Power equipment fault diagnosis method and system based on cloud-edge-end collaborative architecture

Abstract

The present invention discloses a method and system for diagnosing faults of electric power equipment based on a cloud-edge-end collaborative architecture, the method comprising: obtaining the voiceprint vibration data of the target electric power equipment; extracting feature vectors from the voiceprint vibration data using a neural network that incorporates the voiceprint vibration propagation mechanism to obtain multiple feature vectors, and performing weight adaptive calculation and weighted fusion on the multiple feature vectors to obtain a mixed feature vector; using a deep residual network to perform fault identification and classification on the mixed feature vector; and outputting the results of fault identification and classification. The present invention can improve the accuracy and stability of fault diagnosis of electric power equipment while ensuring real-time performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid fault diagnosis, and specifically relates to a method and system for power equipment fault diagnosis based on a cloud-edge-end collaborative architecture. Background Art

[0002] With the continuous advancement of smart grid construction, the power system is gradually developing in the direction of intelligence and automation, and the safe operation of power equipment provides a basic guarantee for the development of the power system. Among many power equipment, transformers, switches and other power equipment are key hubs for power conversion and transmission, and their operating status is of great significance to the safe and reliable operation of the power grid. Affected by many factors such as the environment and material aging, power equipment will inevitably have defects and failures in long-term operation, which can easily lead to power outages in local areas of the power grid and cause serious economic losses. Therefore, how to effectively evaluate the operating status of the transformer and quickly and accurately diagnose faults is important for improving the reliability of transformer operation.

[0003] Traditional power equipment status assessment usually relies on experienced engineers to judge the sound of power equipment. This manual judgment method is usually inefficient and limited by the experience level of the staff, and the accuracy is not high. In addition, some scholars use machine learning models to identify power equipment faults. These methods require the extraction and screening of appropriate feature vectors, which is time-consuming and labor-intensive, and how to select appropriate feature vectors is also a huge challenge. In recent years, deep learning has provided a new solution for power equipment fault diagnosis, but the model usually does not consider the physical mechanism and the diagnostic reliability is not high. Therefore, how to improve the reliability of the model in fault diagnosis has become a key issue that needs to be solved. Summary of the invention

[0004] In view of the above-mentioned shortcomings, the object of the present invention is to provide a method and system for power equipment fault diagnosis based on a cloud-edge-end collaborative architecture. The present invention can improve the accuracy and stability of power equipment fault diagnosis while ensuring real-time performance.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for diagnosing power equipment faults based on a cloud-edge-end collaborative architecture, comprising:

[0007] Acquire the soundprint vibration data of the target power equipment;

[0008] A neural network incorporating the voiceprint vibration propagation mechanism is used to extract feature vectors from the voiceprint vibration data to obtain a plurality of feature vectors, and weight adaptive calculation and weighted fusion are performed on the plurality of feature vectors to obtain a mixed feature vector;

[0009] Using a deep residual network to perform fault identification and classification on the mixed feature vector;

[0010] Output the results of fault identification and classification.

[0011] As a further improvement of the present invention, before the step of extracting feature vectors from the voiceprint vibration data using a neural network incorporating the voiceprint vibration propagation mechanism to obtain a plurality of feature vectors, the step further includes:

[0012] The voiceprint vibration data is denoised by a neural network based on the SE attention mechanism.

[0013] As a further improvement of the present invention, the neural network based on the SE attention mechanism denoises the voiceprint vibration data, including:

[0014] The voiceprint vibration data is input as the time domain signal x into the hidden layer h1 of the neural network to obtain the preliminary feature vector F calculated by the network layer. x1 :

[0015] The SE attention mechanism is introduced to the preliminary feature vector F x1 , the attention coefficient M is obtained through the SENet network:

[0016] The attention coefficient M and the preliminary feature vector F x1 Multiply to get the noise removal base value F t :

[0017] The preliminary feature vector F x1 With the noise removal base value F t The denoised signal F is obtained by subtraction dn ;

[0018] The denoised signal F dn As preprocessed voiceprint vibration data.

