Train axle temperature anomaly recognition method based on generative adversarial network and ensemble learning

Abstract

The application discloses a train axle temperature anomaly recognition method based on a generative adversarial network and ensemble learning, which comprises the following steps: collecting operation data in actual operation of an urban rail train, obtaining a two-class data set with unbalanced categories after preprocessing, and dividing the data set into a training set and a test set; training a constructor and a discriminator of the generative adversarial network by using abnormal data samples in the training set, and realizing automatic network parameter adjustment by using a Bayesian optimization algorithm; synthesizing abnormal samples by using the trained generative adversarial network model, and jointly constructing a training set with balanced categories with the original training set; filtering and screening noise samples by using a cross-committee filtering technology; constructing an axle temperature anomaly recognition classifier by using an AdaBoost method, training the ensemble learning model by using the training set, and inputting the test set to obtain a test result. The application solves the problems of missing of the axle temperature abnormal samples of the urban rail train and data imbalance, and improves the accuracy and correctness of the axle temperature anomaly recognition.

Description

Technical Field

[0001] This invention relates to the fields of data mining and urban rail train axle temperature anomaly detection, specifically to a train axle temperature anomaly identification method based on generative adversarial networks and ensemble learning. Background Technology

[0002] To ensure the safe and reliable operation of trains, online monitoring of key components and their operating status (e.g., axle temperature, pantograph wear, and steering knuckle condition) is a crucial aspect of ensuring the operational safety of urban rail train systems. Utilizing infrared radiation thermometry to dynamically measure bearing temperatures during train operation is one of the important means of guaranteeing railway safety and preventing axle overheating and subsequent axle burnout accidents.

[0003] Bearings are a critical component of urban rail trains, closely related to the safe operation of the train. Axle temperature is an important indicator for monitoring the health of train bearings. Currently, the main method for detecting abnormal axle temperature is to identify the bearing condition by establishing corresponding logical rules and temperature thresholds. However, this method has significant shortcomings. It does not consider the trend of axle temperature changes, does not systematically analyze historical data, and does not take into account the influence of factors such as speed and ambient temperature. It cannot detect and issue early warnings for axle temperature anomalies. This means that once such a temperature alarm occurs, it indicates that the train bearing has suffered significant damage, which may lead to a major accident.

[0004] In recent years, numerous methods have emerged for detecting bearing temperature anomalies using data from urban rail train operation. However, due to the scarcity of accidents in actual urban rail train operation, the field of axle temperature anomaly detection suffers from a severe lack of abnormal samples. This results in highly imbalanced datasets for many axle temperature anomaly detection methods, leading to low accuracy and precision in anomaly identification. Therefore, finding an axle temperature anomaly detection method that can effectively learn from abnormal sample information and significantly improve the accuracy and precision of anomaly identification is crucial. Data analysis from online axle temperature detection systems for urban rail trains reveals that bearing temperature anomalies can be categorized into three types: slight heating, strong heating, and extreme heating. The dataset formed by combining data from normal urban rail train operation with these three anomaly categories constitutes an imbalanced binary classification dataset. Therefore, urban rail train axle temperature anomaly detection can be abstracted as a task of classifying an imbalanced binary classification dataset.

[0005] For the classification problem of imbalanced data, researchers have proposed many solutions, with generating minority class samples considered an effective method to address class imbalance. Among them, Random Oversampling (ROS), Synthetic Minority Oversampling Technique (SMOTE), and Borderline SMOTE are considered the best traditional oversampling algorithms. However, when dealing with extremely imbalanced class problems, the performance of these traditional oversampling algorithms degrades significantly. Summary of the Invention

[0006] The purpose of this invention is to provide a train axle temperature anomaly identification method based on generative adversarial networks and ensemble learning. It combines the diverse features of the train and uses generative adversarial networks to overcome the problem of scarce abnormal samples in the field of axle temperature anomaly detection. At the same time, it adopts the AdaBoost algorithm in ensemble learning for classification training of axle temperature anomaly identification, thereby improving the overall performance of urban rail axle temperature anomaly identification in class imbalance scenarios.

