A high-speed train shock absorber fault detection method combined with auxiliary enhancement

Abstract

The present invention discloses a high-speed train shock absorber fault detection method combined with auxiliary enhancement, and relates to the field of railway transportation technology. The method includes constructing a shock absorber fault detection model, including a main task subnetwork and an auxiliary task subnetwork connected to a feature extraction network respectively; inputting the training samples into the main task subnetwork and the auxiliary task subnetwork respectively through the feature extraction network, calculating the total loss through a learnable weighted coefficient, and training the network parameters of the shock absorber fault detection model; inputting the monitoring data to be detected into the shock absorber fault detection model, obtaining the main task output and thresholding it, and judging whether the corresponding shock absorber is faulty according to the thresholded elements. The high-speed train shock absorber fault detection method can effectively locate the shock absorber with faults, and has the technical advantages of low cost, strong adaptability, high accuracy and high transferability; further, training in combination with auxiliary tasks helps to improve training efficiency and the credibility and accuracy of fault detection.

Description

Technical Field

[0001] The invention relates to the technical field of railway transportation, and in particular to a high-speed train shock absorber fault detection method combined with auxiliary enhancement. Background Art

[0002] In terms of high-speed train shock absorber failures, traditional fault detection methods mainly rely on physical models, empirical rules and signal processing technology. Their limitations are gradually revealed in the complex environment of high-speed train operation, especially when dealing with complex fault types and early fault scenarios. The large number of sensors installed on high-speed trains provides a data basis for shock absorber fault detection, but because different vehicle types and line conditions may cause shock absorbers to exhibit different characteristics, the development and migration cost of detection methods based on physical models is high and their applicability is poor.

[0003] With the rapid development of machine learning technology, single-task learning methods have begun to be introduced into the field of fault detection. Single-task learning methods focus on specific equipment or fault types and rely on a large amount of labeled data to build dedicated models. Since high-speed train suspension systems usually contain multiple shock absorbers, their use environments and working conditions may vary. Building a separate detection model for each shock absorber will increase the cost of data acquisition and annotation, and the generalization ability is poor, making it difficult to adapt to changes under different working conditions, resulting in limited accuracy and efficiency of fault detection.

[0004] In addition, the single-task learning method is often not comprehensive enough in feature extraction, resulting in low information diversity, incomplete features, and lack of migration ability, which reduces the adaptability of the model. When dealing with various faults of high-speed train shock absorbers, the single-task learning method is difficult to fully utilize information of different granularities and angles, which limits the effectiveness of overall fault detection. Summary of the invention

[0005] In order to solve the above-mentioned technical problems existing in the prior art, the present invention aims to provide a shock absorber fault detection method which can adapt to different working conditions, has excellent generalization ability, and has high detection accuracy, detection efficiency and effectiveness.

[0006] The present invention provides a high-speed train shock absorber fault detection method combined with auxiliary enhancement, comprising the following steps:

[0007] Step S1: constructing a shock absorber fault detection model, including a feature extraction network, a main task subnetwork and an auxiliary task subnetwork; the main task subnetwork and the auxiliary task subnetwork are respectively connected to the feature extraction network;

[0008] Step S2: training a shock absorber fault detection model, including:

[0009] The training samples are input into the shock absorber fault detection model, and the cross entropy loss of the main task sub-network and the mean square error loss of the auxiliary task sub-network are calculated respectively;

[0010] The cross entropy loss and the mean square error loss are combined to obtain the combined loss and minimized to optimize the network parameters of the shock absorber fault detection model;

[0011] Step S3: input the multi-channel monitoring data of the suspension system to be detected into the trained shock absorber fault detection model to obtain the main task output and threshold it;

[0012] Among them, if the first element of the thresholded main task output is 1, there is no shock absorber fault in the suspension system; if the non-first element is 1, there is a fault in the corresponding shock absorber.

[0013] Preferably, step S2 specifically includes:

[0014] Step S21: Get the The monitoring dataset of the training suspension system with shock absorbers is divided into training samples;

[0015] Step S22: training samples , extract its feature representation through the feature extraction network and obtain the training feature matrix , the formula is as follows:

[0016] ;

[0017] In the formula, is the activation function of the feature extraction network, and are the weight matrix and bias vector of the feature extraction network respectively;

[0018] Step S23:

[0019] 1) The training feature matrix Input the main task sub-network and get the main task training output , that is, the size is The training probability vector of

[0020] The formula is as follows:

[0021] ;

[0022] In the formula, represents the normalized exponential function, and Represent the weight matrix and bias vector of the main task sub-network respectively;

[0023] Main task training output The first element is the probability that there is no shock absorber failure in the training suspension system. The element is the shock absorber failure located at The probability of a shock absorber; where ;

[0024] Through thresholding, the main task training output The element with the largest median value is set to 1, and the rest of the elements are set to 0;