[0019] As a further improvement of the present invention, the method of extracting feature vectors from the voiceprint vibration data using a neural network incorporating the voiceprint vibration propagation mechanism includes:

[0020] Construct the soundprint vibration equation in different media:

[0021]

[0022]

[0023] In the formula: u is the displacement of the sound wave vibration, t is the time, c is the propagation speed of the sound wave vibration, and s is the square root of the position variables x, y, and z, that is, x, y, z are position variables, γ is the specific heat ratio of the gas; B and A are nonlinear parameters of the liquid; let is the nonlinear coefficient in the fluid;

[0024] According to the propagation characteristics of the voiceprint vibration signal, the boundary conditions are constructed:

[0025] u(x,y1,z1,t)=a1

[0026] u(x1, y, z1, t) = a2

[0027] u(x1, y, z1, t) = a3

[0028] Where: a1, a2, a3 are the calculated boundary condition values ​​of the displacement u of the acoustic vibration under different variables, x1, y1, z1 are the boundary values ​​of the x, y, z position variables respectively;

[0029] According to the voiceprint vibration equation and the boundary conditions, a loss function is constructed:

[0030]

[0031]

[0032]

[0033]

[0034] Among them, L1, L2, L3, and L4 are the four constructed loss functions, N1 and N2 are the number of values ​​randomly sampled within the variable value range, and x i ,y i 、z i are the values ​​of x, y, and z position variables respectively, t i is a value of the time variable, is the expression of the displacement u of the acoustic wave vibration when the independent variable takes a specific value, are the second-order partial derivatives of the displacement u of the acoustic vibration with respect to the x, y, and z position variables, u s is the partial derivative of the displacement u of the acoustic vibration with respect to the variable s;

[0035] According to the loss functions L1, L2, L3, and L4, different weights are assigned to them and summed up to obtain the total loss function Loss:

[0036] Loss=ω1L1+ω2L2+ω3L3+ω4L4+Loss origin

[0037] Among them, Loss originThe loss function for calculating the deviation between the output layer of the traditional neural network and the sample label, ω1, ω2, ω3, and ω4 are different weight coefficients respectively;

[0038] Based on the total loss function, construct a feature vector self-extraction neural network embedded with the voiceprint vibration mechanism;

[0039] By training the neural network, the data calculated by multiple hidden layers in the network layer are output to obtain feature vector information, and the feature vector information is used as multiple feature vectors of the voiceprint vibration data automatically extracted by the neural network.

[0040] As a further improvement of the present invention, the step of performing weight adaptive calculation and weighted fusion on the multiple feature vectors to obtain a mixed feature vector includes: performing constant scaling processing and adaptive fusion processing on the multiple feature vectors to obtain a mixed feature vector;

[0041] The identical scaling process includes: aligning the dimensions of the multiple feature vectors through upsampling and downsampling operations to make the dimensions of the multiple feature vectors consistent; wherein, upsampling is performed on feature vectors whose dimensions are less than a preset threshold, and downsampling is performed on feature vectors whose dimensions are greater than or equal to the preset threshold;

[0042] The adaptive fusion process includes weighted fusion of the multiple feature vectors obtained by the identity scaling process.

[0043] As a further improvement of the present invention, the upsampling of the feature vector whose dimension is smaller than the preset threshold comprises: upsampling the feature vector whose dimension is smaller than the preset threshold by means of interpolation plus convolution;

[0044] The downsampling of the feature vectors whose dimensions are greater than or equal to the preset threshold includes: using a pooling operation to reduce the dimensions of the feature vectors whose dimensions are greater than or equal to the preset threshold, so as to achieve downsampling of the feature vectors whose dimensions are greater than or equal to the preset threshold.

[0045] As a further improvement of the present invention, upsampling the feature vector whose dimension is less than a preset threshold by interpolation plus convolution includes:

[0046] When interpolating between two sampling points of a feature vector whose dimension is less than a preset threshold, obtaining an interpolation point of the two sampling points to obtain a feature vector of an interpolated and expanded dimension;

[0047] Performing convolution calculation on the feature vector of the interpolation-expanded dimension to obtain the feature vector after dimension expansion, thereby achieving upsampling of the feature vector whose dimension is less than a preset threshold;

[0048] The reducing the dimension of the feature vectors whose dimension is greater than or equal to the preset threshold by using a pooling operation includes:

[0049] For feature vectors whose dimensions are greater than or equal to the preset threshold, a random pooling operation is used to process them to obtain feature vectors with reduced dimensions.

[0050] As a further improvement of the present invention, the weighted fusion of the multiple feature vectors obtained by the identity scaling process includes:

[0051] When the deep neural network is trained on the multiple feature vectors, a fusion weight of each feature vector in the multiple feature vectors is obtained;

[0052] According to the fusion weight of each feature vector in the multiple feature vectors, the multiple feature vectors obtained by the identity scaling process are fused to obtain a mixed feature vector of the voiceprint vibration data.