[0007] The technical solution to achieve the purpose of this invention is: a method for identifying train axle temperature anomalies based on generative adversarial networks and ensemble learning, comprising the following steps:

[0008] Step 1: Preprocess the collected data from the actual operation of urban rail trains to obtain samples with normal and abnormal axle temperatures, forming a class-imbalanced binary classification dataset, which will be divided into training and testing sets.

[0009] Step 2: Construct the Generative Adversarial Network (CWGAN-GP) model. The CWGAN-GP model is a Conditional Wasserstein Generative Adversarial Network. It uses Wasserstein distance to measure the difference between the actual sample distribution and the generated sample distribution. The input information of the generator and discriminator is increased with data categories, and gradient penalty is used instead of weight pruning strategy to make the generated data more diverse.

[0010] Step 3: Use the shaft temperature anomaly samples in the training set to train the constructor and discriminator of the CWGAN-GP model, and use the Bayesian optimization algorithm to achieve automatic parameter tuning of the network.

[0011] Step 4: Use the trained CWGAN-GP model to generate anomalous samples and add them to the training set to obtain a class-balanced training set.

[0012] Step 5: Use the Cross-Council Filtering (CVCF) technique to filter and screen noisy samples generated during the sampling process;

[0013] Step 6: Construct an axle temperature anomaly identification classifier using the AdaBoost algorithm in ensemble learning to obtain an ensemble learning model. Train the ensemble learning model using the data-filtered training set. After training, test the classification effect on the test set to obtain various evaluation indicators.

[0014] Furthermore, in step 1, the collected data from the actual operation of the urban rail trains is preprocessed, specifically including:

[0015] Extract multi-data information about train operation, including train number, train type, direction of travel, main control terminal, ambient temperature, train speed, axle box position, axle box temperature rise, peak axle temperature, and average axle box temperature;

[0016] Calculate the ratio of the temperature of this axle box to the average temperature of the axle boxes in the same train, and record it as the same train ratio; calculate the ratio of the temperature of this axle box to the average temperature of all axle boxes in the train, and record it as the same vehicle ratio; extract the temperature values ​​of this axle box of the train in the most recent five times, and record them as the five query values; the same train ratio, same vehicle ratio, and five query values ​​together constitute the axle temperature multivariate information dataset.

[0017] Normal samples in the actual data are labeled as positive, all abnormal samples are labeled as negative, and the classification features are One-Hot encoded. The data features are standardized, and missing and duplicate data are deleted.

[0018] Furthermore, in step 2, the optimization function of the CWGAN-GP model is:

[0019]

[0020] Among them, Loss WGAN-GP Let z represent the optimization function of the model, y represent the label information, E represent the expected value, and p represent the noise. r and p z Representing the distribution of real data and the distribution of noisy data, λ is the penalty coefficient. It is a penalty item. It is real data x and synthetic data Point-by-point interpolation between These are the sampling distributions of these linear interpolations. It is the gradient of the discriminator's output relative to the interpolation;

[0021] The generator and discriminator networks of the CWGAN-GP model are constructed using three fully connected layers. The Batch Normalization (BN) and dropout layers are removed, and Adam is used as the optimization algorithm. The initial learning rate α = 0.0001, the first-order variable β1 = 0.5, and the second-order variable β2 = 0.9 are controlled. Random noise is sampled from a Gaussian distribution with an initial length of 100 and a penalty coefficient λ = 10.

[0022] Furthermore, in step 3, the constructor and discriminator of the CWGAN-GP model are trained using the shaft temperature anomaly samples in the training set. During the training process, the Bayesian optimization algorithm is used to achieve automatic parameter tuning of the network. Given the objective function, based on Bayes' theorem, the posterior distribution of the objective function is estimated by continuously adding sample points. Then, the hyperparameter combination for the next sampling is selected according to the distribution, and the process is iterated repeatedly until the estimated posterior distribution of the objective function approximates the true distribution.

