A 10kV Circuit Breaker Mechanical Fault Detection Method Based on Voiceprint

Abstract

The present invention discloses a method for mechanical fault detection of 10 kV circuit breakers based on voiceprint, which includes the following steps: S1, building a MobileNet_YOLOv4 neural network model; S2, collecting voiceprint through a high-sensitivity sensor; S3, performing convolution operation of convolution kernels for edge-side recognition and early warning; S4, identifying circuit breaker faults; S5, performing an improved S transform on the signal. In the present invention, a resource allocation module embedded in the cloud platform is used to dynamically calculate task allocation, build a lightweight model, accelerate the model operation speed, and build a real-time target recognition and classification system. The Moblienet neural network structure reduces the operating memory of the neural network model and accelerates the operation speed of the neural network model. In order to improve the target recognition speed and solidify the model onto edge devices, the MobileNet network structure applies the technology of depthwise separable convolution, which divides the standard convolution of the convolutional neural network into two steps, namely depth convolution and pointwise convolution, during the calculation process, greatly reducing the computational amount of the network.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical fault detection of circuit breakers, and particularly relates to a method for detecting mechanical faults of 10 kV circuit breakers based on voiceprint. Background Art

[0002] As an important switching device in the power system, when a system fault occurs, the circuit breaker, as an actuator, can timely and accurately cut off the fault part, close and open the line, and plays an important role in the control and protection of the power system. Given that all power equipment has a certain service life cycle, the service life of high-voltage circuit breakers generally does not exceed forty years under normal circumstances, and some power equipment components will age due to long-term use. Therefore, it is very necessary to perform timely and accurate condition-based maintenance and monitoring diagnosis on circuit breaker faults and defects. Condition-based maintenance combines real-time condition monitoring and fault diagnosis of electrical equipment. The signals monitored are the basis for judging whether the equipment is operating normally. Based on this, the staff can make more timely predictions and formulate timely and accurate maintenance plans. As an important type of electrical equipment in the power system, fault diagnosis is achieved by comprehensively analyzing the historical archives, operating conditions, and live test or on-line monitoring data of high-voltage circuit breakers. It can not only timely determine the location of the fault and the severity of the component fault, but also accurately diagnose the cause of the equipment fault. Therefore, it is the general trend to perform condition-based maintenance and fault diagnosis on high-voltage circuit breakers.

[0003] Although the price of a circuit breaker is usually much lower than that of the power equipment it protects, the losses caused by its faults are huge, far exceeding its own price. The opening and closing of the circuit breaker are mainly completed by a mechanical structure. To achieve fast operation, the mechanical structure needs to transmit high power and heavy loads. As the number of operations increases, the mechanical performance of the circuit breaker will inevitably deteriorate. Some research on the reliability of circuit breakers shows that mechanical faults are the main factor for the failure of high-voltage circuit breakers. However, there are still various problems in the mechanical fault detection of various circuit breakers on the market.

[0004] However, although the cost of the circuit breaker in the prior art is not high, the impact of its faults is extremely bad. Edge computing has shown its powerful functions in many fields, but it has not been involved in the mechanical monitoring of circuit breakers. Moreover, with the increasing computing efficiency and better performance of deep learning technology, the lightweight neural network model has higher accuracy under the condition of the same training data set. However, this also means that the structure of the deep learning network model is becoming more and more complex, and the computing resources it occupies are increasing. At the same time, for edge anomaly analysis, it is intolerable that the resources required for model operation are too much and the memory occupied is too large. Therefore, we propose a method for detecting mechanical faults of 10 kV circuit breakers based on voiceprint. Summary of the Invention

[0005] The object of the present invention is to provide a 10kV circuit breaker mechanical fault detection method based on voiceprint to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a 10kV circuit breaker mechanical fault detection method based on voiceprint, comprising the following steps:

[0007] S1. Construction of MobileNet_YOLOv4 neural network model: The MobileNet_YOLOv4 neural network model uses multi-scale feature map mapping, and the clipping neural network obtains feature maps of different scales of image data. After each layer of feature maps is extracted, predictions are made separately. The convolution layer of the MobileNet_YOLOv4 neural network model extracts six different feature maps for prediction, thereby constructing target recognition prior frames of different scales. The feature maps of different layers in the model respectively test the large and small targets of the matching size. Finally, the model makes predictions on feature maps of different scales to obtain prediction values ​​of different scales, which significantly improves the target recognition performance of the model.

[0008] S2. Collecting soundprints through high-sensitivity sensors: The high-sensitivity sensor includes a cone-shaped sound collector, a mounting ring, an energy converter, an upper pressure plate and an isolator. The circuit breaker mechanical fault soundprint is collected through the high-sensitivity sensor. The cone-shaped sound collector uses the principle of gas compression to amplify the original audio signal, and then transmits the data through low-power, long-distance wireless communication;

[0009] S3. Convolution operation of the convolution kernel for edge recognition warning: The standard convolution structure of all convolution layers in Backbone in YOLOv4 is replaced with a depth-separable convolution structure. The MobileNet_YOLOv4 neural network model first completes the convolution operation by treating each neural network node of the input data as a specific convolution kernel through a 3×3 depth convolution.