[0025] 2) The training feature matrix Input the auxiliary task sub-network and get a size of Output of auxiliary task training ,

[0026] The formula is as follows:

[0027] ;

[0028] In the formula, and Represent the weight matrix and bias vector of the auxiliary task sub-network respectively;

[0029] Auxiliary task training output In The value of the element represents the The performance surplus of each shock absorber, ;

[0030] Step S24:

[0031] Calculate the cross entropy loss of the main task sub-network as , the formula is as follows:

[0032] ;

[0033] In the formula, Indicates the fault category index, , is the indicative function, 1 when it is, otherwise 0;

[0034] Calculate the mean squared error loss of the auxiliary task subnetwork , the formula is as follows:

[0035] ;

[0036] In the formula, For training samples The corresponding true label, i.e., the true value of performance degradation;

[0037] Step S25: Calculate the combined loss of the shock absorber fault detection model , the formula is as follows:

[0038] ;

[0039] In the formula, Loss weight for the main task; Loss weight for auxiliary tasks;

[0040] Step S26: Minimize the combined loss , to optimize the network parameters of the shock absorber fault detection model.

[0041] Preferably, step S26 also includes dynamically adjusting the main task loss weight and auxiliary task loss weights The method is as follows:

[0042] Optimizing the combined loss of the shock absorber fault detection model , the formula is as follows:

[0043] ;

[0044] Calculate the combined loss function through the back-propagation algorithm The gradient of the main task loss weight is dynamically adjusted according to the gradient and auxiliary task loss weights .

[0045] Preferably, step S3 specifically includes:

[0046] Step S31: The high-speed train includes The suspension system of the shock absorber samples multi-channel monitoring data through the monitoring system; ;

[0047] Step S32: extracting feature representation of multi-channel monitoring data through a feature extraction network to obtain a feature matrix;

[0048] Step S33: Input the feature matrix into the main task sub-network to obtain the main task output , that is, a size of The probability vector of the main task output The first element is the probability that there is no shock absorber failure in the suspension system. The element is The probability that a shock absorber is faulty; among them, ;

[0049] Through thresholding, the main task output The element with the largest median value is converted to 1, and the remaining elements are converted to 0;

[0050] For the thresholded main task output , if the first element is 1, it is determined that there is no shock absorber fault in the suspension system;

[0051] Jordi The value of the first element is 1, then One shock absorber is faulty, and the rest are normal.

[0052] Furthermore, step S3 also includes:

[0053] Step A: Input the multi-channel monitoring data of the suspension system to be tested into the optimized shock absorber fault detection model, and each element output by the auxiliary task represents the performance surplus of the corresponding shock absorber.

[0054] Preferably, step A specifically includes:

[0055] For high-speed trains, including The suspension system of the shock absorber samples multi-channel monitoring data through the monitoring system; ;

[0056] Extract feature representation of multi-channel monitoring data through feature extraction network to obtain feature matrix;

[0057] Input the feature matrix into the auxiliary task sub-network to obtain a size of Output of auxiliary tasks ;

[0058] Output for auxiliary tasks , No. The value of the element represents the The performance surplus of a shock absorber has a value range of [0, 1]; .

[0059] Preferably, the feature extraction network is a Transformer encoder, a multi-layer one-dimensional convolutional network, a multi-layer recurrent neural network, an autoencoder or a residual network.

[0060] Preferably, the main task subnetwork is a single-layer fully connected network, a multi-layer fully connected network connected in sequence, a single-layer one-dimensional convolutional network, or a multi-layer one-dimensional convolutional network connected in sequence.

[0061] Preferably, the auxiliary task subnetwork is a single-layer fully connected network, a multi-layer fully connected network connected in sequence, a single-layer one-dimensional convolutional network, or a multi-layer one-dimensional convolutional network connected in sequence.

[0062] The technical solution provided by the present invention can realize the detection of shock absorber faults under different line conditions, and can effectively locate the faulty shock absorber in the suspension system. It has the technical advantages of low cost, simple debugging process, high accuracy and good adaptability. Furthermore, combined with auxiliary task training, it helps to improve the training efficiency of the shock absorber fault detection model and the credibility and accuracy of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of experimental setting of training data in one embodiment of the present invention.

[0064] Figure 2 Schematic diagram of a network architecture for shock absorber fault detection in one embodiment of the present invention.

[0065] Figure 3 This is a comparison chart of training results of various detection methods in one embodiment of the present invention.

[0066] Figure 4 FIG. 4 is a curve diagram of the loss weight based on uncertainty in training according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The technical solution provided by the present invention will be further explained in detail below with reference to the accompanying drawings.