[0023] Furthermore, in step 3, the batch size is set to the size of each training batch, and n is the number of iterations. The training process of the CWGAN-GP model is as follows:

[0024] (3.1) Load the dataset and filter out abnormal samples;

[0025] (3.2) Define a generator and a discriminator, each with 3 fully connected layers; the generator outputs data with the same dimension as the dataset as simulated samples; the discriminator outputs a numerical value with dimension 1 to represent the discrimination result;

[0026] (3.3) From the distribution p of the noise data z (z) Extract noise samples z from the real data distribution P r A real sample x is drawn from (x), and a random number ε∈[0,1] is selected;

[0027] (3.4) Obtain the synthesized sample through the generator:

[0028] (3.5) Randomly sample from real and synthetic data based on the penalty coefficient:

[0029] (3.6) Calculate the discriminator loss

[0030] (3.7) Calculate the gradient penalty term

[0031] (3.8) Repeat steps (3.3) to (3.7) until the training of k samples in this batch is completed;

[0032] (3.9) Update discriminator parameters

[0033] (3.10) Repeat steps (3.3) to (3.9) until n iterations;

[0034] (3.11) Take m samples from the noise space: And add conditional information:

[0035] (3.12) Calculate the generator loss and update the generator parameters:

[0036]

[0037]

[0038] (3.13) Repeat steps (3.3) to (3.12) until Nash equilibrium is reached.

[0039] Furthermore, in step 4, the process of constructing the class-balanced training set is as follows:

[0040] Calculate the number of positive and negative samples in the training set, generate abnormal samples using the trained CWGAN-GP model, and add them to the training set so that the difference between the number of normal samples and abnormal samples in the training set is within a set threshold, thus obtaining a class-balanced training set.

[0041] Furthermore, in step 6, the specific steps for constructing the shaft temperature anomaly identification classifier using the AdaBoost algorithm in ensemble learning are as follows:

[0042] (6.1) Initialize the weights for the original dataset;

[0043] (6.2) Train a weak learner using a weighted dataset;

[0044] (6.3) Calculate the weights of weak learning based on the error of the weak learner;

[0045] (6.4) Adjust the weights of the dataset;

[0046] (6.5) Repeat steps (6.2) to (6.4) K-1 times;

[0047] (6.6) The results of the K-1 weak learners are weighted and combined, and the specific formula is as follows:

[0048]

[0049]

[0050] Among them, H finall To ensemble the strong classifiers obtained through learning, C t (x) represents the t-th weak classifier in the decision tree, and the sign function is used to take the weighted sign of the results of the t weak classifiers. α t Let e ​​be the weight of the t-th weak classifier in the decision tree. t Let be the error rate of the weak classifier in the t-th decision tree.

[0051] Compared with the prior art, the present invention has the following significant advantages: (1) It overcomes the defects of traditional threshold-based temperature anomaly detection. Considering the trend of axle temperature change, it systematically analyzes historical data and takes into account the influence of factors such as speed and ambient temperature. It can detect axle temperature anomalies early and issue warnings, thus avoiding major accidents; (2) It combines the multi-features of the train and uses generative adversarial networks to overcome the problem of scarce abnormal samples in the field of axle temperature anomaly detection, thereby improving the accuracy and correctness of anomaly identification; (3) It adopts the AdaBoost algorithm in ensemble learning to perform classification training for axle temperature anomaly identification, thereby improving the overall performance of urban rail axle temperature anomaly identification in class imbalance scenarios. Attached Figure Description

[0052] Figure 1 This is a flowchart of the train axle temperature anomaly identification method based on generative adversarial networks and ensemble learning according to the present invention.

[0053] Figure 2 This is a network structure diagram of the CWGAN-GP generator and discriminator in the train axle temperature anomaly identification method based on generative adversarial networks and ensemble learning of the present invention.

[0054] Figure 3 This is a flowchart illustrating the training process of the CWGAN-GP module in the train axle temperature anomaly identification method based on generative adversarial networks and ensemble learning of this invention.

[0055] Figure 4 This is a Gaussian kernel density estimation plot of the synthetic sample and the real sample. Detailed Implementation

[0056] This invention provides a method for identifying axle temperature anomalies in urban rail trains based on improved generative adversarial networks and ensemble learning, comprising the following steps:

[0057] Step 1: Preprocess the collected data from the actual operation of urban rail trains to obtain samples with normal and abnormal axle temperatures, forming a class-imbalanced binary classification dataset, which will be divided into training and testing sets.