[0010] S4. Circuit breaker fault identification: The accuracy of the target identification and classification process of the normal circuit breaker and the faulty circuit breaker obtained by the MobileNet_YOLOv4 neural network model after training changes with the signal confidence. The neural network model is trained with a data set in PASCAL VOC format, and the pre-trained weights on the data set are loaded when the model is built. The front-layer convolutional neural network is responsible for some low-level image feature extraction of the input image data;

[0011] S5. Improved S transform of the signal: Improve the signal feature extraction method of S transform, construct an improved S transform model, establish a feature parameter recognition mechanism, and realize fast and accurate signal recognition.

[0012] Preferably, the MobileNet_YOLOv4 neural network model in S1 is divided into two independent steps, namely depth convolution operation and 1×1 convolution process; two non-linear activation units and BN layers are concatenated at the output end of the two steps; MobileNet makes extensive use of the alternating use of 1×1 convolution and depth convolution. During the training process of the MobileNet neural network model, 1×1 convolution occupies three-quarters of the number of parameters and 95% of the computing resources.

[0013] Preferably, the width factor and resolution factor of the MobileNet_YOLOv4 neural network model are adjusted;

[0014] Width factor is a hyperparameter in the training and fitting process of the neural network model, and its value range is (0, 1]; simply put, it is regarded as the ratio of the total number of convolution operations of each convolution layer of the adjusted MobileNet neural network model to the total number of convolution operations of each convolution layer of the standard MobileNet neural network model; for the MobileNet neural network model using depthwise separable convolution, after using the width factor, the total computational amount S of the neural network model is shown as follows:

[0015] ,

[0016] The standard MobileNet neural network model is 1. Generally speaking, when reducing the network width factor to achieve the purpose of the computing resources occupied during the model operation, the width factor generally takes these four scales; when training on the same training set ImageNet, different values of the width factor are taken for the total number of parameters, computing resources, and the accuracy of the neural network model.

[0017] Preferably, the resolution factor of the MobileNet_YOLOv4 neural network model takes values between (0, 1]. In short, the resolution factor is the proportional scaling of the input image data of each convolution layer of the MobileNet neural network model. It is regarded as reducing a certain proportion of each feature layer of the model to achieve the purpose of trimming the total computational amount of the network model. The computational amount S of the MobileNet neural network model with depthwise separable convolution as the core is shown as the formula:

[0018] ,

[0019] shows that when the width factor remains unchanged, different resolution factors are taken when training on the same training set ImageNet values of the total number of parameters, computing resources, and the accuracy of the neural network model.

[0020] Preferably, the MobileNet_YOLOv4 neural network model uses the Mish activation function;

[0021] The expression of the Mish activation function is as follows:

[0022] .

[0023] Preferably, during the training process and before and after the model compression process of the MobileNet_YOLOv4 neural network model, the width factor and the resolution factor are both set to 0.75, which not only ensures that the hyperparameters will not have a great impact on the accuracy of the network model, but also greatly reduces the computing resources occupied during the operation of the model;

[0024] During the training process of the MobileNet_YOLOv4 neural network model, the number of categories to be classified is set to 2, and the labeled data sets of normal samples and faulty samples are represented by n7 and n8 respectively; 123 groups are planned as the validation set, 315 groups are the test set, the initial learning rate is set to 0.002, and the number of iterations in the entire training process is 10,000 times; the time taken for the entire training process is 9 hours.

[0025] Preferably, in S3, before performing convolution operation on the breaker fault signal detection to obtain the feature signal, variational mode decomposition of the signal is first performed. The optimal solution of variational mode decomposition is obtained by iterative update through the alternating direction multiplier method, that is, the original signal is reconstructed by analyzing the finite bandwidth and center frequency of each intrinsic mode component; when using variational mode decomposition for signal decomposition, the number of modal components decomposed is controlled by manual setting, and the decomposition scale K is defined to determine how many intrinsic mode components the original signal is decomposed into.

[0026] Preferably, the specific principle of the variational mode decomposition is as follows:

[0027] S301, for each observed signal, assuming it is the original signal superimposed with independent Gaussian noise, first denoise and reconstruct the sampled signal f0; the objective function is expressed as:

[0028] ,

[0029] where f is obtained through a regularization method:

[0030] ,

[0031] S302. Calculate the analytic signal of each mode \(u\) k through Hilbert transform to obtain the unilateral spectrum of the mode components;

[0032] S303. Estimate the central frequency through the mixed mode function \(u\) k and the exponent, and then convert the spectrum of each mode to its respective estimated central frequency;

[0033] S304. Use the Gaussian smoothing method of the demodulated signal to estimate the bandwidth of each mode function. The objective function is:

[0034] ,

[0035] where \(u\) k is a subsequence; \(K\) is the total number of subsequences; \(w\) k is the central frequency; \(f(t)\) is the original signal; \(\delta\) is the Dirac distribution. Introduce the quadratic penalty factor, \(f(\omega)\), \(u(\omega)\), \(u(\omega)\) and the Lagrange multiplier into the above formula to realize the transformation from the constrained variational problem to the unconstrained variational problem. The following is the optimized expression:

[0036] ,

[0037] S305. To further solve the above formula, use the alternating direction multiplier method, and the formula is extended as follows:

[0038] ,

[0039] ,

[0040] In the variational mode decomposition algorithm, first transform the signal decomposition method to make it a variational problem under constrained conditions, and then calculate it to achieve the purpose of decomposing the signal by finding the optimal solution. Continuously repeat this decomposition step to continuously update the central frequency and the corresponding bandwidth in each mode until multiple relatively narrow-band intrinsic mode components are obtained.

[0041] Preferably, after the variational mode decomposition processes the signal, an improved S-transform is performed. The signal feature extraction method of the improved S-transform is used to construct an improved S-transform model and establish a feature parameter recognition mechanism to achieve fast and accurate signal recognition, good time-frequency analysis and feature extraction characteristics, and more intuitive results.

[0042] Preferably, the implementation steps of the improved S-transform are as follows:

[0043] S306. Calculate the N - point FFT spectrum X(k / NT) of the signal x(nT);

[0044] S307. Use the envelope extreme value algorithm for X(k / NT) to determine the characteristic frequency k i , where i = 1, 2,... q, and q is the number of characteristic frequency points;

[0045] S308. According to the frequency band where the frequency point k i is located, determine the value of the window width adjustment scale factor of the Gauss adaptive optimization window. The parameter values are determined based on a large number of experiments: a = 0, c = 1; at the fundamental frequency, b = 1.6; in the 0 - 350 Hz frequency band excluding the fundamental frequency point, b = 0.4; in the 350 - 700 Hz frequency band, b = 0.35; above 700 Hz, b = 0.3; the above values can achieve satisfactory measurement results;

[0046] S309. Calculate the Fourier transform of the Gauss adaptive optimization window at the frequency point k i :

[0047] ,

[0048] S310. For the frequency point k i , translate the spectrum X(k i / NT) to X((k i+r ) / NT);

[0049] S311. Calculate the product A(r, k i+r ) of X((k i / (NT)) and W(r / (NT), k i / (NT)):

[0050] ,

[0051] S312. For the frequency point k i , calculate the inverse fast Fourier transform of A(r, k i ) to obtain the improved S - transform corresponding to k i :

[0052] ;

[0053] S313. Repeat steps S308 - S312 to complete the improved S - transforms corresponding to all characteristic frequency points and complete the processing of edge - side abnormal signals.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] The optimization of dynamic computing task allocation in the present invention is realized by a resource allocation module embedded in a cloud platform. Some Internet of Things application tasks require programmed analysis, processing, and control of global data and resources, and a data center can provide strong computing support for this. Other tasks require efficient and fast real-time response, which are supported by edge computing and terminal devices deployed on the edge side;

[0056] Under the condition of ensuring a certain accuracy and performance of the neural network model as much as possible, a lightweight model is constructed as much as possible to speed up the model operation speed and build a real-time target recognition and classification system.

[0057] Currently, lightweight models mainly focus on the redesign of existing neural network models, constructing a more efficient and simpler neural network structure. While reducing the computing power occupied by the number of neural network parameters, the accuracy of the neural network model is ensured not to decrease significantly. At the same time, during the training process of the neural network, using a more lightweight module structure to splice with the existing model also reduces the operation time;

[0058] The Moblienet neural network structure can greatly reduce the running memory of the neural network model and speed up the running speed of the neural network model. In order to improve the target recognition speed and solidify the model on edge devices, the MobileNet network structure mainly applies the technology of depthwise separable convolution, which can divide the standard convolution of the convolutional neural network into two steps: depthwise convolution and pointwise convolution during the calculation process, greatly reducing the computational amount of the network;

[0059] The accuracy of the model obtained by using standard convolution is about 1% higher than that of depthwise separable convolution. However, the computing resources occupied by the model operation are at least 8 times that of standard convolution. Therefore, the MobileNet network structure can significantly reduce the computing resources and the number of parameters of the model operation while ensuring a considerable accuracy, and can greatly speed up the reaction speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a schematic diagram of the step flow of the present invention;

[0061] Figure 2 It is a schematic diagram of standard convolution and depthwise separable convolution of the present invention;

[0062] Figure 3 It is for different width factors of the present invention

[0063] Figure 4 It is for different resolution factors of the present invention

[0064] Figure 5This is a schematic diagram of edge side recognition warning of the present invention;

[0065] Figure 6 It is a schematic diagram of the neural network model training process of the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] See also Figures 1-6 The present invention provides a technical solution: a 10kV circuit breaker mechanical fault detection method based on voiceprint, comprising the following steps:

[0068] S1. Construction of MobileNet_YOLOv4 neural network model: The MobileNet_YOLOv4 neural network model uses multi-scale feature map mapping, and the clipping neural network obtains feature maps of different scales of image data. After each layer of feature maps is extracted, predictions are made separately. The convolution layer of the MobileNet_YOLOv4 neural network model extracts six different feature maps for prediction, thereby constructing target recognition prior frames of different scales. The feature maps of different layers in the model respectively test the large and small targets of the matching size. Finally, the model makes predictions on feature maps of different scales to obtain prediction values ​​of different scales, which significantly improves the target recognition performance of the model.