[0068] Example 1

[0069] Based on multi-task auxiliary enhancement, this embodiment provides a shock absorber fault detection method, which emphasizes the important role of auxiliary tasks in improving the performance of the main task. Auxiliary task learning is used as a special form of multi-task learning, and the performance of the main task is improved by introducing auxiliary tasks related to the main task. In the neural network model, the shock absorber fault detection task is set as the main task, and the shock absorber performance degradation estimation task is set as the auxiliary task, and the two tasks are optimized simultaneously through the multi-task learning framework. Through the main and auxiliary shared feature extraction layer, the model can effectively learn the shared representation between the two tasks. In order to further optimize the effect of multi-task learning, this embodiment also introduces a method based on uncertainty weighted loss. This method estimates the uncertainty of each task loss and dynamically adjusts the contribution of each task to the overall loss function, ensuring that the model can handle multiple tasks more fairly and effectively. The shock absorber fault detection method based on auxiliary task enhancement can not only improve the accuracy of detection, but also enhance the model's adaptability to changes in shock absorber performance, providing strong support for the maintenance and detection of shock absorbers.

[0070] Specifically, the steps include:

[0071] Step 1: Acquire multi-channel sensor monitoring data from the high-speed train monitoring system.

[0072] Step 2: Use the main neural network as the feature extraction network to extract feature representations from multi-channel monitoring data and obtain the feature matrix , and share feature representations in subsequent steps to optimize feature utilization and improve overall performance. The formula is as follows:

[0073] ;

[0074] In the formula, Represents multi-channel monitoring data, is the activation function of the feature extraction network, and are the weight matrix and bias vector of the feature extraction network respectively.

[0075] Among them, the main neural network is a Transformer encoder, a multi-layer one-dimensional convolutional network, a multi-layer recurrent neural network, an autoencoder or a residual network.

[0076] Step 3: The present invention models the shock absorber fault detection problem as a multi-task supervised learning problem, in which the main task and auxiliary task focus on the fault classification and performance degradation estimation of the shock absorber respectively. The main task is designed as a multi-classification task, aiming to distinguish the normal state from the faults of shock absorbers in different positions. Its sub-network consists of fully connected layers, and a softmax activation function (normalized exponential function) is used in the output layer to generate the probability distribution of each category. The auxiliary task is designed as a multi-output regression task, the goal is to estimate the performance degradation level of each shock absorber, and its sub-network is also composed of fully connected layers. The output layer does not use an activation function and directly generates a continuous value to reflect the performance surplus.

[0077] Feature matrix obtained by feature extraction network from monitoring data The main task subnetwork and the auxiliary task subnetwork are shared. The main task subnetwork can be a single-layer fully connected network, a multi-layer fully connected network connected in sequence, a single-layer one-dimensional convolutional network, or a multi-layer one-dimensional convolutional network connected in sequence. The auxiliary task subnetwork can be a single-layer fully connected network, a multi-layer fully connected network connected in sequence, a single-layer one-dimensional convolutional network, or a multi-layer one-dimensional convolutional network connected in sequence.

[0078] Main task output is a classification probability:

[0079] ;

[0080] In the formula, represents the normalized exponential function, and They represent the weight matrix and bias vector of the main task sub-network respectively.

[0081] Main task output The first element is the probability that there is no shock absorber failure in the training suspension system. The element is the shock absorber failure located at The probability of a shock absorber; where , is the number of shock absorbers in the suspension system to be tested.

[0082] Through thresholding, the main task output The element with the largest median value is set to 1, and the remaining elements are set to 0.

[0083] Auxiliary task output performance degradation estimation, i.e., performance surplus :

[0084] ;

[0085] In the formula, and They represent the weight matrix and bias vector of the auxiliary task sub-network respectively.

[0086] For the thresholded main task output , if the first element is 1, it is determined that there is no shock absorber fault in the suspension system;

[0087] Jordi The value of the first element is 1, then One shock absorber is faulty, and the rest are normal.

[0088] Performance surplus , No. The value of the element represents the The performance surplus of a shock absorber has a value range of [0, 1], which is used to represent the percentage of the shock absorber's damping performance relative to the fully healthy state. .

[0089] The outputs of the main task and the auxiliary task are jointly provided as fault detection results to the on-board monitoring system and to the train monitor, where the main task provides a coarse-grained fault judgment and the auxiliary task provides a fine-grained fault degree judgment. The main task and the auxiliary task share the feature extraction network, and the auxiliary task is used to enhance the feature extraction expression.

[0090] Training the shock absorber fault detection model includes the following steps:

[0091] Step 1: To test the performance of the present invention in shock absorber fault detection, the performance degradation of four lateral shock absorbers distributed on two bogies of the train is simulated. The four lateral shock absorbers are simulated in normal state and 0% to 90% performance degradation state. Train monitoring data with different performance degradation degrees at different positions are obtained, including acceleration sensors installed at different positions of the car body and bogies, and the sensor measurement directions are respectively along the longitudinal, transverse and vertical directions of the train.

