Distribution network hardware fitting defect detection method and system

Abstract

The invention discloses a distribution network fitting defect detection method and system. The method comprises the following steps: preprocessing a collected distribution network fitting vibration signal and dividing a data set; designing a multi-scale time-frequency domain neural network core structure; constructing a model based on a multi-scale time-frequency domain neural network; and obtaining an output result of the test set sample based on the multi-scale time-frequency domain neural network. According to the distribution network hardware defect detection method and system disclosed by the invention, multi-time-frequency scale analysis of nonlinear non-stationary hardware vibration signals is provided, and the characteristics of the signals in the time domain and the frequency domain can be dynamically captured, so that the hardware defect detection precision is improved, and the actual requirements of accuracy and rapidness are met.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network hardware defect detection, and in particular to a distribution network hardware defect detection method and system. Background Art

[0002] Connectors and protective fittings are key components in the distribution network and play a decisive role in the safe operation of the power grid. Connectors are not only the basic components of the power system, but also responsible for the transmission and distribution of electricity. Their functionality is directly related to the stability and safety of the entire power grid. Good connectors can ensure stable connections between power equipment and effectively prevent electrical accidents and line failures. At the same time, as a protective device for the power grid, protective fittings can protect the power grid and its connected equipment from damage in the event of overload, short circuit and ground fault. Therefore, ensuring the reliability and stability of connectors and protective fittings is crucial to the smooth operation of the distribution network and needs to be strictly managed and continuously optimized. By deeply studying and improving these fittings, we can improve the reliability and safety of the power grid and meet the growing demand for stable power supply in modern society.

[0003] Traditional methods for detecting defects in distribution network hardware include manual visual inspection, infrared and ultraviolet detection, sound detection, and physical and electrical performance measurements. These methods usually require professional inspectors to perform them regularly and rely on manual operation and intuitive observation, such as checking with the naked eye whether the hardware has obvious damage such as cracks, rust or deformation. Although these methods are simple and direct, they are usually not accurate enough to provide early warning of faults. In addition, due to the heavy reliance on manpower for inspections, these traditional methods are slow in operation speed and efficiency, and consume a lot of manpower, especially in vast or complex distribution network areas.

[0004] With the advancement of technology, neural networks, especially deep learning methods, have been applied to defect detection of distribution network fittings, mainly through image recognition technology. These methods use deep learning models such as convolutional neural networks to analyze images taken by surveillance cameras or drones to automatically identify various defects in fittings. However, this image-based neural network method faces some challenges, such as the limited number of training samples and the small size of the fitting targets in the image, resulting in low resolution, which affects the detection accuracy. In addition, the image processing speed is slow when processing a large number of high-resolution images, which becomes a limiting factor in application scenarios that require real-time or near real-time response.

[0005] Xie Zhihui and others from the School of Automation and Electronic Engineering at Qingdao University of Science and Technology proposed a small sample recognition algorithm for hardware images based on an improved U-Net. The algorithm expands the data set by generating adversarial networks and introduces foreground enhancement and attention mechanisms to enhance feature extraction. However, this method faces the problems of large differences between generated samples and actual ones, noise introduced by background masks, and increased computational complexity affecting model efficiency and accuracy.

[0006] Zhao Zhenbing from the Department of Electronics and Communication Engineering at North China Electric Power University proposed a method for detecting transmission line fittings and their defects based on contextual structural reasoning. Combining the structural knowledge of transmission line fittings and their defects and using bidirectional gated recurrent units and self-attention mechanisms, the confidence of correct positive samples is increased and the confidence of incorrect positive samples is reduced, thereby improving the overall average accuracy. However, this method faces the problems of high computing resource consumption, slow processing speed, and high difficulty in adjusting and maintaining complex models.

[0007] In the prior art, the invention patent with the publication number CN116992280A discloses a bearing fault diagnosis method and system based on a multi-scale wavelet thresholding network, and inputs the one-dimensional bearing fault vibration signal of the training set into the fault diagnosis model for training. The training process of the fault diagnosis model includes: performing frequency division analysis on the bearing fault vibration signal through a multi-scale convolutional wavelet decomposition network; wherein the multi-scale convolutional wavelet decomposition network includes a single-channel convolutional wavelet decomposition network with three different-scale convolutional kernels; the convolutional wavelet decomposition networks of different channels are used to extract time-frequency features of different characteristics; the output features of the multi-scale convolutional wavelet decomposition network are further extracted by using the parallel network DRSN-BiGRU of the deep residual shrinkage network and BiGRU; and the output features of DRSN-BiGRU are classified by using the fully connected layer and the Dropout layer. However, the wavelet transform cannot be adjusted dynamically, and the parameters of the wavelet decomposition network are preset values. As well as the combination of convolutional wavelet decomposition network and BiGRU network, it focuses more on utilizing the multi-scale capability of wavelet analysis and the bidirectional information flow of BiGRU. It is suitable for applications that require comprehensive analysis of signal history and future trends, and is not suitable for local characterization. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a rapid and accurate detection of hardware defects.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] A method for detecting defects in distribution network hardware, comprising: using a real-time vibration signal of a distribution network hardware as an input of an optimal multi-scale time-frequency domain neural network, and the optimal multi-scale time-frequency domain neural network outputs a defect type of the distribution network hardware;

