A fault diagnosis method for switch machine based on segmented prior feature fusion

Through a method based on segmented prior feature fusion, combined with integrated learning and GRU-TCN model, feature extraction and fault identification of the current time series of the rail transit transfer machine is solved, and the accuracy of fault diagnosis of the existing technology transfer machine is solved. and automation problems, achieving more efficient troubleshooting.

CN118568599BActive Publication Date: 2024-11-29NANJING NRIET IND CORP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411045964.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-29
Estimated Expiration
2044-08-01

AI Technical Summary

Technical Problem

The existing technology relies on traditional indicators in the fault diagnosis of transfer machines in rail transit, lacks effective automated diagnosis methods, and it is difficult to accurately identify the fault curve, resulting in insufficient diagnostic efficiency and accuracy.

Method used

The fault diagnosis method of the switch machine based on segmented prior feature fusion is adopted. Through the feature extraction and integrated learning model of the current time series, combined with the attention mechanism and the GRU-TCN model, the current curve is identified and classified to achieve automatic faults. diagnosis.

Benefits of technology

It improves the accuracy and automation of fault diagnosis of switch machine, can effectively identify fault curves, and improves diagnostic efficiency and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118568599B_ABST
    Figure CN118568599B_ABST
Patent Text Reader

Abstract

The present invention discloses a fault diagnosis method for a switch machine based on segmented prior feature fusion, comprising the following steps: collecting current time series data collected by a maintenance support system (MSS) sensor when the switch machine is in action; extracting features from the current time series, including manual feature engineering and automatic feature extraction. Manual feature engineering classifies the faults of the switch machine according to the causes by analyzing the curve features. Automatic feature extraction uses a denoising autoencoder to detect the unsupervised similarity between the sample to be classified and the ideal curve; uses an attention mechanism to improve the three-stage segmentation algorithm and extract features segment by segment; uses multiple sub-models to identify the current curve during the work of the turnout; and uses the late-fusion method in ensemble learning to fuse the classification results of multiple sub-models. Using the trained model, input data and output a fault diagnosis scheme for railway switch machines.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent operation and maintenance technology, and in particular to a switch machine fault diagnosis method based on segmented priori feature fusion. Background Art

[0002] In recent years, machine learning technologies represented by deep learning have made breakthrough progress, and they have achieved good results in the fields of image recognition (IR), speech recognition (SR), natural language processing (NLP), etc. The application of deep learning in the field of intelligent transportation has also become a research hotspot, such as traffic light recognition, pedestrian trajectory prediction, lane line recognition, traffic sign recognition, etc., but its application in rail transit is still relatively small. Modern rail transit has installed a maintenance support system (MSS) for key equipment. The system will collect a large amount of sensor time series data for the switch machine when it is in action, and this data can be used as a data set for training neural networks (NN). At present, the monitoring of switch machines on the market mainly focuses on indicators such as voltage, current, power, conversion force, oil pressure and acoustic characteristics. Summary of the invention

[0003] In view of the problems existing in the prior art, the present invention provides a switch machine fault diagnosis method based on segmented prior feature fusion, selects the current time series during the switch machine action as the research object, studies the fault curve identification method, and proposes a switch machine fault classification model based on ensemble learning. The core idea is to identify the current curve during the turnout working process through multiple sub-models, and finally fuse the classification results to achieve automatic diagnosis of switch machine faults.

[0004] The purpose of the present invention is achieved through the following technical solutions.

[0005] A switch machine fault diagnosis method based on segmented prior feature fusion includes the following steps:

[0006] Step 1: Collect the current time series data collected by the maintenance support system sensor during the switch machine operation;

[0007] Step 2: Extract features from the current time series, including manual feature engineering and automatic feature extraction. Manual feature engineering classifies the faults of the switch machine according to the causes by analyzing the curve features, and automatic feature extraction uses a denoising autoencoder to detect the unsupervised similarity between the classified samples and the ideal curve.

[0008] Step 3: Use the attention mechanism to improve the three-stage segmentation algorithm and extract features segment by segment;

[0009] Step 4: Use multiple sub-models to identify the current curve during the turnout operation;

[0010] Step 5: Use the late-fusion method in ensemble learning to fuse the classification results of multiple sub-models;

[0011] Step 6: Use the model trained in step 5, input data, and output the railway switch machine fault diagnosis solution.

[0012] In step 1, the original data needs to be processed first. For a time series with a length of n , subtract the previous term from the latter term to get the rate of change sequence The operation is recorded as :

[0013] ;

[0014] Further calculation The rate of change sequence is :

[0015] ;

[0016] at this time, The length is , in order to and Align, choose to fill at the beginning and end so that The length and The same length as:

[0017] ;

[0018] The final sequence , then you can use the sequence To solve the segmentation point and ;

[0019] The dividing point between the start phase and the transition phase It is the turning point after the startup current changes suddenly and then returns to the normal value. It is shown as the second peak in the sequence in D. Since the height of the first two peaks is much larger than the subsequent peaks, an adaptive threshold is used to find , the process is as follows:

[0020] The initialization threshold is half of the maximum value of the time series;

[0021] Find all peaks whose height is higher than the threshold;

[0022] If the number of peak points found is less than 2, two cases are performed:

[0023] The recursive termination condition has been reached, that is, the threshold is 0, indicating that the segmentation point does not exist, and the execution ends directly.

[0024] Otherwise, reduce the threshold and re-execute the search for all peak points with heights higher than the threshold;

[0025] Select the second peak point as , returns the index of the point and the corresponding timestamp;

[0026] In peak finding, the average signed distance of each point is calculated. For each point in the time series, the average signed distance between it and its left and right neighbors is calculated. The calculation formula is as follows:

[0027] ;

[0028] in, Indicates The average signed distance at each time point, Indicates the time series The value of a point, Represents the window size. After the calculation is completed, if the average symbol distance value of a point is greater than the mean plus the standard deviation of the sensitivity times, the point is added to the peak set. For each pair of adjacent peaks in the peak set, if the distance between them is less than or equal to the window size, , then remove the peaks that are lower than the previous peak;

[0029] After that, find the segmentation point between the conversion phase and the indication phase according to the following process :

[0030] convert the original sequence The previous part is cut off to get the modified time series;

[0031] Find peaks with a threshold below 5;

[0032] If no peak is found, it means that the segmentation point does not exist and the execution ends;

[0033] If the number of peaks is 1, jump to return the index of the selected peak and the corresponding timestamp;

[0034] If the number of peaks is greater than or equal to 2, the peaks need to be scored and the one with the highest score is selected. The scoring basis is as follows:

[0035] The peak height must be the highest

[0036] The closer the corresponding values ​​of this peak and all subsequent peaks are to the average of the third stage, the higher the score.