[0058] Step 2: Construct the Generative Adversarial Network (CWGAN-GP) model. The CWGAN-GP model is a Conditional Wasserstein Generative Adversarial Network. It uses Wasserstein distance to measure the difference between the actual sample distribution and the generated sample distribution. The input information of the generator and discriminator is increased with data categories, and gradient penalty is used instead of weight pruning strategy to make the generated data more diverse.

[0059] Step 3: Use the shaft temperature anomaly samples in the training set to train the constructor and discriminator of the CWGAN-GP model, and use the Bayesian optimization algorithm to achieve automatic parameter tuning of the network.

[0060] Step 4: Use the trained CWGAN-GP model to generate anomalous samples and add them to the training set to obtain a class-balanced training set.

[0061] Step 5: Use the Cross-Council Filtering (CVCF) technique to filter and screen noisy samples generated during the sampling process;

[0062] Step 6: Construct an axle temperature anomaly identification classifier using the AdaBoost algorithm in ensemble learning to obtain an ensemble learning model. Train the ensemble learning model using the data-filtered training set. After training, test the classification effect on the test set to obtain various evaluation indicators.

[0063] As a specific example, in step 1, the collected data from the actual operation of the urban rail train is preprocessed, specifically including:

[0064] Extract multi-data information about train operation, including train number, train type, direction of travel, main control terminal, ambient temperature, train speed, axle box position, axle box temperature rise, peak axle temperature, and average axle box temperature;

[0065] Calculate the ratio of the temperature of this axle box to the average temperature of the axle boxes in the same train, and record it as the same train ratio; calculate the ratio of the temperature of this axle box to the average temperature of all axle boxes in the train, and record it as the same vehicle ratio; extract the temperature values ​​of this axle box of the train in the most recent five times, and record them as the five query values; the same train ratio, same vehicle ratio, and five query values ​​together constitute the axle temperature multivariate information dataset.

[0066] Normal samples in the actual data are labeled as positive, all abnormal samples are labeled as negative, and the classification features are One-Hot encoded. The data features are standardized, and missing and duplicate data are deleted.

[0067] As a specific example, in step 2, the Generative Adversarial Network (CWGAN-GP) is a Conditional Wasserstein Generative Adversarial Network, which uses Wasserstein distance to measure the difference between the actual sample distribution and the generated sample distribution. Data categories are added to the input information of both the generator and the discriminator, and gradient penalty is used instead of weight pruning strategy to make the generated data more diverse.

[0068] The optimization function of the CWGAN-GP model is:

[0069]

[0070] Among them, Loss WGAN-GP Let z represent the optimization function of the model, y represent the label information, E represent the expected value, and p represent the noise. r and pz Representing the distribution of real data and the distribution of noisy data, λ is the penalty coefficient. It is a penalty item. It is real data x and synthetic data Point-by-point interpolation between These are the sampling distributions of these linear interpolations. It is the gradient of the discriminator's output relative to the interpolation.

[0071] The generative and discriminative networks of the CWGAN-GP model are constructed using three fully connected layers. Due to the introduction of GP, there is a conflict with the Batch Normalization (BN) layer that is usually added in general GANs and WGANs. Therefore, the BN layer and dropout layer are removed, and Adam is used as the optimization algorithm. The initial learning rate α = 0.0001, the first-order variable β1 = 0.5, and the second-order variable β2 = 0.9 are controlled. The random noise is sampled from a Gaussian distribution with an initial length of 100, and the penalty coefficient λ in WGAN-GP is 10.

[0072] As a concrete example, in step 3, the constructor and discriminator of the CWGAN-GP model are trained using shaft temperature anomaly samples from the training set. Bayesian hyperparameter tuning is faster with fewer iterations, while grid search is slower and prone to dimensionality explosion with many parameters. During training, a Bayesian optimization algorithm is used to automatically tune the network. Given a target function, based on Bayes' theorem, the posterior distribution of the target function is estimated by continuously adding sample points. Then, the next combination of hyperparameters is selected based on the distribution, iterating repeatedly until the estimated posterior distribution of the target function approximates the true distribution.

