Traction substation grounding grid breakpoint fault diagnosis method

Abstract

The invention provides a traction substation grounding grid breakpoint fault diagnosis method, and belongs to the technical field of traction substation grounding grids, and the method comprises the steps: firstly constructing an equivalent simulation circuit grid model of a traction substation grounding grid, and simulating the pilot frequency AC excitation current of traction backflow; grounding grid potential distribution data are collected in different working states, and normalization is carried out to form training samples. And then constructing a fuzzy neural network fault diagnosis model, and training to obtain the fault diagnosis capability. And finally, collecting surface potential data of a to-be-detected traction substation grounding grid, normalizing the surface potential data, and inputting the surface potential data into the diagnosis model to judge whether a breakpoint fault occurs or not and the position of the breakpoint fault. According to the traction substation grounding grid breakpoint fault diagnosis method provided by the invention, the advanced electromagnetic field theory, fuzzy logic and neural network technologies are fully fused, the defects of the existing diagnosis method are overcome, and the technical problems of insufficient diagnosis precision and efficiency in the aspect of grounding grid fault diagnosis in the prior art are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of traction substation grounding network, and in particular relates to a method for diagnosing breakpoint faults of a traction substation grounding network. Background Art

[0002] The traction power supply system is an important part of railway transportation, and its reliability is directly related to the safety and stability of railway transportation. As the core part of the traction power supply system, the health of the grounding grid system of the traction substation is crucial to the operation of the entire power supply system. Failure of the grounding grid system will cause power supply abnormalities and lead to serious safety accidents. Therefore, timely and effective diagnosis of hidden dangers of grounding grid system failures is of great significance to ensure the safe operation of the railway power supply system.

[0003] The existing grounding grid fault diagnosis methods mainly include:

[0004] On-site measurement method. By measuring the grounding resistance, grounding voltage and other parameters of the grounding grid on-site, it is possible to determine whether the grounding grid has a fault. This method is simple and direct, but it can only reflect the state of the grounding grid at the time of measurement, and cannot fully grasp the working conditions of the entire grounding grid system.

[0005] Numerical simulation method. Establish a mathematical model of the grounding grid, use computer simulation to analyze the voltage and current distribution of the grounding grid, and thus diagnose faults. This method can comprehensively analyze the working status of the grounding grid, but it is difficult to establish an accurate mathematical model, requires a large number of input parameters, and has high calculation complexity.

[0006] Intelligent diagnosis method. Artificial intelligence technologies, such as neural networks and fuzzy logic, are used to establish a grounding grid fault diagnosis model. This method has self-learning capabilities and can realize intelligent diagnosis of complex grounding grid faults, but it requires a large amount of measured data for training, and the modeling process is relatively complicated.

[0007] The above-mentioned prior art has some problems in grounding grid fault diagnosis:

[0008] The diagnostic accuracy needs to be improved. The on-site measurement method can only reflect the local state, the numerical simulation method is limited by the accuracy of the mathematical model, and the intelligent diagnosis method depends on the quality of the training data. It is difficult to fully and accurately diagnose the fault state of the grounding grid.

[0009] The diagnostic process is complex. On-site measurement requires a lot of testing work, numerical simulation requires complex modeling and calculation, and intelligent diagnosis requires a lot of training data and model construction. The overall diagnostic process is relatively complex.

[0010] Low diagnostic efficiency. The diagnostic process of the above method is cumbersome and cannot achieve fast and automated fault diagnosis, which makes it difficult to meet the needs of actual operation and maintenance.

[0011] Therefore, there is an urgent need for a new grounding grid fault diagnosis method that can overcome the shortcomings of the above-mentioned existing technologies, improve diagnostic accuracy and efficiency, and provide strong guarantees for the safe operation of the railway power supply system. Summary of the invention

[0012] In view of this, the present invention provides a method for diagnosing breakpoint faults in a traction substation grounding grid, which can solve the technical problem of insufficient diagnostic accuracy and efficiency in the prior art in grounding grid fault diagnosis.

[0013] The present invention is achieved in that:

[0014] The present invention provides a method for diagnosing a breakpoint fault in a traction substation grounding grid, comprising the following steps:

[0015] S10, establishing a mathematical model of the surface potential distribution of the grounding grid of the traction substation, wherein the mathematical model is based on the Green's function of a point current source and a weighted residual method, and takes the traction ground return current as self-excitation;

[0016] S20, constructing an equivalent analog circuit of the traction substation grounding grid, and setting the equivalent analog circuit to a grid circuit model;

[0017] S30, selecting any two points in the grid circuit model as traction return points, and passing a different-frequency AC excitation current to simulate the return current of the traction substation;

[0018] S40, setting different working states such as normal working condition, single breakpoint fault and multiple breakpoint fault, and collecting potential distribution data of the grounding grid of the traction substation;

[0019] S50, normalizing the potential distribution data to form an input set, establishing a corresponding output set, and combining the input set and the output set to form a training sample database;

[0020] S60, constructing a fuzzy neural network model, and inputting the training sample database into the fuzzy neural network for training to obtain a fault diagnosis model;

[0021] S70, collecting the surface potential data of the grounding grid of the traction substation to be tested, and normalizing the surface potential data as a new input set;

[0022] S80, inputting a new input set into the fault diagnosis model, and judging whether a breakpoint fault occurs in the grounding grid of the traction substation and the position of the breakpoint fault according to the output result.

