Voiceprint-based 10kV circuit breaker mechanical fault detection method

Abstract

The invention discloses a voiceprint-based 10 kV circuit breaker mechanical fault detection method. Comprising the following steps: S1, building a MobileNetYOLOv4 neural network model; s2, acquiring voiceprints through a high-sensitivity sensor; s3, performing convolution operation on a convolution kernel of edge side identification early warning; s4, circuit breaker fault identification; s5, carrying out improved S transformation on the signal; according to the invention, task allocation is dynamically calculated by a resource allocation module embedded in a cloud platform, a lightweight model is constructed, the model operation speed is accelerated, and a real-time target identification and classification system is constructed; the Moblienet neural network structure reduces the running memory of a neural network model and accelerates the running speed of the neural network model, and in order to improve the target recognition speed and solidify the model to an edge device, the MobileNet network structure applies a deep separable convolution technology, so that the target recognition efficiency is improved. The standard convolution of the convolutional neural network is divided into two steps of deep convolution and point-by-point convolution in the calculation process, so that the calculation amount of the network is greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of circuit breaker mechanical fault detection, and in particular relates to a 10kV circuit breaker mechanical fault detection method based on voiceprint. Background Art

[0002] As an important switchgear in the power system, when a system fault occurs, the circuit breaker as an actuator can timely and accurately remove the faulty part, close and disconnect the line, and play an important role in the control and protection of the power system. In view of the fact that power equipment has a certain service life cycle, the service life of high-voltage circuit breakers generally does not exceed 40 years under normal circumstances, and some power equipment components will age due to long-term use. Therefore, it is necessary to conduct timely and accurate status maintenance and monitoring diagnosis of circuit breaker faults and defects. Status maintenance combines real-time status monitoring and fault diagnosis of electrical equipment. The signal monitored by the status is 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 to conduct a comprehensive analysis of the historical archives, operating conditions, live testing or online monitoring data of high-voltage circuit breakers, which can not only timely determine the location of the fault and the severity of component failure, but also accurately diagnose the cause of the equipment failure. Therefore, it is a general trend to conduct status maintenance and fault diagnosis of 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 loss caused by its failure is huge, far exceeding its own price. The opening and closing of the circuit breaker is mainly completed by the mechanical structure. In order to achieve fast action, the mechanical structure needs to transmit high power and heavy load. As the number of actions increases, the mechanical properties of the circuit breaker will inevitably deteriorate. Some investigations and studies on the reliability of circuit breakers have shown that mechanical failure is the main factor in the failure of high-voltage circuit breakers. However, there are still various problems with the mechanical fault detection of various circuit breakers on the market.

[0004] However, although the cost of circuit breakers in the existing technology is not high, the impact of the failures caused by them is extremely severe. Edge computing has demonstrated its powerful functions in many fields, but it has not yet been involved in the mechanical monitoring of circuit breakers. In addition, the lightweight neural network model has become more and more efficient with the computational efficiency of deep learning technology, and its performance has become better and better. Under the condition of the same training data set, the accuracy is getting higher and higher. However, this also means that the structure of the deep learning network model is becoming more and more complex, and the computing resources occupied are increasing. At the same time, for edge anomaly analysis, it is difficult to tolerate too many resources and too much memory required for model operation. For this reason, we propose a 10kV circuit breaker mechanical fault detection method 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, deep convolution operation and 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 alternating 1×1 convolution and deep convolution. 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.

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

[0014] Width factor The width factor α is a hyperparameter in the training and fitting process of the neural network model. The value range of α is (0, 1]; in simple terms, α 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, the total computational amount S of the neural network model after using the width factor is as follows:

[0015] S=D K ·D K αM D F ·D F +αM·αN·D F ·D F ,

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

[0017] Preferably, the resolution factor β of the MobileNet_YOLOv4 neural network model is between (0, 1]. In short, the resolution factor β is the scaling of the input image data of each convolutional layer of the MobileNet neural network model, which is regarded as reducing each feature layer of the model by a certain proportion to achieve the purpose of cutting 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:

[0018] S=D K ·D K ·αM·βD F βD F+αM·αN·βD F βD F ,

[0019] It shows the total number of parameters, computing resources, and accuracy of the neural network model when taking different values ​​of resolution factor β when training on the same training set ImageNet while keeping the width factor α unchanged.