[0092] The construction process of the dataset is based on incomplete sample distribution, that is, for the shock absorber training samples, the performance degradation range is randomly distributed in the range of 0% to 90%, and there are cases where there are no samples in a specific range. These missing state information are considered completely unknown, and the model cannot obtain this information during the training process to simulate the incompleteness and uncertainty of data in the real world. The first lateral shock absorber fault (LatB1F) and the second lateral shock absorber fault (LatB2F) on the front bogie of the high-speed train, as well as the first lateral shock absorber fault (LatB1T) and the second lateral shock absorber fault (LatB2T) on the rear bogie, the training data settings of the four lateral shock absorber faults are as follows: Figure 1 The test set (experimental training set 1-experimental training set 4) will use data samples of all performance degradation states to verify the generalization ability of the model.

[0093] Step 2: First, the fault detection of the shock absorber is constructed as a supervised learning task, where the main task is the fault detection of the shock absorber, that is, distinguishing the normal state from the faults at different positions. Therefore, the main task can be constructed as a multi-classification task of class, is the number of categories in the classification task, Is categories, including 1 normal state category of shock absorbers, where: Category Each shock absorber is tested separately There are a number of fault types, each shock absorber corresponds to a fault and different shock absorber faults occur in different locations.

[0094] The auxiliary task is to The performance of each shock absorber is estimated. Estimating the degree of performance degradation helps to understand the health status and remaining service life of the shock absorber. Shock absorbers are located in different positions of the suspension system. Each shock absorber may have unique performance characteristics and degradation rates due to different environmental conditions. The main goal is to estimate the degree of performance degradation of the shock absorber, which can be constructed as Outputs (for shock absorber), that is, the model simultaneously predicts By providing performance degradation information, the auxiliary task helps the main task better understand the working status of each shock absorber, thereby improving the accuracy of fault detection.

[0095] The main task is to classify the faults of the shock absorber, which is a multi-classification task. In this task, the goal is to distinguish the normal state of the shock absorber from the There are different types of failures, so there are kind( type of fault + 1 normal state); the auxiliary task corresponds to The performance degradation estimation of each shock absorber is a multi-output regression task. The goal is to estimate the performance degradation level of each shock absorber. Therefore, outputs (one estimated value of performance degradation for each shock absorber). Indicates that training samples and The input data of features, Is the input data The number of training samples in is the number of features of each training sample. Each training sample is usually composed of Features represent the sensor data of the shock absorber in different states. The training samples will be used to train the model so that it can learn to distinguish different fault states and evaluate the degree of shock absorber performance degradation.

[0096] Input Data Include training samples, among which, training samples , extract its feature representation through the feature extraction network and obtain the training feature matrix , the formula is as follows:

[0097] ;

[0098] The training feature matrix Input the main task sub-network and get the main task training output , that is, the size is The training probability vector is as follows:

[0099] ;

[0100] The first element is the probability that there is no shock absorber failure in the training suspension system, and the second The element is the shock absorber failure located at The probability of a shock absorber; where .

[0101] Through thresholding, the main task training output The element with the largest median value is set to 1, and the remaining elements are set to 0.

[0102] The training feature matrix Input the auxiliary task sub-network and get a size of Output of auxiliary task training , the formula is as follows:

[0103] ;

[0104] Among them, The value of the element represents the The performance surplus of each shock absorber, .

[0105] Calculate the cross entropy loss of the main task sub-network as , the formula is as follows:

[0106] ;

[0107] In the formula, Indicates the fault category index, , is the indicative function, 1 if the value is 0, otherwise it is 0.

[0108] The mean square error loss of the multi-output regression task is calculated by the predicted output of the above auxiliary multi-output regression task. Its purpose is to calculate the square of the error between the true label and the predicted label of each sample, and finally evaluate the prediction performance of the model by the average error of all samples. The formula is as follows:

[0109] ;

[0110] In the formula, For sample The true label of the corresponding shock absorber, that is, the sample The true regression value of , the true value of the shock absorber performance degradation.

[0111] The total loss function can be achieved by weighting the losses of the two tasks and balancing the objectives of the two tasks by taking a weighted summation approach, where: and Represent the loss weights of the main task and auxiliary task respectively, then the combined loss function of multi-task learning is:

[0112] ;

[0113] By minimizing the combined loss , in order to optimize the network parameters of the shock absorber fault detection model, the training of the model can be completed.

[0114] Since different tasks in multi-task learning scenarios are unbalanced and the loss functions behave differently, an uncertainty-based loss weighting method is used to balance the losses of the main task and auxiliary tasks, see step 4.

[0115] Step 3: Use the Transformer model as a feature extraction network to handle the shock absorber fault detection task. Transformer is a neural network model based on the self-attention mechanism. It has been widely used in various sequence data processing tasks and has shown strong capabilities in feature learning.