[0023] On the basis of the above technical solution, the method for diagnosing breakpoint faults in a traction substation grounding grid of the present invention can also be improved as follows:

[0024] Wherein, the fuzzy neural network model includes an input layer, a fuzzification layer, an intermediate layer, a fuzzy decision layer and an output layer;

[0025] The input layer specifically caches input values ​​and transmits the input values ​​to the fuzzy units in the fuzzification layer;

[0026] The fuzzification layer specifically performs fuzzification processing on the input signal, including a fuzzification equation group and a sigmoid function module for determining the fuzziness; the fuzzification equation group includes a maximum membership equation, a minimum membership equation, a center membership equation and a mean membership equation;

[0027] The maximum membership equation is used to calculate the maximum membership of an input value, the input is normalized potential distribution data, and the output is the maximum membership value;

[0028] The minimum membership equation is used to calculate the minimum membership of an input value, the input is normalized potential distribution data, and the output is the minimum membership value;

[0029] The central membership equation is used to calculate the central membership of the input value, the input is the normalized potential distribution data, and the output is the central membership value;

[0030] The mean membership equation is used to calculate the mean membership of an input value, the input is normalized potential distribution data, and the output is the mean membership value;

[0031] The intermediate layer specifically performs nonlinear transformation and feature extraction on the output value of the fuzzification layer;

[0032] The fuzzy decision layer specifically receives the data processed by the intermediate layer and performs defuzzification processing, including a decision equation group, a centroid method module for defuzzification, and a maximum membership method module for classification; the decision equation group includes a fault judgment equation, a fault location equation, and a fault degree assessment equation;

[0033] The fault judgment equation is used to judge whether there is a fault in the grounding grid, the input is the characteristic vector output by the middle layer, and the output is the probability of the fault existing;

[0034] The fault location equation is used to determine the fault location, the input is the feature vector output by the middle layer, and the output is the coordinates of the fault location;

[0035] The fault severity evaluation equation is used to evaluate the severity of the fault, the input is the feature vector output by the middle layer, and the output is the level of the fault severity;

[0036] The output layer specifically outputs the fuzzy neural network operation results and provides the traction substation grounding grid breakpoint fault diagnosis results.

[0037] The sigmoid function module for determining the ambiguity specifically calculates the ambiguity through the following formula:

[0038]

[0039] In the formula, μ(x) is the fuzziness; x is the input value; a is the curve slope parameter, which controls the steepness of the function, a>0; c is the center point parameter, which determines the center position of the function. Parameters a and c are obtained by fitting the training data.

[0040] The maximum membership equation is used to calculate the maximum membership of the input value, which is specifically expressed as follows:

[0041]

[0042] In the formula, μ max (x) is the maximum membership value; x is the normalized potential distribution data; μ i (x) is the membership function of the i-th fuzzy set; n is the number of fuzzy sets.

[0043] The minimum membership equation is used to calculate the minimum membership of the input value, which is specifically expressed as follows:

[0044]

[0045] In the formula, μ min (x) is the minimum membership value; x is the normalized potential distribution data; μ i (x) is the membership function of the i-th fuzzy set; n is the number of fuzzy sets.

[0046] The center membership equation is used to calculate the center membership of the input value, which is specifically expressed as follows:

[0047]

[0048] In the formula, μ center (x) is the central membership value; x is the normalized potential distribution data; c is the central value; σ is the standard deviation, which controls the width of the membership function. Parameters c and σ are obtained by fitting the training data.

[0049] The mean membership equation is used to calculate the mean membership of the input value, which is specifically expressed as follows:

[0050]

[0051] In the formula, μ mean (x) is the mean membership value; x is the normalized potential distribution data; μ i (x) is the membership function of the i-th fuzzy set; n is the number of fuzzy sets.

[0052] The fault judgment equation is used to judge whether there is a fault in the grounding grid, which is specifically expressed as follows:

[0053]

[0054] Where P fault is the probability of failure; f i is the i-th feature component output by the middle layer; w i is the corresponding weight; b is the bias term; m is the dimension of the feature vector. i And the bias term b is obtained by optimizing the training data.

[0055] The fault location equation is used to determine the fault location, which is specifically expressed as follows:

[0056]

[0057] Where, X fault and Y fault are the x- and y-coordinates of the fault location; f i is the i-th feature component output by the middle layer; a i and b i is the corresponding coefficient; ε x and ε y is the error term; m is the dimension of the eigenvector. Coefficient a i and b i Obtained by fitting the training data.

[0058] The fault severity evaluation equation is used to evaluate the severity of the fault, and is specifically expressed as follows:

[0059]

[0060] In the formula, S fault is the level of fault severity; f i is the i-th feature component output by the middle layer; c i and d i is the corresponding coefficient; e is the constant term; ε s is the error term; m is the dimension of the eigenvector. Coefficient c i d i and the constant term e are obtained by fitting the training data.

[0061] The centroid method module for defuzzification specifically calculates the defuzzification result by the following formula:

[0062]

[0063] Where y is the output value after defuzzification; y iis the output value of the ith discrete point; μ(y i ) is the corresponding membership value; n is the number of discrete points.

[0064] The maximum membership method module for classification is specifically implemented by the following steps:

[0065] Step 1: Calculate the membership of each category:

[0066]

[0067] In the formula, μ j (x) is the membership of input x to category j; c j is the center vector of category j; a j and b j To adjust the parameters; ||xc j || represents the Euclidean distance.

[0068] Step 2: Select the category with the largest membership as the classification result:

[0069] class(x)=argmax j {μ j (x)};

[0070] Where class(x) is the classification result of input x.

[0071] Parameter acquisition method:

[0072] 1. For the parameters in the fuzzy equation group, such as a and c of the sigmoid function, c and σ of the central membership equation, they are obtained by fitting the training data. The specific steps are as follows:

[0073] Step 1: Collect training data, including input potential distribution data and corresponding fault labels;

[0074] Step 2: Use the least squares method or gradient descent method to optimize the parameters and minimize the prediction error.