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

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

[0022]

[0023] Preferably, the width factor α and the resolution factor β of the MobileNet_YOLOv4 neural network model are both set to 0.75 during the training process and before and after the model compression process, so as to ensure that the hyperparameters will not have a significant impact on the accuracy of the network model and greatly reduce the computing resources occupied by the model during operation;

[0024] During the training process, the MobileNet_YOLOv4 neural network model sets the number of categories that need to be classified to 2, with n7 and n8 representing the labeled data sets of normal samples and fault samples respectively; 123 groups of validation sets and 315 groups of test sets are planned, the initial learning rate is set to 0.002, and the number of iterations of the entire training process is 10,000; the entire training process takes 9 hours.

[0025] Preferably, in S3, before performing convolution operation on circuit breaker fault signal detection to obtain characteristic signal, variational modal decomposition of the signal is performed first, and the optimal solution of variational modal decomposition is iteratively updated by 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.

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

[0027] S301, for each observed signal, it is assumed that it is an independent Gaussian noise superimposed by the original signal, and the sampled signal f0 is first denoised and reconstructed; the objective function is expressed as:

[0028] f0=f+η,

[0029] Where f is obtained by regularization method:

[0030]

[0031] S302, calculate each mode u by Hilbert transform k The analytical signal of the modal component is obtained by

[0032] S303, through the mixed mode function u k and exponents to estimate the center frequency, and then transform the spectrum of each mode to its respective estimated center frequency;

[0033] S304, using the Gaussian smoothing method of the demodulated signal to estimate the bandwidth of each modal function, the objective function is:

[0034]

[0035] where u k is a 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 multiplication operator 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:

[0036]

[0037] S305, to further solve the above formula, the alternating direction multiplication method is used, and the formula is expanded as follows:

[0038]

[0039] In the variational mode decomposition algorithm, the signal decomposition method is first transformed into a variational problem under constraints, and then it is calculated to achieve the purpose of decomposing the signal by finding the optimal solution. The decomposition steps are repeated continuously so that the center frequency and corresponding bandwidth in each mode are continuously updated until multiple inherent mode components with relatively narrow bandwidth are obtained.

[0040] Preferably, the variational mode decomposition is used to process the signal and then an improved S transform is performed. The signal feature extraction method of the improved S transform constructs an improved S transform model, establishes a feature parameter recognition mechanism, realizes fast and accurate signal recognition, realizes good time-frequency analysis and feature extraction characteristics, and the result is more intuitive.

[0041] Preferably, the improved S transform is implemented in the following steps:

[0042] S306, calculating the N-point FFT spectrum X(k / NT) of the signal x(nT);

[0043] S307, using the envelope extreme value algorithm to determine the characteristic frequency k for X(k / NT) i , where i = 1, 2, ... q, q is the number of characteristic frequency points;

[0044] S308, according to the frequency point k i The frequency band in which the Gauss adaptive optimization window is located determines the value of the window width adjustment scale factor. The parameter values ​​are determined based on a large number of experiments: a = 0, c = 1; b = 1.6 at the base frequency; b = 0.4 in the 0-350Hz frequency band except the base frequency point; b = 0.35 in the 350-700Hz frequency band; b = 0.3 above 700Hz; the above values ​​can achieve satisfactory measurement results;

[0045] S309, calculate frequency point k i The Fourier transform of the Gauss adaptive optimization window at:

[0046]

[0047] S310, for frequency point k i , shift spectrum X(k i / NT) to X((k i+r ) / NT);

[0048] S311, calculate X((k i+r ) / NT) and W(r / (NT), k i / (NT)) product A(r, k i ):