[0116] In the present invention, only the encoder part of the Transformer model is used to extract the feature representation of the monitoring signal, ensuring that this part of the network can be shared in the classification task (fault detection) as the main task and the regression task (performance degradation estimation, i.e., calculating the performance surplus) as the auxiliary task, so as to improve the efficiency of feature utilization. Although the encoder part of the Transformer model is used for feature extraction, in order to adapt to the different requirements of the main task and the auxiliary task, the present invention uses the fully connected layer as the subnetwork structure for the main task subnetwork and the auxiliary task subnetwork. The difference is that the main task subnetwork uses the softmax activation function to achieve multi-classification, while the auxiliary task is a regression task, the goal is to estimate the performance degradation degree of the shock absorber, the output of the regression task is a continuous value, and the output value does not need to be normalized, so the activation function is not used in the output layer of the auxiliary task subnetwork. This not only ensures that the main task and the auxiliary task can share the same feature extraction network, but also improves the computational efficiency and task adaptability of the model.

[0117] The network architecture of the present invention is as follows Figure 2 As shown in the figure, the main task sub-network is mainly used for the shock absorber fault detection task, which is a multi-classification task whose goal is to distinguish different fault types and normal states. The output layer of the main task uses a softmax activation function to represent each category probabilistically for classification. The auxiliary task is a regression task to estimate the degree of performance degradation of the shock absorber. The output layer of the regression task has no activation function because the output of the regression task is a continuous value and the output value does not need to be normalized. Through the shared feature extraction network, these features are fed into the network layers of the main task and the auxiliary task for fault classification and performance degradation estimation, respectively.

[0118] Step 4: Construct a multi-task learning model for shock absorber fault detection through step 2, where the auxiliary task is the estimation of shock absorber performance degradation, and the main task is shock absorber fault detection. The main task is usually a classification task that outputs class probabilities, while the auxiliary task is a regression task that outputs continuous values. The loss functions of these two tasks (such as classification loss and regression loss) may differ in numerical scales, targets, and training dynamics. The numerical range and importance of the loss functions of these two tasks may differ during the optimization process. If the losses are summed directly without weighting, the auxiliary task may dominate the training process, resulting in a decrease in the training effect of the main task and affecting the performance of shock absorber fault detection. In order to optimize the influence of the auxiliary task, the total loss function is a weighted combination of the losses of the two tasks. The purpose of the weighted loss function is to balance the influence of different tasks on model training, so that each task can be properly valued and avoid deviations between tasks; however, different tasks in multi-task learning scenarios are unbalanced. Each task in multi-task learning may face different uncertainties, and a simple fixed weight weighting method may not be able to dynamically adapt to changes between tasks.

[0119] The uncertainty-based loss weighting method is introduced to balance the loss of the main task and the auxiliary task, and the weight of the loss function is dynamically adjusted according to the uncertainty of each task, so as to ensure that the two tasks are balanced and optimized during the training process, so as to improve the accuracy of shock absorber fault detection. The uncertainty-based loss weighting method is used to balance the loss of the main task and the auxiliary task, and the weight of the loss function of each task is dynamically adjusted by introducing the homoscedastic uncertainty in Bayesian modeling. Homoscedastic uncertainty is an important concept in Bayesian modeling, which describes the uncertainty associated with the task itself, which remains constant in all input data but varies between different tasks. In multi-task learning, different tasks may have different complexities or uncertainties, so homoscedastic uncertainty can be used as a basis for dynamically adjusting the task loss weight, so as to ensure that the loss of each task is reasonably balanced during the optimization process.

[0120] The introduction of homoscedastic uncertainty in Bayesian modeling enables the loss weighting method to not only consider the prediction error of the task, but also reflect the uncertainty of the task. Specifically, homoscedastic uncertainty can dynamically adjust the weight according to the difficulty or predictability of each task, so that the model can more effectively balance the influence between tasks and avoid some tasks from dominating or being ignored in the training process due to too high or too low weights. In the multi-task learning model of shock absorber fault detection, the main task is more important or more complex than the auxiliary task, so it is hoped that more attention should be paid to the main task in the loss function. Through homoscedastic uncertainty in Bayesian modeling, appropriate weights can be assigned to each task according to its characteristics, thereby improving the accuracy of fault detection and optimizing the training process of the model. The uncertainty-based loss weighting method models uncertainty in multi-task learning by maximizing the Gaussian likelihood with homoscedastic uncertainty. For the multi-classification task corresponding to the main task, the likelihood is modeled by softmax on the output, while for the auxiliary regression task, the likelihood is modeled as a Gaussian distribution with the model output as the mean. The core multi-task likelihood function combines the outputs of different tasks to form a loss function based on data-dependent uncertainty, which can adaptively weight the losses of each task.

