A data modeling method and system based on reliability analysis of railway power supply equipment

The data of railway powered equipment is cleaned and labeled through stack noise reduction automatic encoder network and BP neural network. Combined with fast Fourier transform and convolutional neural network, the health factors and fault characteristics of the equipment are automatically extracted, and data pollution is solved and aging process describe problems, improving the accuracy and efficiency of equipment failure monitoring.

CN112231975BActive Publication Date: 2024-11-01NANJING POWER SUPPLY SECTION OF CHINA RAILWAY SHANGHAI BUREAU GRP CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011090779.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-10-13
Publication Date
2024-11-01
Estimated Expiration
2040-10-13

AI Technical Summary

Technical Problem

The fault monitoring data of railway power supply equipment is contaminated by environmental noise and coupled component signals, and cannot be directly used for deep learning network training. Traditional methods rely on expert experience to make it difficult to comprehensively and accurately describe the aging process of equipment, resulting in inefficient equipment reliability evaluation.

Method used

The stack noise reduction automatic encoder network and BP neural network are used to denoising and cleaning and labeling the historical data of railway power supply equipment. Combined with fast Fourier transform and convolutional neural network, it automatically extracts the health factor diagram of the equipment and deep faults. feature.

Benefits of technology

It improves the accuracy and data processing efficiency of equipment failure status identification, can describe the equipment aging failure process more comprehensively and accurately, and realizes automation and accuracy of the reliability evaluation of railway power supply equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112231975B_ABST
    Figure CN112231975B_ABST
Patent Text Reader

Abstract

The present invention proposes a data modeling method and system based on the reliability analysis of railway power supply equipment, and the system includes a training function module and a testing function module. The training function module uses a stacked noise reduction encoder to perform noise reduction processing on the historical operation data of the railway power supply equipment and extract the main features of the data; the processed data information is input into the BP neural network to extract the health factor (HI) diagram of the railway power supply equipment. The testing function module inputs the operating status data of the railway power supply equipment into the trained stacked noise reduction autoencoder to verify the noise reduction effect of the stacked noise reduction autoencoder; the main characteristic data of the operating status of the railway power supply equipment is extracted to automatically identify the fault type of the railway power supply equipment. Through the coordinated use of the training function module and the testing function module, a deep-level characteristic data set of railway power supply equipment faults can be constructed, which is of great significance for the subsequent completion of online monitoring of railway power supply equipment faults.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a data modeling method and system based on reliability analysis of railway power supply equipment, and relates to the technical field of intelligent detection. Background Art

[0002] With the increasing demand for national economic development and the advancement of high-speed rail technology, the scale of my country's railways is showing a trend of continuous expansion. The safety and stability of railway power supply equipment is the basis for the reliable operation of railways. Once a railway power supply equipment fails, it will cause railway delays at the least, affecting people's normal production and life; at the worst, it will cause safety accidents, causing major economic losses and adverse social impacts. Therefore, real-time monitoring of the health status of railway power supply equipment, early warning of potential equipment failures, and timely maintenance and replacement of damaged equipment are of great significance to improving the reliable operation of railways.

[0003] For a long time, the regular maintenance mechanism implemented by power companies for railway power supply equipment has problems such as "insufficient maintenance" and "excessive maintenance", which not only causes significant waste of resources, but also affects the reliability of equipment power supply to a certain extent. Therefore, it is imperative to carry out status maintenance of railway power supply equipment based on the equipment history and current operating status, using data such as online monitoring and offline experiments. At present, with the advancement of communication, computer and control technology, fault monitoring systems have been widely used in railway power supply equipment, accumulating massive data. How to deeply analyze these data and maintain the safe operation of railway power supply equipment has become a difficult problem that urgently needs to be solved at this stage.

[0004] Railway power supply equipment has many types, complex parameters, large amount of monitoring data, and diverse operating environments. The equipment information collected by sensors may be "contaminated" by environmental noise and signals from other coupled components, covering up key information reflecting the status of the equipment. The data collected by sensors are all unlabeled data (excluding the health status of the equipment). These data cannot be directly entered into the deep learning network for learning. How to obtain the health factor (health indicator-HI) diagram of the entire life cycle of railway power supply equipment is a prerequisite for deep learning network training. In addition, the aging and failure process of railway power supply equipment is affected by many factors. At this time, if the equipment failure process is physically modeled through expert experience and manual feature extraction, it will inevitably lead to inefficiency and may not be able to make a comprehensive and accurate description of the aging process of the equipment. How to automatically extract the fault characteristics of railway power supply equipment is also a difficult problem faced by the reliability evaluation of railway power supply equipment. Summary of the invention

[0005] Purpose of the invention: One purpose is to propose a data modeling method based on reliability analysis of railway power supply equipment to solve the above problems existing in the prior art. A further purpose is to propose a system for implementing the above method.