[0053] As a further improvement of the present invention, the method of using a deep residual network to perform fault identification and classification on the mixed feature vector includes:

[0054] Performing fault diagnosis on the mixed feature vector through a ResNet network to obtain a fault identification result; performing fault diagnosis on the mixed feature vector through a ResNet network includes:

[0055] The input mixed feature vector x passes through the convolution layer and then undergoes ReLU calculation, and then passes through the convolution layer and ReLU layer again to obtain the output F(x), and the identity transformation of the input mixed feature vector x is added to obtain the final output as the fault recognition result;

[0056] Fault classification is performed on the fault identification result.

[0057] A power equipment fault diagnosis system based on cloud-edge-end collaborative architecture, comprising:

[0058] An acquisition module, used to acquire the voiceprint vibration data of the target power equipment;

[0059] A feature extraction module, for extracting feature vectors from the voiceprint vibration data using a neural network that incorporates the voiceprint vibration propagation mechanism to obtain a plurality of feature vectors, and performing weight adaptive calculation and weighted fusion on the plurality of feature vectors to obtain a mixed feature vector;

[0060] An identification and classification module, used for performing fault identification and classification on the mixed feature vector using a deep residual network;

[0061] The output module is used to output the results of fault identification and classification.

[0062] An electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture are implemented.

[0063] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for diagnosing faults of electric power equipment based on a cloud-edge-end collaborative architecture.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] The present invention aims at fault diagnosis of power equipment. It is based on adaptive fusion and fault diagnosis of acoustic-vibration feature vectors. The soundprint vibration data of the power equipment is first obtained, and a neural network with SE attention mechanism is used to reduce noise on the soundprint vibration data. A neural network that incorporates the soundprint vibration propagation mechanism is used to extract feature vectors. After that, the dimensions of multiple feature vectors are aligned, and adaptive weight calculation and weighted fusion of multiple feature vectors are performed. Then, a deep residual network is used in the cloud to identify and classify the fused feature vectors for faults. This cloud-edge-end collaborative diagnosis method can maximize the reliability of diagnosis while ensuring the accuracy of diagnosis. A deep residual network is used to identify and classify the fused feature vectors for faults. It can improve the accuracy and stability of fault diagnosis of power equipment while ensuring real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of a method for diagnosing faults of electric power equipment based on a cloud-edge-end collaborative architecture according to the present invention;

[0067] Figure 2 This is the signal denoising flow chart based on SE attention mechanism;

[0068] Figure 3 A flowchart for automatically extracting feature vectors through a deep neural network that incorporates the propagation mechanism of acoustic vibrations;

[0069] Figure 4 A neural network for self-extraction of feature vectors embedded in the vibration mechanism of voiceprints;

[0070] Figure 5 A flow chart for obtaining fused feature vectors based on an adaptive fusion method;

[0071] Figure 6 The flowchart of the network results for calculating the fusion weights of multiple feature vectors;

[0072] Figure 7 It is a schematic diagram of the fused feature vector;

[0073] Figure 8 A flowchart for diagnosing fault types using ResNet in the cloud;

[0074] Fig. 9 Schematic diagram of conventional neural network versus ResNet network;

[0075] Fig.10 This is a comparison chart of five categories output by the ResNet model;

[0076] Fig.11 A comparison chart before and after using the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture of the present invention;

[0077] Fig.12 The present invention is a power equipment fault diagnosis system based on a cloud-edge-end collaborative architecture that takes into account node network loss;

[0078] Fig.13 The figure is a schematic diagram of an electronic device according to the present invention. DETAILED DESCRIPTION

[0079] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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 should fall within the scope of protection of the present invention.

[0080] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0081] The present invention relates to an adaptive fusion and fault diagnosis method of acoustic and vibration feature vectors of electric power equipment based on edge-cloud collaboration, and in particular to a fault diagnosis technology for electric power equipment based on a cloud-edge-end collaborative architecture, specifically to an online monitoring and fault diagnosis technology for transformers.

[0082] like Figure 1As shown, the first object of the present invention is to provide a method for diagnosing faults of power equipment based on a cloud-edge-end collaborative architecture, comprising:

[0083] Acquire the soundprint vibration data of the target power equipment;

[0084] A neural network incorporating the voiceprint vibration propagation mechanism is used to extract feature vectors from the voiceprint vibration data to obtain a plurality of feature vectors, and weight adaptive calculation and weighted fusion are performed on the plurality of feature vectors to obtain a mixed feature vector;

[0085] Using a deep residual network to perform fault identification and classification on the mixed feature vector;

[0086] Output the results of fault identification and classification.

