Rail transit DC switch trip fault diagnosis method
Through the combination of wavelet packet transformation and convolutional neural network, the problems of strong subjectivity and insufficient high-frequency signal decomposition in fault diagnosis of rail transit DC switches are solved, and high-accuracy fault identification and diagnosis are achieved, making full use of the signal time Frequency characteristics.
Patent Information
- Application Number
- CN202411146394.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing technology has problems such as strong subjectivity, lack of quantization standards and difficulty in making full use of historical data in the fault diagnosis of rail transit DC switches. In addition, wavelet transformation does not decompose high-frequency signals, and convolutional neural networks are easily introduced in fault recording processing. Artificial error.
Wavelet packet transformation is used to perform multi-layer signal decomposition, calculate the energy spectrum distribution of wavelet packets, establish an energy feature matrix, and combine it with convolutional neural network for deep feature self-extraction and fault identification to reduce feature losses and errors in human operations.
The accurate diagnosis of rail transit DC switch faults has been achieved, the recognition accuracy rate has reached 98.74%, and the time-frequency localization analysis capability and model generalization capability of fault diagnosis have been improved.
Smart Images

Figure CN119025903A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of rail transit, and in particular relates to a rail transit DC switch tripping fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network. Background Art
[0002] Fault recording is a dispatching-end power grid fault diagnosis system based on fault recording information. When a fault occurs in the power system, the fault recorder can automatically and accurately record the changes in various electrical quantities before and after the fault. The changes in these electrical quantities play an important role in analyzing and handling accidents, judging whether the protection is operating correctly, and improving the safe operation level of the power system.
[0003] As the core system to ensure the normal operation of rail transit, the traction DC power supply system involves many key devices. Once a fault occurs, on-site maintenance personnel need to manually read the fault recording data and analyze and diagnose based on experience. This method has problems such as strong subjectivity, lack of quantitative standards, and difficulty in making full use of historical data.
[0004] The wavelet transform only further decomposes the low-frequency part of the signal, but no longer decomposes the high-frequency part, which is the detail part of the signal. Therefore, the wavelet transform can well characterize a large class of signals with low-frequency information as the main component, but cannot well decompose and represent signals containing a large amount of detail information (fine edges or textures).
[0005] At the same time, convolutional neural network (CNN) is a deep learning model that is widely used in image processing, speech recognition, natural language processing and other fields. However, according to the characteristics of fault recording, there are still problems of feature loss or human error caused by human operation. Summary of the invention
[0006] To achieve the above object, the present invention provides a rail transit DC switch tripping fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network.
[0007] Wavelet packet transform can decompose both the low-frequency part of the signal and the high-frequency part, and this decomposition has neither redundancy nor omission, so it can perform better time-frequency localization analysis on signals containing a large amount of medium and high-frequency information.
[0008] The present invention provides a rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network, comprising the following steps:
[0009] S1, extraction and classification of fault recording data;
[0010] S2, using wavelet packet transform to decompose the fault signal into multiple layers;
[0011] S3, calculate the wavelet packet energy, obtain the wavelet packet energy spectrum distribution when the equipment fails, and establish the energy characteristic matrix;
[0012] S4, deep feature extraction is performed on the input samples to establish a CNN fault recognition model;
[0013] S5, judging the fault type according to the obtained characteristic representation and performing fault diagnosis.
[0014] Step S1: Extraction and classification of fault recording data;
[0015] Preprocess the recorded DC switch action event fault recording data, use the same sampling frequency and signal length, and ensure that the number of samples of each type is as similar as possible. The sampling frequency is preferably 1kHz, and the signal length is preferably 4096.
[0016] The training data set and the test data set are classified and labeled according to the fault type, which includes short circuits in DC power supply switchgear, short circuits in overhead line equipment, and short circuit faults of downstream loads.
[0017] The DC switch integrated protection system can adopt a multi-CPU structure to realize DC switch protection, monitoring, control, measurement, communication and other functions, such as Sitras PRO SW.
[0018] Step S2: Package
[0019] The wavelet basis function suitable for DC system fault feature extraction is used to decompose the DC switch historical tripping fault recording obtained by the DC switch integrated protection system for multiple times, and only the wavelet packet coefficients of the first multiple frequency bands are taken.