[0069] The standard convolution structure of all convolution layers in Backbone in YOLOv4 is replaced by the deep separable convolution structure, and the fully connected layers and Softmax layers in the backbone network structure of CSP-Darknet53 in Backbone are cut off, because the MobileNet model has its own fully connected layers and Softmax layers. The deep separable convolution will greatly reduce the amount of network calculation. It decomposes the standard convolution into two steps: deep convolution and point-wise convolution. The MobileNet_YOLOv4 neural network model first completes the convolution operation by treating each neural network node of the input data as a specific convolution kernel through 3×3 deep convolution, and uses 1×1 convolution technology to complete the integration of the image data processed in the previous step, preventing each neuron node from repeatedly integrating all the signals input to the previous neural network layer. The total number of parameters of the MobileNet_YOLOv4 neural network model is much smaller than that of YOLOv4, and the computing resources occupied will also be significantly reduced. At the same time, the memory occupied by the model becomes extremely small;

[0070] S2. Collect voiceprint through a high-sensitivity sensor: The high-sensitivity sensor includes a microphone with a conical structure, a mounting ring, an energy converter, an upper pressing piece, and an isolator. The mechanical fault voiceprint of the circuit breaker is collected through the high-sensitivity sensor. The microphone with a conical structure amplifies the original audio signal using the gas compression principle, and then data is transmitted through low-power and long-distance wireless communication.

[0071] S3. Convolution operation of the convolution kernel for edge-side recognition and early warning: Use the depthwise separable convolution structure to replace the standard convolution structure of all convolution layers in the Backbone of YOLOv4. The MobileNet_YOLOv4 neural network model first completes the convolution operation by regarding the neural network nodes of each input data as specific convolution kernels through 3×3 depth convolution.

[0072] S4. Circuit breaker fault recognition: The accuracy of the target recognition and classification process model of normal and faulty circuit breakers obtained after the MobileNet_YOLOv4 neural network model is trained changes with the signal confidence. The PASCAL VOC format dataset is used to train the neural network model, and the pre-trained weights on the dataset are loaded when the model is built. The front-layer convolutional neural network is responsible for extracting some low-level image features from the input image data.

[0073] S5. Improved S transform for the signal: Improve the signal feature extraction method of the S transform, construct an improved S transform model, establish a feature parameter recognition mechanism, and achieve fast and accurate signal recognition.

[0074] The ImageNet image data training set is used for comparative testing using standard convolution and depthwise separable convolution. The accuracy of the model obtained using standard convolution is about 1% higher than that of depthwise separable convolution. However, the computing resources occupied by the model operation are at least 8 times that of standard convolution. Therefore, the MobileNet network structure can ensure a relatively high accuracy while significantly reducing the computing resources and the number of parameters required for model operation. At the same time, it can greatly accelerate the response speed of the model. In this embodiment, preferably, the MobileNet_YOLOv4 neural network model in S1 is divided into two independent steps, namely depth convolution operation and 1×1 convolution process; a non-linear activation unit and a BN layer are concatenated at the output end of the two steps; MobileNet makes extensive use of the alternating use of 1×1 convolution and depth convolution. During the training process of the MobileNet neural network model, 1×1 convolution accounts for three-quarters of the number of parameters and 95% of the computing resources.

[0075] MobileNet has an absolute advantage over traditional neural network models in terms of the number of parameters and computational complexity. However, for embedded edge computing devices, this is far from enough. Therefore, it is also necessary to adjust the width factor and the resolution factor of the MobileNet neural network model to meet the extremely stringent requirements for the computational resources occupied by the neural network model on edge devices. In this embodiment, preferably, the width factor and the resolution factor of the MobileNet_YOLOv4 neural network model are adjusted;

[0076] Width factor The width factor is a hyperparameter in the process of neural network model training and fitting, and its value range is (0, 1]; simply put it can be regarded as the ratio of the total number of convolution operations in each convolutional layer of the adjusted MobileNet neural network model to the total number of convolution operations in each convolutional layer of the standard MobileNet neural network model; for the MobileNet neural network model using depthwise separable convolutions, the total computational complexity S of the neural network model after using the width factor is shown as follows:

[0077] ,

[0078] The standard MobileNet neural network model is 1. Generally speaking, when reducing the network width factor to achieve the purpose of the computational resources occupied by the model during operation, the width factor generally takes these four scales;

[0079] When training on the same training set ImageNet, take different width factors values for the total number of parameters, computational resources, and the accuracy of the neural network model;

[0080] The resolution factor of the MobileNet_YOLOv4 neural network model takes values between (0, 1]. In short, the resolution factor is the proportional scaling of the input image data for each convolutional layer of the MobileNet neural network model, and can be regarded as reducing the scale of each feature layer of the model to achieve the purpose of trimming the total computational complexity of the network model. The computational complexity S of the MobileNet neural network model with depthwise separable convolutions as the core is shown as the formula:

[0081] ,

[0082] represents that at the width factor When it remains unchanged, different resolution factors are taken during training on the same training set ImageNet The total amount of parameters, computing resources, and the accuracy of the neural network model occupied by the values.