[0049]

[0050] S312, for frequency point k i , calculate A(r, k i ) is transformed by the inverse fast Fourier transform to obtain k i The corresponding improved S transform:

[0051]

[0052] S313, repeating steps S308-S312, completing the improved S transform corresponding to all characteristic frequency points, and completing the abnormal signal processing at the edge side.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention uses a resource allocation module embedded in the cloud platform to achieve optimized dynamic computing task allocation. Some IoT application tasks require programmed analysis, processing, and management of global data and resources, and data centers can provide strong computing support for this. Other tasks require efficient and fast real-time responses, which are supported by edge computing and terminal devices deployed on the edge side.

[0055] While ensuring that the neural network model has a certain level of accuracy and performance, we try to build a lightweight model, speed up the model operation, and build a real-time target recognition and classification system.

[0056] At present, the lightweight model mainly aims at redesigning the existing neural network model, building a more efficient and simpler neural network structure, while reducing the computing power occupied by the neural network parameters, and trying to ensure that the accuracy of the neural network model will not decrease significantly. At the same time, in the training process of the neural network, using a lighter module structure to splice with the existing model can also reduce the computing time;

[0057] The MobileNet 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 the edge device, the MobileNet network structure mainly uses the technology of deep separable convolution, which can divide the standard convolution of the convolutional neural network into two steps: deep convolution and point-by-point convolution in the calculation process, greatly reducing the amount of network calculation;

[0058] The model accuracy 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 significantly reduce the computing resources and parameters of the model operation while ensuring a relatively considerable accuracy rate, and can also greatly speed up the response speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of the steps of the present invention;

[0060] Figure 2 Schematic diagram of standard convolution and depth-wise separable convolution of the present invention;

[0061] Figure 3 It is a schematic diagram of the model accuracy of different width factors α of the present invention;

[0062] Figure 4 It is a schematic diagram of the model accuracy of different resolution factors β of the present invention;

[0063] Figure 5 This is a schematic diagram of edge side recognition warning of the present invention;

[0064] Figure 6 It is a schematic diagram of the neural network model training process of the present invention. DETAILED DESCRIPTION

[0065] 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.

[0066] See also Figure 1-Figure 6 The present invention provides a technical solution: a 10kV circuit breaker mechanical fault detection method based on voiceprint, comprising the following steps:

[0067] 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.

[0068] The standard convolution structure of all convolution layers in Backbone in YOLOv4 is replaced by the deep separable convolution structure, and the fully connected layer and Softmax layer in the backbone network structure of CSP-Darknet53 in Backbone are cut off, because the MobileNet model has its own fully connected layer and Softmax layer. 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;

[0069] 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;

[0070] 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.

[0071] 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;

[0072] 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.

[0073] The ImageNet image data training set uses standard convolution and depth separable convolution for comparative testing. The model accuracy obtained using standard convolution is about 1% higher than that of depth 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 parameter amount of the model operation while ensuring a relatively considerable accuracy rate. At the same time, the reaction speed of the model can be greatly accelerated. In this embodiment, preferably, the MobileNet_YOLOv4 neural network model in S1 is divided into two independent steps, namely, the depth convolution operation and the 1×1 convolution process; the two steps splice the nonlinear activation unit and the BN layer at the output end; MobileNet uses a large number of 1×1 convolutions and depth convolutions alternately. In the training process of the MobileNet neural network model, 1×1 convolution occupies three quarters of the parameters and 95% of the computing resources.