[0020] The decomposition and reconstruction of wavelet packets uses the db6 (Daubechies 6) wavelet basis function suitable for DC system fault feature extraction. The DC switch historical tripping fault recording obtained by the DC switch integrated protection system is decomposed into several frequency segments at different scales, namely Decomposed and Wavelet packet decomposition can fully extract the detail signal of the high-frequency part and improve the resolution of the high-frequency band. Its specific expression is as follows:
[0021]
[0022] The wavelet packet reconstruction object is the relatively important frequency band in the decomposition sequence, so the reconstructed signal frequency band is simpler than the original signal, but the signal length does not change. and Back to
[0023]
[0024] From the wavelet packet operation process, we can know that the wavelet packet technology can adapt to the signal characteristics and obtain the frequency band that matches its spectrum distribution, thereby realizing the effective extraction of features.
[0025] The DC switch tripping historical recording data is processed in a diversity manner. The fault recording historical data obtained from the DC switch integrated protection system is processed in a diversity manner, divided into training set data and test set data according to a specific ratio and input into the fault identification network.
[0026] Taking the short-circuit fault of overhead line equipment as an example, the DC switch historical trip fault recording obtained by the DC switch integrated protection system is decomposed 8 times, and the wavelet packet coefficients of the first 16 frequency bands are taken. Other fault types have similar situations, but the specific decomposition times and the wavelet packet coefficients of the first frequency band are different.
[0027] Step S3 includes:
[0028] According to the results obtained in S2, the wavelet packet energy is calculated to obtain the wavelet packet energy spectrum distribution when the overhead line equipment is short-circuited, thereby establishing the energy characteristic matrix.
[0029] Calculate the wavelet packet energy and obtain the last layer of the decomposition tree structure 2 N The wavelet packet energy E1, E2, ..., E j , j represents the number of frequency bands contained in the last layer of wavelet packet decomposition, j = 1, 2, 3, .., 2 N . Its wavelet packet energy can be expressed as:
[0030]
[0031] Where: x j represents the discrete data obtained by batch acquisition; n represents the number of sampling points in the batch acquisition process of input data; S j (t) represents the original input electrical quantity data of the DC switch tripping fault; E j Represents the wavelet packet energy of the jth node in the last layer of wavelet packet decomposition.
[0032] Construct an energy characteristic matrix. Since the energy corresponding to each frequency band will change differently when a DC switch trips, the nature of the fault can be represented by constructing an energy characteristic vector T, which is expressed as follows:
[0033]
[0034]
[0035] Where: E represents the energy value of the jth frequency band in the last layer of wavelet packet reconstruction; E j,min 、E j,max Respectively represent the minimum and maximum energy in each frequency band of the last layer; E j,m represents the normalized processing result of the energy value of the jth frequency band in the last layer; T represents the energy feature vector of a single sample.
[0036] Furthermore, in order to avoid feature loss or human error caused by human operation during feature extraction, step S4 includes:
[0037] CNN is used to extract deep features from input samples. The specific steps are as follows:
[0038] 1) Use one-hot encoding to label samples;
[0039] 2) Input samples are self-extracted through two convolution operations;
[0040] 3) In the second convolutional layer, random dropout technique is used to extract some more prominent features;
[0041] 4) Use Sigmoid function as the output layer activation function to achieve classification;
[0042] 5) Use the cross entropy loss function to evaluate the difference between the predicted value and the actual value;
[0043] 6) Select Adam optimizer to optimize the model.
[0044] Step S4 includes using one-hot encoding to uniformly label the fault-free samples and samples under the three fault conditions. The fault-free samples are represented by 0, and the three different types of fault samples are represented by 1 to 3. The codes of each type of samples after one-hot encoding are shown in Table 1.
[0045] Table 1: Data encoding
[0046]
[0047] The input sample is the energy feature matrix T, and the features are extracted by self-extraction through two convolution operations. The convolution kernel size in the convolution layer is 3×3, the first convolution layer has 32 convolution kernels, and the second convolution layer has 64 convolution kernels. The convolution layer activation function selects Relu. The Relu function is widely used due to its advantages such as simple calculation, fast convergence speed and no gradient disappearance problem. Its expression is:
[0048] f(x)=max(0,x)
[0049] To optimize the structure of the convolutional neural network (CNN), the pooling layer is replaced and the random dropout technique is introduced in the second convolutional layer to improve the generalization ability and performance of the model. The pooling layer is replaced by the convolutional layer, which helps to retain more feature information and reduce computational complexity without significantly affecting the accuracy of the model.