[0073] Furthermore, in step 3, the batch size is set to the size of each training batch, and n is the number of iterations. The training process of the CWGAN-GP model is as follows:

[0074] (3.1) Load the dataset and filter out abnormal samples;

[0075] (3.2) Define a generator and a discriminator, each with 3 fully connected layers; the generator outputs data with the same dimension as the dataset as simulated samples; the discriminator outputs a numerical value with dimension 1 to represent the discrimination result;

[0076] (3.3) From the distribution p of the noise data z (z) Extract noise samples z from the real data distribution P r A real sample x is drawn from (x), and a random number ε∈[0,1] is selected;

[0077] (3.4) Obtain the synthesized sample through the generator:

[0078] (3.5) Randomly sample from real and synthetic data based on the penalty coefficient:

[0079] (3.6) Calculate the discriminator loss

[0080] (3.7) Calculate the gradient penalty term

[0081] (3.8) Repeat steps (3.3) to (3.7) until the training of k samples in this batch is completed;

[0082] (3.9) Update discriminator parameters

[0083] (3.10) Repeat steps (3.3) to (3.9) until n iterations;

[0084] (3.11) Take m samples from the noise space: And add conditional information:

[0085] (3.12) Calculate the generator loss and update the generator parameters:

[0086]

[0087]

[0088] (3.13) Repeat steps (3.3) to (3.12) until Nash equilibrium is reached.

[0089] As a specific example, the process of constructing the class-balanced training set in step 4 is as follows:

[0090] Calculate the number of positive and negative samples in the training set, generate abnormal samples using the trained CWGAN-GP model, and add them to the training set so that the difference between the number of normal samples and abnormal samples in the training set is within a set threshold, thus obtaining a class-balanced training set.

[0091] As a specific example, in step 5, the purpose of using CVCF for noise filtering is that: negative class samples are very scarce in the original shaft temperature dataset. After some negative class samples are synthesized by CWGAN-GP, noisy class samples and class overlap samples may appear, which will affect the classification effect of the classifier.

[0092] As a specific example, in step 6, the specific steps for constructing a shaft temperature anomaly identification classifier using the AdaBoost algorithm in ensemble learning are as follows:

[0093] (6.1) Initialize the weights for the original dataset;

[0094] (6.2) Train a weak learner using a weighted dataset;

[0095] (6.3) Calculate the weights of weak learning based on the error of the weak learner;

[0096] (6.4) Adjust the weights of the dataset;

[0097] (6.5) Repeat steps (6.2) to (6.4) K-1 times;

[0098] (6.6) The results of the K-1 weak learners are weighted and combined, and the specific formula is as follows:

[0099]

[0100]

[0101] Among them, H finall To ensemble the strong classifiers obtained through learning, C t (x) represents the t-th weak classifier in the decision tree, and the sign function is used to take the weighted sign of the results of the t weak classifiers. α t Let e ​​be the weight of the t-th weak classifier in the decision tree. t Let be the error rate of the weak classifier in the t-th decision tree.

[0102] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0103] Example

[0104] This embodiment provides a method for identifying axle temperature anomalies in urban rail transit trains based on improved generative adversarial networks and ensemble learning. The method includes:

[0105] Step 1: Preprocess the collected axle temperature data from the actual operation of the urban rail train to obtain samples with normal and abnormal axle temperatures, forming a non-balanced binary classification dataset, and divide it into training and test sets.

[0106] Step 2: Use axis temperature anomaly samples to train the constructor and discriminator of the improved generative adversarial network model (CWGAN-GP);

[0107] Step 3: Use the trained CWGAN-GP model to generate anomalous samples and add them to the training set so that the number of anomalous samples among the normal samples in the training set is similar, thus obtaining a class-balanced training set.

[0108] Step 4: Use Cross-Committee Filtering (CVCF) to filter and screen noisy samples generated during the sampling process;

[0109] Step 5: Train AdaBoost using the cleaned training set. After training, test the classification performance on the test set to obtain various evaluation metrics.

[0110] In step 1, the dataset used in this invention comes from data collected by sensors on urban rail trains during actual operation, including both abnormal and normal samples. Preprocessing operations include extracting axle temperature multivariate features and removing duplicate or missing rows. Since the dataset contains categorical features and is discrete and unordered, the categorical feature columns are One-Hot encoded. Continuous number columns are standardized, scaling them to a mean of 0 and a variance of 1. The dataset is then divided into training and test sets according to a certain ratio.