[0037] All peaks before this peak should be in the second stage

[0038] Finally, the formula is:

[0039] ;

[0040] Among them, i is the index of the peak, Indicates the value corresponding to the peak in T, represents the average of this stage, represents the height of the wave crest, is a discount factor used to reduce the impact of future peaks on the score, making the score more focused on the present while considering the overall situation;

[0041] Returns the index of the selected peak and the corresponding timestamp.

[0042] In step 2, representative feature indicators are manually selected based on the time domain and value domain to construct the sequence of each stage The candidate feature set is finally composed of feature vectors according to their contribution to the classification results to form a vector space ,remember Represents the i-th candidate feature index, and the sampling interval is , then the candidate feature set The composition is as follows:

[0043] Maximum value, which represents the maximum value in the sequence:

[0044] ;

[0045] Minimum, which represents the minimum value in the sequence:

[0046] ;

[0047] The mean reflects the central tendency of the sequence elements:

[0048] ;

[0049] Mean amplitude / rectified mean value, meaning the average of the absolute values:

[0050] ;

[0051] The median reflects the middle level of all data points:

[0052] ;

[0053] ;

[0054] The standard deviation reflects the stability of the current time series:

[0055] ;

[0056] RMS value / effective value reflects the energy intensity and stability of the sequence:

[0057] ;

[0058] Peak-to-peak value reflects the variation range of the maximum and minimum values ​​of the current data:

[0059] ;

[0060] Skewness, reflecting the asymmetry of the current curve:

[0061] ;

[0062] Kurtosis / kurtosis reflects the degree to which the current value deviates from the average value:

[0063] ;

[0064] The impulse factor is an indicator used to describe the impact of a signal:

[0065] ;

[0066] The crest factor is used to detect the presence of shocks in the sequence:

[0067] ;

[0068] Crest factor, which indicates the strength of the impulses in the signal:

[0069] ;

[0070] The margin factor describes the degree of impact characteristics of the signal:

[0071] ;

[0072] Time span, reflecting the time taken for each stage:

[0073] ;

[0074] When there are less than two segmentation points, all eigenvalues ​​corresponding to the missing stages are set to -1. In addition, when some indicators have "division by zero" anomalies, they are directly set to 0;

[0075] The feature set constructed using the above formula contains singular values ​​that are too large or too small. The existence of these data may cause the model to overfit a certain type of feature or even fail to converge. In order to eliminate the dimensional influence between different feature indicators, data normalization is required. The maximum and minimum normalization is used to scale the feature values ​​to the [0,1] interval. The formula is as follows:

[0076] ;

[0077] Generate candidate feature sets for 50 different types of samples and visualize them.

[0078] In step 3, the attention mechanism based on the gated recurrent unit network is used to score the confidence of each point in the sequence as a segmentation point, distinguish which parts of the sequence are important for the identification of the segmentation point, and use this as an aid to the numerical segmentation algorithm; the core of the gated recurrent unit is the gating mechanism, which controls the propagation of information by resetting the gate and updating the gate. The formula is as follows:

[0079] ;

[0080] ;

[0081] in, and are the calculated reset and update thresholds, W is the weight matrix, is the hidden state of the previous moment, is the input at time t, b is the bias, represents the sigmoid function, represents matrix multiplication;

[0082] Reset door control to previous state Remain in Candidate Status The proportion of :

[0083] ;

[0084] in, represents dot product (multiplying elements at corresponding positions), Equivalent to a temporary variable, used to calculate the new state later ;

[0085] Update the door control status at the last moment and candidate status Keep in new state The proportion of , for the previous state Selectively "forget" the current candidate state Selective "Memory":

[0086] ;

[0087] The final model structure consists of a gated recurrent unit layer, a multi-layer perception structure including three fully connected layers, and an activation function. The three-phase currents A, B, and C are selected to correspond to the three channel input models respectively. The gated recurrent unit layer is responsible for feature extraction of the input sequence, and the fully connected layer is responsible for converting the extracted features into dimensions and potential meanings. The final result is then scaled to Within the interval, the input sequence of the model is controlled at an appropriate length of 240;

[0088] The unique hot encoding corresponding to the fault is changed to the attention score sequence to obtain the dataset for training. The construction process is as follows:

[0089] Initialize all elements in G to 0;

[0090] In G and The corresponding position is assigned the highest score v ( ), if it does not exist, skip it;

[0091] Expand forward and backward in a sliding window manner, and decrease the element value in the window by exponential w until the window boundary k: .

[0092] In step 4, a fault classification subnetwork based on GRU-TCN is provided. GRU-TCN is a hybrid model that combines a gated recurrent unit widely used in time series related work and a time domain convolutional neural network.

[0093] Unlike the two-dimensional and three-dimensional convolutions used to process images, TCN performs convolution by sliding a convolution kernel with a fixed height of 1 on the input sequence. The one-dimensionality of this convolution enables it to learn long-term and short-term dependencies and patterns in time series data. Let the input sequence T = { }, convolution kernel , TCN layer The value of the moment is As a one-dimensional fully convolutional neural network dedicated to sequence data, TCN adds the following improvements on the basis of traditional CNN:

[0094] Dilated causal convolution: Dilation refers to introducing holes in the convolution to increase the receptive field, thereby capturing longer dependencies. This operation is equivalent to downsampling the input. The sampling rate is controlled by the dilation coefficient. When the dilation coefficient is 1, the network degenerates into a traditional CNN. Reasonable setting of the dilation coefficient can obtain a receptive field much larger than that of the traditional CNN on the basis of a stack with no more than 3 layers. The dilated convolution generates The process can be expressed as:

[0095] ;

[0096] Causality means that there is a one-way time constraint, which restricts the reading of future data. The result should only depend on ;

[0097] Residual connection: The output of the dilated causal convolution is added to the input data through the residual connection as the final output , which aims to solve the gradient vanishing problem in the process of parameter transfer. The network can be built deeper through cross-layer transfer of weights. The specific implementation of residual connection in TCN is as follows:

[0098] ;

[0099] in, Represents 1x1 convolution;

[0100] Spatial Dropout: A regularization method specifically for sequence data. Unlike traditional dropout, which randomly sets elements to zero, Spatial Dropout sets the elements in a region to zero.