[0011] Among them, the optimal multi-scale time-frequency domain neural network is obtained through the following steps:

[0012] Construct a training set of distribution network hardware vibration signals;

[0013] Establish an initial multi-scale time-frequency domain neural network, including an adaptive filter bank decomposition layer, a multi-time-frequency scale encoder layer, a frequency domain graph convolution layer, and a hardware defect feature detection layer;

[0014] The kth distribution network fitting defect x at the tth sampling time in the training set k,t Input into the adaptive filter bank decomposition layer to obtain the kth distribution network hardware defect data x k,t The intrinsic information F of the distribution network fittings defect at the tth time step k,t ;

[0015] The intrinsic information of the defects of the distribution network hardware F k,t Input to the multi-time-frequency scale Encoder layer to obtain the modulation feature R of the distribution network hardware defect k,t ;

[0016] The kth data x at the tth sampling time in the training set k,t Input the mid-frequency domain graph convolution layer to obtain the spectrum feature H l ;

[0017] The modulation feature R of the distribution network hardware defect k,t And the spectrum characteristics H l Input into the hardware defect feature detection layer, for the kth data x at the tth sampling time k,t Identify and classify defects in distribution network fittings;

[0018] Based on the distribution network hardware vibration signal training set and combined with the optimization algorithm, the initial multi-scale time-frequency domain neural network is optimized to obtain the optimal multi-scale time-frequency domain neural network.

[0019] Advantages: The present invention utilizes a deep learning network framework and comprehensively considers the defect characteristics of distribution network hardware. It uses an adaptive filter bank decomposition layer to preliminarily process the input distribution network hardware vibration signal data, and extracts the intrinsic information and coding features of distribution network hardware defects through a multi-time-frequency scale encoder layer and a frequency domain graph convolution layer. Then, the dynamic gating information of different scales is analyzed through a multi-head wavelet attention mechanism layer, and the learning ability of the model is enhanced by combining a residual connection feedforward neural network layer. Finally, the final fault type prediction is output through a Softmax layer.

[0020] The input of the model is the defect data of distribution network fittings. After preliminary processing by the adaptive filter bank decomposition layer, the data enters the multi-time-frequency scale Encoder layer, which uses wavelet transforms of different time and frequency scales to extract deep fault features. Then, these features are sent to the frequency domain graph convolution layer to construct a spectrum graph and perform graph convolution operations to obtain more refined spectrum features. Next, the features of different scales are dynamically adjusted and integrated through the multi-head wavelet attention mechanism layer. Finally, the residual connection feedforward neural network layer further modulates these features, and the probability values ​​of all fault types corresponding to the input fault data are calculated through the Softmax layer to obtain the final fault label.

[0021] In one embodiment of the present invention, a distribution network hardware vibration signal training set is constructed and obtained by the following steps:

[0022] Collect vibration signal data of different distribution network hardware and construct a hardware defect classification set, denoted as X, and X = {X 1 …X k …X K}, X k represents the vibration signal of the kth hardware defect data, and X k ={x k,1 …x k,t …x k,T}, x k,t Indicates the vibration amplitude of the kth hardware defect data at the tth sampling time; 1≤k≤K, K represents the total number of hardware defects; 1≤t≤T, T represents the total sampling time;

[0023] Construct the defect category set Y of distribution network fittings, denoted as Y = {y 1 ,…,y k ,…,y K}, where y k represents the label value of the kth distribution network hardware defect data, and y k Belongs to [1,N], where N is the number of types of hardware defects;

[0024] The labeled distribution network fitting defect data set P = (X, Y) is randomly shuffled and used as the distribution network fitting vibration signal training set.

[0025] In one embodiment of the present invention, the adaptive filter bank decomposition layer uses (1)-(5) to perform adaptive filtering processing to obtain the kth distribution network hardware defect data x k,t The intrinsic information F of the distribution network fittings defect at the tth time step k,t :

[0026] P k,t =|x k,t | 2 #(1)

[0027]

[0028] α k,t =Sigmoid(W p ·P k,t +W ω ·ω k,t +b α )#(4)

[0029]

[0030] Where ω is the frequency, which usually ranges from 0 to 2π; M is the length of the window function, m belongs to [-M,M]; α k,t is the adjustment function of the defect data of the kth distribution network hardware at the tth moment; P k,t is the kth distribution network fitting defect data x k,t The instantaneous energy at the tth sampling moment; Hilbert represents the Hilbert transform; represents the result of Hilbert transform; ∠ represents the phase angle; ω k,t is the instantaneous frequency; W p YesP k,t The weight matrix W ω Yes k,t The weight matrix of α is the bias; Sigmoid is a function, e is a natural constant, and j is an imaginary unit.