[0121] The goal of a multi-output model is to minimize a combined loss function that contains both continuous and discrete outputs. The task-specific noise parameters control the relative weight of each task loss (i.e., the noise parameters include the main task loss weight and auxiliary task loss weights ), which enables the model to learn the relative importance of each task loss during training. This method also introduces the idea of ​​predicting logarithmic variance, which is more numerically stable and avoids the problem of division by zero. This method effectively prevents task weights from converging to zero and ensures that each task can make a balanced contribution during the optimization process. The final loss function is smooth, differentiable, and applicable to a combination of continuous and discrete outputs.

[0122] In the shock absorber fault classification task, the main task sub-network can generate the predicted probability distribution of each shock absorber fault state, indicating the possibility of the shock absorber being in different fault states. The formula is as follows:

[0123] ;

[0124] in, ;

[0125] In the formula, Output for the main task, i.e. classification output; is the mapping of the main task sub-network, represents the weight matrix of the main task sub-network, Represents the bias vector of the main task sub-network.

[0126] is the categorical output of the target variable The predicted probability distribution of Represents the probability distribution of the shock absorber failure state output by the main task subnetwork.

[0127] In the auxiliary task, when the output is one or more vectors, each output can be regarded as an independent Gaussian distribution, and the joint probability density of these distributions can be expressed as the product of Gaussian distributions. This method is suitable for multi-task learning or multi-output regression problems, where each output (or task) has an independent noise model, as follows:

[0128] ;

[0129] in, ;

[0130] In the formula, Output for auxiliary tasks, i.e. regression output; is the mapping of the auxiliary task sub-network, represents the weight matrix of the auxiliary task sub-network, Represents the bias vector of the auxiliary task sub-network. is the regression output of the target variable The predicted probability distribution of Indicates the output of the given auxiliary task subnetwork The probability distribution of the shock absorber performance surplus under the condition of . is a Gaussian distribution, and the regression output The probability distribution is a is the mean, is a normal distribution with a variance of .

[0131] The minimization objective of the multi-output model in the regression task is adjusted, and the model is optimized by maximizing the log-likelihood function. The noise and uncertainty between tasks are considered during the optimization process, which not only optimizes the prediction accuracy of the model, but also automatically adjusts the model's estimate of the uncertainty of each task. In the regression task, the log-likelihood can be written as:

[0132] ;

[0133] For a The Gaussian likelihood of the model, the observation noise parameter Capture the amount of noise in the output and then maximize the observation noise parameter The log-likelihood of .

[0134] Minimize the loss weight of the main task and auxiliary task loss weights The last goal is to adaptively learn the cross entropy loss based on the data as and mean square error loss The relative weight of the main task. With the variable When the noise parameter increases, the cross entropy loss is On the other hand, when the noise decreases, the weight of the corresponding target increases. The last term acts as a regularizer for the noise term, suppressing the situation where the noise is too large. This construction can be extended to multi-class output. The scaled version of the model output is compressed by adjusting the classification probability through the softmax function:

[0135] ;

[0136] In the formula, the positive scalar is the Boltzmann distribution in the probability distribution, where the input is scaled to This scalar can be fixed or learned, and its size determines the "uniformity" (flatness) of the discrete distribution.

[0137] The main task and auxiliary tasks are relatively independent. The classification output of the main task is , the fault state classification of the shock absorber; the regression output of the auxiliary task is , the performance surplus of the shock absorber is jointly modeled, and the formula is as follows:

[0138] =

[0139] ;

[0140] In the formula, It represents the probability distribution of the fault state classification of the shock absorber and the performance surplus of the shock absorber.

[0141] Then the minimization objective of the shock absorber fault detection model, that is, the combined loss function, is as follows:

[0142] =

[0143] =

[0144] =

[0145] ≈

[0146] ;

[0147] The goal of the auxiliary task is to minimize the combined loss function , including both continuous and discrete outputs. Task-specific noise parameters ( , ) controls the relative weight of each task loss, so that the model can learn the relative importance of each task loss during training. This method also introduces the idea of ​​predicting logarithmic variance, which is more numerically stable and avoids the problem of division by zero. This method can effectively prevent the task weights from converging to zero and ensure that each task can make a balanced contribution during the optimization process. The final loss function is smooth, differentiable, and applicable to the combination of continuous and discrete outputs. To avoid negative values ​​in the loss function, the form is slightly modified, and the formula is as follows:

[0148] ;

[0149] The above minimization loss function is used as part of network training, and the combined loss function is calculated through the back propagation algorithm. For auxiliary task parameters and noise parameters ( , ). During the training process, the noise parameter ( , ) will be adjusted adaptively, thus affecting the relative weight of task loss. By adjusting the noise parameter ( , ), the network can dynamically balance the attention paid to each task (shock absorber fault detection and performance degradation estimation) to optimize the loss of each task. After multiple iterations of optimization, the network will converge to a better parameter value, which minimizes the combined loss function and ensures that the losses of the two tasks are balanced during training.