[0111] In step 2, the generator and discriminator networks are constructed using three fully connected layers, removing the batch normalization (BN) and dropout layers. This is because the generator network requires one input to one output for the discriminator, while BN normalizes samples in a batch, meaning one batch of inputs corresponds to one batch of outputs. Therefore, BN cannot accurately calculate the gradient of the discriminator for each input sample. Furthermore, BN introduces dependencies between different samples within the same batch. The activation functions used in constructing the neural network include leaky ReLU and tanh. A sigmoid activation function is added to the discriminator to obtain class probabilities. Their mathematical expressions are:

[0112]

[0113]

[0114]

[0115] In addition, Adam is used as the optimization algorithm (with learning rate α = 0.0001, β1 = 0.5, β2 = 0.9). Random noise is sampled from a Gaussian distribution with a length of 100. The penalty coefficient λ in WGAN-GP is 10.

[0116] The specific training steps for CWGAN-GP are as follows, where n is the batch size for each training round and n is the number of iterations.

[0117] (1) Load the dataset and filter out abnormal samples;

[0118] (2) Define a generator and a discriminator, each with 3 fully connected layers. The generator outputs data with the same dimensions as the dataset as simulated samples. The discriminator outputs a 1-dimensional value representing its discrimination result;

[0119] (3) From the noise data distribution p z (z) Extract noise samples z from the real data distribution P r A real sample x is drawn from (x), and a random number ε∈[0,1] is selected;

[0120] (4) Obtain the synthesized sample through the generator:

[0121] (5) Randomly sample from real and synthetic data based on the penalty coefficient:

[0122] (6) Calculate the discriminator loss

[0123] (7) Calculate the gradient penalty term

[0124] (8) Repeat steps (3) to (7) until the training of k samples in this batch is completed;

[0125] (9) Update the discriminator parameters

[0126] (10) Repeat steps (3) to (9) until n iterations;

[0127] (11) Take m samples from the noise space: And add conditional information:

[0128] (12) Calculate the generator loss and update the generator parameters:

[0129]

[0130]

[0131] (13) Repeat steps (3) to (12) until Nash equilibrium is reached;

[0132] Meanwhile, since generative adversarial networks are quite sensitive to hyperparameters, Bayesian optimization is used to achieve automatic hyperparameter tuning. The core of Bayesian optimization is the probabilistic model and the acquisition function. The probabilistic model is used to represent the unknown objective function of the model to be evaluated. By iteratively updating the parameters of the probabilistic model, the estimated objective function distribution can gradually approach the true distribution.

[0133] Step 3 involves constructing a class-balanced training set, specifically including: first, calculating the number of positive and negative class samples in the training set, calculating the imbalance rate, generating outlier data samples using the trained CWGAN-GP model, and finally adding the synthesized outlier data samples to the training set to ensure that the number of positive and negative class samples in the training set is equal. Figure 4The figure shows the Gaussian kernel density estimation results for the generated samples and the real samples;

[0134] In step 4, the Cross-Council Filtering (CVCF) technique, as an effective noise filtering technique, can effectively identify noisy samples using a cross-validation strategy and remove them during the final classification process.

[0135] In step 5, the specific steps of AdaBoost are as follows:

[0136] (1) Initialize the weight distribution of the training data. Each sample is initially assigned the same weight, thus initializing the weight distribution.

[0137] (2) Select the weak classifier h with the lowest current error rate as the t-th basic classifier H. t And calculate the weak classifier h t :X→{-1,1}, this weak classifier is in distribution D t The error is:

[0138] (3) Calculate the weight of the weak classifier in the final classifier:

[0139] (4) Update the weight distribution of the training samples: Z t Normalization constant

[0140] (5) Repeat (2)~(4) iteratively, t=1,…T;

[0141] (6) According to the weight α of the weak classifier t By combining the various weak classifiers and applying the sign function, a strong classifier is obtained.

[0142] This embodiment uses an improved generative adversarial network and ensemble learning-based method for identifying axle temperature anomalies in urban rail transit trains. The following simulations are performed: A binary classification dataset for identifying axle temperature anomalies in urban rail transit trains is constructed. This dataset includes 2000 axle temperature data points, of which 256 are fault data points and 1744 are normal data points. After preprocessing, SMOTEBagging, SMOTEBoost, EasyEnsemble, BalanceCascade, and the proposed CWGAN-AdaBoost algorithm are used for classification training on the dataset. Performance metrics are compared on the training set. Accuracy, Recall, AUC, G-means, and F1 score are selected as evaluation metrics, and the average of 5-fold cross-validation is used as the final result for each evaluation metric. The confusion matrix constructed for classification is shown in Table 1.