[0031] In one embodiment of the present invention, the modulation characteristic R of the distribution network fitting defect k,t Obtained through:

[0032] The multi-time-frequency scale Encoder layer includes a multi-scale wavelet attention mechanism layer and a residual connection feedforward neural network layer connected in sequence;

[0033] In the multi-scale wavelet attention mechanism layer:

[0034] Extracting intrinsic information of distribution network hardware defects k,t The multi-scale time-frequency characteristics of the k-th hardware defect data x are obtained k,t Distribution network fitting defect coding features at the s-th scale at the t-th time step

[0035] According to the s-th scale hardware defect coding characteristics Get the kth distribution network hardware defect data x k,t The dynamic gating G at time step t k,t ;

[0036] Combine all scales of hardware defect coding features to obtain the kth hardware defect data x k,tAggregation characteristics M of distribution network fitting defects at the t-th time step k,t ;

[0037] Residual connection feedforward neural network layer:

[0038] Input the aggregation characteristics M of distribution network fitting defects k,t into the residual connection feedforward neural network layer to obtain the k-th fitting defect data x k,t Modulation characteristics R of distribution network fitting defects at the t-th time step k,t .

[0039] In an embodiment of the present invention, the k-th fitting defect data x k,t Fitting defect coding characteristics of the s-th scale at the t-th time step are obtained using equations (6)-(9):

[0040] γ′ s =γ s ·(1 + η s ·sigmoid(W γ ·F k,t +b γ ))#(6)

[0041] λ′ s =λ s ·(1 + ξ s ·sigmoid(W λ ·F k,t +b λ ))#(7)

[0042]

[0043] In the formula, u is the wavelet time factor, 1 < u ≤ U; s is the wavelet scale factor; x k,t+u is the distribution fitting defect data at the k-th and (t + u)-th moments; ψ′ s (t, u; γ′ s , λ′ s ) is the dynamically adjusted wavelet basis function; ω s is the corresponding scale angular frequency; γ′ s is the amplitude adjustment parameter; λ′ s is the scale adjustment parameter; γ s is the initial amplitude parameter; λ s is the initial scale adjustment parameter; W γ and b γ are the amplitude weight and amplitude bias; η s is the amplitude adjustment factor; ξ s is the scale adjustment factor, Sigmoid is the function, e is the natural constant, j is the imaginary unit, and ψ is the wavelet basis function.

[0044] In one embodiment of the present invention, the kth distribution network fitting defect data x k,t The dynamic gating G at time step t k,t , using equations (10)-(11) to obtain:

[0045]

[0046] In the formula, is the kth distribution network fitting defect data x k,t The signal energy of the sth scale at the tth time step, S is the maximum decomposition scale; log is a logarithmic function; β s and σ s Trainable weights to control the effects of energy and coefficient of variation; yes The variance of .

[0047] In one embodiment of the present invention, the kth distribution network fitting defect data x k,t The distribution network hardware defect aggregation feature M at the tth time step k,t , using formula (12) to obtain:

[0048]

[0049] In the formula, Concat is the concatenation function; Attention represents the dot product scaling attention mechanism; Q s ,K s ,V s The query value, key value and true value in the attention mechanism are obtained by linear transformation; S is the maximum decomposition scale.

[0050] In one embodiment of the present invention, the spectrum feature H is obtained. l include:

[0051] According to the distribution network hardware defect data x k,t , obtain the spectrum amplitude S of the distribution network hardware defect k,f ;

[0052] According to the frequency spectrum amplitude S of the distribution network hardware defects k,f , construct the normalized adjacency matrix of the defect spectrum of distribution network hardware

[0053] According to the normalized adjacency matrix Get the spectrum feature H l .

[0054] In one embodiment of the present invention, for the kth data x at the tth sampling time k,t The defects of the distribution network fittings are identified and classified, and the following equations (19) and (20) are used to obtain:

[0055] Fcom =Concat(H L ,R k,t )#(19)

[0056]

[0057] In the formula, H L represents the feature matrix of the L-th layer node; F com It is the defect fusion feature of distribution network hardware; W d is the weight matrix of the defect feature detection layer of the distribution network hardware; b d is the bias term of the predicted distribution network fitting defect type label; Concat is the concatenation function; Softmax is the activation function; It is the label of the predicted defect type of the distribution network fittings.

[0058] The present invention also provides a system according to the above-mentioned distribution network hardware defect detection method, comprising:

[0059] Execution module: used to use the real-time vibration signal of distribution network hardware as the input of the optimal multi-scale time-frequency domain neural network, and the optimal multi-scale time-frequency domain neural network outputs the defect type of distribution network hardware;

[0060] Building module: used to build the best multi-scale time-frequency domain neural network, including the following steps:

[0061] Construct a training set of distribution network hardware vibration signals;

[0062] Establish an initial multi-scale time-frequency domain neural network, including an adaptive filter bank decomposition layer, a multi-time-frequency scale encoder layer, a frequency domain graph convolution layer, and a hardware defect feature detection layer;

[0063] The kth distribution network fitting defect x at the tth sampling time in the training set k,t Input into the adaptive filter bank decomposition layer to obtain the kth distribution network hardware defect data x k,t The intrinsic information F of the distribution network fittings defect at the tth time step k,t ;