[0101] The proposed GRU-TCN model structure, the model input is a three-channel time series ,remember Indicates channel Previous The value corresponding to the moment, for example ; The TCN block extracts multi-scale features of the input sequence through two sets of dilated causal convolutions with different receptive fields. The role of the squeeze and activation block is to provide channel attention support for the previous TCN block. The principle of the squeeze and activation block is as follows;

[0102] First, use global average pooling to reduce the dimension of the feature map of each channel:

[0103] ;

[0104] The reduced features are then passed to the fully connected layer, and the value of each channel is multiplied by the importance coefficient given by the fully connected layer as the output:

[0105] ;

[0106] in, is the input feature tensor, and Represent the Sigmoid and ReLU activation functions respectively, and is a learnable weight matrix, Represents element-wise multiplication;

[0107] Since TCN and GRU differ in input order, the one-dimensional convolution operation slides along the sequence length seq_len, so the input shape is It is more in line with the habit of convolution operation, and RNN usually regards the features at each time step as a vector, so the expected input is , the permute operation is introduced to swap the last two dimensions as needed,

[0108] In the GRU-TCN model, the TCN and GRU parts first extract features from time series data from different angles at the same time, and then splice their output feature vectors together to obtain a richer and more accurate feature representation; finally, the spliced ​​vectors are input into the fully connected layer and the Softmax layer to obtain the classification result.

[0109] In step 5, a decision-level fusion scheme based on fuzzy logic is selected: the membership function is used to reflect the membership of a value in the fuzzy set, and the domain describes the range of values ​​to which the membership function is applied. Each element in the classifier result, that is, the confidence corresponding to each category, is expressed by a fuzzy set: , the domain is described by fuzzy logic composed of all fault types. If the confidence value of a classifier for a certain type of sample is higher than 0.7, the "true judgment" and "false judgment" parts of the membership function on the corresponding domain should present right-skewed and left-skewed S-shaped curves respectively, and the steeper the slope, the higher the classifier's discrimination of the fault type; on the contrary, if the confidence value of the classifier for the sample of this type is between 0.3 and 0.5, the corresponding membership function is moderate and gentle;

[0110] remember Representation classifier Determine the input sample as Confidence of type, classifier In the domain The membership function on is , the process of late fusion of the classification results obtained after the samples are input into each classifier is as follows:

[0111] The classification confidence of each fault type output by the classifier is mapped through the corresponding membership function to obtain the membership:

[0112] ;

[0113] in, Representation classifier Output The confidence level of the type is the membership of "true" and "false";

[0114] The new comprehensive membership is synthesized by subtracting the "false" membership from the "true" membership. According to the previous analysis, the larger the comprehensive membership, the more reliable the classifier's recognition of the category:

[0115] ;

[0116] Summarize the comprehensive membership vectors of all classifiers to obtain the final classification result Y:

[0117] ;

[0118] .

[0119] In step 6, after the text generation model is trained, each time a time series is input into the model, the model will determine whether the switch machine is faulty based on the training content, thereby achieving the desired effect.

[0120] The autoencoder is an unsupervised neural network-based algorithm for distinguishing the similarity between the input value and the target value. Its structure consists of three parts: encoder, decoder, and bottleneck layer. The sequence to be encoded is :

[0121] Encoder: Encodes the raw data into feature representations at the bottleneck layer :

[0122] ;

[0123] Bottleneck layer: Also known as the compression layer, it connects the encoder and decoder at both ends. The bottleneck layer is responsible for compressing the feature dimension and limiting the number of features entering the decoder ( ), which is equivalent to further abstracting the original data, thus forcing the network to learn only the most significant features;

[0124] Decoder: Finally, the feature representation from the bottleneck layer Reconstruct the original data:

[0125] ;

[0126] Ideally, the output of the self-decoder should be as consistent as possible with the input sequence ( ), using the mean absolute error to measure and The difference between:

[0127] ;

[0128] The encoder and decoder are implemented with multi-layer LSTM and GRU recurrent neural networks, and the sequence after the controlled length is As the model input, the DRAE corresponding to each type of sample is trained, and the data set used for training is composed entirely of samples of this type. During the training process, DRAE will minimize the reconstruction error by continuously adjusting the parameters of the encoder and decoder. ,Finally, a series of DRAEs that can reconstruct the corresponding types of samples are obtained;

[0129] Follow the following process Identify the fault type:

[0130] Create a separate dataset for each type of sample and train DRAE;

[0131] For the sample sequence to be identified To split;

[0132] After the division Enter DRAE;

[0133] Compute the reconstruction loss set , Each element in is composed of 3 stages of losses;

[0134] Each type of sample behaves significantly differently in one or more stages. We assign weights to each stage according to its contribution to the result, and obtain a new loss set by weighting. ,by For example, the calculation formula is:

[0135] ;

[0136] The larger the loss, the more obvious the morphological difference is, and the lower the confidence that the sample belongs to this category. Flip the values ​​to generate a confidence set :

[0137] ;

[0138] Among them, m is used to adjust the scaling ratio in the x-axis direction;

[0139] Use the SoftMax function to scale the elements in C so that each element The range and the sum is 1;

[0140] Output the confidence set C corresponding to all categories.

[0141] Compared with the prior art, the advantages of the present invention are: the present invention selects the current time series during the operation of the switch as the research object, studies the identification method of the fault curve, and proposes a switch fault classification model based on ensemble learning. The core idea is to identify the current curve during the working process of the turnout through multiple sub-models, and finally fuse the classification results to achieve automatic diagnosis of switch faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0142] Figure 1 It is a flow chart of the overall working steps of the present invention.

[0143] Figure 2 This is a decision-level fusion solution based on fuzzy logic in the present invention. DETAILED DESCRIPTION

[0144] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0145] like Figure 1 As shown, a switch machine fault diagnosis method based on segmented prior feature fusion includes the following steps:

[0146] Step 1: Collect the current time series data collected by the maintenance support system (MSS) sensor during the switch machine operation.

[0147] Step 2: Extract features from the current time series, including manual feature engineering and automatic feature extraction. Manual feature engineering classifies the faults of the switch machine according to their causes by analyzing the curve features. Automatic feature extraction uses a denoising autoencoder to detect the unsupervised similarity between the samples to be classified and the ideal curve.

[0148] Step 3: Use the attention mechanism to improve the three-stage segmentation algorithm and extract features segment by segment.

[0149] Step 4: Use multiple sub-models to identify the current curve during the turnout operation.

[0150] Step 5: Use the late-fusion method in ensemble learning to fuse the classification results of multiple sub-models.

[0151] Step 6: Use the model trained in step 5, input data, and output the railway switch machine fault diagnosis solution.

[0152] In step 1, the current time series data of the switch machine during operation collected by the maintenance support system (MSS) sensor is collected. Since many fixed thresholds are inevitably used in the subsequent time series processing and analysis, and the models of the maintenance support systems (MSS) deployed on the line are different, the data collected at different sampling rates will affect the accuracy of the scheme recognition. In addition, if the sampling rate of the sequence is low, it may cause information loss, and if the sampling rate is high, it may cause a waste of computing resources and is not convenient for subsequent debugging and maintenance. Therefore, resampling is required before the data is input into the model to ensure that the sampling rate of the sequence is consistent with the specified value. In order to eliminate the influence of sensor noise and mutation values, so that the scheme can better focus on analyzing the trends and laws of the time series, the present invention needs to filter the input data.