[0143] Table 1 shows the confusion matrix constructed by the classification.

[0144]

[0145] Here, AUC corresponds to the classifier predicting a higher probability of a randomly selected positive class sample than a randomly selected negative class sample. This fundamentally avoids the impact of class imbalance on the evaluation metric. G-means evaluates the accuracy of the majority and minority classes in a balanced way. Another important feature of G-means is its non-linearity; misclassification has a greater impact on the G-mean metric of the minority class, thus avoiding losses caused by ignoring important minority class samples. F1 score is an average, a result of averaging precision and recall, taking both into account. Their calculation formulas are:

[0146]

[0147]

[0148]

[0149]

[0150]

[0151]

[0152] Table 2 Comparison of Experimental Results

[0153]

[0154] The experimental results of this invention are shown in Table 2. As can be seen from Table 2, the method proposed in this invention outperforms SMOTEBagging, SMOTEBoost, EasyEnsemble, and BalanceCascade in all evaluation metrics on the urban rail train axle temperature anomaly identification and classification dataset.

[0155] In summary, this invention overcomes the shortcomings of traditional threshold-based shaft temperature anomaly identification. It combines multi-dimensional shaft temperature information to form a dataset, and uses a generative adversarial network to generate scarce anomaly samples to reduce the imbalance of the dataset. At the same time, it employs the CVCF algorithm to mitigate the impact of noise on classification performance, and uses AdaBoost ensemble to classify the dataset. Experimental results show that the method of this invention improves the overall recognition performance of shaft temperature anomaly identification.

Claims

1. A train axle temperature anomaly identification method based on a generative adversarial network and ensemble learning, characterized in that, The method comprises the following steps: Step 1, preprocessing the collected data in the actual operation of the urban rail train to obtain normal and abnormal axle temperature samples, forming a class-unbalanced binary classification data set, and dividing into a training set and a test set; Step 2, constructing a generative adversarial network model, namely a CWGAN-GP model, the CWGAN-GP model being a conditional Wasserstein generative adversarial network, adopting a Wasserstein distance to measure the difference between an actual sample distribution and a generated sample distribution, increasing data classes in input information of the generator and the discriminator, and adopting a gradient penalty instead of a weight pruning strategy to make the generated data more diversified; Step 3, training the constructor and the discriminator of the CWGAN-GP model by using the abnormal axle temperature samples in the training set, and realizing automatic network parameter adjustment by using a Bayesian optimization algorithm; Step 4, generating abnormal samples by using the trained CWGAN-GP model and adding the abnormal samples to the training set to obtain a class-balanced training set; Step 5, filtering noise samples generated in the sampling process by using a cross committee filtering technology, CVCF; Step 6, constructing an axle temperature abnormality recognition classifier by using an AdaBoost algorithm in ensemble learning to obtain an ensemble learning model, training the ensemble learning model by using the training set after data filtering, verifying the classification effect on the test set after the training is completed, and obtaining various evaluation indexes.

2. The train wheel-temperature anomaly identification method based on a generative adversarial network and ensemble learning according to claim 1, wherein, In step 1, the collected data in the actual operation of the urban rail train is preprocessed, specifically including: Extracting multivariate data information of train operation, including train number, train type, running direction, main control end, environmental temperature, train speed, axle box position, axle box temperature rise, axle temperature peak value, and axle box temperature average value; Calculating a ratio of the axle box to the average temperature of the same axle box of the train, which is called a same-column ratio; calculating a ratio of the axle box to the average temperature of all axle boxes of the train, which is called a same-vehicle ratio; extracting the temperature values of the axle box of the train for the last five times, which are called five query values; the same-column ratio, the same-vehicle ratio and the five query values jointly constitute an axle temperature multivariate information data set; Marking normal samples in the actual data as positive classes, marking all abnormal class samples as negative classes, and performing One-Hot coding on the classification features, standardizing the data features, and deleting missing and repeated data.