[0064] The intrinsic information of the defects of the distribution network hardware F k,t Input to the multi-time frequency scale Encoder layer to obtain the modulation feature T of the distribution network hardware defect k,t ;

[0065] The kth data x at the tth sampling time in the training set k,t Input the mid-frequency domain graph convolution layer to obtain the spectrum feature H l ;

[0066] The modulation feature R of the distribution network hardware defect k,t And the spectrum characteristics Hl Input into the hardware defect feature detection layer, for the kth data x at the tth sampling time k,t Identify and classify defects in distribution network fittings;

[0067] Based on the distribution network hardware vibration signal training set combined with the optimization algorithm, the initial multi-scale time-frequency domain neural network is optimized to obtain the optimal multi-scale time-frequency domain neural network.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] The present invention aims to improve the processing capability of vibration signals of hardware defects in different environments and types by adaptively adjusting the filter group and the multi-time-frequency scale encoder layer by learning the characteristics of the vibration signals of the hardware in the distribution network; a multi-head wavelet attention mechanism layer is proposed to enhance the multi-scale time-frequency feature modeling capability. In addition, a frequency domain graph convolution layer is proposed to realize the construction of hardware defect features from the spectrum graph by calculating the graph convolution of the frequency domain features of the hardware vibration signal. Dynamic gating is proposed to further screen the complex hardware vibration signal features to obtain accurate hardware defect features. The efficiency and accuracy of hardware defect detection are significantly improved through the above-mentioned precise positioning and construction of the time-frequency features.

[0070] Adaptive filter banks and multi-time-frequency scale encoder layers are used to automatically adjust the filtering and encoding process according to the characteristics of the vibration signal, which improves the ability to handle various field environments and different types of hardware defects. In addition, the frequency domain graph convolution layer further enhances the extraction of frequency domain features and optimizes the identification of defect features in the spectrum graph. The dynamic gating and multi-head wavelet attention mechanism layer improves the detection speed and accuracy of the system in complex signal environments, and significantly improves the efficiency and accuracy of defect detection by accurately locating and emphasizing key time-frequency features.

[0071] It is able to dynamically capture the characteristics of hardware vibration signals at different time-frequency scales, which enables more accurate analysis and identification of defects in nonlinear and non-stationary vibration signals. This multi-scale analysis provides richer information than traditional single-scale analysis, thereby improving the accuracy and reliability of detection.

[0072] The present invention can dynamically adjust the wavelet basis according to the specific characteristics of the signal, so that the signal decomposition is more adapted to the current signal state. This method optimizes the local representation of the signal by adjusting the wavelet parameters, providing higher flexibility and adaptability. By dynamically adjusting the combination of wavelet transform and attention mechanism, more attention is paid to the adaptive processing of the signal and the extraction of key features, which is suitable for scenes with large dynamic changes and the need to emphasize specific signal features.

[0073] The present invention focuses on the frequency domain characteristics, converts the signal to the frequency domain through fast Fourier transform, and processes the spectrum data. This method can intuitively capture the periodicity and frequency distribution characteristics of the signal, and is particularly suitable for analyzing data with strong periodicity such as vibration, sound or electrical signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 The present invention is a schematic flow chart of a method for detecting defects in distribution network fittings according to an embodiment of the present invention.

[0075] Figure 2 The present invention is a block diagram of a distribution network fittings defect detection system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to facilitate those skilled in the art to understand the technical solution of the present invention, the technical solution of the present invention is further described in conjunction with the accompanying drawings of the specification.

[0077] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0078] See also Figure 1 As shown, the present invention provides a method for detecting defects in distribution network fittings, including using real-time distribution network fitting vibration signals as input of an optimal multi-scale time-frequency domain neural network, and the optimal multi-scale time-frequency domain neural network outputs the type of defects in the distribution network fittings.

[0079] In one embodiment of the present invention, the optimal multi-scale time-frequency domain neural network is obtained by the following steps:

[0080] Construct a training set of distribution network hardware vibration signals.

[0081] An initial multi-scale time-frequency domain neural network is established, including an adaptive filter bank decomposition layer, a multi-time-frequency scale Encoder layer, a frequency domain graph convolution layer, and a hardware defect feature detection layer.

[0082] The kth distribution network fitting defect x at the tth sampling time in the training set k,t Input into the adaptive filter bank decomposition layer to obtain the kth distribution network hardware defect data x k,t The intrinsic information F of the distribution network fittings defect at the tth time step k,t .

[0083] The intrinsic information of the defects of the distribution network hardware F k,t Input to the multi-time-frequency scale Encoder layer to obtain the modulation feature R of the distribution network hardware defectk,t .

[0084] The kth data x at the tth sampling time in the training set k,t Input the mid-frequency domain graph convolution layer to obtain the spectrum feature H l .

[0085] The modulation feature R of the distribution network hardware defect k,t And the spectrum characteristics H l Input into the hardware defect feature detection layer, for the kth data x at the tth sampling time k,t Identify and classify defects in distribution network fittings.