[0006] Technical solution: A data modeling method based on reliability analysis of railway power supply equipment, comprising the following steps:

[0007] Step 1: construct a stacked denoising autoencoder network to perform denoising and cleaning on the original historical data of railway power supply equipment;

[0008] Step 2: Add health tags to the historical data of the railway power supply equipment that has undergone noise reduction and cleaning;

[0009] Step 3: Clean the original historical abnormal data through the stacked denoising autoencoder network and the linear regression of the BP neural network to obtain clean, low-dimensional data with healthy labels;

[0010] Step 4: Input the railway power supply equipment test set into the stacked noise reduction autoencoder network for noise reduction to obtain clean equipment operation status data;

[0011] Step 5: Perform fast Fourier transform on the processed data and input it into the BP neural network to verify the accuracy of the fault characteristics of the railway power supply equipment extracted by the trained convolutional neural network.

[0012] In a further embodiment, a stacked denoising autoencoder network is constructed to perform denoising and cleaning on the original historical data of railway power supply equipment. The historical data information of railway power supply equipment is divided into a training set and a test set, the data of the training set is input into the constructed network for training, the sum of the squares of the difference between the input data and the output result is set as the penalty function, and the weight parameters of the network are updated using the stochastic gradient descent algorithm; the test set data is input into the trained network to verify the denoising effect of the stacked denoising autoencoder network. The above process is repeated until the test accuracy reaches more than 95%. The trained network can realize denoising processing of the status data of railway power supply equipment.

[0013] In a further embodiment, a health label is added to the historical data of the railway power supply equipment after noise reduction and cleaning. The initial 10% of the data is identified as a healthy state, and the health label is set to 1. The last 10% of the data is identified as a failed state, and the health label is set to 0. The labeled data is input into the constructed BP neural network, and the square sum of the difference between the output value of the BP neural network and the label value is used as a penalty function. The weight parameters of the BP neural network are updated using the stochastic gradient descent algorithm. The training process is repeated until the test accuracy reaches more than 95%. Then the remaining 80% of the data is input into the BP neural network, and the result obtained is the health label corresponding to the data. All the data and the corresponding health labels are plotted into a graph, that is, the health factor (HI) graph of the railway power supply equipment is obtained.

[0014] In a further embodiment, the stacked noise reduction autoencoder network cleans the original historical abnormal data and the BP neural network linear regression to obtain clean, low-dimensional data with healthy labels. The cleaned historical data of the railway power supply equipment is divided according to the time length L, and a fast Fourier transform is performed. The obtained spectrum is input into the convolutional neural network for training, the deep fault characteristics of the railway power supply equipment are extracted, and the weight parameters in the convolutional network are assigned.

[0015] In a further embodiment, the railway power supply equipment test set is input into the stacked noise reduction autoencoder network trained in step 1 for noise reduction processing to obtain clean equipment operation status data. The processed data is fast Fourier transformed and input into the convolutional neural network trained in step 3 to verify the accuracy of the railway power supply equipment fault features extracted by the trained convolutional neural network.

[0016] In a further embodiment, a stacked denoising autoencoder network is constructed. For the denoising autoencoder, the historical data of the railway power supply equipment is constructed into a matrix X, X = {x(1), x(2), x(3), ... x(N)}, x(i)∈R M , and add a certain amount of "damage noise" to the matrix X to obtain the noisy data χ, satisfying χ~qD(χ|X), qD is a form of noise distribution, that is, "damage noise" is added according to the qD distribution. Encode χ through the autoencoder to obtain y, hoping that y can approach or reconstruct the original input X. In this process, the joint distribution is defined:

[0017]

[0018] Among them, when f θ When (χ)≠y, Set to 0, q 0(X) is the empirical distribution determined based on N sets of input data. Therefore, y is a deterministic function of χ with noisy inputs, and let θ be the joint distribution q 0 If the deterministic parameters of (X, χ, y) are connected, then θ is the connection parameter between χ and y, and the cost function optimized by the gradient descent method is:

[0019]

[0020] Where L(X, X') is the "reconstruction error" to determine the degree of network reconstruction. The cost function is optimized step by step through the stochastic gradient descent algorithm to maximize the reconstruction of the original input X from the damaged input χ. The specific implementation process of the denoising autoencoder can be described as follows: randomly set some elements of the input matrix X to zero according to a certain ratio to become χ, and then encode χ to obtain the hidden layer expression y, y = f θ (χ)=sigmoid(W χ +b), θ = {W, b}. The hidden layer y reconstructs the input to obtain the output X', X' = g θ′ (y) = sigmoid (W1y + b1), θ′ = {W1, b1}, and then calculate the reconstruction error L(X, X') through the values ​​of X' and X. Finally, the encoder cost function is optimized step by step through the stochastic gradient descent algorithm to maximize the reconstruction of the original input X from the damaged input χ.