[0075] 2. For the parameters in the decision equation group, such as w in the fault judgment equation i and b, a of the fault location equation i and b i , c of the fault degree assessment equation i d i and e are obtained through the neural network training process, the specific steps are as follows:

[0076] Step 1: Initialize the parameters to random small values;

[0077] Step 2: Use the back propagation algorithm to calculate the parameter gradient;

[0078] Step 3: Update the parameters using an optimization algorithm such as stochastic gradient descent.

[0079] Step 4: Repeat steps 2 and 3 until the preset number of iterations or convergence condition is reached.

[0080] Specifically, the step S10 specifically includes:

[0081] Step 101: According to Kirchhoff's current law of the circuit, establish the leakage current equation in the grounding grid The B matrix represents the self-conductance of the middle node of each conductor segment, the C matrix represents the mutual conductance of the middle node and the end node of the conductor, and the E matrix represents the self-conductance of the end node of each conductor segment.

[0082] Step 102: According to the total potential on the conductor and leakage current I b =R -1 φ b ,Construct a mathematical model to describe the surface potential distribution of the grounding grid;

[0083] Step 103: Calculate the equivalent resistance between each conductor according to the point current source image method. Improve the mathematical model. Through these steps, a mathematical model that can describe the surface potential distribution of the actual grounding grid is established, providing a basis for subsequent fault diagnosis.

[0084] Specifically, the specific steps of step S20 include:

[0085] Step 201, simplifying the entire grounding grid system into a grid-shaped circuit model with m rows and n columns, where each grid node represents a position of the actual grounding grid, and adjacent grids are coupled through resistors;

[0086] Step 202: This grid-shaped circuit model can well reflect the topological structure and electrical parameter characteristics of the actual grounding grid, and provide a simulation platform for subsequent fault diagnosis. By constructing such an equivalent simulation circuit, the working state of the grounding grid can be better analyzed, and data support can be provided for subsequent fault diagnosis.

[0087] Specifically, the step S30 specifically includes:

[0088] Step 301: In the constructed grid circuit model, select any two nodes as traction return points, representing the positions receiving traction return in the actual grounding grid;

[0089] Step 302: Pass the hetero-frequency AC excitation current through the two selected traction return points to simulate the actual traction return process, obtain the potential distribution data of the grounding grid under different working conditions, and provide the necessary data basis for subsequent fault diagnosis. By applying this hetero-frequency AC excitation, the potential distribution data of the grounding grid under different working conditions can be obtained, providing more realistic working conditions for subsequent fault diagnosis.

[0090] Specifically, the step S40 specifically includes:

[0091] Step 401, setting different working states such as normal working condition, single breakpoint fault and multiple breakpoint fault to simulate various fault conditions that may occur in the actual grounding grid;

[0092] Step 402: Under each working state, collect the surface potential distribution data of each node of the grounding grid to form a potential distribution matrix X, and record the potential distribution characteristics of the grounding grid under different working conditions; by setting different working states and collecting corresponding potential distribution data, provide necessary training samples for the subsequent establishment of a fault diagnosis model, reflecting the difference characteristics of the potential distribution of the grounding grid under normal working conditions and fault conditions.

[0093] Specifically, the step S50 specifically includes:

[0094] Step 501: normalize each element in the potential distribution matrix X to obtain a normalized input set X. The specific method is: in is the average value of the elements in the jth column, and then the normalized data is linearly normalized:

[0095] Step 502: Based on the known working state, construct the corresponding output set Y to be suitable for subsequent model training. Through the above processing, the original potential distribution data is converted into an input and output sample database suitable for model training, which is beneficial to subsequent model training.

[0096] Specifically, the step S60 specifically includes:

[0097] Step 601, construct a five-layer fuzzy neural network model, including an input layer, a fuzzification layer, an intermediate layer, a fuzzy decision layer and an output layer;

[0098] Step 602: Input the training sample database established above into the fuzzy neural network for training, which includes two main steps: 1) For the parameters in the fuzzy equation group, such as a and c of the sigmoid function, c and σ of the central membership equation, they are obtained by fitting the training data; 2) For the parameters in the decision equation group, such as w of the fault judgment equation, iand b, a of the fault location equation i and b i ,c of the fault degree evaluation equation i d i and e, are obtained through the neural network training process. Through the training of this fuzzy neural network model, an intelligent model that can accurately diagnose grounding grid faults can be obtained.

[0099] Specifically, the step S70 specifically includes:

[0100] Step 701, collecting the surface potential data of the grounding grid of the traction substation to be detected, and forming a new potential distribution matrix X;

[0101] Step 702: Perform the same normalization processing as step S50 on the collected potential distribution data to ensure that the data to be detected and the training sample data have the same data characteristics, providing appropriate input for subsequent fault diagnosis. Through this normalization processing, it can be ensured that the data to be detected and the training sample data have the same data characteristics.

[0102] Specifically, the step S80 specifically includes:

[0103] Step 801, inputting the potential distribution data matrix X processed in step S70 into the trained fuzzy neural network model;

[0104] Step 802: The fuzzy neural network model gives the fault diagnosis result at the output layer according to the input data through the operation of the fuzzification layer, the intermediate layer, and the fuzzy decision layer. ij =1, it means that a breakpoint fault has occurred in the i-th row and j-th column of the grounding grid;

[0105] Step 803: In this way, it is possible to determine whether there is a breakpoint fault in the grounding grid and determine the specific location of the fault. This fault diagnosis process utilizes the powerful learning and reasoning capabilities of the fuzzy neural network, and can well identify the difference characteristics of the potential distribution of the grounding grid under normal working conditions and fault conditions.