[0050] The random dropout technique is applied in the second convolutional layer, and the dropout rate is set to 50% to extract some more prominent features and improve the model performance. Finally, the extracted feature map is converted into classification results through the fully connected layer and the output layer. Fault state recognition can be regarded as a classification problem. All samples are divided into 4 categories as described in the table. Therefore, the number of neurons in the output layer is 4. The activation function uses the Sigmoid function to achieve classification. The expression of the Sigmoid function is:
[0051]
[0052] When the convolutional neural network classifies the fault state, the loss function is used to measure the inconsistency between the predicted value and the label value. During the training process, the loss function value is expected to be as small as possible, while the judgment accuracy is expected to be as large as possible. The cross entropy loss function can better balance the above two points, so that the model has better classification performance. Therefore, the present invention chooses to use the cross entropy loss function to evaluate the difference between the predicted value and the actual value. The cross entropy loss function can be expressed as:
[0053]
[0054] Where n is the number of samples, y is the sample label value, Predict values for the convolutional neural network.
[0055] Choosing Adam to optimize the convolutional neural network model can improve model performance. Adam combines the advantages of both AdaGrad and RMSProp optimizers, has higher computational efficiency, and can achieve better performance by adjusting a small number of hyperparameters.
[0056] Step S5, using the data extracted in S1 to form a fault recording database, and then forming an energy feature matrix through wavelet packet decomposition and wavelet packet energy spectrum distribution calculation, which is divided into training set data and test set data in a 7:3 ratio. The training set model is used for training, and then the test set network is input into the trained model for prediction accuracy test.
[0057] After calculation, the recognition accuracy of the model reached 98.74%, and the recognition time was 0.12ms. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:
[0059] Figure 1 is a flow chart of the present invention;
[0060] Figure 2 It is a schematic diagram of wavelet packet decomposition of the present invention;
[0061] Figure 3 It is the 8-layer wavelet packet decomposition diagram of the short circuit of the contact network equipment of the present invention;
[0062] Figure 4 It is a schematic diagram of the energy spectrum distribution of the short-circuit wavelet packet of the overhead contact network equipment of the present invention;
[0063] Figure 5 It is a schematic diagram of the CNN fault identification structure of the present invention. DETAILED DESCRIPTION
[0064] In order to make the purpose, technical solution and advantages of the implementation of the present invention clearer, the technical solution in the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings in the embodiment of the present invention.
[0065] The present invention provides a technical solution: a rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network, including extraction and classification of fault recording data, multi-layer decomposition of fault signals, establishment of energy feature matrix, establishment of fault identification model, and fault diagnosis. The rail transit DC switch trip fault diagnosis method based on waveform feature analysis includes the following steps:
[0066] S1, extraction and classification of fault recording data;
[0067] S2, using wavelet packet transform to decompose the fault signal into multiple layers;
[0068] S3, calculate the wavelet packet energy, obtain the wavelet packet energy spectrum distribution when the equipment fails, and establish the energy characteristic matrix;
[0069] S4, deep feature extraction is performed on the input samples to establish a CNN fault recognition model;
[0070] S5, judging the fault type according to the obtained characteristic representation and performing fault diagnosis.
[0071] Wherein, step S1 comprises:
[0072] The DC switch integrated protection system is designed for DC power supply systems of urban rail and trunk railways to protect DC switchgear and overhead line systems from extreme operating conditions. The system can detect the occurrence of a short circuit before the current rises to the maximum short-circuit current and record the fault measurement value in the event record storage area.
[0073] Since the DC switch integrated protection system only stores the recorded data of faults and lacks analysis and classification of the causes of faults, it is necessary to extract the recorded data of DC switch action events from the DC switch integrated protection system for preprocessing, unify the sampling frequency (generally 1kHz) and signal length (generally 4096), and ensure that the number of samples of each type is as similar as possible. The training data set and the test data set are labeled according to the three short-circuit fault types: 1) short circuit occurring in the DC power supply switch equipment; 2) short circuit in the contact network equipment; and 3) short circuit of the lower load.
[0074] Step S2 includes:
[0075] The db6 wavelet basis function suitable for DC system fault feature extraction is used to decompose the DC switch historical trip fault recording obtained by the DC switch integrated protection system into several frequency bands at different scales. Specifically, the wavelet packet decomposition filters the input waveform through high-pass filter and low-pass filter to obtain approximate components and detail components, and finally converts the original signal into 2 j In the (j is the number of decomposition layers) frequency bands, wavelet packets are constructed by the double-scale equation, which can be written as:
[0076]
[0077] where h k , g k Represent the low-pass and high-pass filter coefficients respectively; k is the decomposition node number.