[0087] This method first uses the edge to collect voiceprint vibration data, and uses a neural network with an SE attention mechanism to reduce the noise of the voiceprint vibration data. The cloud uses a neural network that incorporates the voiceprint vibration propagation mechanism to extract feature vectors, and then aligns the dimensions of multiple feature vectors, and then performs weight adaptive calculation and weighted fusion of multiple feature vectors. The deep residual network is then used in the cloud to identify and classify the fused feature vectors for fault identification. This cloud-edge-end collaborative diagnosis method can maximize the reliability of diagnosis while ensuring the accuracy of diagnosis.

[0088] This cloud-edge-device collaborative diagnosis method can improve the reliability of diagnosis as much as possible while ensuring the accuracy of diagnosis. It also uses a deep residual network to identify and classify faults in the fused feature vectors. It can improve the accuracy and stability of power equipment fault diagnosis while ensuring real-time performance.

[0089] Furthermore, the present invention uses a neural network that incorporates an attention mechanism to denoise the voiceprint vibration data, thereby improving the efficiency of denoising; uses a neural network that incorporates the acoustic vibration propagation mechanism to automatically extract feature vectors of the acoustic vibration signal, so that the feature vector extraction follows the constraints of physical laws and improves the quality of the feature vector; and performs weighted adaptive calculation and weighted fusion of multiple categories of feature vectors of the transformer voiceprint vibration data to enhance the deep feature vectors of the voiceprint vibration data, thereby providing higher quality data for the fault classification algorithm.

[0090] The following is a detailed description of the present invention using transformer online monitoring and fault diagnosis technology as an example:

[0091] Example

[0092] The present invention relates to an adaptive fusion and fault diagnosis method of acoustic and vibration feature vectors of power equipment based on edge-cloud collaboration, and in particular to online monitoring and fault diagnosis technology for transformers. The present invention can improve the accuracy and stability of fault diagnosis of power equipment such as transformers while ensuring real-time performance.

[0093] Taking the transformer acoustic vibration fault data collected by a 500kV substation as an example, the present invention achieves the above-mentioned purpose through the technical scheme of the following steps:

[0094] 1. First, the neural network based on the SE attention mechanism is used to denoise the voiceprint vibration data. The specific steps are:

[0095] Based on the attention mechanism, the edge preprocesses the voiceprint vibration data. After the edge terminal device collects the time-domain voiceprint vibration data from the power equipment, the neural network based on the SE attention mechanism sets the noise threshold of the transformer voiceprint vibration data, thereby realizing the voiceprint vibration data denoising based on the attention mechanism.

[0096] like Figure 2 The signal denoising flow chart based on SE attention mechanism is shown in Figure 1. The specific process is as follows. The collected time domain signal x is input into the hidden layer h1 of the neural network to obtain the signal F calculated by the network layer. x1 :

[0097] F x1 =f(W h ·x+b h ) (1)

[0098] F x1 is the calculation result of the network layer, f is the calculation function of the network layer, W is the weight coefficient matrix, h is the sequence number of a network layer, and b is the bias value.

[0099] Then, the SE attention mechanism is introduced. For the initial feature vector F x1 , after passing through the SENet network, the attention coefficient M is obtained. The SENet network first contains a global average pool layer, and then passes through the fully connected layer, ReLU layer, fully connected layer, and Sigmoid layer in sequence. After passing through the SENet network, the SE attention coefficient M can be obtained.

[0100] M=SE(F x1 ) (2)

[0101] The attention coefficient M and the time domain signal feature vector F x1 Multiplying together gives the noise removal base value.

[0102] F t =M·F x1 (3)

[0103] Then F x1 With F t The denoised signal F can be obtained by subtraction dn :

[0104] F dn =F x1 -F t (4)

[0105] 2. A neural network incorporating the voiceprint vibration propagation mechanism is used to extract feature vectors from the voiceprint vibration data to obtain multiple feature vectors.