[0106] Compared with the prior art, the beneficial effects of the method for diagnosing breakpoint faults of a traction substation grounding grid provided by the present invention are:

[0107] The diagnostic accuracy is significantly improved. By establishing a mathematical model and an equivalent circuit model of the grounding grid and combining measured data to perform fuzzy neural network training, the difference characteristics of the grounding grid potential distribution under normal working conditions and fault conditions can be effectively learned, thereby accurately judging whether the grounding grid has a fault and the fault location. Compared with existing diagnostic methods, the diagnostic accuracy of the method of the present invention is greatly improved.

[0108] The diagnostic process is automated. The method of the present invention does not require tedious on-site measurements and complex mathematical modeling. It only needs to collect the surface potential data of the grounding grid, and can automatically perform fault diagnosis through the trained fuzzy neural network model. The entire diagnostic process is automated, greatly improving the diagnostic efficiency.

[0109] Strong adaptability. The method of the present invention is based on fuzzy neural network and has strong self-learning ability. By continuously supplementing training samples, the model can be continuously optimized, thereby improving the diagnostic ability of various complex grounding network faults and having wider adaptability.

[0110] High application convenience. The method of the present invention only needs to collect the surface potential data of the grounding grid, without other complicated on-site tests, and is simple and easy to operate, and is convenient for promotion and application in actual operation and maintenance.

[0111] To sum up, the traction substation grounding grid breakpoint fault diagnosis method proposed in the present invention fully integrates advanced electromagnetic field theory, fuzzy logic and neural network technology, overcomes the shortcomings of existing diagnostic methods, and solves the technical problems of insufficient diagnostic accuracy and efficiency in grounding grid fault diagnosis in existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] Figure 1 A flow chart of the method provided by the present invention;

[0113] Figure 2 It is the topological structure diagram of the equivalent model of the grounding grid;

[0114] Figure 3 This is a simulation diagram of the traction substation grounding grid breakpoint fault experiment. DETAILED DESCRIPTION

[0115] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0116] like Figure 1 FIG. 1 is a flow chart of a method for diagnosing a breakpoint fault in a traction substation grounding grid provided by the present invention. The method comprises the following steps:

[0117] S10, establish a mathematical model of the surface potential distribution of the traction substation grounding grid, the mathematical model is based on the Green's function of the point current source and the weighted residual method, and takes the traction ground return current as the self-excitation;

[0118] S20, constructing an equivalent analog circuit of the traction substation grounding grid, and setting the equivalent analog circuit to a grid circuit model;

[0119] S30, selecting any two points in the grid circuit model as traction return points, and passing a different-frequency AC excitation current to simulate the return current of the traction substation;

[0120] S40, setting different working states such as normal working condition, single breakpoint fault and multiple breakpoint fault, and collecting potential distribution data of the grounding grid of the traction substation;

[0121] S50, normalizing the potential distribution data to form an input set, establishing a corresponding output set, and combining the input set and the output set to form a training sample database;

[0122] S60, constructing a fuzzy neural network model, and inputting a training sample database into the fuzzy neural network for training to obtain a fault diagnosis model;

[0123] S70, collecting the surface potential data of the grounding grid of the traction substation to be tested, and normalizing the surface potential data as a new input set;

[0124] S80, inputting the new input set into the fault diagnosis model, and judging whether a breakpoint fault occurs in the grounding grid of the traction substation and the position of the breakpoint fault according to the output result.

[0125] The specific implementation methods of the above steps are described in detail below:

[0126] Step S10: Establishing a mathematical model of the surface potential distribution of the traction substation grounding grid.

[0127] First, according to Kirchhoff's current law of the circuit, the leakage current equation in the grounding grid can be obtained: The B matrix represents the self-conductance of the nodes in the middle of each conductor segment, where Where S i is the cross-sectional area of ​​the conductor (m 2 ),ρ c is the conductor resistivity (Ω·m), L i is the length of the conductor segment (m). The C matrix represents the mutual conductance of the middle node and the end node of the conductor. When the i branch is connected to the j branch, On the contrary, C ij = 0. The E matrix represents the self-conductance of each conductor end node. represents the total potential on conductor j, and the leakage current I b =R -1 φ b The R matrix represents the equivalent resistance between conductors, where It is calculated based on the point current source image method. Here, ρ is the soil resistivity, k is the number of grounding grid nodes, and r ijk and r′ijk They represent the actual distance and mirror distance between conductor i and conductor j respectively.

[0128] By establishing such a mathematical model, the surface potential distribution of each point in the grounding grid can be described, providing a basis for subsequent fault diagnosis. This mathematical model is based on circuit theory and electromagnetic field theory. By analyzing the current and potential distribution in the conductor, a mathematical expression that can describe the surface potential distribution of the entire grounding grid is obtained.

[0129] Step S20: construct an equivalent analog circuit of the traction substation grounding grid.

[0130] The entire grounding grid system is set as a grid circuit model with m rows and n columns. Each grid node represents a position in the actual grounding grid, and adjacent grids are coupled through resistors to form a circuit that can simulate the potential distribution of the actual grounding grid. This grid circuit model can well reflect the topological structure and electrical parameter characteristics of the actual grounding grid, providing a simulation platform for subsequent fault diagnosis.

[0131] Step S30: Select a traction return point and introduce a different-frequency AC excitation current.

[0132] In the constructed grid circuit model, any two nodes are selected as traction return points. These two nodes represent the locations in the actual grounding grid that receive traction return. Then, a different-frequency AC excitation current is introduced into these two selected traction return points. This different-frequency AC signal can better excite the current distribution in the grounding grid, making the entire grounding grid in an electromagnetic environment similar to the actual working state.

[0133] Step S40: Setting different working states and collecting potential distribution data.