[0078] Let u0t=Φ(t), u1t=ψ(t), we can get the following recursive relationship:
[0079]
[0080] Wavelet packet decomposition can fully extract the detail signal of the high-frequency part and improve the resolution of the high-frequency band. The calculation expression of the wavelet packet coefficient of the k-th frequency band of the j-th layer is as follows:
[0081]
[0082] By comparing the experimental results, the input signal is decomposed into 8 layers of wavelet packets. The wavelet packet decomposition diagram is shown in Figure 2As shown in the figure, S is the input fault recording data. Taking the overhead short circuit fault as an example, the first 16 frequency bands of the 8-layer wavelet packet are selected, and the original waveform and the wavelet packet coefficients decomposed into each frequency band are processed by wavelet packet decomposition as shown in the figure. Figure 3 As shown in the figure, it can be seen that after the fault signal is decomposed, fault signal components of different scales are obtained. Each component corresponds to a different scale, which has more sufficient information than the original signal. Among them, node 8.0 is the low-frequency component of the signal, and nodes 8.1-8.15 are the high-frequency components obtained by decomposing other low-frequency components of the signal. The decomposition results are used to calculate the energy distribution of each frequency band.
[0083] Among them, step S3 includes when a short circuit fault occurs in the primary circuit of the DC system, the DC switch trips, and the short circuit energy will change significantly due to the impact of the short circuit current, and the energy characteristics of each frequency band will also be different. Therefore, the nature of the fault can be identified by the energy distribution characteristics of each frequency band.
[0084] Each frequency band in the energy spectrum is orthogonal and independent. At the same time, the total energy of all frequency band signals is constant, satisfying the law of conservation of energy.
[0085] Assume that the discrete signal x k,m The sampling length of (i) is N, and its energy can be expressed as:
[0086]
[0087] Where: k is the number of decompositions; m is the position number of the decomposition frequency band (m = 0, 1, ..., 2 k -1).
[0088] According to the law of conservation of energy, there is the following relationship:
[0089]
[0090] The energy percentage of the mth frequency band signal and the sum of the energy of each frequency band is 1, which has the following relationship:
[0091]
[0092]
[0093] Since the DC switch trip fault signal is divided into multiple frequency bands by wavelet packets, and the energy distribution of each frequency band can effectively reflect the short-circuit fault of the DC primary circuit, different fault types can be effectively distinguished by the energy spectrum. N The wavelet packet energy E1, E2, ..., E j , j represents the number of frequency bands contained in the last layer of wavelet packet decomposition, j = 1, 2, 3, .., 2N . Its wavelet packet energy can be expressed as:
[0094]
[0095] Where: x j represents the discrete data obtained by batch collection; n represents the number of sampling points of the data; S j (t) represents the original input electrical quantity data of the DC switch tripping fault; E j Represents the wavelet packet energy of the jth node in the last layer of wavelet packet decomposition.
[0096] Taking the overhead line short circuit fault as an example, the energy spectrum distribution after 8-layer wavelet packet decomposition is as follows: Figure 4 As shown in the figure, since the energy percentage of the 17th frequency band is 0, we only need to consider the energy distribution of the first 16 frequency bands.
[0097] Since the energy corresponding to each frequency band will vary differently due to different fault types when the DC switch trips, the energy characteristic vector T is constructed to represent the fault nature, and its expression is as follows:
[0098]
[0099]
[0100] Where: E represents the energy value of the jth frequency band in the last layer of wavelet packet reconstruction; E j,min 、E j,max Respectively represent the minimum and maximum energy in each frequency band of the last layer; E j,m represents the normalized processing result of the energy value of the jth frequency band in the last layer; T represents the energy feature vector of a single sample.
[0101] Among them, step S4 includes using one-hot encoding to uniformly label the fault-free samples and samples under the three fault conditions. The fault-free samples are represented by 0, and the three different types of fault samples are represented by 1 to 3. The codes of each type of samples after one-hot encoding are shown in Table 1.