[0150] Step 5: After the training is completed, the performance is evaluated to verify the performance of the proposed method on the main task and the auxiliary task. The evaluation indicators of the classification task (such as accuracy) and the evaluation indicators of the regression task (such as mean square error, mean absolute error, etc.) are used to evaluate the performance of the model on the two tasks. Through the performance on the test set, it is confirmed whether the model can perform well on both tasks at the same time. In order to verify the effectiveness of the method of the invention, a series of comparative experiments are conducted to compare the performance of the method of the invention with the single-task model and fixed-weight multi-task learning. Ensure that the proposed method can effectively solve the task balance problem in multi-task learning and achieve good performance on both tasks.

[0151] In the single-task model, only the classification of normal state and four lateral shock absorber faults is considered to ensure the balance of data samples in each category, so as to avoid the problem of data imbalance. In the multi-task learning model, the multi-task learning with fixed weights takes shock absorber fault detection as the main task and shock absorber performance degradation estimation task as the auxiliary task, but only uses fixed weights to balance the losses of the two tasks, and sets the weight ratio of the main task to the auxiliary task to 0.7:0.3. In contrast, the method of the invention dynamically adjusts the loss weights of the main task and the auxiliary task through the uncertainty-based loss weighting method to adapt to the uncertainty of the task. The experimental results are shown in Table 1. The method of the invention is superior to the single-task model and the multi-task learning with fixed weights in terms of accuracy, mean absolute error (MAE) and mean square error (MSE), and achieves better results in fault detection problems.

[0152] Table 1 Comparison of different methods

[0153] .

[0154] Figure 3 FIG. 1 is a comparison chart of the training results of various detection methods in one embodiment of the present invention, and also shows the changes in the training curves and verification indicators of different comparative experiments. Figure 3 It can be seen that the method of the invention also performs better in the training process. Not only does it converge faster, but it also has smaller fluctuations and shows better stability.

[0155] To further demonstrate the uncertainty-based loss weighting adaptive process, Figure 4 The dynamic changes of the loss weights of the main task and the auxiliary task during the training process are shown. As the training progresses, the loss weights are dynamically adjusted according to the uncertainty of the task, ensuring the balanced optimization of the two tasks during the training process, thereby improving the overall performance of the model.

[0156] In summary, the technical solution provided by the present invention can realize the detection of shock absorber faults under different line conditions, and can effectively locate the faulty shock absorber in the suspension system. It has the technical advantages of low cost, simple debugging process, high accuracy and good adaptability. Furthermore, combined with auxiliary task training, it helps to improve the training efficiency of the shock absorber fault detection model and the credibility and accuracy of fault detection.

[0157] Furthermore, the main task and auxiliary tasks share the feature extraction layer, which helps to improve detection efficiency and model adaptability on the basis of reducing model complexity and improving feature extraction efficiency; multi-task learning can improve information richness and feature comprehensiveness, and has a certain knowledge transfer ability. When dealing with various faults of high-speed train shock absorbers, it can realize comprehensive analysis and improve the effectiveness and accuracy of fault detection; the loss weight is dynamically adjusted according to the uncertainty of each task to ensure that the main task and auxiliary tasks remain balanced during the training process, improve the stability and robustness of the model, and enable it to maintain a high fault detection accuracy in various situations; the multi-task learning method can better cope with different degrees of performance degradation by dynamically evaluating the loss of each task, and effectively enhance the flexibility of the model in complex environments.