[0134] First, different working states such as normal working condition, single breakpoint fault and multiple breakpoint fault are set. These working states are set to simulate various fault conditions that may occur in the actual grounding grid. Then, under each working state, the surface potential distribution data of each node of the grounding grid is collected. These data will form a potential distribution matrix X, which records the potential distribution characteristics of the grounding grid under different working conditions.

[0135] Step S50: normalize and form a training sample database.

[0136] The potential distribution matrix X can be expressed as:

[0137]

[0138] First, normalize each element in X:

[0139]

[0140] in, is the average of the elements in the jth column.

[0141] Then, the normalized data is linearly normalized:

[0142]

[0143] In this way, the normalized input set X is obtained. At the same time, according to the known working status, the corresponding output set Y is constructed. For example, if the node in the i-th row and j-th column fails, then y ij =1, otherwise y ij =0.

[0144] Through the above processing, the original potential distribution data is converted into an input and output sample database suitable for model training. This normalization process can eliminate the dimension effect and make the sample data distribution more concentrated, which is conducive to subsequent model training.

[0145] Step S60: construct a fuzzy neural network model and train it.

[0146] First, a five-layer fuzzy neural network model is constructed, including input layer, fuzzification layer, intermediate layer, fuzzy decision layer and output layer. Among them, the fuzzification layer performs fuzzification processing on the input signal, including the following fuzzification equations:

[0147] Maximum membership equation:

[0148]

[0149] Among them, μ max (x) is the maximum membership value, x is the normalized potential distribution data, μ i (x) is the membership function of the i-th fuzzy set, and n is the number of fuzzy sets.

[0150] Minimum membership equation:

[0151]

[0152] Among them, μ min (x) is the minimum membership value, x is the normalized potential distribution data, μ i (x) is the membership function of the i-th fuzzy set, and n is the number of fuzzy sets.

[0153] Central membership equation:

[0154]

[0155] Among them, μ center(x) is the central membership value, x is the normalized potential distribution data, c is the central value, and σ is the standard deviation, which controls the width of the membership function.

[0156] Mean membership equation:

[0157]

[0158] Among them, μ mean (x) is the mean membership value, x is the normalized potential distribution data, μ i (x) is the membership function of the i-th fuzzy set, and n is the number of fuzzy sets.

[0159] The fuzzy decision layer includes the following decision equations:

[0160] Fault judgment equation:

[0161]

[0162] Among them, P fault is the probability of failure, f i is the i-th feature component output by the intermediate layer, w i is the corresponding weight, b is the bias term, and m is the dimension of the feature vector.

[0163] Fault location equation:

[0164]

[0165] Among them, X fault and Y fault are the x- and y-coordinates of the fault location, f i is the i-th feature component output by the intermediate layer, a i and b i is the corresponding coefficient, ε x and ε y is the error term, and m is the dimension of the feature vector.

[0166] Fault severity evaluation equation:

[0167]

[0168] Among them, S fault is the fault severity level, f i is the i-th feature component output by the intermediate layer, c i and d i is the corresponding coefficient, e is the constant term, ε s is the error term, and m is the dimension of the feature vector.

[0169] Then, the training sample database established earlier is input into the fuzzy neural network for training. The training process includes two main steps:

[0170] 1) The parameters in the fuzzy equation group, such as a and c of the sigmoid function, c and σ of the central membership equation, are obtained by fitting the training data. The specific method is to use the least squares method or gradient descent method to optimize the parameters to minimize the prediction error.

[0171] 2) For the parameters in the decision equation group, such as w in the fault judgment equation i and b, a of the fault location equation i and b i ,c of the fault degree evaluation equation i d i and e are obtained through the neural network training process. That is, they are initialized to random small values, and then the gradient is calculated using the back propagation algorithm, and then the optimization algorithm (such as stochastic gradient descent) is used to update the parameters until the convergence condition is reached.

[0172] By training this fuzzy neural network model, an intelligent model that can accurately diagnose grounding grid faults can be obtained.

[0173] Step S70: Collect the potential data of the grounding grid to be detected and normalize it.

[0174] First, the surface potential data of the grounding grid of the traction substation to be detected is collected to form a new potential distribution matrix X. Then, the collected potential distribution data is normalized in the same way as step S50:

[0175]

[0176] in, is the average of the elements in the jth column.

[0177] Then perform linear normalization on the normalized data:

[0178]

[0179] Through this normalization process, it can be ensured that the data to be tested has the same data characteristics as the training sample data, providing appropriate input for subsequent fault diagnosis.

[0180] Specifically, the principle of the present invention is:

[0181] 1. Mathematical modeling of grounding grid based on electromagnetic field theory

[0182] As a typical electromagnetic field problem, the grounding grid system can be modeled by circuit theory and electromagnetic field theory. First, according to Kirchhoff's current law, a set of equations describing the distribution of leakage current in the grounding grid can be established. This set of equations reflects the relationship between current and potential between nodes in the grounding grid.

[0183] Then, according to the total potential on the conductor and leakage current I b =R -1 φ b , a mathematical model can be further established to describe the surface potential distribution of the entire grounding grid. ij It is the equivalent resistance between conductors calculated based on the point current source image method.

[0184] By establishing such a mathematical model, the electromagnetic field characteristics of each node in the grounding grid can be fully described, providing a basis for subsequent fault diagnosis.

[0185] 2. Grounding grid simulation based on equivalent circuit

[0186] In order to better analyze the working state of the grounding grid, the present invention simplifies the entire grounding grid into a grid-shaped circuit model with m rows and n columns. Each grid node represents a position of the actual grounding grid, and adjacent grids are coupled through resistors to form a circuit that can simulate the potential distribution of the actual grounding grid.