[0106] 3. Use a neural network that incorporates the voiceprint vibration propagation mechanism to extract feature vectors from the voiceprint vibration data to obtain multiple feature vectors, and perform weight adaptive calculation and weighted fusion on the multiple feature vectors to obtain a mixed feature vector. The specific steps are:

[0107] Step 1: Based on the deep neural network that incorporates the propagation mechanism of sound wave vibration, multiple feature vectors of voiceprint vibration data are automatically extracted.

[0108] like Figure 3 As shown, it is a flow chart of automatically extracting feature vectors through a deep neural network that incorporates the sound wave vibration propagation mechanism.

[0109] Get the mixed feature vector as time series data;

[0110] Construct the soundprint vibration equation in different media:

[0111]

[0112] Where: u is the displacement of the sound wave vibration, t is the time, c is the propagation speed of the sound wave vibration, and s is the square root of the position variables x, y, and z, that is, x, y, z are position variables, γ is the specific heat ratio of the gas; B and A are nonlinear parameters of the liquid. is the nonlinear coefficient in the fluid.

[0113] In addition, according to the propagation characteristics of the voiceprint vibration signal, the boundary conditions are constructed:

[0114] u(x,y1,z1,t)=a1 (6)

[0115] u(x1, y, z1, t) = a2 (7)

[0116] u(x1, y, z1, t) = a3 (8)

[0117] According to the voiceprint vibration equation (5) and boundary conditions (6) to (8), the loss function is constructed:

[0118]

[0119]

[0120]

[0121]

[0122] Among them, L1, L2, L3, and L4 are the four constructed loss functions, N1 and N2 are the number of values ​​randomly sampled within the variable value range, and x i ,y i 、z i are the values ​​of x, y, and z position variables respectively, t i is a value of the time variable, is the expression of the displacement u of the acoustic wave vibration when the independent variable takes a specific value, are the second-order partial derivatives of the displacement u of the acoustic vibration with respect to the x, y, and z position variables, u s is the partial derivative of the displacement u of the acoustic vibration with respect to the variable s;

[0123] Then, the total loss function is as follows:

[0124] Loss=ω1L1+ω2L2+ω3L3+ω4L4+Loss origin (13)

[0125] Among them, Loss origin The loss function for calculating the deviation between the output layer of the traditional neural network and the sample label is constructed. ω1, ω2, ω3, and ω4 are different weight coefficients. Based on the above loss function, Figure 4 The feature vector self-extraction neural network embedded in the voiceprint vibration mechanism shown in the figure extracts the feature vector information calculated by multiple hidden layers in the network layer after training the neural network, that is, multiple feature vectors automatically extracted by the neural network are recorded as fe1, fe2...fe n .

[0126] Step 2: For the multiple feature vectors fe1, fe2, ...fe obtained by the neural network in Step 1 n Perform feature vector fusion. Here, the adaptive fusion method is used to adaptively learn the fusion weight of each feature vector, which is divided into two steps: constant scaling and adaptive fusion.

[0127] like Figure 5 Shown is a flow chart for obtaining a fused feature vector based on an adaptive fusion method.

[0128] (1) Constant scaling:

[0129] Through upsampling and downsampling operations, the sizes of different feature vectors are aligned to make the feature vector dimensions consistent, which is convenient for fusion operations. Among them, the feature vectors with smaller dimensions are upsampled, and the feature vectors with larger dimensions are downsampled.

[0130] 1) Upsampling.

[0131] Upsampling is achieved by interpolation plus convolution. As shown in the figure below, when interpolating between (m0, n0) and (m1, n1), the following method is used:

[0132]

[0133] The obtained (m, n) is the interpolation point, and the convolution calculation is performed on the feature vector of the interpolation expansion dimension. The calculation formula is as follows:

[0134]

[0135] In the formula, out(fe i ,C out ) is the convolution solution value, bias(C out ) is the bias, fe i is the feature vector, Cout is the output dimension, C in is the dimension of the input, weight is the weight function, input is the input function, and k is the counting variable.

[0136] After the difference plus convolution operation, the feature vector with expanded dimension can be obtained.

[0137] 2) Downsampling.

[0138] Downsampling uses pooling operation to reduce the dimension. In order to solve the problem that the maximum pooling loses detail information and the average pooling easily blurs the main feature vector, a random pooling operation is used here.

[0139]

[0140] Among them, out is the pooling calculation result, λ is a random value of 0 or 1, indicating the selection of maximum pooling or average pooling, R is the value range, and x p is a eigenvector element, and p is a value in the value range.

[0141] After the pooling operation, the feature vectors fe1, fe2, ..., fen with reduced dimensions can be obtained.