[0083] YOLOv4 improves the feature extraction network compared to the traditional YOLO model. After the Backbone, algorithmic means are used to increase the receptive field of data features. At the same time, the activation function is changed to the Mish activation function. Since the MobileNet_YOLOv4 neural network model used replaces the Backbone of YOLOv4 with MobileNet, and the Mish activation function has no upper bound, there is no gradient saturation phenomenon caused by the reduction of training speed during the training of the neural network model. At the same time, the Mish activation function has no lower bound, which can make the model regularization stronger and the fitting effect better. At the same time, the Mish activation function has a larger computational cost compared to the ReLU used in the previous YOLO series models, but the results obtained under the same training set conditions are better when tested in the same neural network structure. In this embodiment, preferably, the MobileNet_YOLOv4 neural network model uses the Mish activation function;

[0084] The expression of the Mish activation function is as follows:

[0085] .

[0086] Before the training starts, the hyperparameters of MobileNet in the MobileNet_YOLOv4 network structure need to be calibrated first. When the values of α and β in the MobileNet network structure gradually decrease, the accuracy of the MobileNet_YOLOv4 neural network model for object recognition and classification gradually decreases, and the trend becomes more and more significant. Especially when the width factor and the resolution factor decrease from 0.5 to 0.25, the function of the neural network model even reaches the point of direct disappearance. Therefore, while considering the computing resources occupied by the model during operation, a certain amount of accuracy also needs to be ensured, because subsequent model pruning can also significantly reduce the computing resources occupied by the model. In this embodiment, preferably, the width factor and the resolution factor are both set to 0.75 during the training process and before and after the model compression of the MobileNet_YOLOv4 neural network model, which greatly reduces the computing resources occupied by the model during operation while ensuring that the hyperparameters do not have a large impact on the accuracy of the network model;

[0087] During the training process of the MobileNet_YOLOv4 neural network model, the number of categories to be classified is set to 2. The labeled datasets of normal samples and faulty samples are represented by n7 and n8 respectively. 123 groups are planned as the validation set, 315 groups as the test set, the initial learning rate is set to 0.002, and the number of iterations in the entire training process is 10,000 times. The entire training process takes 9 hours.

[0088] After the MobileNet_YOLOv4 neural network model is trained, the accuracy of the target recognition and classification process models of normal and faulty circuit breakers changes with the signal confidence. Among them, n7 is the dataset label of normal samples when using the LabelImg visual image annotation tool for annotation, n8 is the dataset label of faulty samples, and the hyperparameter of the YOLOv4 neural network model is set to 0.5;

[0089] Trainloss represents the loss of the model on the training set, representing the model fitting ability of the neural network model for the data of this experimental training set. Valloss represents the loss of the model on the test set, representing the model fitting ability of the neural network model for the data of this experimental test set. This ability truly reflects the model's fitting ability for the test data in the experiment and can also be regarded as the performance of the model, collectively referred to as the generalization ability of the model. After the loss function decreases by a certain amount with the increase in the number of iterations, it gradually stabilizes, indicating that the model training process has been completed. The loss function converges quickly and the final value is low, indicating that the model training effect is good, the data fitting ability is strong, and it has a good fitting effect on both the training set and the test set.

[0090] By training a large number of neural network models on the same training set, the performance analysis of each neural network model on the same training set is completed. It is found that there are a large number of convolutional kernel weights in the convolutional layer network structure of various neural network models. After these weight parameters are operated through the activation function on the neuron nodes, the output value of the deep convolutional neural network is approximately zero. Therefore, in order to minimize the impact on the final performance of the neural network model after training, the neuron nodes with the above properties can be selectively pruned, so that the convolutional calculation amount required by the neural network model during testing will be greatly reduced, and the computing resources occupied will also be reduced accordingly.

[0091] Repeatedly prune and train the MobileNet_YOLOv4 neural network model to reduce the connection channels between neurons in each convolutional layer and fully connected layer. First, analyze the importance of each convolutional layer, and sort the output weights of the convolutional kernels in each convolutional layer. After pruning, the output channel number of the convolutional layer in each convolutional layer can be adjusted to the most appropriate value to achieve the purpose of network simplification. Generally, during the process of pruning the convolutional layer channels, a pruning factor for each channel in each layer needs to be set. Then, sort the weights of each pruning factor, and finally, the pruning of the convolutional layer in the conditional network model can be carried out. Thus, a lightweight neural network model after pruning is obtained. At the same time, the model also needs to be retrained and optimized according to the performance and accuracy decline trends of the model before and after pruning.