[0021] In a further embodiment, a BP neural network is used to perform linear regression on the historical data of the railway power supply equipment after noise reduction to extract a health factor graph of the railway power supply equipment: the historical data of the railway power supply equipment is divided, the initial 10% of the data is identified as a healthy state, the health label is set to 1, and the last 10% of the data is identified as a failure state, and the health label is set to 0; the above data is further divided into a training set and a test set, a BP neural network is constructed, the network learning rate ε is set, and the model connection weight W and the offset b are randomly initialized; the batch training number, the number of iterations, etc. in the forward propagation algorithm are set, the training set data is input into the BP neural network, the forward propagation algorithm is executed, and the cost function is calculated using the output of the BP neural network:

[0022]

[0023] Among them, h w,b (x(i)) is the output of the BP neural network, and y(i) is the health label of the device;

[0024] The stochastic gradient descent algorithm is used to perform back propagation calculations, and the test set data is input into the trained BP neural network to verify the effect of the BP neural network; the remaining 80% of the historical data of the railway power supply equipment is input into the verified BP neural network, and the output value of the BP neural network is the health value corresponding to the railway power supply equipment, and finally the health factor graph of the railway power supply equipment is obtained.

[0025] In a further embodiment, the original historical abnormal data is cleaned by the stacked denoising autoencoder network, and the linear regression of the BP neural network can obtain clean, low-dimensional data with healthy labels. The cleaned historical data of the railway power supply equipment is divided according to the time length L, and the divided data is subjected to fast Fourier transform to obtain a spectrum diagram, which is input into the convolutional neural network for training, extracting the deep fault characteristics of the railway power supply equipment, and assigning the weight parameters in the convolutional neural network. The specific implementation process is as follows:

[0026] Convolutional neural network was built. Considering the complexity of the operating environment of railway power supply equipment and the sample training scale, a stacked convolutional neural network was used, the network learning rate ε was set, and the model connection weights W and offset b were randomly initialized.

[0027] The cleaned data is recorded as X = (X 1 , X 2 , X 3 , …X n ), and divided into test set and training set, the training set data is input into the convolutional neural network to perform the forward propagation algorithm, the convolutional neural network contains convolutional layer and fully connected layer, the calculation of the convolutional layer is as follows:

[0028]

[0029] Where D is the depth, F is the filter size, and w d,m,n represents the weight of the filter in the dth layer, row m, and column n, w b represents the bias term of the filter, a i,j represents the element of the dth layer, row i, and column j of the image, and f is the activation function (Relu). After the convolutional layer is calculated, the fully connected layer is calculated, and the cost function is calculated based on the output of the fully connected layer:

[0030]

[0031] Among them, g w,b (x(i)) is the output of the convolutional neural network, and y(i) is the health label of the device.

[0032] The stochastic gradient descent algorithm is used to perform back propagation calculations and update the weight coefficients of each layer of the convolutional neural network;

[0033] The test set data is input into the trained convolutional neural network to verify the effect of the convolutional neural network training. If the accuracy of the convolutional neural network output result is low, the parameters are readjusted to repeat the above process until the accuracy reaches more than 95%.

[0034] A data modeling system for implementing the above method includes two functional modules, a training functional module and a testing functional module. The training functional module uses a stacked noise reduction encoder to perform noise reduction processing on the historical operation data of railway power supply equipment and extract the main features of the data; the processed data information is input into the BP neural network to extract the health factor (HI) diagram of the railway power supply equipment; the spectrum diagram of the processed data is obtained by fast Fourier transform, combined with the health factor (HI) diagram, and input into the constructed convolutional neural network to automatically extract the deep-level fault characteristics of the railway power supply equipment. The testing functional module inputs the operation status data (test set) of the railway power supply equipment into the trained stacked noise reduction autoencoder to verify the noise reduction effect of the stacked noise reduction autoencoder; extracts the main feature data of the operation status of the railway power supply equipment, performs fast Fourier transform on the main feature data, and then inputs it into the trained convolutional neural network to automatically identify the fault type of the railway power supply equipment. Through the coordinated use of the training functional module and the testing functional module, the deep-level characteristics of the railway power supply equipment fault can be extracted.

[0035] In a further embodiment, the training function module further uses a stacked noise reduction encoder to construct a matrix X of the railway power supply equipment historical data:

[0036] X={x(1),x(2),x(3),…x(N)},x(i)∈R M ;

[0037] A predetermined amount of damage noise is added to the matrix X to obtain noisy data χ, satisfying x~qD(x|X), where qD represents the noise distribution form;

[0038] The autoencoder encodes x to get y. In this process, the joint distribution is defined as:

[0039]

[0040] Among them, when f θ When (χ)≠y, Set to 0, q 0 (X) is the empirical distribution determined based on N sets of input data;

[0041] Therefore, y is a deterministic function of x with noisy input, and let θ be the joint distribution q 0If the deterministic parameters of (X, χ, y) are connected, then θ is the connection parameter between χ and y, and the cost function optimized by the gradient descent method is:

[0042]

[0043] Where L(X, X') represents the reconstruction error, which determines the degree of network reconstruction. The cost function is optimized by the stochastic gradient descent algorithm to maximize the reconstruction of the original input X from the damaged input χ.