[0086] Based on the distribution network hardware vibration signal training set and combined with the optimization algorithm, the initial multi-scale time-frequency domain neural network is optimized to obtain the optimal multi-scale time-frequency domain neural network.

[0087] In this embodiment, constructing a distribution network hardware vibration signal training set includes:

[0088] The vibration signal data of different distribution network hardware are collected, and the vibration signals of the distribution network hardware collected by the installed distribution network hardware vibration signal sensor are normalized, and a distribution network hardware defect classification set is constructed, denoted as X, and X = {X 1 …X k …X K}, X k represents the vibration signal of the kth distribution network hardware defect data, and X k ={x k,1 …x k,t …x k,T}, x k,t It represents the vibration amplitude of the kth hardware defect data at the tth sampling time; 1≤k≤K, K represents the total number of hardware defects; 1≤t≤T, T represents the total sampling time.

[0089] Construct a set of distribution network hardware defect categories Y, and set labels for the vibration signals corresponding to the distribution network hardware defect types according to prior knowledge, denoted as Y = {y 1 ,…,y k ,…,y K}, where y k represents the label value of the kth distribution network hardware defect data, and y k Belongs to [1,N], where N is the number of types of defects in distribution network fittings.

[0090] The labeled distribution network fittings defect dataset P = (X, Y) is randomly shuffled and used as the training set.

[0091] In one embodiment of the present invention, in the adaptive filter bank decomposition layer:

[0092] The kth distribution network fitting defect data x k,t The t-th sampling time data x k,t Input the adaptive filter bank decomposition layer, which is used to perform frequency decomposition on the vibration signal data of the distribution network hardware, dynamically adjust the filter parameters to optimize the signal processing, and provide accurate basic data for subsequent defect feature extraction and analysis. Use equations (1)-(5) to perform adaptive filtering processing to obtain the kth hardware defect data x k,t The intrinsic information F of the hardware defect at the tth time step k,t :

[0093] P k,t =|x k,t | 2 #(1)

[0094]

[0095] α k,t =Sigmoid(W p ·P k,t +W ω ·ω k,t +b α )#(4)

[0096]

[0097] Where ω is the frequency, which usually ranges from 0 to 2π; M is the length of the window function, m belongs to [-M,M]; α k,t is the adjustment function of the defect data of the kth distribution network hardware at the tth moment; P k,t is the kth distribution network fitting defect data x k,t The instantaneous energy at the tth sampling moment; Hilbert represents the Hilbert transform; represents the result of Hilbert transform; ∠ represents the phase angle; ω k,t is the instantaneous frequency; W p YesP k,t The weight matrix W ε Yes k,t The weight matrix of α is the bias; Sigmoid is a function, e is a natural constant, and j is an imaginary unit.

[0098] In one embodiment of the present invention, the multi-time-frequency scale Encoder layer includes a multi-scale wavelet attention mechanism layer and a residual connection feedforward neural network layer connected in sequence.

[0099] In this embodiment, the multi-scale wavelet attention mechanism layer effectively extracts and strengthens the multi-scale time-frequency features of the hardware defect data, enhances the model's ability to identify key defect features, and uses equations (6)-(9) to obtain the kth hardware defect data xk,t The fitting defect coding feature of the s-th scale at the t-th time step

[0100] γ′ s =γ s ·(1 + η s ·sigmoid(W γ ·F k,t +b γ ))#(6)

[0101] λ′ s =λ s ·(1 + ξ s ·sigmoid(W λ ·F k,t +b λ ))#(7)

[0102]

[0103] In the formula, u is the wavelet time factor, 1 < u ≤ U; s is the wavelet scale factor; x k,t+u is the distribution fitting defect data of the k-th at the (t + u)-th moment; ψ′ s (t, u; γ′ s , λ′ s ) is the dynamically adjusted wavelet basis function; ω s is the corresponding scale angular frequency; γ′ s is the amplitude adjustment parameter; λ′ s is the scale adjustment parameter; γ s is the initial amplitude parameter; λ s is the initial scale adjustment parameter; W γ and b γ are the amplitude weight and amplitude bias; η s is the amplitude adjustment factor; ξ s is the scale adjustment factor, Sigmoid is the function, e is the natural constant, j is the imaginary unit, and ψ is the wavelet basis function.

[0104] The multi-head wavelet attention mechanism layer uses equations (10)-(11) to obtain the k-th fitting defect data x k,t The dynamic gating G at the t-th time step k,t :

[0105]

[0106] In the formula, is the signal energy of the k-th distribution fitting defect data x k,t at the s-th scale at the t-th time step, S is the maximum decomposition scale; log is the logarithmic function; β s and σ sTrainable weights to control the effects of energy and coefficient of variation; yes In this embodiment, the multi-head wavelet attention mechanism layer can be understood as multiple wavelet attention mechanism layers, each corresponding to a wavelet scale.