[0153] Feature extraction of current time series, including manual feature engineering and automatic feature extraction. Manual feature engineering classifies switch machine faults according to their causes by analyzing curve features. Automatic feature extraction uses a denoising autoencoder to detect the unsupervised similarity between the sample to be classified and the ideal curve:

[0154] The present invention mentions that in addition to quantifiable features, the shape of the current curve is also an important basis for determining the type of fault. Through the analysis of the current curve shape, it is possible to distinguish small fluctuations, curve trends, extreme point distribution, phase shift and other features that are difficult to describe in terms of dimensions, thereby making the fault diagnosis of the present invention more accurate.

[0155] Auto Encoder (AE) is an unsupervised (input value is the same as target value) morphological similarity discrimination algorithm based on neural network. Its structure consists of three parts: encoder, decoder, and bottleneck layer. The sequence to be encoded is :

[0156] Encoder: Encodes the raw data into feature representations at the bottleneck layer :

[0157] ;

[0158] Bottleneck layer: Also known as the compression layer, it connects the encoder and decoder at both ends. As the name suggests, the bottleneck layer is responsible for compressing the feature dimension and limiting the number of features entering the decoder ( ), which is equivalent to further abstracting the original data, thus forcing the network to learn only the most significant features;

[0159] Decoder: Finally, the feature representation from the bottleneck layer Reconstruct the original data:

[0160] ;

[0161] Ideally, the output of the self-decoder should be as consistent as possible with the input sequence ( ). The present invention uses Mean Absolute Error (MAE) to measure and The difference between:

[0162] ;

[0163] The Deep Recurrent Auto-Encoder of the present invention is based on the traditional auto-encoder, and the encoder and decoder are implemented with multi-layer LSTM, GRU and other recurrent neural networks (RNN). Since the current sequence generated by each action of the switch machine is non-periodic and has a long-term dependency, the present invention abandons the traditional method of sliding the sequence slice, and instead slices the sequence after controlling the length at one time. As model input. After that, the DRAE corresponding to each type of sample is trained, and the dataset used for training consists entirely of samples of that type. Since it is unsupervised training, only the samples themselves are needed without any additional labels or target values. During the training process, DRAE will minimize the reconstruction error by continuously adjusting the parameters of the encoder and decoder. ,Finally, a series of DRAEs that can reconstruct the corresponding types of samples are obtained.

[0164] Based on this characteristic, the present invention can be carried out according to the following process Identify the fault type:

[0165] Step 1: Create a separate dataset for each type of sample and train DRAE

[0166] Step 2: Segment the sample sequence T to be identified;

[0167] Step 3: Enter DRAE;

[0168] Step 4 Calculate the reconstruction loss set , Each element in consists of three stages of losses. , which can be regarded as the error between the input sample and the ideal H1 sample at each stage.

[0169] Step 5 Each type of sample has significantly different performance in one or more stages. We assign weights according to the contribution of each stage to the result and obtain a new loss set by weighting. ,by For example, the calculation formula is:

[0170] ;

[0171] Step 6 The greater the loss, the more obvious the morphological difference, and the lower the confidence that the sample belongs to this category. Therefore, the present invention uses the following function to Flip the values ​​to generate a confidence set :

[0172] ;

[0173] Among them, m is used to adjust the scaling ratio in the x-axis direction and can take different values ​​as needed (here it is 103).

[0174] Step 7 Use the SoftMax function to scale the elements in C so that each element The range and the sum is 1;

[0175] Step 8 Output the confidence set corresponding to all categories .

[0176] Before starting, you need to process the original data. For a time series of length n , subtract the previous term from the latter term to get the rate of change sequence The operation is recorded as :

[0177] ;

[0178] Further calculation The rate of change sequence is :

[0179] ;

[0180] at this time, The length is , in order to and Align, choose to fill at the beginning and end so that The length and The same length as:

[0181] ;

[0182] The final sequence , then you can use the sequence To solve the segmentation point and .

[0183] The dividing point between the start phase and the transition phase It is the turning point after the startup current changes suddenly and then returns to the normal value. It is shown as the second peak in the sequence in D. Since the height of the first two peaks is much larger than the subsequent peaks, an adaptive threshold is used to find , the process is as follows:

[0184] Step 1 Initialize the threshold to half of the maximum value of the time series

[0185] Step 2 Find all peaks with heights above the threshold

[0186] Step 3 If the number of peak points found is less than 2, execute the following two cases:

[0187] The recursive termination condition has been reached (the threshold is 0), indicating that the segmentation point does not exist, and the execution ends directly

[0188] Otherwise, reduce the threshold and execute Step 2 again.

[0189] Step 4 Select the second peak point as , returns the index of the point and the corresponding timestamp

[0190] The peak search function is to find the highest peaks in a given sequence of data. Suspicious points are first identified by calculating the average difference between each point and its adjacent values, and then further screening is performed to determine the true peaks, thereby avoiding false positives caused by too dense peaks.

[0191] In peak finding, the average signed distance of each point is calculated. For each point in the time series, the average signed distance between it and its left and right neighbors is calculated. The calculation formula is as follows:

[0192] ;

[0193] in, Indicates The average signed distance at each time point, Indicates the time series The value of a point, Represents the window size. After the calculation is completed, if the average symbol distance value of a point is greater than the mean plus the standard deviation of the sensitivity times, the point is added to the peak set. For each pair of adjacent peaks in the peak set, if the distance between them is less than or equal to the window size, , then remove the peaks that are lower than the previous peak;

[0194] After that, find the segmentation point between the conversion phase and the indication phase according to the following process :

[0195] Step 1: Convert the original sequence The previous part is cut off to get the modified time series

[0196] Step 2: Find peaks with a threshold value below 5

[0197] Step 3 If no peak is found, it means that the segmentation point does not exist and the execution ends.

[0198] Step 4 If the number of peaks is 1, jump to Step 6

[0199] Step 5 If the number of peaks is greater than or equal to 2, you need to score the peaks and select the one with the highest score. The scoring is based on the following:

[0200] The peak height must be the highest

[0201] The closer the corresponding values ​​of this peak and all subsequent peaks are to the average of the third stage, the higher the score.

[0202] All peaks before this peak should be in the second stage

[0203] Finally, the formula is:

[0204] ;

[0205] in, is the index of the peak, Indicates the value corresponding to the peak in T, represents the average of this stage, represents the height of the wave crest, It is a discount factor used to reduce the impact of future peaks on the score, making the score more focused on the present while considering the overall situation.

[0206] Step 6 Return the index of the selected peak and the corresponding timestamp.