3. The train wheel-axle temperature anomaly identification method based on a generative adversarial network and ensemble learning according to claim 1, wherein, In step 2, the optimization function of the CWGAN-GP model is: where Loss WGAN-GP represents the optimization function of the model, z represents noise, y represents label information, E represents expectation value, p r and p z represent the real data distribution and the noise data distribution, λ is a penalty coefficient, is a penalty term, is a point-by-point interpolation between the real data x and the synthesized data , is the sampling distribution of these linear interpolations, is the gradient of the output of the discriminator with respect to the interpolation; A three-layer fully connected layer is adopted to construct the generation network and the discrimination network of the CWGAN-GP model, the BN layer and the dropout layer are removed, Adam is used as the optimization algorithm, the initial learning rate α is 0.0001, the first-order variable β1 is 0.5, the second-order variable β2 is 0.9; random noise is sampled from a Gaussian distribution, the initial length is 100, and the penalty coefficient λ is 10.

4. The train wheel-axle temperature anomaly identification method based on the generative adversarial network and the ensemble learning according to claim 3, wherein, In step 3, the constructor and discriminator of the CWGAN-GP model are trained using the shaft temperature anomaly samples in the training set. During the training process, the Bayesian optimization algorithm is used to realize the automatic parameter tuning of the network. In the given objective function, based on Bayesian theorem, the posterior distribution of the objective function is estimated by continuously adding sample points. Then, the next sample parameter combination is selected according to the distribution, and the iteration is repeated until the estimated posterior distribution of the objective function approximates the true distribution.

5. The train wheel-axle temperature anomaly identification method based on the generative adversarial network and the ensemble learning according to claim 4, wherein, In step 3, set the batch size as the batch size of each training, and n as the number of iterations. The training process of the CWGAN-GP model is as follows: (3.1) Load the data set and filter the abnormal samples; (3.2) Define the generator and discriminator, each with 3 fully connected layers. The generator finally outputs data with the same dimension as the data set as simulated samples. The discriminator outputs a value with a dimension of 1 representing the discrimination result; (3.3) draw a noise sample z from the noise data distribution p z (z) draw a real sample x from the real data distribution P r (x) draw a real sample x from the real data distribution P (3.4) Obtain synthetic samples by generator: (3.5) Random sampling of real data versus synthetic data according to the penalty coefficient: (3.6) Compute discriminator loss (3.7) Computing the gradient penalty term (3.8) Repeat steps (3.3) to (3.7) until the training of the k samples in this batch is completed; (3.9) updating the discriminator parameters (3.10) Repeat steps (3.3) to (3.9) until the iteration is n times; (3.11) Take m samples from the noise space: and add conditional information: (3.12) Calculate the generator loss and update the generator parameters: (3.13) Repeat steps (3.3) to (3.12) until the Nash equilibrium is reached.

6. The train wheel-axle temperature anomaly identification method based on a generative adversarial network and ensemble learning according to claim 1, wherein, In step 4, the class-balanced training set construction process is as follows: Calculate the number of positive and negative samples in the training set, generate abnormal samples using the trained CWGAN-GP model, and add them to the training set to make the difference between the number of normal samples and abnormal samples in the training set within a certain threshold, and obtain a class-balanced training set.

7. The train wheel-axle temperature anomaly identification method based on a generative adversarial network and ensemble learning according to claim 1, wherein, In step 6, the specific steps of constructing the shaft temperature anomaly recognition classifier using the AdaBoost algorithm in ensemble learning are as follows: (6.1) Initialize the weight of the original data set; (6.2) Train the weak learner with the weighted data set; (6.3) Calculate the weight of the weak learner according to the error of the weak learner; (6.4) Adjust the weight of the data set; (6.5) Repeat steps (6.2) to (6.4) K-1 times; (6.6) Weighted combination of the results of K-1 weak learners, the specific formula is: where H finall is the strong classifier obtained by ensemble learning, C t (x) is the tth decision tree weak classifier, the sign function is used to take the positive and negative of the weighted results of the t weak classifiers, a t is the weight of the tth decision tree weak classifier, e t is the error rate of the tth decision tree weak classifier.

Images