[0107] In this embodiment, the multi-head wavelet attention mechanism layer uses formula (12) to obtain the k-th distribution network hardware defect data x k,t The distribution network hardware defect aggregation feature M at the tth time step k,t :

[0108]

[0109] In the formula, Concat is the concatenation function; Attention represents the dot product scaling attention mechanism; Q s ,K s ,V s are query value, key value and true value respectively, which are obtained by linear transformation; S is the maximum decomposition scale.

[0110] In this embodiment, the residual connection feedforward neural network layer helps the model maintain stable training when the depth increases, improves the efficiency and accuracy of feature processing, and uses formula (13) to obtain the kth distribution network hardware defect data x k,t The modulation characteristic R of the defect of the distribution network hardware at the tth time step k,t :

[0111] R k,t =M k,t +W 2 ReLU(W 1 ·M k,t +b 1 )+b 2 #(13)

[0112] Where W 1 is the first layer modulation weight; b 1 is the first layer modulation bias; W 2 is the second layer modulation weight; b 2 is the second layer modulation bias; ReLU is a nonlinear activation function.

[0113] See also Figure 1 As shown, in one embodiment of the present invention, the frequency domain graph convolution layer includes: hardware defect spectrum calculation, spectrum graph construction, and spectrum graph convolution.

[0114] In this embodiment, the hardware defect spectrum calculation reveals the frequency components and characteristics of the signal, providing key frequency domain information for subsequent defect analysis. Specifically, the hardware defect spectrum amplitude S is obtained using formula (14): k,f :

[0115] S k,f =FFT(x k,t )#(14)

[0116] In formula (14), FFT stands for fast Fourier transform.

[0117] In this embodiment, the spectrum graph construction converts the hardware defect spectrum data into a graph structure and captures and learns the nonlinearity and nonlocal dependencies in the hardware defect data. Specifically, the normalized adjacency matrix of the hardware defect spectrum is constructed using equations (15)-(17):

[0118]

[0119] In the formula, ∥·∥ 2 represents the matrix 2-norm; σ represents the mean; D ii is the degree matrix; S k,i and S k,j They represent the spectrum amplitudes of the kth distribution network fitting defect data at frequencies i and j respectively, and exp is an exponential function.

[0120] In this embodiment, the spectrum graph convolution aggregates and extracts local and global information in the spectrum data and comprehensively considers the structural characteristics and frequency associations in the spectrum graph. Specifically, the spectrum feature H is extracted using formula (18): l :

[0121]

[0122] In the formula, H l is the node feature matrix of the lth layer, that is, the spectrum feature, initially S k,f ; W l and b l are the graph convolution weight and bias of the lth layer respectively; l is the number of graph convolution layers, 1 <l≤L。

[0123] In one embodiment of the present invention, the distribution network fitting defect feature detection layer integrates the features extracted by the previous network layer, and uses a classifier to accurately identify and classify fitting defects, thereby enhancing the diagnostic efficiency and accuracy of the model. The fitting defect type label of the kth fitting defect data is obtained using equations (19)-(20):

[0124] F com =Concat(H L ,R k,t )#(19)

[0125]

[0126] In the formula, H Lrepresents the feature matrix of the L-th layer node; F com It is the defect fusion feature of distribution network hardware; W d is the weight matrix of the defect feature detection layer of the distribution network hardware; b d is the bias term of the predicted distribution network fitting defect type label; Concat is the concatenation function; Softmax is the activation function; It is the label of the predicted defect type of the distribution network fittings.

[0127] In one embodiment of the present invention, based on the distribution network fittings vibration signal training set, the initial multi-scale time-frequency domain neural network is trained using back propagation or gradient descent method. When the training rounds reach the maximum training rounds, or the loss function reaches the minimum, the training is stopped to obtain the trained optimal multi-scale time-frequency domain neural network, which is used to map the input distribution network fittings defect data set to the corresponding distribution network fittings defect category label.

[0128] See also Figure 2 As shown, the present invention also provides a system for a method for detecting defects in distribution network hardware, comprising:

[0129] Execution module: used to use the real-time vibration signal of distribution network hardware as the input of the optimal multi-scale time-frequency domain neural network, and the optimal multi-scale time-frequency domain neural network outputs the defect type of distribution network hardware.

[0130] Building blocks: Used to obtain optimal multi-scale time-frequency domain neural networks.

[0131] Among them, obtaining the optimal multi-scale time-frequency domain neural network includes the following steps:

[0132] Construct a training set of distribution network hardware vibration signals;

[0133] Establish an initial multi-scale time-frequency domain neural network, including an adaptive filter bank decomposition layer, a multi-time-frequency scale encoder layer, a frequency domain graph convolution layer, and a hardware defect feature detection layer;

[0134] The kth distribution network fitting defect x at the tth sampling time in the training set k,t Input into the adaptive filter bank decomposition layer to obtain the kth distribution network hardware defect data x k,t The intrinsic information F of the distribution network fittings defect at the tth time step k,t ;

[0135] The intrinsic information of the defects of the distribution network hardware F k,t Input to the multi-time frequency scale Encoder layer to obtain the modulation feature T of the distribution network hardware defect k,t ;

[0136] The kth data x at the tth sampling time in the training set k,tInput the mid-frequency domain graph convolution layer to obtain the spectrum feature H l ;

[0137] The modulation feature R of the distribution network hardware defect k,t And the spectrum characteristics H l Input into the hardware defect feature detection layer, for the kth data x at the tth sampling time k,t Identify and classify defects in distribution network fittings;

[0138] Based on the distribution network hardware vibration signal training set combined with the optimization algorithm, the initial multi-scale time-frequency domain neural network is optimized to obtain the optimal multi-scale time-frequency domain neural network.