[0207] However, this numerical segmentation point detection algorithm also has many disadvantages. Specifically, when the curve jitters violently (caused by switch machine failure or sensor noise), it is easy to have false alarms or missed alarms. A typical example is that the algorithm returns the wrong segmentation point when H2, H4 and H5 failures occur. Therefore, it is necessary to improve the algorithm so that it has high robustness even in complex environments.

[0208] In step 2, feature extraction is performed on the current time series, including manual feature engineering and automatic feature extraction. Manual feature engineering classifies the faults of the switch machine according to the causes by analyzing the curve characteristics. Automatic feature extraction uses a denoising autoencoder to detect the unsupervised similarity between the samples to be classified and the ideal curve.

[0209] The present invention manually selects some representative characteristic indicators based on the time domain and value domain to construct each stage sequence The candidate feature set is finally composed of feature vectors according to their contribution to the classification results to form a vector space ,remember Represents the i-th candidate feature index, and the sampling interval is , then the candidate feature set The composition is as follows:

[0210] Maximum value (Max), which indicates the maximum value in the sequence:

[0211] Minimum value (Min), represents the minimum value in the sequence:

[0212] Mean reflects the central tendency of sequence elements:

[0213] Mean Absolute / Average rectified value, which means the average of the absolute values:

[0214] Median reflects the middle level of all data points: ;

[0215] ;

[0216] Standard Deviation reflects the stability of the current time series:

[0217] Root Mean Square (RMS) / effective value, reflecting the energy intensity and stability of the sequence:

[0218] Peak-to-Peak reflects the range of maximum and minimum current data values:

[0219] Skewness reflects the asymmetry of the current curve:

[0220] Kurtosis reflects the degree to which the current value deviates from the average value:

[0221] Impulse Factor is an indicator used to describe the impact of a signal:

[0222] Form Factor, used to detect whether there is a shock in the sequence:

[0223] Crest Factor, which indicates the strength of the impulse in the signal:

[0224] The Clearance Factor describes the degree of impact characteristics (high-amplitude transients or mutations) of the signal:

[0225] Time Span reflects the time taken for each stage:

[0226] It should be noted that when there are less than two segmentation points, all eigenvalues ​​corresponding to the missing stages are set to -1. In addition, when some indicators have "division by zero" anomalies, they are directly set to 0.

[0227] The feature set constructed using the above formula may contain singular values ​​that are too large or too small. The presence of these values ​​may cause the model to overfit a certain type of feature or even fail to converge. In order to eliminate the dimensionality effect between different feature indicators, data normalization is required. This article uses Min-Max Normalization to scale the feature values ​​to the [0,1] interval. The formula is as follows:

[0228] ;

[0229] Generate candidate feature sets for 50 different types of samples and visualize them.

[0230] In step 3, the attention mechanism is used to improve the three-stage segmentation algorithm and extract features segment by segment.

[0231] In order to solve the problem of poor accuracy of traditional numerical segmentation point search algorithms in complex environments, this paper proposes an attention mechanism based on the GRU network, which scores the confidence of each point in the sequence as a segmentation point, distinguishes which parts of the sequence are important for the recognition of segmentation points, and uses this as an auxiliary to the numerical segmentation algorithm. In this way, the model can focus on processing the key parts of the input sequence, reduce the impact of other unnecessary parts on the recognition results, and thus improve the model performance.

[0232] The Gated Recurrent Unit (GRU) is a recurrent neural network (RNN) structure that is suitable for processing serialized data such as sensor time series and text. It can quickly respond to short-term changes while retaining long-term memory. As an improvement of the Long Short-term Memory (LSTM), GRU combines the forget gate and input gate into a single update gate, and also integrates the cell state and hidden state, making its structure simpler and the reasoning speed faster.

[0233] The core of GRU is the gating mechanism, which controls the propagation of information through the reset gate and update gate. The formula is as follows:

[0234] ;

[0235] ;

[0236] in, and are the calculated reset and update thresholds, W is the weight matrix, is the hidden state of the previous moment, is the input at time t, b is the bias, represents the sigmoid function, Represents matrix multiplication.

[0237] Reset door control to previous state Remain in Candidate Status The proportion of :

[0238] ;

[0239] in, represents dot product (multiplying elements at corresponding positions), Equivalent to a temporary variable, used to calculate the new state later ;

[0240] Update the door control status at the last moment and candidate status Keep in new state The proportion of , for the previous state Selectively "forget" the current candidate state Selective "Memory":

[0241] ;

[0242] The final model structure consists of a GRU layer, a multi-layer perceptron (MLP) structure containing three fully connected (FC) layers, and an activation function. The present invention selects three-phase currents A, B, and C to correspond to three channel input models respectively. The GRU layer is responsible for feature extraction of the input sequence, and the fully connected layer is responsible for converting the extracted features into dimensions and potential meanings. The final result is then scaled to It should be noted that the input sequence of the model is controlled at an appropriate length of 240. In addition to increasing the training time and resource usage, excessively long sequence input may also cause gradient vanishing or gradient explosion problems due to weight decay and gradient accumulation, resulting in the failure of the model to converge.

[0243] Since the present invention has determined the segmentation points in the process of simulating and generating the data set in the previous step and , so we only need to modify the label set (change the one-hot encoding corresponding to the fault to the attention score sequence) to get the dataset for training. The construction process is as follows:

[0244] Step 1 Initialize all elements in G to 0

[0245] Step 2: and The corresponding position is assigned the highest score v ( ), if it does not exist, skip it

[0246] Step 3: Expand forward and backward in a sliding window manner, and decrease the element value in the window by exponential w until the window boundary k: .

[0247] In step 4, multiple sub-models are used to identify the current curve during the turnout operation.

[0248] The present invention proposes a fault classification subnetwork based on GRU-TCN. GRU-TCN is a hybrid model that combines the gated recurrent unit (GRU) and the temporal convolutional neural network (TCN), which are currently widely used in time series related work. Different from the original two methods, GRU-TCN benefits from the powerful fitting and feature extraction capabilities of deep learning, and automatically learns useful features from the current curve. It is a solution based on automatic feature extraction.