[0092] In the MobileNet_YOLOv4 neural network model, the fully connected method is still the main way of connecting the network layers. Neurons with relatively large weights and high sensitivity can be retained, and neurons with relatively small weights and low sensitivity can be pruned, which can greatly increase the sparsity of the convolutional layer channels and reduce the total number of parameters of the MobileNet_YOLOv4 neural network model. This is beneficial to the full utilization of limited computing resources on edge computing devices. Since the PASCAL VOC format dataset is used to train the neural network model, and the pre-trained weights on the dataset are loaded when the model is built, the front-layer convolutional neural network is responsible for extracting some low-level image features from the input image data. At this time, the model already has a very good fitting effect on the feature information of the edge contour of the input image data. Therefore, the channel information of the front-layer neural network is far less important than the image information of the deep-layer neural network, so the weight is relatively small. The number of channels to be pruned decreases successively with the deepening of the complexity of the neural network model, and the weights increase successively. This pruning method can theoretically retain a considerable accuracy. At the same time, the sum of the absolute values of each parameter on each convolutional kernel can be used to compare and analyze the contribution of each convolutional kernel channel to the entire neural network. The summation formula for each convolutional kernel is shown as follows:

[0093] ;

[0094] Compare the size of each channel corresponding to each convolutional kernel with a set threshold. The channels below this threshold in each layer can be pruned. This can not only greatly reduce the total number of parameters of the model, but also minimize the loss of the average accuracy of the model. As the channels of each layer of the neural network are pruned, the computing resources occupied during the operation of the model gradually decrease.

[0095] Before detecting the characteristic signal of the circuit breaker fault signal, the variational mode decomposition of the signal is first performed. In this embodiment, preferably, in S3, before performing the convolution operation on the circuit breaker fault signal detection to obtain the characteristic signal, the variational mode decomposition of the signal is first performed. The optimal solution of the variational mode decomposition is obtained by the alternating direction multiplier method through iterative update, that is, the original signal is reconstructed by analyzing the finite bandwidth and center frequency of each intrinsic mode component; when using the variational mode decomposition to decompose the signal, the number of decomposed mode components is controlled by manual setting, and the original signal is decomposed into a certain number of intrinsic mode components by defining the decomposition scale K.

[0096] The original signal is reconstructed by analyzing the finite bandwidth and center frequency of each intrinsic mode component. In this embodiment, preferably, the specific principle of the variational mode decomposition is as follows:

[0097] S301. For each observed signal, assuming that it is the independent Gaussian noise superimposed on the original signal, first, the sampled signal f0 is denoised and reconstructed; the objective function is expressed as:

[0098] ,

[0099] where f is obtained by the regularization method:

[0100] ,

[0101] S302. Calculate the analytic signal of each mode u k through the Hilbert transform to obtain the unilateral spectrum of the mode component;

[0102] S303. Estimate the center frequency by the mixed mode function u k and the exponent, and then convert the spectrum of each mode to its respective estimated center frequency;

[0103] S304. Use the Gaussian smoothing method of the demodulated signal to estimate the bandwidth of each mode function, and the objective function is:

[0104] ,

[0105] where u k is the subsequence; K is the total number of subsequences; w k is the center frequency; f(t) is the original signal; is the Dirac distribution. The quadratic penalty factor, f(w), u(w), u(w) and the Lagrange multiplier are introduced into the above formula to realize the transformation from the constrained variational problem to the unconstrained variational problem. The following is the optimized expression:

[0106] ,

[0107] S305. To further solve the above formula, the alternating direction multiplier method is adopted, and the formula is extended as follows:

[0108] ,

[0109] ,

[0110] In the variational mode decomposition algorithm, first, the signal decomposition method is transformed into a variational problem under constraints, and then it is calculated to decompose the signal by finding the optimal solution. This decomposition step is continuously repeated to continuously update the central frequency and the corresponding bandwidth in each mode until multiple intrinsic mode components with relatively narrow bandwidths are obtained.

[0111] After the above processing of the signal, an improved S-transform is performed. As an extension of the short-time Fourier transform and the wavelet transform, the S-transform has good time-frequency analysis and feature extraction characteristics, and the result is more intuitive than the wavelet transform. However, the window function of the S-transform cannot be adjusted according to specific application requirements, lacks flexibility, and has a large amount of calculation, making it not suitable for embedded system applications. Therefore, this implementation scheme proposes a signal feature extraction method based on the improved S-transform, constructs an improved S-transform model, establishes a feature parameter recognition mechanism, and realizes fast and accurate signal recognition. In this embodiment, preferably, the variational mode decomposition is performed on the signal and then the improved S-transform is performed. For the signal feature extraction method of the improved S-transform, an improved S-transform model is constructed, a feature parameter recognition mechanism is established, fast and accurate signal recognition is realized, good time-frequency analysis and feature extraction characteristics are achieved, and the result is more intuitive.