[0187] This grid circuit model can well reflect the topological structure and electrical parameter characteristics of the actual grounding grid, providing a simulation platform for subsequent fault diagnosis. On this basis, by applying heterofrequency AC excitation at the selected traction return point, the potential distribution data of the grounding grid under different working conditions can be obtained, providing necessary training samples for fault diagnosis.

[0188] 3. Fault diagnosis based on fuzzy neural network

[0189] The method of the present invention adopts fuzzy neural network technology to realize intelligent diagnosis of grounding grid faults. Fuzzy neural network combines the fuzzy reasoning ability of fuzzy logic and the learning ability of neural network, and can well handle complex fault diagnosis problems such as grounding grid.

[0190] A specific embodiment 1 of the present invention is provided below:

[0191] Firstly, based on the Green's function of point current source and weighted residual method, a calculation model of the surface potential of the grounding grid is established, and the distribution characteristics of the surface potential of the traction substation grounding grid are analyzed.

[0192] Under the self-excitation of the traction ground return of the electrified railway, the grounding grid will generate potential on the surface. When the grounding grid fails, the distribution of the potential will change, which can reflect the operating status of the grounding grid. Based on this, a Figure 2 The mathematical model of the surface potential distribution of the 3*3 grounding grid shown.

[0193] The excitation current is injected from the node into the ground grid. When the current flows in the ground grid, it causes leakage current. According to Kirchhoff's current law of the circuit, the leakage current can be described as follows:

[0194]

[0195] Where: The B matrix represents the self-conductance of the nodes in the middle of each conductor segment, Where S i is the cross-sectional area of ​​the conductor (m2), ρ c is the conductor resistivity (Ω.m), L i is the length of the conductor segment (m); C represents the mutual conductance matrix of the middle and end nodes of the conductor. When the i branch is connected to the j branch, On the contrary, C ij =0; The E matrix represents the self-conductance of each conductor end node,

[0196] The leakage current of each conductor will generate a potential on the surface of all conductors. According to the Green's function of the point current source and the weighted residual method, the total potential on conductor j can be expressed as:

[0197]

[0198] Therefore, the leakage current I b It can be described as follows:

[0199]

[0200] Based on the point current source mirror method, R ij It can be described as follows:

[0201]

[0202] Combining equations (1), (3), and (5) we can get the leakage current I b Substituting it into formula (2), the potential of any point on the ground surface can be obtained, and thus the surface potential of the traction substation grounding grid can be calculated.

[0203] Establish as Figure 2The 3*3 grounding grid equivalent simulation circuit of the traction substation shown in the figure is simply set as a 3-row 3-column grid circuit model without considering the capacitance. Select any two points as traction return points, and pass the different-frequency AC excitation current to simulate the return current of the traction substation, as shown in Figure 3 Set up a traction substation grounding grid breakpoint fault simulation experiment. According to the surface potential distribution mathematical model of the traction substation grounding grid in step 1, collect the potential distribution data of the traction substation grounding grid, normalize it, and organize the normalized data as the input set X. mn = 0, when a fault occurs at row m and column n, y mn = 1. The input set X and its corresponding output set Y form a test set, and multiple test sets are sorted to form a training sample database.

[0204]

[0205]

[0206] Where: X is the distribution matrix of the measured data of the ground potential of the traction substation grounding grid, x mn is the measured value of the ground potential at the mth row and nth column just above the intersection node of the traction substation grounding grid, max(x j ) is the maximum value of the jth column, min(x j ) is the minimum value of the j-th column (i=1,2,3; j=1,2,3).

[0207] The five-layer fuzzy neural network model includes input layer 1, fuzzification layer 2, intermediate layer 3, fuzzy decision layer 4 and output layer 5. The number of neurons in the fuzzy neural network input layer 1 is initialized to 9, the number of neurons in the fuzzification layer 2 is 18, the number of neurons in the fuzzy decision layer 4 is 9, and the number of neurons in the output layer 5 is 9. The training set sample database in step 2 is input into the fuzzy neural network to train the fuzzy neural network to identify and locate the breakpoint fault of the traction substation grounding network.

[0208] For the traction substation grounding grid to be tested, the traction ground return current is used as self-excitation, and the surface potential data of the traction substation grounding grid to be tested is collected as the input set X. The input set X is input into the trained fuzzy neural network to obtain the output set Y. Determine the output set Y. 11 =1, it means that a breakpoint fault occurs at the node in the first row and first column of the traction substation grounding grid.

[0209] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: There is a hidden danger of breakpoint fault in the grounding network system of a traction substation under the jurisdiction of a certain railway bureau. In order to diagnose the fault situation in a timely and effective manner, the railway bureau decides to adopt the traction substation grounding network breakpoint fault diagnosis method proposed in the present invention for detection.

[0210] First, the railway bureau organized relevant technical personnel to conduct a detailed on-site survey and parameter measurement of the grounding grid system. According to the topological structure and electrical parameter characteristics of the actual grounding grid, a 20-row 20-column grid circuit model was established, with a total of 400 grid nodes. Each grid node represents a position in the actual grounding grid, and adjacent grids are coupled through resistance. Through measurement, the cross-sectional area S of each conductor segment was obtained. i and length L i , and parameters such as soil resistivity ρ.

[0211] According to Kirchhoff's current law, a set of equations describing the leakage current distribution in the grounding grid is established:

[0212]

[0213] in, The B matrix represents the self-conductance of the nodes in the middle of each conductor segment; The C matrix represents the mutual conductance of the nodes in the middle and end of the conductor; The E matrix represents the self-conductance of the end nodes of each conductor segment.