[0249] Unlike the two-dimensional and three-dimensional convolutions used to process images, TCN performs convolution by sliding a convolution kernel with a fixed height of 1 on the input sequence. The one-dimensionality of this convolution enables it to learn long-term and short-term dependencies and patterns (Paradigm) in time series data. }, convolution kernel , TCN layer The value of the moment is As a one-dimensional fully convolutional neural network (FCN) dedicated to sequence data, TCN adds the following improvements on the basis of traditional CNN:

[0250] Causal Dilated Convolution: Dilation refers to the introduction of holes in the convolution to increase the receptive field, thereby capturing longer dependencies. This operation is equivalent to downsampling the input. The sampling rate is controlled by the dilation coefficient. When the dilation coefficient is 1, the network degenerates to the traditional CNN. Reasonable setting of the dilation coefficient can obtain a receptive field much larger than that of the traditional CNN on the basis of a stack with no more than 3 layers. By using a dilation coefficient of The dilated convolution generates The process can be expressed as:

[0251] ;

[0252] Causality means that there is a one-way time constraint, which restricts the reading of future data. The results should only depend on ;

[0253] Residual Connection: The output of the dilated causal convolution is added to the input data through the residual connection as the final output , which aims to solve the gradient vanishing problem in the parameter transfer process, and make the network deeper by transferring weights across layers. The specific implementation of residual connection in TCN is as follows:

[0254] ;

[0255] in, Represents a 1x1 convolution. Its uses are: (1) solving the channel number mismatch problem by setting the channel for the convolution kernel; (2) solving the input and output length inconsistency problem by setting the stride of the convolution kernel.

[0256] Spatial Dropout: A regularization method specifically for sequence data. Unlike traditional dropout, which randomly sets elements to zero, Spatial Dropout sets the elements in a region to zero. This reduces the damage to the relative position relationship between elements and prevents the network from using adjacent elements to restore information, which can effectively avoid overfitting and improve the performance of the model.

[0257] The proposed GRU-TCN model structure, the model input is a three-channel time series ,remember Indicates channel Previous The value corresponding to the moment, for example ; The TCN block performs multi-scale feature extraction on the input sequence through two sets of dilated causal convolutions with different receptive fields. The role of the squeeze and activation block is to provide channel attention support for the previous TCN block. The principle of the squeeze and activation block is as follows.

[0258] First, use Global Average Pooling (GAP) to reduce the dimension of the feature map of each channel:

[0259] The reduced features are then passed to the fully connected layer, and the value of each channel is multiplied by the importance coefficient given by the fully connected layer as the output:

[0260] in, is the input feature tensor, and Represent the Sigmoid and ReLU activation functions respectively, and is a learnable weight matrix, Represents element-wise multiplication.

[0261] Since TCN and GRU differ in input order, the one-dimensional convolution operation slides along the sequence length seq_len, so the input shape is is more in line with the habit of convolution operation. RNN usually regards the features at each time step as a vector, so the expected input is In order to solve the whole problem, the present invention introduces the permute operation to swap the last two dimensions as needed.

[0262] In the GRU-TCN model proposed in this paper, TCN and GRU first extract features from time series data from different angles at the same time, and then concatenate their output feature vectors together to obtain a richer and more accurate feature representation. Finally, the concatenated vectors are input into the fully connected layer and the Softmax layer to obtain the classification result.

[0263] In step 5, the classification results of multiple sub-models are fused using the late-fusion method in ensemble learning.

[0264] The present invention decides to use the decision-level fusion solution for the following reasons:

[0265] Performance: Compared with other fusion methods, decision-level fusion solutions usually do not involve feature concatenation or superposition, avoiding excessively high feature space dimensions and consuming less resources during training and inference.

[0266] Solution coupling degree: Feature-level fusion and data-level fusion destroy the isolation between different inputs. In decision-level fusion, different sub-models perform their respective functions and do not affect each other.

[0267] Interpretability: Each sub-model can output intuitive classification results

[0268] Maintainability: Since each sub-model is relatively independent, system expansion and performance tuning are also relatively convenient.

[0269] Fuzzy logic is a method for processing fuzzy or uncertain information. In fuzzy logic, a value can belong to multiple fuzzy sets, and each fuzzy set corresponds to a membership degree. Fuzzy logic can be used to deal with uncertain classification situations, and fuzzy logic is used to fuzzify the results of different classifiers. Therefore, the present invention proposes the following Figure 2 The decision-level fusion scheme based on fuzzy logic is shown.

[0270] The membership function is used to reflect the membership of a value in the fuzzy set, and the domain describes the range of values ​​to which the membership function is applied. In this problem, each element in the classifier result (the confidence level corresponding to each category) can be represented by a fuzzy set: , the domain is described by fuzzy logic consisting of all fault types. If the confidence value of a classifier for a certain category of samples is higher than 0.7, the "true" and "false" parts of the membership function on the corresponding domain should present right-skewed and left-skewed S-shaped curves respectively, and the steeper the slope, the higher the classifier's discrimination of the fault type. On the contrary, if the confidence value of the classifier for this type of sample is between 0.3 and 0.5, the corresponding membership function is centered and gentle.

[0271] The classification of membership functions is described as follows:

[0272] Poor category:

[0273] The membership function curve is maximum in the low confidence region (close to 0), indicating that at low confidence, the sample is classified as “poor” with a higher membership.

[0274] As the confidence increases, the membership gradually decreases, indicating that at high confidence levels, the likelihood of a sample being classified as “poor” decreases.

[0275] Good Category:

[0276] The membership function curve reaches a peak in the medium confidence region (around 5), indicating that at medium confidence, the sample is classified as "good" with a high membership.

[0277] In the low and high confidence regions, the membership is low, indicating that the sample has a low probability of being classified as “good” at these confidence levels.

[0278] Excellent Category:

[0279] The membership function curve is the largest in the high confidence region (close to 10), indicating that at high confidence, the sample is classified as “excellent” with a high membership.

[0280] As the confidence level decreases, the membership level decreases, indicating that at low confidence levels, the likelihood of a sample being classified as “excellent” decreases.

[0281] The membership function can accurately adapt the classifier performance to each output category. Therefore, compared with traditional fusion methods such as weighted average, the solution of the present invention has a more detailed granularity. At the same time, compared with the method based on neural network fitting, the solution of the present invention is simple to implement, has low resource consumption, and has good interpretability.

[0282] remember Representation classifier Determine the input sample as Confidence of type, classifier In the domain The membership function on is , the process of late fusion of the classification results obtained after the samples are input into each classifier is as follows:

[0283] Step 1 Map the classification confidence of each fault type output by the classifier through the corresponding membership function to obtain the membership degree:

[0284] ;

[0285] in, Representation classifier Output The confidence of the type is the membership of "true" and "false"

[0286] Step 2: Subtract the "false" membership from the "true" membership to synthesize a new comprehensive membership. According to the previous analysis, the larger the comprehensive membership, the more reliable the classifier's recognition of the category:

[0287] ;

[0288] Step 3 summarizes the comprehensive membership vectors of all classifiers to obtain the final classification result Y:

[0289] ;

[0290] .

[0291] In step 6, the model trained in step 5 is used to input data and output a railway switch machine fault diagnosis solution.

[0292] After the text generation model is trained, each time a time series is input into the model, the model will determine whether the switch machine has a fault based on the training content to achieve the desired effect.

