The data includes ten types of faults: polyurethane thermal fault, silicone rubber thermal fault, XLPE thermal fault, EVA thermal fault, EPDM thermal fault, polyurethane discharge fault, silicone rubber discharge fault, XLPE discharge fault, EVA discharge fault and EPDM discharge fault. The number of each type of fault in the experimental data and simulation data is shown in Table 1.

[0092] Table 1 The number of various types of faults in experimental data and simulation data

[0093]

[0094] In order to verify the performance of the algorithm, the simulation data set is first divided into a training set and a test set in a ratio of 8:2. The above fault diagnosis model is trained using the training set, and the model parameters are optimized by the gradient optimization algorithm according to the obtained loss function until the loss function tends to be stable. Then the trained fault diagnosis model is tested using the test set. At the same time, the classification accuracy of this method is compared with that of KNN and SVM. The accuracy comparison of the test set is shown in Table 2.

[0095] Table 2 Comparison of fault identification accuracy of different methods.

[0096]

[0097] Compared with KNN and SVM, the fully connected network proposed in this invention has the highest accuracy of 87.5%. The superior performance of the fully connected neural network compared with KNN and SVM can be attributed to its ability to autonomously learn the complex nonlinear relationship between gas concentration and fault type. Although KNN relies on Euclidean distance and SVM uses hyperplanes for classification, the multi-layer structure of the neural network allows it to capture the complex interactions between features, thereby achieving higher diagnostic accuracy. This shows that this method is superior to several other methods in performance.

[0098] Figure 5 The t-SNE visualization results of the simulated test set samples before and after classification using the method of the embodiment of the present invention. After classification, similar data are clustered into adjacent areas.

[0099] The previously trained fault diagnosis model is tested on the experimental basic data set, and the results are shown in Table 3.

[0100] Table 3 Comparison of fault recognition accuracy of different methods.

[0101]

[0102] It can be seen from Table 3 that, on the experimental basic data set, the present invention still has a relatively high classification accuracy rate, reaching 82.5%, which exceeds several other classification methods.

[0103] The above experiments prove the effectiveness and generalization of the data enhancement and fault diagnosis model proposed in the present invention. This shows that the fully connected neural network can automatically extract and learn the relevant features between the characteristic gas concentration data and use them for the diagnosis and classification of fault types.

[0104] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0105] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0107] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cable joint early fault diagnosis model training method, characterized in that: The following steps are involved: S1 obtains cable joint gas concentration data related to early failure types of multiple cable joints; S2 performs data enhancement processing on the cable joint gas concentration data related to the early fault type of each cable joint based on the mean shift algorithm; the data enhancement processing is performed in the following steps: S21 initializes the position of each sample in the cable joint gas concentration dataset; S22 takes any position as the cluster center position of the data set and performs kernel density estimation; S23 calculates the average displacement vector; S24 updates the cluster center position according to the average displacement vector; S25 determines whether the modulus of the average displacement vector is less than the set convergence threshold. If so, the iteration ends and several process point positions are obtained. If not, return to step S22 and recalculate the kernel density based on the updated cluster center position. S26 sets the cluster center position to the position at the end of the iteration; S27 calculates the distance between the iterative process point position and the final cluster center position, determines whether it is less than the set sample threshold, and uses the process point position less than the sample threshold as the generated data sample to complete the data enhancement process; S3 trains the fault diagnosis model based on the cable joint gas concentration data after data enhancement processing.

2. The cable joint early fault diagnosis model training method according to claim 1 is characterized in that: In step S1, the early fault types of the cable joint are several of polyurethane thermal fault, silicone rubber thermal fault, XLPE thermal fault, EVA thermal fault, EPDM thermal fault, polyurethane discharge fault, silicone rubber discharge fault, XLPE discharge fault, EVA discharge fault, and EPDM discharge fault.

3. The cable joint early fault diagnosis model training method according to claim 1 or 2, characterized in that: The cable joint gas concentration data are the concentrations of several gases including carbon dioxide (CO2), carbon monoxide (CO), methane (CH4), ethylene (C2H4), ethane (C2H6), and hydrogen (H2).

4. The cable joint early fault diagnosis model training method according to claim 1 is characterized in that: In step S21, a training data set is constructed based on the cable joint gas concentration data; the data set includes a number of sample data D = {x1, x2, x3, ..., x N }, each sample x i Represents a d-dimensional vector; for the i-th sample, its position is C i =x i ,i=1,2,…,N; at the same time, the convergence threshold ε, number of iterations T and bandwidth parameter h must be initialized.

5. The cable joint early fault diagnosis model training method according to claim 1 or 4, characterized in that: In step S22, the kernel density estimate g(C t ): In the formula, C t Indicates the location where the density value is calculated in the current iteration process, C i represents the location of each sample in the data set, and h represents the bandwidth parameter of the kernel density estimation; In step S23, the average displacement vector is calculated according to the following formula: Where, ΔC t represents the average drift vector of the t-th iteration; In step S24, the current cluster center position is updated according to the average displacement vector according to the following formula: C t+1 =C t +ΔC t ; In the formula, C t+1 Indicates the updated position.

6. The cable joint early fault diagnosis model training method according to claim 5 is characterized in that: In step S25, according to |ΔC t |<ε to judge, where |ΔC t | represents the displacement vector ΔC t The modulus of , ε represents the convergence threshold.

7. The cable joint early fault diagnosis method according to claim 5, characterized in that: In step S27, according to ∥C t -C C ∥<λ, where ∥C t -C C ∥ represents the location of the cluster center process point C of each category t and the final cluster center position C C The L2 distance between them, λ represents the set sample threshold, and the cluster center position is set to the position at the end of the iteration, denoted as C C ; Process points near the final cluster center are selected and used as generated data samples to achieve data augmentation.

8. The cable joint early fault diagnosis model training method according to claim 1 is characterized in that: In step S3, a fully connected neural network is used as a fault diagnosis model.

9. Use the diagnostic model described in any one of claims 1 to 8 to diagnose early-stage faults of cable joints, characterized in that: The concentration data of the cable joint to be tested is input into the trained fault diagnosis model to determine the early fault type of the cable joint to be tested.

10. A processing device, characterized in that: It comprises at least one processor and at least one memory in communication with the processor, wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute any method according to claims 1 to 8.

Images