[0044] The training function module further performs linear regression on the noise-reduced historical data of the railway power supply equipment through the BP neural network to extract the health factor graph of the railway power supply equipment:

[0045] The historical data of railway power supply equipment is divided into the following categories: the initial 10% of the data is considered healthy, and the health label is set to 1; the last 10% of the data is considered failed, and the health label is set to 0;

[0046] The above data is further divided into training set and test set, BP neural network is constructed, the network learning rate ε is set, and the model connection weight W and offset b are randomly initialized;

[0047] Set the batch training number and iteration number in the forward propagation algorithm, input the training set data into the BP neural network, execute the forward propagation algorithm, and use the output of the BP neural network to calculate the cost function:

[0048]

[0049] Among them, h w,b (x(i)) is the output of the BP neural network, and y(i) is the health label of the device;

[0050] The stochastic gradient descent algorithm is used to perform back propagation calculations, and the test set data is input into the trained BP neural network to verify the effect of the BP neural network;

[0051] The remaining 80% of the historical data of the railway power supply equipment is input into the verified BP neural network. The output value of the BP neural network is the health value corresponding to the railway power supply equipment, and finally the health factor graph of the railway power supply equipment is obtained;

[0052] The test function module further cleans the original historical abnormal data through a stacked denoising autoencoder network and linear regression of a BP neural network to obtain clean, low-dimensional data with healthy labels; divides the cleaned historical data of the railway power supply equipment according to the time length L, performs fast Fourier transform on the divided data to obtain a spectrum, inputs the spectrum into a convolutional neural network for training, extracts deep fault features of the railway power supply equipment, and assigns weight parameters in the convolutional neural network; builds a convolutional neural network, and adopts a stacked convolutional neural network in consideration of the complexity of the operating environment of the railway power supply equipment and the sample training scale, sets the network learning rate ε, and randomly initializes the model connection weights W and the offset b;

[0053] The cleaned data is recorded as X = (X 1 , X 2 , X 3 , …X n ), and divided into test set and training set, the training set data is input into the convolutional neural network to perform the forward propagation algorithm, the convolutional neural network contains convolutional layer and fully connected layer, the calculation of the convolutional layer is as follows:

[0054]

[0055] Where D is the depth, F is the filter size, and w d,m,n represents the weight of the filter in the dth layer, row m, and column n, w b represents the bias term of the filter, a i,j represents the element of the dth layer, row i, and column j of the image, and f is the activation function (Relu);

[0056] After the convolutional layer is calculated, the fully connected layer is calculated, and the cost function is calculated based on the output of the fully connected layer:

[0057]

[0058] Among them, g w,b (x(i)) is the output of the convolutional neural network, and y(i) is the health label of the device;

[0059] The stochastic gradient descent algorithm is used to perform back propagation calculations and update the weight coefficients of each layer of the convolutional neural network;

[0060] The test set data is input into the trained convolutional neural network to verify the effect of the convolutional neural network training. If the accuracy of the convolutional neural network output result is low, the parameters are readjusted to repeat the above process until the accuracy reaches more than 95%.

[0061] The beneficial effects of the present invention are:

[0062] 1. A stacked sparse autoencoder is used to clean the railway power supply equipment data. While retaining the most essential features of the data, it reduces the interference of environmental noise or other coupling components on the data signal and improves the accuracy of fault status identification of railway power supply equipment.

[0063] 2. The BP neural network is used to perform linear regression processing on the historical data of railway power supply equipment, which can directly label the operating status data of the railway power supply equipment with a health status label and obtain the health factor (HI) diagram of the railway power supply equipment. There is no need to use experimental or simulation methods to obtain the label data of the railway power supply equipment, which improves the efficiency of data processing and the authenticity of the data.

[0064] 3. The aging and failure process of railway power supply equipment is a very complex process, which is affected by many factors such as the operating environment and working status. At this time, it is difficult to physically model the failure process of railway power supply equipment by manually extracting features. However, convolutional neural networks can automatically extract deep-level fault features of railway power supply equipment, and can describe the aging and failure process of railway power supply equipment more comprehensively and accurately.

[0065] 4. By inputting the railway power supply equipment test set data into the trained convolutional network, the accuracy of the railway power supply equipment fault characteristics extracted by the trained convolutional neural network can be automatically verified, and the test results can be fed back to the training stage to adjust the training process of the convolutional neural network, and finally a complete railway power supply equipment fault characteristic data set can be constructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 The invention discloses a data modeling method based on reliability analysis of railway power supply equipment.

[0067] Figure 2 This is a flow chart of the stacked denoising autoencoder constructed by the present invention.

[0068] Figure 3 It is a railway power supply equipment health factor (HI) diagram extracted by the present invention through the BP neural network. DETAILED DESCRIPTION

[0069] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known in the art are not described.