[0074] MobileNet has an absolute advantage in parameter quantity and computational complexity over traditional neural network models. However, for embedded edge computing devices, this is far from enough. Therefore, it is also necessary to adjust the width factor α and resolution factor β of the MobileNet neural network model to meet the extremely stringent requirements on the computing resources occupied by the neural network model on the edge device. In this embodiment, preferably, the width factor α and resolution factor β of the MobileNet_YOLOv4 neural network model are adjusted;

[0075] Width factor The width factor α is a hyperparameter in the training and fitting process of the neural network model. The value range of α is (0, 1]; in simple terms, α 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, the total computational amount S of the neural network model after using the width factor is as follows:

[0076] S=D K ·D K αM D F ·D F +αM·αN·D F ·D F ,

[0077] The standard MobileNet neural network model has an α of 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 α is generally taken as these four scales; the total number of parameters, computing resources and accuracy of the neural network model when taking different width factor α values ​​when training on the same training set ImageNet;

[0078] The resolution factor β of the MobileNet_YOLOv4 neural network model takes values ​​between (0, 1]. In short, the resolution factor β is the scaling of the input image data of each convolutional layer of the MobileNet neural network model, which is regarded as reducing each feature layer of the model by a certain proportion to achieve the purpose of cutting 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 in the formula:

[0079] S=D K ·D K ·αM·βD F βD F +αM·αN·βD F βD F ,

[0080] It shows the total number of parameters, computing resources, and accuracy of the neural network model when taking different values ​​of resolution factor β when training on the same training set ImageNet while keeping the width factor α unchanged.

[0081] Compared with the traditional YOLO model, YOLOv4 improves the feature extraction network, uses algorithmic means after Backbone to increase the data feature receptive field, and changes the activation function to the Mish activation function. Since the MobileNet_YOLOv4 neural network model used replaces the Backbon of YOLOv4 with MobileNet, the Mish activation function has no upper bound, so the neural network model does not have the gradient saturation phenomenon caused by the reduction in training speed during training. At the same time, the Mish activation function has no lower bound, which can make the model more regularized and have a better fitting effect. At the same time, the Mish activation function has a larger amount of calculation than the ReLU used in the previous YOLO series models, but the results obtained under the same training set conditions in the same neural network structure are better. In this embodiment, preferably, the MobileNet_YOLOv4 neural network model adopts the Mish activation function;

[0082] The Mish activation function expression is as follows:

[0083]

[0084] Before the training begins, the MobileNet hyperparameters in the MobileNet_YOLOv4 network structure must be calibrated. When the values ​​of α and β in the MobileNet network structure gradually decrease, the accuracy of the target recognition and classification completed by the MobileNet_YOLOv4 neural network model gradually decreases, and the trend becomes increasingly significant, especially when the width factor α and the resolution factor β values ​​decrease from 0.5 to 0.25, the function of the neural network model even reaches the point of disappearing directly. Therefore, while considering the computing resources occupied by the model during operation, a certain amount of accuracy must also be guaranteed, because the model will be pruned later, and these operations can also greatly reduce the computing resources occupied by the model. In this embodiment, preferably, the width factor α and the resolution factor β of the MobileNet_YOLOv4 neural network model are set to 0.75 during the training process and before and after the model compression process, ensuring that the hyperparameters will not have a significant impact on the accuracy of the network model, while greatly reducing the computing resources occupied by the model during operation;

[0085] During the training process, the MobileNet_YOLOv4 neural network model sets the number of categories that need to be classified to 2, with n7 and n8 representing the labeled data sets of normal samples and fault samples respectively; 123 groups of validation sets and 315 groups of test sets are planned, the initial learning rate is set to 0.002, and the number of iterations of the entire training process is 10,000; the entire training process takes 9 hours.

[0086] The accuracy of the target recognition and classification process model 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, where n7 is the dataset label of the normal sample when the LabelImg visual image annotation tool is used to annotate the dataset, n8 is the dataset label of the faulty sample, and the hyperparameter of the YOLOv4 neural network model is set to 0.5;

[0087] Trainloss represents the loss of the model on the training set, and represents the model fitting ability of the neural network model for the experimental training set data. Valloss represents the loss of the model on the test set, and represents the model fitting ability of the neural network model for the experimental test set data. This ability truly reflects the model's ability to fit the test data, and can also be regarded as the performance of the model, collectively referred to as the generalization ability of the model. The loss function gradually stabilizes after decreasing by a certain amount as the number of iterations increases, proving that the model training process has been completed. The loss function converges quickly and the final value is low, proving that the model training effect is good, the ability to fit data is strong, and it has a good fitting effect on both the training set and the test set.