[0214] Then, according to the total potential on the conductor and leakage current I b =R -1 φ b , a mathematical model describing the surface potential distribution of the entire grounding grid was established, in which is the equivalent resistance between the conductors.

[0215] In order to obtain the potential distribution data of the grounding grid under different working conditions, the railway bureau selected two nodes numbered (5,5) and (15,15) as traction return points in the constructed grid circuit model, and passed an AC current with a frequency of 50 Hz and a current amplitude of 100 A through these two nodes. On this basis, three conditions were set: normal working condition, single breakpoint fault condition, and multiple breakpoint fault condition.

[0216] For normal working conditions, the surface potential distribution data of each node of the grounding grid is shown in Table 1:

[0217] Table 1 Grounding grid surface potential distribution data under normal working conditions

[0218] Node coordinates Potential (V) (1,1) 2.13 (1,2) 2.15 … … (20,20) 1.98

[0219] For the single-breakpoint fault condition, a fault point is set at the node numbered (10,10), and the surface potential distribution data of each node of the grounding grid are shown in Table 2:

[0220] Table 2 Ground potential distribution data of grounding grid under single breakpoint fault condition

[0221] Node coordinates Potential (V) (1,1) 2.21 (1,2) 2.24 … … (10,10) 3.15 … … (20,20) 2.06

[0222] For the multi-breakpoint fault condition, the fault points are set at two nodes numbered (5, 10) and (15, 15). The surface potential distribution data of each node of the grounding grid are shown in Table 3:

[0223] Table 3 Surface potential distribution data of grounding grid under multi-breakpoint fault conditions

[0224] Node coordinates Potential (V) (1,1) 2.26 (1,2) 2.29 … … (5,10) 3.41 (15,15) 3.57 … … (20,20) 2.12

[0225] Through the potential distribution data under the above three working conditions, the railway bureau built a training sample database. First, the potential distribution data matrix X is normalized:

[0226]

[0227] in, is the average value of the elements in the jth column. Then, the normalized data is linearly normalized:

[0228]

[0229] Get the normalized input set X. At the same time, according to the known fault conditions, construct the corresponding output set Y. If the node in the i-th row and j-th column fails, then y ij =1, otherwise y ij =0.

[0230] Next, the railway bureau built a five-layer fuzzy neural network model, including input layer, fuzzification layer, intermediate layer, fuzzy decision layer and output layer. In the fuzzification layer, the following four fuzzification equations were used:

[0231] Maximum membership equation:

[0232] Minimum membership equation:

[0233] Central membership equation:

[0234] Mean membership equation:

[0235] In the fuzzy decision layer, the following three decision equations are used:

[0236] Fault judgment equation:

[0237] Fault location equation:

[0238]

[0239] Fault severity evaluation equation:

[0240]

[0241] The training sample database is input into the fuzzy neural network for training. The parameters in the fuzzy equation group, such as a and c of the sigmoid function, c and σ of the central membership equation, are obtained by fitting the training data using the least squares method. The parameters in the decision equation group, such as w of the fault judgment equation, are obtained by fitting the training data using the least squares method. i and b, a of the fault location equation i and b i ,c of the fault degree evaluation equation i d i and e, are obtained through the neural network training process. After multiple iterations of training, a fuzzy neural network model that can accurately diagnose grounding grid faults is finally obtained.

[0242] In the actual detection, the railway bureau collected the surface potential distribution data of the grounding grid to be detected to form a new potential distribution matrix X. X was also normalized to obtain a normalized input set. Then, the input set was input into the trained fuzzy neural network model, and fault diagnosis was performed based on the output results.

[0243] In a specific test, the surface potential distribution data of the grounding grid to be tested collected by the railway bureau is shown in Table 4:

[0244] Table 4 Surface potential distribution data of the grounding grid to be tested

[0245] Node coordinates Potential (V) (1,1) 2.24 (1,2) 2.27 … … (9,9) 3.02 … … (20,20) 2.09

[0246] First, the collected potential distribution data is normalized to obtain the normalized input set X. Then, x is input into the trained fuzzy neural network model. According to the output results of the fuzzy neural network, the following diagnostic results can be obtained:

[0247] 1. Fault judgment: fuzzy neural network output P fault =0.87, indicating that the probability of a fault in the grounding grid is very high.

[0248] 2. Fault location: Fuzzy neural network output X fault =9.2,Y fault=9.1, indicating that the fault is located near the grid coordinate (9,9).

[0249] 3. Fault degree assessment: Fuzzy neural network output S fault =4, indicating that the severity of the fault is medium and needs to be dealt with as soon as possible.

[0250] Through the above diagnostic results, the railway bureau promptly discovered that there was a moderate breakpoint fault in the grounding grid, located near the grid coordinates (9,9). Subsequently, they immediately organized personnel to conduct on-site maintenance, troubleshoot and repair the fault point.

[0251] After the inspection, the railway bureau collected the surface potential distribution data of the grounding grid again and found that the potential of each node had returned to normal level. The diagnosis results showed that the grounding grid had no hidden faults. Through this practice, the railway bureau fully realized the superiority of the method of the present invention, which can not only accurately diagnose grounding grid faults, but also make the diagnosis process automated and efficient, providing a strong guarantee for the safe operation of the railway power supply system.