[0070] like Figure 1 As shown, the present invention proposes a data modeling method based on railway power supply equipment reliability analysis, comprising the following steps:

[0071] (1) Construct a stacked denoising autoencoder network to denoise and clean the original historical data of railway power supply equipment. The historical data information of railway power supply equipment is divided into a training set and a test set. The data of the training set is input into the constructed network for training. The sum of the squares of the difference between the input data and the output result is set as the penalty function, and the weight parameters of the network are updated using the stochastic gradient descent algorithm; the test set data is input into the trained network to verify the denoising effect of the stacked denoising autoencoder network. The above process is repeated until the test accuracy reaches more than 95%. The trained network can realize denoising processing of railway power supply equipment status data.

[0072] (2) Add health labels to the historical data of railway power supply equipment that has been denoised and cleaned. The initial 10% of the data is identified as healthy, and the health label is set to 1. The last 10% of the data is identified as failed, and the health label is set to 0. The labeled data is input into the constructed BP neural network, and the square sum of the difference between the output value of the BP neural network and the label value is used as the penalty function. The weight parameters of the BP neural network are updated using the stochastic gradient descent algorithm. The training process is repeated until the test accuracy reaches more than 95%. Then the remaining 80% of the data is input into the BP neural network, and the result is the health label corresponding to the data. All the data and the corresponding health labels are plotted into a graph, and the health factor (HI) graph of the railway power supply equipment is obtained.

[0073] (3) Cleaning the original historical abnormal data through the stacked denoising autoencoder network and the linear regression of the BP neural network can obtain clean, low-dimensional data with healthy labels. The cleaned historical data of the railway power supply equipment is divided according to the time length L and subjected to fast Fourier transform. The obtained spectrum graph is input into the convolutional neural network for training, the deep fault characteristics of the railway power supply equipment are extracted, and the weight parameters in the convolutional network are assigned.

[0074] (4) The railway power supply equipment test set is input into the stacked denoising autoencoder network trained in step (1) for denoising to obtain clean equipment operation status data. The processed data is fast Fourier transformed and input into the convolutional neural network trained in step (3) to verify the accuracy of the railway power supply equipment fault features extracted by the trained convolutional neural network.

[0075] Based on the above method, the present invention further proposes a data modeling system, which includes two functional modules, a training functional module and a testing functional module. The training functional module uses a stacked noise reduction encoder to perform noise reduction processing on the historical operation data of the railway power supply equipment and extract the main features of the data; the processed data information is input into the BP neural network to extract the health factor (HI) diagram of the railway power supply equipment; the spectrum diagram of the processed data is obtained by fast Fourier transform, combined with the health factor (HI) diagram, and input into the constructed convolutional neural network to automatically extract the deep-level fault characteristics of the railway power supply equipment. The testing functional module inputs the operating status data (test set) of the railway power supply equipment into the trained stacked noise reduction autoencoder to verify the noise reduction effect of the stacked noise reduction autoencoder; extracts the main feature data of the operating status of the railway power supply equipment, performs fast Fourier transform on the main feature data, and then inputs it into the trained convolutional neural network to automatically identify the fault type of the railway power supply equipment. Through the coordinated use of the training functional module and the testing functional module, the deep-level characteristics of the railway power supply equipment fault can be extracted.

[0076] like Figure 2 As shown in the figure, the constructed stacked denoising autoencoder network. For the denoising autoencoder, the historical data of railway power supply equipment is constructed into a matrix X, X = {x(1), x(2), x(3), ... x(N)}, x(i)∈R M , and add a certain amount of "damage noise" to the matrix X to obtain the noisy data χ, satisfying x~qD(x|X), qD is a form of noise distribution, that is, "damage noise" is added according to the qD distribution. Encode χ through the autoencoder to obtain y, hoping that y can approach or reconstruct the original input X. In this process, the joint distribution is defined:

[0077]

[0078] Among them, when f θ When (χ)≠y, Set to 0, q 0 (X) is the empirical distribution determined based on N sets of input data. Therefore, y is a deterministic function of χ with noisy inputs, and let θ be the joint distribution q 0 If the deterministic parameters of (X, χ, y) are connected, then θ is the connection parameter between χ and y, and the cost function optimized by the gradient descent method is:

[0079]

[0080] Where L(X, X') is the "reconstruction error" to determine the degree of network reconstruction. The cost function is optimized step by step through the stochastic gradient descent algorithm to maximize the reconstruction of the original input X from the damaged input χ. The specific implementation process of the denoising autoencoder can be described as follows: randomly set some elements of the input matrix X to zero according to a certain ratio to become χ, and then encode χ to obtain the hidden layer expression y, y = f θ (χ)=sigmoid(W χ +b), θ = {W, b}. The hidden layer y reconstructs the input to obtain the output X', X' = g θ′ (y) = sigmoid (W1y + b1), θ′ = {W1, b1}, and then calculate the reconstruction error L(X, X') through the values ​​of X' and X. Finally, the encoder cost function is optimized step by step through the stochastic gradient descent algorithm to maximize the reconstruction of the original input X from the damaged input χ.