[0088] 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 was completed. It was found that there are a lot of convolution kernel weights in the convolution layer network structure in various neural network models. After these weight parameters are parameterized by the activation function on the neuron node, 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 amount of convolution calculations required for the neural network model during testing will be greatly reduced, and the computing resources occupied will also be reduced.

[0089] Repeated pruning training is performed on the MobileNet_YOLOv4 neural network model to reduce the connection channels between each convolutional layer and each neuron node in the fully connected layer. First, the importance of each convolutional layer should be analyzed, and the output weights of the convolution kernels in each convolutional layer should be sorted; after pruning, the number of output channels of the convolutional layer in each convolutional layer can be adjusted to the most appropriate value to achieve the purpose of network simplification. Generally, in the process of pruning the convolutional layer channels, it is necessary to set the pruning factors for each channel in each layer. Then the weights of each pruning factor are sorted, and finally the pruning of the convolutional layer in the conditional network model is possible. Thus, a lightweight neural network model is obtained after pruning; at the same time, the model should be retrained and optimized according to the performance and accuracy decline trend of the model before and after model pruning.

[0090] In the MobileNet_YOLOv4 neural network model, the fully connected mode is still the main mode of connection between each layer of the network. It can retain neurons with relatively large weights and high sensitivity, and cut off neurons with small weights and low sensitivity, which can greatly increase the sparsity of the convolution layer channels and reduce the overall parameter volume of the MobileNet_YOLOv4 neural network model. This is conducive 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 of the convolutional neural network is responsible for some low-level image feature extraction of the input image data. At this time, the model 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 of the neural network is far less important than the image information of the deep neural network, so the weight is small. The number of channel cuts decreases with the increasing complexity of the neural network model, and the weight increases. This cutting method can theoretically retain a relatively considerable accuracy. At the same time, the absolute value box of each parameter on each convolution kernel can be used to compare and analyze the contribution of each convolution kernel channel to the entire neural network. The summation formula of each convolution kernel is shown as follows:

[0091]

[0092] The size of the channel corresponding to each convolution kernel is compared with a certain threshold. Channels below the threshold in each layer can be pruned. This can not only greatly reduce the total number of model parameters, but also minimize the average accuracy loss of the model. With the pruning of channels in each layer of the neural network, the computing resources occupied by the model are gradually reduced.

[0093] Before performing circuit breaker fault signal detection characteristic signal, variational modal decomposition of the signal is performed. In this embodiment, preferably, in S3, variational modal decomposition of the signal is performed before performing convolution operation on circuit breaker fault signal detection to obtain characteristic signal. The variational modal decomposition is used to obtain the optimal solution by iteratively updating through alternating direction multiplier method, that is, by analyzing the finite bandwidth and center frequency of each eigenmode component 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.

[0094] By analyzing the finite bandwidth and center frequency of each eigenmode component, the original signal is reconstructed. In this embodiment, preferably, the specific principle of the variational mode decomposition is as follows:

[0095] S301, for each observed signal, it is assumed that it is an independent Gaussian noise superimposed by the original signal, and the sampled signal f0 is first denoised and reconstructed; the objective function is expressed as:

[0096] f0=f+η,

[0097] Where f is obtained by regularization method:

[0098]

[0099] S302, calculate each mode u by Hilbert transform k The analytical signal of the modal component is obtained by

[0100] S303, through the mixed mode function u k and exponents to estimate the center frequency, and then transform the spectrum of each mode to its respective estimated center frequency;

[0101] S304, using the Gaussian smoothing method of the demodulated signal to estimate the bandwidth of each modal function, the objective function is:

[0102]

[0103] where u k is a 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 multiplication operator 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:

[0104]

[0105] S305, to further solve the above formula, the alternating direction multiplication method is used, and the formula is expanded as follows:

[0106]

[0107] In the variational mode decomposition algorithm, the signal decomposition method is first transformed into a variational problem under constraints, and then it is calculated to achieve the purpose of decomposing the signal by finding the optimal solution. The decomposition steps are repeated continuously so that the center frequency and corresponding bandwidth in each mode are continuously updated until multiple inherent mode components with relatively narrow bandwidth are obtained.