[0112] Construct an improved S-transform model and establish a feature parameter recognition mechanism. In this embodiment, preferably, the implementation steps of the improved S-transform are as follows:

[0113] S306. Calculate the N-point FFT spectrum X(k / NT) of the signal x(nT);

[0114] S307. Use the envelope extreme value algorithm for X(k / NT) to determine the characteristic frequency k i , where i = 1, 2,... q, and q is the number of characteristic frequency points;

[0115] S308. According to the frequency band where the frequency point k i is located, determine the value of the window width adjustment scale factor of the Gauss adaptive optimization window. The parameter values are determined based on a large number of experiments: a = 0, c = 1; at the fundamental frequency, b = 1.6; in the 0 - 350 Hz frequency band excluding the fundamental frequency point, b = 0.4; in the 350 - 700 Hz frequency band, b = 0.35; above 700 Hz, b = 0.3; the above values can achieve satisfactory measurement results;

[0116] S309, Calculate the frequency point k i at the Fourier transform of the Gauss adaptive optimization window:

[0117] ,

[0118] S310, For the frequency point k i , shift the spectrum X(k i / NT) to X((k i+r ) / NT);

[0119] S311, Calculate the product A(r, k i+r ) / NT) and W(r / (NT), k i / (NT)): i ):

[0120] ,

[0121] S312, For the frequency point k i , calculate the inverse fast Fourier transform of A(r, k i ), and obtain the improved S transform corresponding to k i :

[0122] ;

[0123] S313, Repeat steps S308 - S312 to complete the improved S transforms corresponding to all characteristic frequency points, and complete the processing of abnormal signals on the edge side.

[0124] The working principle and usage process of the present invention:

[0125] Step 1, Construction of the MobileNet_YOLOv4 neural network model: The MobileNet_YOLOv4 neural network model uses multi-scale feature map mapping. The neural network is clipped to obtain feature maps of different scales of image data, and predictions are made separately after each layer of feature map is extracted; The convolutional layer of the MobileNet_YOLOv4 neural network model extracts six different feature maps for prediction, thereby constructing prior boxes for object recognition of different scales; Feature maps of different layers in the model respectively detect target objects of corresponding sizes; Finally, the model makes separate predictions on feature maps of different scales to obtain prediction values of different scales, significantly improving the object recognition performance of the model;

[0126] Replace the standard convolution structure of all convolutional layers in the Backbone of YOLOv4 with a depthwise separable convolution structure, and clip the fully connected layer and Softmax layer in the CSP-Darknet53 backbone network structure in the Backbone, because the MobileNet model comes with a fully connected layer and a Softmax layer. The depthwise separable convolution will greatly reduce the computational complexity of the network. It decomposes the standard convolution into two steps: depthwise convolution and pointwise convolution. The MobileNet_YOLOv4 neural network model first completes the convolution operation by treating the neural network nodes of each input data as specific convolution kernels through a 3×3 depthwise convolution, and at the same time uses the 1×1 convolution technology to complete the integration of the image data processed in the previous step, preventing the repeated operation of each neuron node from integrating all the signals input from the previous layer of the neural network. The total number of parameters of the MobileNet_YOLOv4 neural network model is greatly reduced compared to YOLOv4, the computational resources occupied will also decrease significantly, and the memory occupied by the model also becomes extremely small;

[0127] Step 2: Collect voiceprint through a high-sensitivity sensor: The high-sensitivity sensor includes a cone-shaped collector, a mounting ring, an energy converter, an upper pressing piece, and an isolator. The mechanical fault voiceprint of the circuit breaker is collected through the high-sensitivity sensor. The cone-shaped collector amplifies the original audio signal using the gas compression principle, and then conducts data transmission through low-power and long-distance wireless communication;

[0128] Step 3: Convolution operation of the convolutional kernel for edge-side recognition and early warning: Replace the standard convolution structure of all convolutional layers in the Backbone of YOLOv4 with a depthwise separable convolution structure. The MobileNet_YOLOv4 neural network model first completes the convolution operation by treating the neural network nodes of each input data as specific convolution kernels through a 3×3 depthwise convolution;

[0129] Step 4: Circuit breaker fault recognition: The accuracy of the target recognition and classification process model of normal and faulty circuit breakers obtained after the MobileNet_YOLOv4 neural network model is trained changes with the signal confidence. Train the neural network model with the PASCAL VOC format dataset, and load the pre-trained weights on the dataset when building the model. The front-layer convolutional neural network is responsible for extracting some low-level image features from the input image data;

[0130] Step 5: Perform an improved S transform on the signal: Improve the signal feature extraction method of the S transform, construct an improved S transform model, establish a feature parameter recognition mechanism, and achieve fast and accurate signal recognition.