[0081] Label data is a prerequisite for training deep learning networks, and it is also the key to improving the prediction accuracy of deep learning networks. The historical data of railway power supply equipment usually only contains the operating status of each device, not its corresponding health status. These original unlabeled data cannot be used for training deep learning networks, namely convolutional neural networks. Traditional methods use experiments or simulations to extract label data of each device (i.e., the operating status of the device and its corresponding health status), but the laboratory environment or simulation environment is very different from the actual operating environment of the equipment. The diagnosis accuracy of the deep learning network trained by laboratory or simulation data is low and cannot meet the application requirements.

[0082] like Figure 3 As shown in the figure, the process of obtaining the health factor (HI) diagram of railway power supply equipment. The specific implementation process is as follows:

[0083] 1) The historical data of railway power supply equipment is divided into the following categories: the initial 10% of the data is identified as healthy, and the health label is set to 1; the last 10% of the data is identified as failed, and the health label is set to 0;

[0084] 2) The above data is further divided into training set and test set;

[0085] 2) Construct a BP neural network, set the network learning rate ε, and randomly initialize the model connection weights W and offset b;

[0086] 3) Set the batch training number and iteration number in the forward propagation algorithm, input the training set data into the BP neural network, execute the forward propagation algorithm, and use the output of the BP neural network to calculate the cost function:

[0087]

[0088] Among them, h w,b(x(i)) is the output of the BP neural network, and y(i) is the health label of the equipment.

[0089] 4) Use the stochastic gradient descent algorithm to perform back propagation calculations and update the network weight parameters according to the following formula:

[0090]

[0091]

[0092] 5) Input the test set data into the trained BP neural network to verify the effect of the BP neural network;

[0093] 6) The remaining 80% of the historical data of the railway power supply equipment is input into the verified BP neural network. The output value of the BP neural network is the health value corresponding to the railway power supply equipment, so that the health factor (HI) diagram of the railway power supply equipment is obtained.

[0094] By cleaning the original historical abnormal data through the stacked denoising autoencoder network and the linear regression of the BP neural network, clean, low-dimensional data with healthy labels can be obtained. The cleaned historical data of the railway power supply equipment is divided according to the time length L, and the divided data is subjected to fast Fourier transform to obtain a spectrum graph, which is input into the convolutional neural network for training, extracting the deep fault characteristics of the railway power supply equipment, and assigning the weight parameters in the convolutional neural network. The specific implementation process is as follows.

[0095] 1) Construct a convolutional neural network. Considering the complexity of the operating environment of railway power supply equipment and the sample training scale, a stacked convolutional neural network is used, the network learning rate ε is set, and the model connection weights W and offset b are randomly initialized;

[0096] 2) The cleaned data is recorded as X = (X 1 , X 2 , X 3 , …X n ), and divided into test set and training set, the training set data is input into the convolutional neural network to perform the forward propagation algorithm, the convolutional neural network contains convolutional layer and fully connected layer, the calculation of the convolutional layer is as follows:

[0097]

[0098] Where D is the depth, F is the filter size, and w d,m,n represents the weight of the filter in the dth layer, row m, and column n, w b represents the bias term of the filter, a i,jrepresents the element of the dth layer, row i, and column j of the image, and f is the activation function (Relu). After the convolutional layer is calculated, the fully connected layer is calculated, and the cost function is calculated based on the output of the fully connected layer:

[0099]

[0100] Among them, g w,b (x(i)) is the output of the convolutional neural network, and y(i) is the health label of the device.

[0101] 3) Use the stochastic gradient descent algorithm to perform back propagation calculations and update the weight coefficients of each layer of the convolutional neural network;

[0102] 4) Input the test set data into the trained convolutional neural network to verify the effect of the convolutional neural network training. If the accuracy of the convolutional neural network output result is low, readjust the parameters and perform the above process until the accuracy reaches more than 95%.

[0103] Through the above process, the training phase and the testing phase of the algorithm have been completed. Through the coordinated use of the two phases, a fault feature set of historical data of railway power supply equipment is established, which is of great significance for the subsequent online monitoring of railway power supply equipment faults.