[0108] After the above-mentioned processing is performed on the signal, an improved S transform is performed. As an extension of short-time Fourier transform and 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, which is not suitable for embedded system applications. Therefore, this implementation scheme proposes a signal feature extraction method based on an improved S transform, constructs an improved S transform model, establishes a feature parameter identification mechanism, and realizes rapid and accurate signal recognition. In this embodiment, preferably, the variational mode decomposition is used to process the signal before the improved S transform is performed. The signal feature extraction method of the improved S transform constructs an improved S transform model, establishes a feature parameter identification mechanism, realizes rapid and accurate signal recognition, and realizes good time-frequency analysis and feature extraction characteristics, and the result is more intuitive.

[0109] An improved S transform model is constructed to establish a characteristic parameter identification mechanism. In this embodiment, preferably, the improved S transform is implemented in the following steps:

[0110] S306, calculating the N-point FFT spectrum X(k / NT) of the signal x(nT);

[0111] S307, using the envelope extreme value algorithm to determine the characteristic frequency k for X(k / NT) i , where i = 1, 2, ... q, q is the number of characteristic frequency points;

[0112] S308, according to the frequency point k i The frequency band in which the Gauss adaptive optimization window is located determines the value of the window width adjustment scale factor. The parameter values ​​are determined based on a large number of experiments: a = 0, c = 1; b = 1.6 at the base frequency; b = 0.4 in the 0-350Hz frequency band except the base frequency point; b = 0.35 in the 350-700Hz frequency band; b = 0.3 above 700Hz; the above values ​​can achieve satisfactory measurement results;

[0113] S309, calculate frequency point k iThe Fourier transform of the Gauss adaptive optimization window at:

[0114]

[0115] S310, for frequency point k i , shift spectrum X(k i / NT) to X((k i+r ) / NT);

[0116] S311, calculate X((k i+r ) / NT) and W(r / (NT), k i / (NT)) product A(r, k i ):

[0117]

[0118] S312, for frequency point k i , calculate A(r, k i ) is transformed by the inverse fast Fourier transform to obtain k i The corresponding improved S transform:

[0119]

[0120] S313, repeating steps S308-S312, completing the improved S transform corresponding to all characteristic frequency points, and completing the abnormal signal processing at the edge side.

[0121] The working principle and use process of the present invention:

[0122] Step 1: Build the 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 convolutional 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 targets of different sizes that match them. 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.

[0123] The standard convolution structure of all convolution layers in Backbone in YOLOv4 is replaced by the deep separable convolution structure, and the fully connected layer and Softmax layer in the backbone network structure of CSP-Darknet53 in Backbone are cut off, because the MobileNet model has its own fully connected layer and Softmax layer. 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;

[0124] Step 2: Collect 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 is subjected to mechanical fault soundprint collection 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.

[0125] Step 3: Convolution operation of the convolution kernel for edge-side 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.

[0126] Step 4. Circuit breaker fault identification: After the training of the MobileNet_YOLOv4 neural network model, the target recognition and classification process model accuracy of normal circuit breakers and faulty circuit breakers varies with the signal confidence. The neural network model is trained with a data set in PASCAL VOC format, and the model is loaded with pre-trained weights on the data set when it is built. The front-layer convolutional neural network is responsible for some low-level image feature extraction of the input image data;

[0127] 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.