[0252] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for diagnosing breakpoint faults in a traction substation grounding grid, characterized in that: The following steps are involved: S10, establishing a mathematical model of the surface potential distribution of the grounding grid of the traction substation, wherein the mathematical model is based on the Green's function of a point current source and a weighted residual method, and takes the traction ground return current as self-excitation; S20, constructing an equivalent analog circuit of the traction substation grounding grid, and setting the equivalent analog circuit to a grid circuit model; S30, selecting any two points in the grid circuit model as traction return points, and passing a different-frequency AC excitation current to simulate the return current of the traction substation; S40, setting different working states such as normal working condition, single breakpoint fault and multiple breakpoint fault, and collecting potential distribution data of the grounding grid of the traction substation; S50, normalizing the potential distribution data to form an input set, establishing a corresponding output set, and combining the input set and the output set to form a training sample database; S60, constructing a fuzzy neural network model, and inputting the training sample database into the fuzzy neural network for training to obtain a fault diagnosis model; S70, collecting the surface potential data of the grounding grid of the traction substation to be tested, and normalizing the surface potential data as a new input set; S80, inputting a new input set into the fault diagnosis model, and judging whether a breakpoint fault occurs in the grounding grid of the traction substation and the position of the breakpoint fault according to the output result.

2. A method for diagnosing a breakpoint fault in a traction substation grounding grid according to claim 1, characterized in that: The fuzzy neural network model includes an input layer, a fuzzification layer, an intermediate layer, a fuzzy decision layer and an output layer; The input layer specifically caches input values ​​and transmits the input values ​​to the fuzzy units in the fuzzification layer; The fuzzification layer specifically performs fuzzification processing on the input signal, including a fuzzification equation group and a sigmoid function module for determining the fuzziness; the fuzzification equation group includes a maximum membership equation, a minimum membership equation, a center membership equation and a mean membership equation; The maximum membership equation is used to calculate the maximum membership of an input value, the input is normalized potential distribution data, and the output is the maximum membership value; The minimum membership equation is used to calculate the minimum membership of an input value, the input is normalized potential distribution data, and the output is the minimum membership value; The central membership equation is used to calculate the central membership of the input value, the input is the normalized potential distribution data, and the output is the central membership value; The mean membership equation is used to calculate the mean membership of an input value, the input is normalized potential distribution data, and the output is the mean membership value; The intermediate layer specifically performs nonlinear transformation and feature extraction on the output value of the fuzzification layer; The fuzzy decision layer specifically receives the data processed by the intermediate layer and performs defuzzification, including a decision equation group, a centroid method module for defuzzification, and a maximum membership method module for classification; the decision equation group includes a fault judgment equation, a fault location equation, and a fault degree assessment equation; The fault judgment equation is used to judge whether there is a fault in the grounding grid, the input is the characteristic vector output by the middle layer, and the output is the probability of the fault existing; The fault location equation is used to determine the fault location, the input is the feature vector output by the middle layer, and the output is the coordinates of the fault location; The fault severity evaluation equation is used to evaluate the severity of the fault, the input is the feature vector output by the middle layer, and the output is the level of the fault severity; The output layer specifically outputs the fuzzy neural network operation results and provides the traction substation grounding grid breakpoint fault diagnosis results.

3. A method for diagnosing a breakpoint fault in a traction substation grounding grid according to claim 2, characterized in that: The sigmoid function module for determining the ambiguity specifically calculates the ambiguity through the following formula: Where μ(x) is the fuzziness; x is the input value; a is the curve slope parameter, which controls the steepness of the function, a>0; c is the center point parameter, which determines the center position of the function.

4. A method for diagnosing a breakpoint fault in a traction substation grounding grid according to claim 3, characterized in that: The maximum membership equation is used to calculate the maximum membership of the input value, which is specifically expressed as follows: In the formula, μ max (x) is the maximum membership value; x is the normalized potential distribution data; μ i (x) is the membership function of the i-th fuzzy set; n is the number of fuzzy sets.

5. A method for diagnosing a breakpoint fault in a traction substation grounding grid according to claim 4, characterized in that: The minimum membership equation is used to calculate the minimum membership of the input value, which is specifically expressed as follows: Where, μm in (x) is the minimum membership value; x is the normalized potential distribution data; μ i (x) is the membership function of the i-th fuzzy set; n is the number of fuzzy sets.

6. A method for diagnosing breakpoint faults in a traction substation grounding grid according to claim 5, characterized in that: The center membership equation is used to calculate the center membership of the input value, which is specifically expressed as follows: In the formula, μ center (x) is the central membership value; x is the normalized potential distribution data; c is the central value; σ is the standard deviation, which controls the width of the membership function.

7. A method for diagnosing a breakpoint fault in a traction substation grounding grid according to claim 6, characterized in that: The mean membership equation is used to calculate the mean membership of the input value, which is specifically expressed as follows: In the formula, μ mean (x) is the mean membership value; x is the normalized potential distribution data; μ i (x) is the membership function of the i-th fuzzy set; n is the number of fuzzy sets.

8. A method for diagnosing a breakpoint fault in a traction substation grounding grid according to claim 7, characterized in that: The fault judgment equation is used to judge whether there is a fault in the grounding grid, which is specifically expressed as follows: Where P fault is the probability of failure; f i is the i-th feature component output by the intermediate layer; w i is the corresponding weight; b is the bias term; m is the dimension of the feature vector.

9. A method for diagnosing a breakpoint fault in a traction substation grounding grid according to claim 8, characterized in that: The fault location equation is used to determine the fault location, which is specifically expressed as follows: In the formula, X fault and Y fault are the x- and y-coordinates of the fault location; f i is the i-th feature component output by the intermediate layer; a i and b i is the corresponding coefficient; ε x and ε y is the error term; m is the dimension of the feature vector.

10. A method for diagnosing breakpoint faults in a traction substation grounding grid according to claim 9, characterized in that: The fault severity evaluation equation is used to evaluate the severity of the fault, and is specifically expressed as follows: In the formula, S fault is the level of fault severity; f i is the i-th feature component output by the intermediate layer; c i and d i is the corresponding coefficient; e is the constant term; ε s is the error term; m is the dimension of the feature vector.

Images