[0104] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A data modeling method based on reliability analysis of railway power supply equipment, characterized by The following steps are involved: Step 1: construct a stacked denoising autoencoder network to perform denoising and cleaning on the original historical data of railway power supply equipment; Step 1-1, divide the historical data information of railway power supply equipment into a training set and a test set, input the data of the training set into the constructed network for training, set the sum of the squares of the difference between the input data and the output result as the penalty function, and use the stochastic gradient descent algorithm to update the weight parameters of the network; Step 1-2: Input the test set data into the trained network to verify the denoising effect of the stacked denoising autoencoder network; Step 1-2a, the stacked noise reduction encoder constructs the matrix X with the historical data of the railway power supply equipment: X={x(1),x(2),x(3),…x(N)},x(i)∈R M ; And add a predetermined amount of damage noise to the matrix X to obtain noisy data χ, satisfying , where qD represents the noise distribution form; Step 1-2b, encode x through the autoencoder to get y. In this process, define the joint distribution: ; Among them, when hour, Set to 0, q 0 (X) is the empirical distribution determined based on N sets of input data; Therefore, y is a deterministic function of x with noisy input, and let θ be the joint distribution q 0 (X, χ, y), then θ is the connection parameter between χ and y, and the cost function optimized by the gradient descent method is: ; Where L(X,X') represents the reconstruction error, which determines the degree of network reconstruction. The cost function is optimized by the stochastic gradient descent algorithm to maximize the reconstruction of the original input X from the damaged input χ. Step 1-3: Repeat step 1 to step 2 until the test accuracy reaches more than 95%; Step 2: Add health tags to the historical data of the railway power supply equipment that has undergone noise reduction and cleaning; Step 2-1: The initial 10% of data is identified as healthy, and the health label is set to 1. The last 10% of data is identified as invalid, and the health label is set to 0. Step 2-2: Input the labeled data into the constructed BP neural network, and use the sum of the squares of the difference between the output value of the BP neural network and the label value as the penalty function. Use the stochastic gradient descent algorithm to update the weight parameters of the BP neural network. Repeat the training process until the test accuracy reaches more than 95%; Step 2-3: Input the remaining 80% of the data into the BP neural network, and the result is the health label corresponding to the data; perform linear regression on the noise-reduced historical data of the railway power supply equipment through the BP neural network to extract the health factor graph of the railway power supply equipment: Step 2-3a: divide the historical data of railway power supply equipment, the initial 10% of the data is identified as healthy, the health label is set to 1, and the last 10% of the data is identified as failed, the health label is set to 0; Step 2-3b, further divide the above data into training set and test set, construct BP neural network, set network learning rate ε, randomly initialize model connection weight W and offset b; Step 2-3c, set the batch training number and iteration number in the forward propagation algorithm, input the training set data into the BP neural network, execute the forward propagation algorithm, and use the output of the BP neural network to calculate the cost function: ; Among them, h w,b (x(i)) is the output of the BP neural network, and y(i) is the health label of the device; Step 2-3d, use the stochastic gradient descent algorithm to perform back propagation calculation, input the test set data into the trained BP neural network, and verify the effect of the BP neural network; Step 2-3e, input the remaining 80% of the historical data of the railway power supply equipment into the verified BP neural network, the output value of the BP neural network is the health value corresponding to the railway power supply equipment, and finally obtain the health factor graph of the railway power supply equipment; Step 3: Clean the original historical abnormal data through the stacked denoising autoencoder network and the linear regression of the BP neural network to obtain clean, low-dimensional data with healthy labels; Step 4: Input the railway power supply equipment test set into the stacked noise reduction autoencoder network for noise reduction to obtain clean equipment operation status data; Step 5: Perform fast Fourier transform on the processed data and input it into the BP neural network to verify the accuracy of the fault characteristics of the railway power supply equipment extracted by the trained convolutional neural network.

2. A data modeling method based on railway power supply equipment reliability analysis according to claim 1, characterized in that: The step 3 further comprises: dividing the cleaned historical data of the railway power supply equipment according to the time length L, and performing a fast Fourier transform, inputting the obtained spectrum graph into a convolutional neural network for training, extracting the deep-level fault characteristics of the railway power supply equipment, and assigning weight parameters in the convolutional network.

3. The data modeling method based on reliability analysis of railway power supply equipment according to claim 1 is characterized in that: The step 4 is further as follows: Step 4-1, input the railway power supply equipment test set into the stacked noise reduction autoencoder network trained in step 1-1 to perform noise reduction processing to obtain clean equipment operation status data; Step 4-2: Perform a fast Fourier transform on the processed data and input it into the convolutional neural network trained in step 3 to verify the accuracy of the fault characteristics of the railway power supply equipment extracted by the trained convolutional neural network.

4. The data modeling method based on reliability analysis of railway power supply equipment according to claim 1 is characterized in that: Step 1-1 further includes: Step 1-1a, build a convolutional neural network. Considering the complexity of the operating environment of railway power supply equipment and the sample training scale, a stacked convolutional neural network is used, the network learning rate ε is set, and the model connection weight W and offset b are randomly initialized; Step 1-1b: record the cleaned data as X=(X 1 , X 2 , X 3 , …X n ), and divided into test set and training set, the training set data is input into the convolutional neural network to execute the forward propagation algorithm. The convolutional neural network contains convolutional layer and fully connected layer. The calculation of convolutional layer is as follows: ; Where D is the depth, F is the filter size, and w d,m,n represents the weight of the filter in the dth layer, row m, column n, w b represents the bias term of the filter, a i,j represents the element of the dth layer, row i, and column j of the image, and f is the activation function (Relu); Step 1-1c, after the convolutional layer is calculated, the fully connected layer is calculated, and the cost function is calculated based on the output of the fully connected layer: ; Among them, g w,b (x(i)) is the output of the convolutional neural network, and y(i) is the health label of the device; Step 1-1d, use the stochastic gradient descent algorithm to perform back propagation calculations to update the weight coefficients of each layer of the convolutional neural network; Step 1-1e: Input the test set data into the trained convolutional neural network to verify the effect of the convolutional neural network training. If the accuracy of the convolutional neural network output result is low, readjust the parameters and repeat the above process until the accuracy reaches more than 95%.