[0142] (2) Adaptive Fusion:

[0143] The multiple feature vectors fe1, fe2, ... fen obtained by constant scaling are fused, and the feature vector adaptive fusion is achieved by weighted fusion of multiple feature vectors. The weights of each feature vector are obtained when the deep neural network is trained on the voiceprint vibration feature vector. Figure 6 As shown, the network results of calculating the fusion weights of multiple feature vectors fe1, fe2...fen are as follows:

[0144] α1, α2...α n fe1, fe2...fe n The fusion weight is α i For example, the calculation method is:

[0145]

[0146] In the formula, j is the count variable, L is the loss function variable, and y j is a variable in L, f ei is the ith eigenvector.

[0147] After obtaining the fusion weights of different feature vectors, the different feature vectors can be fused by the following formula to obtain the mixed feature vector F of the transformer voiceprint vibration data: m ,The mixed feature vector can better reflect the characteristics of the transformer and lay the foundation for the subsequent diagnosis.

[0148] F m =α1·fe1+α2·fe2+......+α n f n (18)

[0149] The fused feature vector is as follows Figure 7 shown.

[0150] Step 3: Figure 8 This is a flowchart for diagnosing fault types using ResNet in the cloud, and obtaining F after fusion of feature vectors m Finally, fault diagnosis is performed in the cloud through the ResNet network. The ResNet network introduces a direct connection from input to output, that is, the convolution fit is the difference between output and input:

[0151] H(x)=F(x)+x (19)

[0152] When the ResNet network determines that there is a fault, the fault type is output and the process ends; when the ResNet network determines that there is no fault, the process also ends.

[0153] Fig. 9Schematic diagram of conventional neural network vs. ResNet network, Fig. 9 The ResNet network on the right has an additional step of long-distance error transmission compared to conventional neural networks, which can alleviate the problem of gradient disappearance and thus greatly increase the scale of the model.

[0154] The processing process of the ResNet basic unit is as follows: the input mixed feature vector x passes through the convolution layer and then performs ReLU calculation, and then passes through the convolution layer and ReLU layer again to obtain the output F(x), and then the identity transformation of the input mixed feature vector x is added to obtain the final output.

[0155] The fault identification results are classified into fault categories. Fig.10 As shown in the figure, 1 to 5 represent classification numbers, corresponding to five states, namely: there are five categories of ResNet model output, namely normal, winding deformation, DC bias magnetization, partial discharge, and other abnormalities.

[0156] After the cloud completes the diagnosis, the results are transmitted back to the edge, where abnormal status management and fault handling are performed.

[0157] like Fig.11 As shown, there is a comparison chart before and after the use of the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture of the present invention, in which features are extracted from the voiceprint vibration signal and the fusion feature process is performed after fusion processing.

[0158] It can be seen that the present invention has the following advantages:

[0159] (1) A neural network that incorporates the voiceprint vibration propagation mechanism is used to extract feature vectors, so that the feature vector extraction meets the constraints of physical laws and improves the quality and effectiveness of the extracted feature vectors;

[0160] (2) By performing weighted adaptive calculation and weighted fusion on the multi-category feature vectors of the voiceprint and vibration data of power equipment, the deep feature vectors of the voiceprint and vibration data are enhanced, providing higher quality data for the state assessment and fault classification algorithms.

[0161] like Fig.12 As shown, the present invention also provides a power equipment fault diagnosis system based on a cloud-edge-end collaborative architecture, comprising:

[0162] An acquisition module, used to acquire the voiceprint vibration data of the target power equipment;

[0163] A feature extraction module, for extracting feature vectors from the voiceprint vibration data using a neural network that incorporates the voiceprint vibration propagation mechanism to obtain a plurality of feature vectors, and performing weight adaptive calculation and weighted fusion on the plurality of feature vectors to obtain a mixed feature vector;

[0164] An identification and classification module, used for performing fault identification and classification on the mixed feature vector using a deep residual network;

[0165] The output module is used to output the results of fault identification and classification.

[0166] The present invention also provides an edge-cloud collaborative online monitoring and fault diagnosis platform, comprising:

[0167] At the edge, the edge terminal device at the edge collects the original voiceprint vibration data in the power equipment and performs preprocessing to obtain the voiceprint vibration data;

[0168] And the cloud, the cloud executes the steps of the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture.

[0169] like Fig.13 As shown, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture are implemented.