[0139] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0140] The above-described embodiments merely represent implementation methods of the invention. The protection scope of the present invention is not limited to the above-described embodiments. For those skilled in the art, several modifications and improvements may be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for detecting defects in distribution network hardware, characterized in that: include: The real-time vibration signal of the distribution network hardware is used as the input of the optimal multi-scale time-frequency domain neural network, and the optimal multi-scale time-frequency domain neural network outputs the defect type of the distribution network hardware; Among them, the optimal multi-scale time-frequency domain neural network is obtained through the following steps: Construct a training set of distribution network hardware vibration signals; Establish an initial multi-scale time-frequency domain neural network, including an adaptive filter bank decomposition layer, a multi-time-frequency scale encoder layer, a frequency domain graph convolution layer, and a hardware defect feature detection layer; The kth distribution network fitting defect x at the tth sampling time in the training set k，t Input into the adaptive filter bank decomposition layer to obtain the kth distribution network hardware defect data x k,t The intrinsic information F of the distribution network fittings defect at the tth time step k，t ; The intrinsic information of the defects of the distribution network hardware F k，t Input to the multi-time-frequency scale Encoder layer to obtain the modulation feature R of the distribution network hardware defect k ,t; The kth data x at the tth sampling time in the training set k，t Input the mid-frequency domain graph convolution layer to obtain the spectrum feature H l ; The modulation feature R of the distribution network hardware defect k,t And the spectrum characteristics H l Input into the hardware defect feature detection layer, for the kth data x at the tth sampling time k，t Identify and classify defects in distribution network fittings; Based on the distribution network hardware vibration signal training set and combined with the optimization algorithm, the initial multi-scale time-frequency domain neural network is optimized to obtain the optimal multi-scale time-frequency domain neural network.

2. The method for detecting defects in distribution network fittings according to claim 1, characterized in that: Construct a distribution network hardware vibration signal training set and obtain it through the following steps: Collect vibration signal data of different distribution network hardware and construct a hardware defect classification set, denoted as X, where X = {X1…X k …X K }, X k represents the vibration signal of the kth hardware defect data, and X k ={x k，1 …x k，t ...x k，T }, x k,t Indicates the vibration amplitude of the kth hardware defect data at the tth sampling time; 1≤k≤K, K represents the total number of hardware defects; 1≤t≤T, T represents the total sampling time; Construct the defect category set Y of distribution network fittings, denoted as Y = {y1, ..., y k , ..., y K }, where y k represents the label value of the kth distribution network hardware defect data, and y k Belongs to [1, N], where N is the number of types of hardware defects; The labeled distribution network hardware defect data set P = (X, Y) is randomly shuffled and used as the distribution network hardware vibration signal training set.

3. The method for detecting defects in distribution network fittings according to claim 1, characterized in that: The adaptive filter bank decomposition layer uses (1)-(5) to perform adaptive filtering processing to obtain the kth distribution network hardware defect data x k , tThe intrinsic information F of the distribution network hardware defect at the tth time step k，t : P k，t =|x k,t | 2 #(1) α k，t =Sigmoid(W p ·P k，t +W ω ·ω k，t +b α )#(4) Where ω is the frequency, which usually ranges from 0 to 2π; M is the length of the window function, m belongs to [-M,M]; α k，t is the adjustment function of the defect data of the kth distribution network hardware at the tth moment; P k，t is the kth distribution network hardware defect data x k,t The instantaneous energy at the tth sampling moment; Hilbert represents the Hilbert transform; represents the result of Hilbert transform; ∠ represents the phase angle; ω k，t is the instantaneous frequency; W p YesP k，t The weight matrix W ω Yes k，t The weight matrix of α is the bias; Sigmoid is a function, e is a natural constant, and j is an imaginary unit.

4. The method for detecting defects in distribution network fittings according to claim 1, characterized in that: Distribution network hardware defect modulation characteristics R k , t is obtained by: The multi-time-frequency scale Encoder layer includes a multi-scale wavelet attention mechanism layer and a residual connection feedforward neural network layer connected in sequence; In the multi-scale wavelet attention mechanism layer: Extract intrinsic information of distribution network hardware defects k，t The multi-scale time-frequency characteristics of the k-th hardware defect data x are obtained k，t Distribution network fitting defect coding features at the s-th scale at the t-th time step According to the coding characteristics of hardware defects of the sth scale Get the kth distribution network hardware defect data x k，t The dynamic gating G at the tth time step k , t; Combine all scales of hardware defect coding features to obtain the kth hardware defect data x k，t The distribution network hardware defect aggregation feature M at the tth time step k , t; Residual connection feed-forward neural network layer: The distribution network hardware defect aggregation feature M k , t is input into the residual connection feedforward neural network layer to obtain the kth hardware defect data x k，t The modulation characteristic R of the defect of the distribution network hardware at the tth time step k，t .