Claims

1. A fault diagnosis method for switch machine based on segmented prior feature fusion, characterized in that: The following steps are involved: Step 1: Collect the current time series data collected by the maintenance support system sensor during the switch machine operation; Step 2: Extract features from the current time series, including manual feature engineering and automatic feature extraction. Manual feature engineering classifies the faults of the switch machine according to the causes by analyzing the curve features, and automatic feature extraction uses a denoising autoencoder to detect the unsupervised similarity between the classified samples and the ideal curve. Step 3: Use the attention mechanism to improve the three-stage segmentation algorithm and extract features segment by segment; Step 4: Use multiple sub-models to identify the current curve during the turnout operation; Step 5: Use the late-fusion method in ensemble learning to fuse the classification results of multiple sub-models; Step 6: Use the model trained in step 5, input data, and output the railway switch machine fault diagnosis solution; In step 1, the original data needs to be processed first. For a time series with a length of n , subtract the previous term from the latter term to get the rate of change sequence The operation is recorded as : ; Further calculation The rate of change sequence is : ; at this time, The length is , in order to and Align, choose to fill at the beginning and end so that The length and The same length as: ; The final sequence , through the sequence To solve the segmentation point and ; The dividing point between the start phase and the transition phase It is the turning point after the startup current changes suddenly and then returns to the normal value. It is shown as the second peak in the sequence in D. Since the height of the first two peaks is much larger than the subsequent peaks, an adaptive threshold is used to find , the process is as follows: The initialization threshold is half of the maximum value of the time series; Find all peaks whose height is higher than the threshold; If the number of peak points found is less than 2, two cases are performed: The recursive termination condition has been reached, that is, the threshold is 0, indicating that the segmentation point does not exist, and the execution ends directly. Otherwise, reduce the threshold and re-execute the search for all peak points with heights higher than the threshold; Select the second peak point as , returns the index of the point and the corresponding timestamp; In peak finding, the average signed distance of each point is calculated. For each point in the time series, the average signed distance between it and its left and right neighbors is calculated. The calculation formula is as follows: ; in, Indicates The average signed distance at each time point, Indicates the time series The value of a point, Represents the window size. After the calculation is completed, if the average symbol distance value of a point is greater than the mean plus the standard deviation of the sensitivity times, the point is added to the peak set. For each pair of adjacent peaks in the peak set, if the distance between them is less than or equal to the window size, , then remove the peaks that are lower than the previous peak; After that, find the segmentation point between the conversion phase and the indication phase according to the following process : convert the original sequence The previous part is cut off to obtain the modified time series; Find peaks with a threshold below 5; If no peak is found, it means that the segmentation point does not exist and the execution ends; If the number of peaks is 1, jump to return the index of the selected peak and the corresponding timestamp; If the number of peaks is greater than or equal to 2, the peaks need to be scored and the one with the highest score is selected. The scoring basis is as follows: The peak height must be the highest The closer the corresponding values ​​of this peak and all subsequent peaks are to the average of the third stage, the higher the score. All peaks before this peak should be in the second stage Finally, the formula is: ; in, is the index of the peak, Indicates the value corresponding to the peak in T, represents the average of this stage, represents the height of the wave crest, is a discount factor used to reduce the impact of future peaks on the score, making the score more focused on the present while considering the overall situation; Returns the index of the selected peak and the corresponding timestamp.

2. A method for fault diagnosis of switch machine based on segmented priori feature fusion according to claim 1, characterized in that: In step 2, representative feature indicators are manually selected based on the time domain and value domain to construct the sequence of each stage The candidate feature set is finally composed of feature vectors according to their contribution to the classification results to form a vector space ,remember Represents the i-th candidate feature index, and the sampling interval is , then the candidate feature set The composition is as follows: Maximum value, which represents the maximum value in the sequence: ; Minimum, which represents the minimum value in the sequence: ; The mean reflects the central tendency of the sequence elements: ; Mean amplitude / rectified mean value, meaning the average of the absolute values: ; The median reflects the middle level of all data points: , ; The standard deviation reflects the stability of the current time series: ; RMS value / effective value reflects the energy intensity and stability of the sequence: ; Peak-to-peak value reflects the variation range of the maximum and minimum values ​​of the current data: ; Skewness, reflecting the asymmetry of the current curve: ; Kurtosis / kurtosis reflects the degree to which the current value deviates from the average value: ; The impulse factor is an indicator used to describe the impact of a signal: ; The crest factor is used to detect the presence of shocks in the sequence: ; Crest factor, which indicates the strength of impulses in a signal: ; The margin factor describes the degree of impact characteristics of the signal: ; Time span, reflecting the time taken for each stage: ; When there are less than two segmentation points, all eigenvalues ​​corresponding to the missing stages are set to -1. In addition, when some indicators have "division by zero" anomalies, they are directly set to 0; The feature set constructed using the above formula contains singular values ​​that are too large or too small. Data normalization is required. The maximum and minimum normalization is used to scale the feature values ​​to the [0,1] interval. The formula is as follows: ; Generate candidate feature sets for 50 different types of samples and visualize them.

3. A method for fault diagnosis of switch machine based on segmented priori feature fusion according to claim 2, characterized in that: In step 3, the attention mechanism based on the gated recurrent unit network is used to score the confidence of each point in the sequence as a segmentation point, distinguish which parts of the sequence are important for the identification of the segmentation point, and use this as an aid to the numerical segmentation algorithm; the core of the gated recurrent unit is the gating mechanism, which controls the propagation of information by resetting the gate and updating the gate. The formula is as follows: ; ; in, and are the calculated reset and update thresholds, W is the weight matrix, is the hidden state of the previous moment, is the input at time t, b is the bias, represents the sigmoid function, represents matrix multiplication; Reset door control to previous state Remain in Candidate Status The proportion of : ; in, represents dot product, Equivalent to a temporary variable, used to calculate the new state later ; Update the door control status at the last moment and candidate status Keep in new state The proportion of , for the previous state Selectively "forget" the current candidate state Selective "memory": ; The final model structure consists of a gated recurrent unit layer, a multi-layer perception structure including three fully connected layers, and an activation function. The three-phase currents A, B, and C are selected to correspond to the three channel input models respectively. The gated recurrent unit layer is responsible for feature extraction of the input sequence, and the fully connected layer is responsible for converting the extracted features into dimensions and potential meanings. The final result is then scaled to Within the interval, the input sequence of the model is controlled at an appropriate length of 240; The unique hot encoding corresponding to the fault is changed to the attention score sequence to obtain the dataset for training. The construction process is as follows: Initialize all elements in G to 0; In G and The corresponding position is assigned the highest score v ( ), if it does not exist, skip it; Expand forward and backward in a sliding window manner, and decrease the element value in the window by exponential w until the window boundary k: .