[0128] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that 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 α and the resolution factor β of the MobileNet_YOLOv4 neural network model are adjusted; Width factor The width factor α is a hyperparameter in the training and fitting process of the neural network model. The value range of α is (0, 1]; in simple terms, α 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, the total computational amount S of the neural network model after using the width factor is as follows: S=D K ·D K ·αM·D F ·D F +αM·αN·D F ·D F , The α of the standard MobileNet neural network model is 1. Generally speaking, when reducing the network width factor to achieve the purpose of reducing the computing resources occupied when the model is running, the width factor α generally takes the following four scales; when training on the same training set ImageNet, the total number of parameters, computing resources, and accuracy of the neural network model occupied by taking different width factor α values.

4. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 3, characterized in that: The resolution factor β of the MobileNet_YOLOv4 neural network model takes values ​​between (0, 1]. In short, the resolution factor β is the scaling of the input image data of each convolutional layer of the MobileNet neural network model, which is regarded as reducing each feature layer of the model by a certain proportion to achieve the purpose of cutting 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 in the formula: S=D K ·D K ·αM·βD F ·βD F +αM·αN·βD F ·βD F , It shows the total number of parameters, computing resources, and accuracy of the neural network model when taking different values ​​of resolution factor β when training on the same training set ImageNet while keeping the width factor α unchanged.

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 α and resolution factor β of the MobileNet_YOLOv4 neural network model are both set to 0.75 during the training process and before and after the model compression process; 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.

8. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 7, characterized in that: The specific principle of the variational mode decomposition is as follows: S301, for each observed signal, it is assumed that it is an independent Gaussian noise superimposed by the original signal, and the sampled signal f0 is first denoised and reconstructed; the objective function is expressed as: f0=f+η, Where f is obtained by regularization method: S302, calculate each mode u by Hilbert transform k The analytical signal of the modal component is obtained by S303, through the mixed mode function u k and exponents to estimate the center frequency, and then transform the spectrum of each mode to its respective estimated center frequency; S304, using the Gaussian smoothing method of the demodulated signal to estimate the bandwidth of each modal function, the objective function is: where u k is a 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 multiplication operator 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: S305, to further solve the above formula, the alternating direction multiplication method is used, and the formula is expanded as follows: In the variational mode decomposition algorithm, the signal decomposition method is first transformed into a variational problem under constraints, and then it is calculated to achieve the purpose of decomposing the signal by finding the optimal solution. The decomposition steps are repeated continuously so that the center frequency and corresponding bandwidth in each mode are continuously updated until multiple inherent mode components with relatively narrow bandwidth are obtained.

9. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 8, characterized in that: The variational mode decomposition is used to process the signal and then an improved S transform is performed, a signal feature extraction method of the improved S transform is used, an improved S transform model is constructed, and a feature parameter recognition mechanism is established.

10. A 10kV circuit breaker mechanical fault detection method based on voiceprint according to claim 9, characterized in that: The improved S transform is implemented as follows: S306, calculating the N-point FFT spectrum X(k / NT) of the signal x(nT); S307, using the envelope extreme value algorithm to determine the characteristic frequency k for X(k / NT) i , where i = 1, 2, ... q, q is the number of characteristic frequency points; S308, according to the frequency point k i The frequency band in which the Gauss adaptive optimization window is located determines the value of the window width adjustment scale factor. The parameter values ​​are determined based on a large number of experiments: a = 0, c = 1; b = 1.6 at the base frequency; b = 0.4 in the 0-350Hz frequency band except the base frequency point; b = 0.35 in the 350-700Hz frequency band; b = 0.3 above 700Hz; The above values ​​can achieve satisfactory measurement results; S309, calculate frequency point k i The Fourier transform of the Gauss adaptive optimization window at: S310, for frequency point k i , shift spectrum X(k i / NT) to X((k i+r ) / NT); S311, calculate X((k i+r ) / NT) and W(r / (NT), k i / (NT)) product A(r, k i ): S312, for frequency point k i , calculate A(r, k i ) is transformed by the inverse fast Fourier transform to obtain k i The corresponding improved S transform: S313, repeating steps S308-S312, completing the improved S transform corresponding to all characteristic frequency points, and completing the abnormal signal processing at the edge side.

Images