[0102] Table 1: Data encoding
[0103]
[0104] The input sample is the energy feature matrix T, and the features are extracted by self-extraction through two convolution operations. The convolution kernel size in the convolution layer is 3×3, the first convolution layer has 32 convolution kernels, and the second convolution layer has 64 convolution kernels. The convolution layer activation function selects Relu. The Relu function is widely used due to its advantages such as simple calculation, fast convergence speed and no gradient disappearance problem. Its expression is:
[0105] f(x)=max(0,x)
[0106] In order to optimize the structure of the convolutional neural network (CNN), the pooling layer is replaced and the random dropout technology is introduced in the second convolution layer to improve the generalization ability and performance of the model. The pooling layer is replaced by the convolution layer, which helps to retain more feature information and reduce the computational complexity without significantly affecting the accuracy of the model. The random dropout technology is applied in the second convolution layer, and the dropout rate is set to 50% to extract some more prominent features and improve the performance of the model. Finally, the extracted feature map is converted into a classification result through the fully connected layer and the output layer. Fault state recognition can be regarded as a classification problem. All samples are divided into 4 categories as described in the table. Therefore, the number of neurons in the output layer is 4, and the activation function uses the Sigmoid function to implement classification. The expression of the Sigmoid function is:
[0107]
[0108] When the convolutional neural network classifies the fault state, the loss function is used to measure the inconsistency between the predicted value and the label value. During the training process, the loss function value is expected to be as small as possible, while the judgment accuracy is expected to be as large as possible. The cross entropy loss function can better balance the above two points, so that the model has better classification performance. Therefore, the present invention chooses to use the cross entropy loss function to evaluate the difference between the predicted value and the actual value. The cross entropy loss function can be expressed as:
[0109]
[0110] Where n is the number of samples, y is the sample label value, Predict values for the convolutional neural network.
[0111] Choosing Adam to optimize the convolutional neural network model can improve model performance. Adam combines the advantages of both AdaGrad and RMSProp optimizers, has higher computational efficiency, and can achieve better performance by adjusting a small number of hyperparameters.
[0112] Step S5 is characterized in that: a fault recording database is formed through the data extracted by S1, an energy feature matrix is formed through wavelet packet decomposition and wavelet packet energy spectrum distribution calculation, and the data is divided into training set data and test set data in a ratio of 7:3. The training set model is used for training, and then the test set network is input into the trained model for prediction accuracy test.
[0113] According to calculation, the recognition accuracy of the model reaches 98.74%, and the recognition time is 0.12ms. The present invention makes full use of the time-frequency characteristics of the signal to improve the diagnosis accuracy.
[0114] The above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. It is only a further detailed description of the purpose, technical scheme and beneficial effects of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network, characterized in that: The following steps are involved: S1, extraction and classification of fault recording data; S2, using wavelet packet transform to decompose the fault signal into multiple layers; S3, calculate the wavelet packet energy, obtain the wavelet packet energy spectrum distribution when the equipment fails, and establish the energy characteristic matrix; S4, deep feature extraction is performed on the input samples to establish a CNN fault recognition model; S5, judging the fault type according to the obtained characteristic representation and performing fault diagnosis.
2. According to claim 1, a rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network is characterized in that: The step S1 comprises: The fault recording data is preprocessed, using the same sampling frequency and signal length, and ensuring that the number of samples of each type is as similar as possible.
3. According to claim 1, a rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network is characterized in that: The step S1 comprises: The training data set and the test data set are classified and labeled according to the fault type, and the fault type includes a short circuit in a DC power supply switch device, a short circuit in an overhead line device, and a short circuit fault of a lower load.
4. The rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network according to claim 1 is characterized in that: The step S2 comprises: The db6 wavelet basis function suitable for DC system fault feature extraction is used to decompose the DC switch historical tripping fault recording into several frequency segments at different scales.
5. The rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network according to claim 1 is characterized in that: The step S3 comprises: The wavelet packet energy is calculated to obtain the wavelet packet energy spectrum distribution, and only the wavelet packet coefficients of the first several frequency bands are taken.
6. A rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network according to claim 1, characterized in that: The step S4 comprises: CNN is used to extract deep features from input samples.
7. A rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network according to claim 6, characterized in that: The deep feature self-extraction includes: Use one-hot encoding to label samples; The input sample extracts features through two convolution operations; The random dropout technique is used to extract some more prominent features; The Sigmoid function is used as the output layer activation function to achieve classification; The cross entropy loss function is used to evaluate the difference between the predicted value and the actual value; Select Adam optimizer to optimize the model.
8. The rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network according to claim 7 is characterized in that: The convolution kernel size in the convolution layer is 3×3, the first convolution layer has 32 convolution kernels, and the second convolution layer has 64 convolution kernels.
9. The rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network according to claim 7, characterized in that: The random dropout technique is introduced in the second convolutional layer, and the dropout rate is set to 50%.
10. The rail transit DC switch trip fault diagnosis method based on wavelet packet energy spectrum and convolutional neural network according to claim 7, characterized in that: The data is divided into training set data and test set data in a ratio of 7:
3. The training set model is used for training, and then the test set network is input into the trained model for prediction accuracy test.
Citation Information
Cited By
Fault diagnosis method and system for power transmission line
CN119646625A