[0170] The power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture includes the following steps:

[0171] Acquire the soundprint vibration data of the target power equipment;

[0172] A neural network incorporating the voiceprint vibration propagation mechanism is used to extract feature vectors from the voiceprint vibration data to obtain a plurality of feature vectors, and weight adaptive calculation and weighted fusion are performed on the plurality of feature vectors to obtain a mixed feature vector;

[0173] Using a deep residual network to perform fault identification and classification on the mixed feature vector;

[0174] Output the results of fault identification and classification.

[0175] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture.

[0176] The power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture includes the following steps:

[0177] Acquire the soundprint vibration data of the target power equipment;

[0178] A neural network incorporating the voiceprint vibration propagation mechanism is used to extract feature vectors from the voiceprint vibration data to obtain a plurality of feature vectors, and weight adaptive calculation and weighted fusion are performed on the plurality of feature vectors to obtain a mixed feature vector;

[0179] Using a deep residual network to perform fault identification and classification on the mixed feature vector;

[0180] Output the results of fault identification and classification.

[0181] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0182] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0183] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0184] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for fault diagnosis of electric power equipment based on cloud-edge-end collaborative architecture, characterized in that: include: Acquire the soundprint vibration data of the target power equipment; A neural network incorporating the voiceprint vibration propagation mechanism is used to extract feature vectors from the voiceprint vibration data to obtain a plurality of feature vectors, and weight adaptive calculation and weighted fusion are performed on the plurality of feature vectors to obtain a mixed feature vector; Using a deep residual network to perform fault identification and classification on the mixed feature vector; Output the results of fault identification and classification; The extracting feature vectors of the voiceprint vibration data using a neural network incorporating the voiceprint vibration propagation mechanism includes: Construct the soundprint vibration equation in different media: Where: u is the displacement of the sound wave vibration, t It's time. c is the propagation speed of sound wave vibration, s For position variables x , y , z The square root of , x , y , z is a position variable, γ is the specific heat ratio of the gas; B. A is the nonlinear parameter of the liquid; let , , is the nonlinear coefficient in the fluid; According to the propagation characteristics of the voiceprint vibration signal, the boundary conditions are constructed: u(x,y1,z1,t)=a1 u(x1,y,z1,t)=a2 u(x1,y,z1,t)=a3 Where: a1 , a2 , a3 are the calculated values ​​of the displacement u of the acoustic vibration under different boundary conditions, x1 , y1 , z1 They are x , y , z Boundary values ​​of position variables; According to the voiceprint vibration equation and the boundary conditions, a loss function is constructed: in, L1 , L2 , L3 , L4 They are the four loss functions constructed respectively. N1 , N2 is the number of values ​​randomly sampled within the variable value range, xi , yi , zi They are x , y , z A value of the position variable, ti is a value of the time variable, is the displacement of the sound wave vibration u An expression when the independent variable takes a specific value, , , , are the displacements of the acoustic vibrations u right x , y , z The second-order partial derivative of the position variable, us is the displacement of the sound wave vibration u For variables s The partial derivative of According to the loss function L1 , L2 , L3 , L4 , assign different weights to them and add them up to get the total loss function Loss : in, The loss function for calculating the deviation from the sample label for the traditional neural network output layer, ω1 , ω2 , ω 3 , ω4 They are different weight coefficients respectively; Based on the total loss function, construct a feature vector self-extraction neural network embedded with the voiceprint vibration mechanism; By training the neural network, the data calculated by multiple hidden layers in the network layer are output to obtain feature vector information, and the feature vector information is used as multiple feature vectors of the voiceprint vibration data automatically extracted by the neural network.

2. The power equipment fault diagnosis method based on cloud-edge-end collaborative architecture according to claim 1 is characterized in that: Before the step of extracting feature vectors from the voiceprint vibration data using a neural network incorporating the voiceprint vibration propagation mechanism to obtain a plurality of feature vectors, the step further includes: The voiceprint vibration data is denoised by a neural network based on the SE attention mechanism.

3. The power equipment fault diagnosis method based on cloud-edge-end collaborative architecture according to claim 2 is characterized in that: The neural network based on the SE attention mechanism denoises the voiceprint vibration data, including: The voiceprint vibration data is used as a time domain signal Input to the hidden layer of the neural network In the network layer, we get the initial feature vector : The SE attention mechanism is introduced to the preliminary feature vector , the attention coefficient M is obtained through the SENet network: The attention coefficient and the preliminary feature vector Multiply to get the noise removal base value : The preliminary feature vector With the noise removal base value Subtract to get the denoised signal ; The denoised signal As preprocessed voiceprint vibration data.