5. The method for detecting defects in distribution network fittings according to claim 4, characterized in that: The kth hardware defect data x k，t The coding features of the hardware defects at the s-th scale at the t-th time step Using equations (6)-(9) we can obtain: c′ s =c s ·(1+n s ·sigmoid(W γ ·F k，t +b γ ))#(6) λ′ s =λ s ·(1+ξ s ·sigmoid(W λ ·F k，t +b λ ))#(7) Where u is the wavelet time factor, 1<u≤U; s is the wavelet scale factor; x k，t+u is the defect data of the kth distribution fitting at the time t+u; ψ′ s (t,u;γ′ s ,λ′ s ) is the dynamically adjusted wavelet basis function; ω s Corresponding scale angular frequency; γ′ s is the amplitude adjustment parameter; λ′ s is the scale adjustment parameter; γ s is the initial amplitude parameter; λ s is the initial scale adjustment parameter; W γ and b γ are the amplitude weight and amplitude bias; η s is the amplitude adjustment factor; ξ s is the scale adjustment factor, Sigmoid is the function, e is the natural constant, j is the imaginary unit, and ψ is the wavelet basis function.

6. The method for detecting defects in distribution network fittings according to claim 4, characterized in that: The kth distribution network hardware defect data x k，t The dynamic gating G at the tth time step k，t , using equations (10)-(11) to obtain: In the formula, is the kth distribution network hardware defect data x k，t The signal energy of the sth scale at the tth time step, S is the maximum decomposition scale; log is a logarithmic function; β s and σ s Trainable weights to control the effects of energy and coefficient of variation; yes The variance of .

7. The method for detecting defects in distribution network fittings according to claim 4, characterized in that: The kth distribution network fitting defect data x k，t The distribution network hardware defect aggregation feature M at the tth time step k，t , using formula (12) to obtain: In the formula, Concat is the concatenation function; Attention represents the dot product scaling attention mechanism; Q s , K s , V s The query value, key value and true value in the attention mechanism are obtained by linear transformation; S is the maximum decomposition scale.

8. The method for detecting defects in distribution network fittings according to claim 1, characterized in that: Get the spectrum feature H l include: According to the distribution network hardware defect data x k，t , obtain the spectrum amplitude S of the distribution network hardware defect k，f ; According to the frequency spectrum amplitude S of the distribution network hardware defects k，f , construct the normalized adjacency matrix of the defect spectrum of distribution network hardware According to the normalized adjacency matrix Get the spectrum feature H l .

9. The method for detecting defects in distribution network fittings according to claim 1, characterized in that: For the kth data x at the tth sampling time k , the defect identification and classification of the distribution network fittings of t are carried out, and the following equations (19)-(20) are used to obtain: F com =Concat(H L ,R k,t )#(19) In the formula, H L Represents the L-th layer node feature matrix; F com It is the defect fusion feature of distribution network hardware; W d is the weight matrix of the defect feature detection layer of the distribution network hardware; b d is the bias term of the predicted distribution network fitting defect type label; Concat is the concatenation function; Softmax is the activation function; It is the label of the predicted defect type of the distribution network fittings.

10. A system according to any one of claims 1 to 9, characterized in that: include: Execution module: used to use the real-time vibration signal of distribution network hardware as the input of the optimal multi-scale time-frequency domain neural network, and the optimal multi-scale time-frequency domain neural network outputs the defect type of distribution network hardware; Building module: used to build the best multi-scale time-frequency domain neural network, including the following steps: Construct a training set of distribution network hardware vibration signals; Establish an initial multi-scale time-frequency domain neural network, including an adaptive filter bank decomposition layer, a multi-time-frequency scale encoder layer, a frequency domain graph convolution layer, and a hardware defect feature detection layer; The kth distribution network fitting defect x at the tth sampling time in the training set k , t is input into the adaptive filter bank decomposition layer to obtain the kth distribution network hardware defect data x k,t The intrinsic information F of the distribution network fittings defect at the tth time step k，t ; The intrinsic information of the defects of the distribution network hardware F k，t Input to the multi-time-frequency scale Encoder layer to obtain the modulation feature R of the distribution network hardware defect k，t ; The kth data x at the tth sampling time in the training set k，t Input the mid-frequency domain graph convolution layer to obtain the spectrum feature H l ; The modulation feature R of the distribution network hardware defect k，t And the spectrum characteristics H l Input into the hardware defect feature detection layer, for the kth data x at the tth sampling time k，t Identify and classify defects in distribution network fittings; Based on the distribution network hardware vibration signal training set combined with the optimization algorithm, the initial multi-scale time-frequency domain neural network is optimized to obtain the optimal multi-scale time-frequency domain neural network.

Images