Claims

1. A high-speed train shock absorber fault detection method combined with auxiliary enhancement, characterized in that: The following steps are involved: Step S1: construct a shock absorber fault detection model, including a feature extraction network, a main task sub-network and an auxiliary task sub-network; The main task sub-network and auxiliary task sub-network are connected to the feature extraction network respectively; Step S2: training a shock absorber fault detection model, including: The training samples are input into the shock absorber fault detection model, and the cross entropy loss of the main task sub-network and the mean square error loss of the auxiliary task sub-network are calculated respectively; The cross entropy loss and the mean square error loss are combined to obtain the combined loss and minimized to optimize the network parameters of the shock absorber fault detection model; Step S3: input the multi-channel monitoring data of the suspension system to be detected into the trained shock absorber fault detection model to obtain the main task output and threshold it; If the first element of the thresholded main task output is 1, there is no shock absorber fault in the suspension system; if the non-first element is 1, there is a fault in the corresponding shock absorber; step S2 specifically includes: Step S21: obtaining a monitoring data set of a training suspension system including N shock absorbers, and dividing it into n training samples; Step S22: For the i-th training sample x i , extract its feature representation through the feature extraction network and obtain the training feature matrix The formula is as follows: In the formula, f (s) (·) is the activation function of the feature extraction network, W (s) and b (s) are the weight matrix and bias vector of the feature extraction network respectively; Step S23: 1) The training feature matrix Input the main task sub-network and get the main task training output A training probability vector of size 1×(N+1), The formula is as follows: In the formula, softmax(·) represents the normalized exponential function, and Represent the weight matrix and bias vector of the main task sub-network respectively; Main task training output The first element of is the probability that there is no shock absorber fault in the training suspension system, and the k3th element is the probability that the shock absorber fault is located in the k3-1th shock absorber; wherein k3=2, 3, ..., N+1; Through thresholding, the main task training output The element with the largest median value is set to 1, and the rest of the elements are set to 0; 2) The training feature matrix Input the auxiliary task subnetwork to obtain the auxiliary task training output of size 1×N The formula is as follows: In the formula, and Represent the weight matrix and bias vector of the auxiliary task sub-network respectively; Auxiliary task training output In the equation, the value of the k4th element represents the performance surplus of the k4th shock absorber, k4=1, 2, ..., N; Step S24: Calculate the cross entropy loss of the main task sub-network as L cls , The formula is as follows: Where c represents the fault category index, c = 1, 2, ..., N + 1, 1 (y i =c) is the indicative function, y i =c is 1, otherwise it is 0; Calculate the mean square error loss L of the auxiliary task subnetwork reg , The formula is as follows: In the formula, y reg，i is the true label corresponding to the training sample i and the true value of performance degradation; Step S25: Calculate the combined loss L of the shock absorber fault detection model MTL , The formula is as follows: L MTL =s cls L cls +s reg L reg ; In the formula, σ cls is the loss weight of the main task; reg Loss weight for auxiliary tasks; Step S26: Minimize the combined loss L MTL , to optimize the network parameters of the shock absorber fault detection model; Step S26 also includes dynamically adjusting the main task loss weight σ cls and auxiliary task loss weight σ reg The method is as follows: Optimizing the combined loss L of the shock absorber fault detection model MTL , The formula is as follows: Calculate the combined loss function L through the back propagation algorithm MTL The gradient of the main task loss weight σ is dynamically adjusted according to the gradient cls and auxiliary task loss weight σ reg .

2. A high-speed train shock absorber fault detection method combined with auxiliary enhancement as claimed in claim 1, characterized in that: Step S3 specifically includes: Step S31: for a suspension system including m shock absorbers in a high-speed train, sampling multi-channel monitoring data through a monitoring system; wherein m=1, 2, 3, ...; Step S32: extracting feature representation of multi-channel monitoring data through a feature extraction network to obtain a feature matrix; Step S33: Input the feature matrix into the main task sub-network to obtain the main task output A probability vector of size 1×(m+1); the main task output The first element of is the probability that there is no shock absorber failure in the suspension system, and the k1th element is the probability that the k1-1th shock absorber has a failure; where k1=2, 3, ..., m+1; Through thresholding, the main task output The element with the largest median value is converted to 1, and the remaining elements are converted to 0; For the thresholded main task output If the first element is 1, it is determined that there is no shock absorber fault in the suspension system; If the value of the k1th element is 1, it is determined that the k1-1th shock absorber is faulty and the remaining shock absorbers are normal.

3. A high-speed train shock absorber fault detection method combined with auxiliary enhancement as claimed in claim 1, characterized in that: Step S3 also includes: Step A: Input the multi-channel monitoring data of the suspension system to be tested into the optimized shock absorber fault detection model, and each element output by the auxiliary task represents the performance surplus of the corresponding shock absorber.

4. A high-speed train shock absorber fault detection method combined with auxiliary enhancement as claimed in claim 3, characterized in that: Step A specifically includes: For a suspension system including m shock absorbers in a high-speed train, multi-channel monitoring data is sampled through a monitoring system; wherein m = 1, 2, 3, ...; Extract feature representation of multi-channel monitoring data through feature extraction network to obtain feature matrix; Input the feature matrix into the auxiliary task subnetwork to obtain an auxiliary task output of size 1×m Output for auxiliary tasks The value of the k2th element represents the performance surplus of the k2th shock absorber, and its value range is [0, 1]; where k2 = 1, 2, ..., m.

5. A high-speed train shock absorber fault detection method combined with auxiliary enhancement as claimed in any one of claims 1 to 4, characterized in that: The feature extraction network is a Transformer encoder, a multi-layer one-dimensional convolutional network, a multi-layer recurrent neural network, an autoencoder or a residual network.

6. A high-speed train shock absorber fault detection method combined with auxiliary enhancement as claimed in any one of claims 1 to 4, characterized in that: The main task subnetwork is a single-layer fully connected network, a multi-layer fully connected network connected in sequence, a single-layer one-dimensional convolutional network, or a multi-layer one-dimensional convolutional network connected in sequence.

7. A high-speed train shock absorber fault detection method combined with auxiliary enhancement as claimed in any one of claims 1 to 4, characterized in that: The auxiliary task subnetwork is a single-layer fully connected network, a multi-layer fully connected network connected in sequence, a single-layer one-dimensional convolutional network, or a multi-layer one-dimensional convolutional network connected in sequence.

Images