[0131] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A 10kV circuit breaker mechanical fault detection method based on voiceprint, characterized in that: The following steps are included: S1. Construction of MobileNet_YOLOv4 neural network model: The MobileNet_YOLOv4 neural network model uses multi-scale feature map mapping, and the neural network is clipped to obtain feature maps of different scales of image data. After each layer of feature maps is extracted, predictions are made separately. The convolution layer of the MobileNet_YOLOv4 neural network model extracts six different feature maps for prediction, thereby constructing target recognition prior frames of different scales. The feature maps of different layers in the model respectively test the large and small targets of the matching size. Finally, the model makes predictions on feature maps of different scales to obtain prediction values ​​of different scales. S2. Collecting soundprints through high-sensitivity sensors: The high-sensitivity sensor includes a cone-shaped sound collector, a mounting ring, an energy converter, an upper pressure plate and an isolator. The circuit breaker mechanical fault soundprint is collected through the high-sensitivity sensor. The cone-shaped sound collector uses the principle of gas compression to amplify the original audio signal, and then transmits the data through low-power, long-distance wireless communication; S3. Convolution operation of the convolution kernel for edge recognition warning: The standard convolution structure of all convolution layers in Backbone in YOLOv4 is replaced with a depth-separable convolution structure. The MobileNet_YOLOv4 neural network model first completes the convolution operation by treating each neural network node of the input data as a specific convolution kernel through a 3×3 depth convolution. S4. Circuit breaker fault identification: The accuracy of the target identification and classification process of the normal circuit breaker and the faulty circuit breaker obtained by the MobileNet_YOLOv4 neural network model after training changes with the signal confidence. The neural network model is trained with a data set in PASCAL VOC format, and the pre-trained weights on the data set are loaded when the model is built. The front-layer convolutional neural network is responsible for some low-level image feature extraction of the input image data; S5. Improved S transform of the signal: Improve the signal feature extraction method of S transform, construct an improved S transform model, establish a feature parameter recognition mechanism, and realize fast and accurate signal recognition.

2. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 1, characterized in that: The MobileNet_YOLOv4 neural network model in S1 is divided into two independent steps, namely the deep convolution operation and the 1×1 convolution process; the two steps are spliced ​​with nonlinear activation units and BN layers at the output end; MobileNet makes extensive use of 1×1 convolution and deep convolution alternately. During the training process of the MobileNet neural network model, 1×1 convolution occupies three-quarters of the parameters and 95% of the computing resources.

3. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 1, characterized in that: The width factor of the MobileNet_YOLOv4 neural network model and resolution factor Make adjustments; Width Factor Width Factor It is a hyperparameter in the training and fitting process of the neural network model. The value range is (0, 1]; in simple terms It is regarded as the ratio of the total number of convolution operations per convolution layer of the adjusted MobileNet neural network model to the total number of convolution operations per convolution layer of the standard MobileNet neural network model; for the MobileNet neural network model using depthwise separable convolution, the total computational effort S of the neural network model after using the width factor is as follows: ， Standard MobileNet neural network model 1. Generally speaking, when reducing the network width factor to achieve the purpose of computing resources occupied when the model is running, the width factor The general values ​​are these four scales; different width factors are used when training on the same training set ImageNet The value of occupies the total number of parameters, computing resources, and accuracy of the neural network model.

4. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 3, characterized in that: Resolution factor of the MobileNet_YOLOv4 neural network model The value is between (0, 1]. In short, the resolution factor Scaling the input image data of each convolutional layer of the MobileNet neural network model is regarded as reducing each feature layer of the model by a certain proportion to achieve the purpose of reducing the total amount of calculation of the network model. The calculation amount S of the MobileNet neural network model with depthwise separable convolution as the core is shown as follows: ， Indicates the width factor When the resolution factor is different, the training is performed on the same training set ImageNet. The value of occupies the total number of parameters, computing resources, and accuracy of the neural network model.

5. The 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 1 is characterized in that: The MobileNet_YOLOv4 neural network model adopts the Mish activation function; The Mish activation function expression is as follows: 。 6. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 3, characterized in that: The width factor of the MobileNet_YOLOv4 neural network model during training and before and after the model compression process and resolution factor The values ​​are all set to 0.75; The MobileNet_YOLOv4 neural network model sets the number of categories to be classified to 2 during the training process, with n7 and n8 representing the labeled data set of normal samples and the labeled data set of fault samples respectively; The validation set is planned to be 123 groups, the test set is 315 groups, the initial learning rate is set to 0.002, the number of iterations of the entire training process is 10,000 times; the entire training process takes 9 hours.

7. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 1, characterized in that: In the S3, before performing convolution operation on the circuit breaker fault signal detection to obtain the characteristic signal, variational modal decomposition of the signal is first performed. The optimal solution of variational modal decomposition is iteratively updated by the alternating direction multiplier method, that is, the finite bandwidth and center frequency of each eigenmode component are analyzed to reconstruct it to obtain the original signal; when using variational modal decomposition to decompose the signal, the number of decomposed modal components is controlled by artificial setting, and the number of eigenmode components into which the original signal is decomposed is determined by defining the decomposition scale K.

Images