5. A data modeling system for implementing the method described in any one of claims 1 to 4, characterized in that: It includes a training function module and a test function module. The training function module uses a stacked noise reduction encoder to perform noise reduction processing on the historical operation data of the railway power supply equipment to extract the main features of the data; The processed data information is input into the BP neural network to extract the health factor graph of the railway power supply equipment; the spectrum of the processed data is obtained by fast Fourier transform, combined with the health factor graph, and input into the constructed convolutional neural network to automatically extract the deep fault characteristics of the railway power supply equipment; The test function module inputs the operating status data of the railway power supply equipment into the trained stacked denoising autoencoder to verify the denoising effect of the stacked denoising autoencoder; extracts the main feature data of the operating status of the railway power supply equipment, performs fast Fourier transform on the main feature data, and inputs it into the trained convolutional neural network to automatically identify the fault type of the railway power supply equipment.

6. The data modeling system according to claim 5, characterized in that: The training function module further uses a stacked noise reduction encoder to construct a matrix X from the historical data of the railway power supply equipment: X={x(1),x(2),x(3),…x(N)},x(i)∈R M ; And add a predetermined amount of damage noise to the matrix X to obtain noisy data χ, satisfying , where qD represents the noise distribution form; The autoencoder encodes x to get y. In this process, the joint distribution is defined as: ; Among them, when hour, Set to 0, q 0 (X) is the empirical distribution determined based on N sets of input data; Therefore, y is a deterministic function of x with noisy input, and let θ be the joint distribution q 0 (X, χ, y), then θ is the connection parameter between χ and y, and the cost function optimized by the gradient descent method is: ; Where L(X,X') represents the reconstruction error, which determines the degree of network reconstruction. The cost function is optimized by the stochastic gradient descent algorithm to maximize the reconstruction of the original input X from the damaged input χ. The training function module further performs linear regression on the noise-reduced historical data of the railway power supply equipment through the BP neural network to extract the health factor graph of the railway power supply equipment: The historical data of railway power supply equipment is divided into the following categories: the initial 10% of the data is considered healthy, with the health label set to 1, and the last 10% of the data is considered failed, with the health label set to 0. The above data is further divided into training set and test set, BP neural network is constructed, the network learning rate ε is set, and the model connection weight W and offset b are randomly initialized; Set the batch training number and iteration number in the forward propagation algorithm, input the training set data into the BP neural network, execute the forward propagation algorithm, and use the output of the BP neural network to calculate the cost function: ; Among them, h w,b (x(i)) is the output of the BP neural network, and y(i) is the health label of the device; The stochastic gradient descent algorithm is used to perform back propagation calculations, and the test set data is input into the trained BP neural network to verify the effect of the BP neural network; The remaining 80% of the historical data of the railway power supply equipment is input into the verified BP neural network. The output value of the BP neural network is the health value corresponding to the railway power supply equipment, and finally the health factor graph of the railway power supply equipment is obtained; The test function module further cleans the original historical abnormal data through a stacked denoising autoencoder network and linear regression of a BP neural network to obtain clean, low-dimensional data with healthy labels; divides the cleaned historical data of the railway power supply equipment according to the time length L, performs fast Fourier transform on the divided data to obtain a spectrum, inputs the spectrum into a convolutional neural network for training, extracts deep fault features of the railway power supply equipment, and assigns weight parameters in the convolutional neural network; builds a convolutional neural network, and adopts a stacked convolutional neural network in consideration of the complexity of the operating environment of the railway power supply equipment and the sample training scale, sets the network learning rate ε, and randomly initializes the model connection weights W and the offset b; The cleaned data is recorded as X=(X 1 , X 2 , X 3 , …X n ), and divided into test set and training set, the training set data is input into the convolutional neural network to execute the forward propagation algorithm. The convolutional neural network contains convolutional layer and fully connected layer. The calculation of convolutional layer is as follows: ; Where D is the depth, F is the filter size, and w d,m,n represents the weight of the filter in the dth layer, row m, column n, w b represents the bias term of the filter, a i,j represents the element of the dth layer, row i, and column j of the image, and f is the activation function (Relu); After the convolutional layer is calculated, the fully connected layer is calculated, and the cost function is calculated based on the output of the fully connected layer: ; Among them, g w,b (x(i)) is the output of the convolutional neural network, and y(i) is the health label of the device; The stochastic gradient descent algorithm is used to perform back propagation calculations and update the weight coefficients of each layer of the convolutional neural network; Input the test set data into the trained convolutional neural network to verify the effect of the convolutional neural network training. If the accuracy of the convolutional neural network output result is low, readjust the parameters and repeat the above process until the accuracy reaches more than 95%.