4. The power equipment fault diagnosis method based on cloud-edge-end collaborative architecture according to claim 1 is characterized in that: The performing weight adaptive calculation and weighted fusion on the multiple feature vectors to obtain a mixed feature vector includes: performing constant scaling and adaptive fusion processing on the multiple feature vectors to obtain a mixed feature vector; The identical scaling process includes: aligning the dimensions of the multiple feature vectors through upsampling and downsampling operations to make the dimensions of the multiple feature vectors consistent; wherein, upsampling is performed on feature vectors whose dimensions are less than a preset threshold, and downsampling is performed on feature vectors whose dimensions are greater than or equal to the preset threshold; The adaptive fusion process includes weighted fusion of the multiple feature vectors obtained by the identity scaling process.

5. The power equipment fault diagnosis method based on cloud-edge-end collaborative architecture according to claim 4 is characterized in that: The up-sampling of the feature vector whose dimension is smaller than the preset threshold comprises: up-sampling the feature vector whose dimension is smaller than the preset threshold by means of interpolation plus convolution; The downsampling of the feature vectors whose dimensions are greater than or equal to the preset threshold includes: using a pooling operation to reduce the dimensions of the feature vectors whose dimensions are greater than or equal to the preset threshold, so as to achieve downsampling of the feature vectors whose dimensions are greater than or equal to the preset threshold.

6. The power equipment fault diagnosis method based on cloud-edge-end collaborative architecture according to claim 5 is characterized in that: The method of upsampling the feature vector whose dimension is less than a preset threshold by interpolation plus convolution includes: When interpolating between two sampling points of a feature vector whose dimension is less than a preset threshold, obtaining an interpolation point of the two sampling points to obtain a feature vector of an interpolated and expanded dimension; Performing convolution calculation on the feature vector of the interpolation-expanded dimension to obtain the feature vector after dimension expansion, thereby achieving upsampling of the feature vector whose dimension is less than a preset threshold; The reducing the dimension of the feature vectors whose dimension is greater than or equal to the preset threshold by using a pooling operation includes: For feature vectors whose dimensions are greater than or equal to the preset threshold, a random pooling operation is used to process them to obtain feature vectors with reduced dimensions.

7. The power equipment fault diagnosis method based on cloud-edge-end collaborative architecture according to claim 4 is characterized in that: The step of performing weighted fusion on the plurality of feature vectors obtained by the constant scaling process comprises: When the deep neural network is trained on the multiple feature vectors, a fusion weight of each feature vector in the multiple feature vectors is obtained; According to the fusion weight of each feature vector in the multiple feature vectors, the multiple feature vectors obtained by the identity scaling process are fused to obtain a mixed feature vector of the voiceprint vibration data.

8. The power equipment fault diagnosis method based on cloud-edge-end collaborative architecture according to claim 1 is characterized in that: The method of using a deep residual network to perform fault identification and classification on the mixed feature vector includes: Performing fault diagnosis on the mixed feature vector through a ResNet network to obtain a fault identification result; performing fault diagnosis on the mixed feature vector through a ResNet network includes: Input mixed feature vector After the convolution layer, the ReLU calculation is performed, and then the output is obtained by passing through the convolution layer and the ReLU layer again. , plus the mixed eigenvector The final output is obtained by the identity transformation as the fault identification result; Fault classification is performed on the fault identification result.

9. A power equipment fault diagnosis system based on a cloud-edge-end collaborative architecture, which implements the power equipment fault diagnosis method based on a cloud-edge-end collaborative architecture as described in any one of claims 1 to 8; characterized in that: include: An acquisition module, used to acquire the voiceprint vibration data of the target power equipment; A feature extraction module, for extracting feature vectors from the voiceprint vibration data using a neural network that incorporates the voiceprint vibration propagation mechanism to obtain a plurality of feature vectors, and performing weight adaptive calculation and weighted fusion on the plurality of feature vectors to obtain a mixed feature vector; An identification and classification module, used for performing fault identification and classification on the mixed feature vector using a deep residual network; The output module is used to output the results of fault identification and classification.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture as described in any one of claims 1 to 8 when executing the computer program.

11. A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the power equipment fault diagnosis method based on the cloud-edge-end collaborative architecture as described in any one of claims 1 to 8 are implemented.

Images