4. A method for fault diagnosis of switch machine based on segmented priori feature fusion according to claim 3, characterized in that: In step 4, a fault classification subnetwork based on GRU-TCN is provided. GRU-TCN is a hybrid model that combines a gated recurrent unit widely used in time series related work and a time domain convolutional neural network. Unlike the two-dimensional and three-dimensional convolutions used to process images, TCN performs convolution by sliding a convolution kernel with a fixed height of 1 on the input sequence. The one-dimensionality of this convolution enables it to learn long-term and short-term dependencies and patterns in time series data. Let the input sequence T = { }, convolution kernel , TCN layer The value of the moment is As a one-dimensional fully convolutional neural network dedicated to sequence data, TCN adds the following improvements on the basis of traditional CNN: Dilated causal convolution: Dilation refers to introducing holes in the convolution to increase the receptive field, thereby capturing longer dependencies. This operation is equivalent to downsampling the input. The sampling rate is controlled by the dilation coefficient. When the dilation coefficient is 1, the network degenerates into a traditional CNN. Reasonable setting of the dilation coefficient can obtain a receptive field much larger than that of the traditional CNN on the basis of a stack with no more than 3 layers. The dilated convolution generates The process can be expressed as: ; Causality means that there is a one-way time constraint, which restricts the reading of future data. The results should only depend on ; Residual connection: The output of the dilated causal convolution is added to the input data through the residual connection as the final output , which aims to solve the gradient vanishing problem in the process of parameter transfer. The network can be built deeper through cross-layer transfer of weights. The specific implementation of residual connection in TCN is as follows: ; in, Represents 1x1 convolution; Spatial Dropout: A regularization method specifically for sequence data. Unlike traditional dropout, which randomly sets elements to zero, Spatial Dropout sets the elements in a region to zero. The proposed GRU-TCN model structure, the model input is a three-channel time series ,remember Indicates channel Previous The value corresponding to the moment; the TCN block extracts multi-scale features of the input sequence through two sets of dilated causal convolutions with different receptive fields. The role of the squeeze and activation block is to provide channel attention support for the previous TCN block. The principle of the squeeze and activation block is as follows; First, use global average pooling to reduce the dimension of the feature map of each channel: ; The reduced features are then passed to the fully connected layer, and the value of each channel is multiplied by the importance coefficient given by the fully connected layer as the output: ; in, is the input feature tensor, and Represent the Sigmoid and ReLU activation functions respectively, and is a learnable weight matrix, Represents element-wise multiplication; Since TCN and GRU differ in input order, the one-dimensional convolution operation slides along the sequence length seq_len, so the input shape is is more in line with the habit of convolution operation, and RNN regards the features at each time step as a vector, so the expected input is , the permute operation is introduced to swap the last two dimensions as needed, In the GRU-TCN model, the TCN and GRU parts first extract features from time series data from different angles at the same time, and then splice their output feature vectors together to obtain a richer and more accurate feature representation; finally, the spliced ​​vectors are input into the fully connected layer and the Softmax layer to obtain the classification result.

5. A method for fault diagnosis of switch machine based on segmented priori feature fusion according to claim 4, characterized in that: In step 5, a decision-level fusion scheme based on fuzzy logic is selected: the membership function is used to reflect the membership of a value in the fuzzy set to which it belongs, and the domain describes the range of values ​​to which the membership function is applied. Each element in the classifier result, that is, the confidence corresponding to each category, is expressed by a fuzzy set: , the domain is described by fuzzy logic composed of all fault types. If the confidence value of a classifier for a certain type of sample is higher than 0.7, the "true" and "false" parts of the membership function on the corresponding domain should present right-skewed and left-skewed S-shaped curves respectively, and the steeper the slope, the higher the classifier's discrimination of the fault type; conversely, if the confidence value of the classifier for a certain type of sample is between 0.3 and 0.5, the corresponding membership function is moderate and gentle; remember Representation classifier Determine the input sample as Confidence of type, classifier In the domain The membership function on is , the process of late fusion of the classification results obtained after the samples are input into each classifier is as follows: The classification confidence of each fault type output by the classifier is mapped through the corresponding membership function to obtain the membership: ; in, Representation classifier Output The confidence level of the type is the membership of "true" and "false"; The new comprehensive membership is synthesized by subtracting the "false" membership from the "true" membership. According to the previous analysis, the larger the comprehensive membership, the more reliable the classifier is in identifying a certain type of sample: ; Summarize the comprehensive membership vectors of all classifiers to obtain the final classification result Y: ; 。 6. A method for fault diagnosis of switch machine based on segmented priori feature fusion according to claim 5, characterized in that: In step 6, after the text generation model is trained, each time a time series is input into the model, the model will determine whether the switch machine is faulty based on the training content, thereby achieving the desired effect.

7. The method for fault diagnosis of switch machine based on segmented priori feature fusion according to claim 1 is characterized in that: The autoencoder is an unsupervised neural network-based algorithm for distinguishing the similarity between the input value and the target value. Its structure consists of three parts: encoder, decoder, and bottleneck layer. The sequence to be encoded is : Encoder: Encodes the raw data into feature representations at the bottleneck layer : ; Bottleneck layer: Also known as the compression layer, it connects the encoder and decoder at both ends. The bottleneck layer is responsible for compressing the feature dimension and limiting the number of features entering the decoder ( ), which is equivalent to further abstracting the original data, thus forcing the network to learn only the most significant features; Decoder: Finally, the feature representation from the bottleneck layer Reconstruct the original data: ; The mean absolute error is used to measure and The difference between: ; The encoder and decoder are implemented with multi-layer LSTM and GRU recurrent neural networks, and the sequence after the controlled length is As the model input, the DRAE corresponding to each type of sample is trained, and the data set used for training is composed entirely of samples of this type. During the training process, DRAE will minimize the reconstruction error by continuously adjusting the parameters of the encoder and decoder. ,Finally, a series of DRAEs that can reconstruct the corresponding types of samples are obtained; Follow the following process Identify the fault type: Create a separate dataset for each type of sample and train DRAE; For the sample sequence to be identified To split; After the division Enter DRAE; Compute the reconstruction loss set , Each element in is composed of 3 stages of losses; Each type of sample behaves significantly differently in one or more stages. We assign weights to each stage according to its contribution to the result, and obtain a new loss set by weighting. ; The larger the loss, the more obvious the morphological difference, and the lower the confidence that the sample belongs to this type. Flip the values ​​to generate a confidence set : ; Among them, m is used to adjust the scaling ratio in the x-axis direction; Use the SoftMax function to scale the elements in C so that each element The sum is 1; Output the confidence set corresponding to all categories .

Citation Information

Patent Citations

  • Method and device for predicting fault of turnout switch machine

    CN109501834A

  • Fault diagnosis method and device based on action current curve of switch machine

    CN110441629A