Full-automatic coupler multi-sensor state feedback and dynamic characteristic analysis system

Abstract

The invention relates to the technical field of coupler state detection, in particular to a full-automatic coupler multi-sensor state feedback and dynamic characteristic analysis system, which comprises a data acquisition module, a configuration sensor group, a real-time acquisition module and a dynamic characteristic analysis module, the original data of the full-automatic car coupler comprises tension data, displacement data, temperature data and vibration data; and the data preprocessing module is used for preprocessing the original data, and the preprocessing comprises time synchronization, noise filtering and normalization processing. According to the full-automatic coupler multi-sensor state feedback and dynamic characteristic analysis system provided by the invention, through multi-sensor fusion and machine learning prediction, high-precision, real-time and intelligent full-automatic coupler health management is realized, the safety, reliability and maintenance efficiency are improved, and a more efficient and safer technical guarantee is provided for railway transportation.

Description

Technical Field

[0001] The present invention relates to the technical field of coupler state detection, and more specifically, to a multi-sensor state feedback and dynamic characteristic analysis system for a full-automatic coupler. Background Art

[0002] In the field of railway and rail transit, the coupler is a key component for train connection, and its performance is directly related to the safety and stability of train operation.

[0003] A patent application with the publication number CN115077770A discloses a coupler traction monitoring sensor and a coupler. The coupler traction monitoring sensor includes a tail pin. A counterbore is provided on the side wall of the tail pin, and an outlet is provided at one end of the tail pin; a strain gauge is installed in the counterbore; a connecting wire is located inside the tail pin. One end of the connecting wire is electrically connected to the strain gauge, and the other end of the connecting wire extends out from the outlet. The tail pin directly bears the traction force transmitted by the vehicle and then feeds it back to the strain gauge, which can improve the service life. The present invention can directly replace the tail pin in the original coupler, serving both as a component of the coupler and as a traction force sensor, without changing the mechanical structure of the original coupler, and without additional steps such as grinding, pasting, and sealing, which is convenient for installation. The tail pin directly bears the traction force transmitted by the vehicle, and the force application point of the sensor can more truly reflect the magnitude of the coupler traction force, so the measured traction force data is more accurate, improving the force measurement accuracy.

[0004] As in the above application, the coupler monitoring systems widely used in the market mainly rely on a single sensor (such as a displacement sensor or a force sensor) to monitor the working state of the coupler. Although these data can provide certain operation state information, there are obvious limitations. Specifically, the single-sensor monitoring scheme often cannot comprehensively reflect the multi-dimensional dynamic characteristics of the coupler under complex working conditions, such as temperature changes, vibration conditions, wear degree, etc., resulting in untimely and inaccurate fault warnings, high maintenance costs, and it is difficult to achieve in-depth analysis and optimization of the coupler performance. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a multi-sensor state feedback and dynamic characteristic analysis system for a full-automatic coupler.

[0006] The present invention adopts the following technical solutions. The multi-sensor state feedback and dynamic characteristic analysis system for a full-automatic coupler includes:

[0007] A data acquisition module, configured with a sensor group to collect the original data of the full-automatic coupler in real time. The original data of the full-automatic coupler includes tensile force data, displacement data, temperature data, and vibration data;

[0008] Data preprocessing module, which preprocesses the original data. The preprocessing includes time synchronization, noise filtering, and normalization processing;

[0009] Feature extraction module, which extracts the maximum value, minimum value, mean value, and root mean square of the tensile data, obtains the mean value and standard deviation of the displacement data, obtains the change rate and maximum value of the temperature data, and obtains the peak value and mean value of the vibration data;

[0010] Feature fusion module, which fuses the extracted feature data, constructs a high-dimensional feature vector, and performs dimensionality reduction on the high-dimensional feature vector through linear discriminant analysis to generate multimodal data;

[0011] State prediction module, which inputs the collected multimodal data into a pre-constructed machine learning model of the coupler state and outputs the health state of the fully automatic coupler.

[0012] As a further description of the above technical solution: The method for preprocessing the original data includes:

[0013] S01: Through the time interpolation method for the obtained original data, unify all data to the same time reference;

[0014] S02: Use low-pass filtering or wavelet transform to remove high-frequency noise;

[0015] S03: Use the Z-score normalization formula to convert the data of each sensor to the same scale.

[0016] As a further description of the above technical solution: The method for performing dimensionality reduction on the high-dimensional feature vector through linear discriminant analysis to generate multimodal data includes:

[0017] Calculate the within-class scatter matrix SW;

[0018] Calculate the between-class scatter matrix SB;

[0019] By maximizing the between-class scatter matrix SB and minimizing the within-class scatter matrix SW, find the optimal projection matrix W. The calculation formula of the optimal projection matrix is: where WT is the transpose of W. By solving the eigenvalues and eigenvectors of, obtain the projection matrix W;

[0020] Generate a dimensionality reduction matrix Y based on the optimal projection matrix W, and denote the dimensionality reduction matrix Y as multimodal data. The generation method of the dimensionality reduction matrix Y is: Y = WTX.

[0021] As a further description of the above technical solution: The calculation method of the within-class scatter is as follows. First, calculate the center point of each data category, that is, the average value. Then calculate the distance from each sample to the center point of the category, which is denoted as the within-class difference distance. Add up the within-class difference distances of all categories to obtain the within-class scatter SW.

[0022] As a further description of the above technical solution: The calculation method of the between-class scatter is as follows: Calculate the overall average value of all data. Then calculate the distance between the average value of each category and the overall average value, which is denoted as the between-class difference distance. Add up the between-class difference distances of all categories to obtain the between-class scatter SB.

[0023] As a further description of the above technical solution: The characteristic data are the maximum value, minimum value, mean value, and root mean square of the tensile force data, the mean value and standard deviation of the displacement data, the change rate and maximum value of the temperature data, and the peak value and mean value of the vibration data.

[0024] As a further description of the above technical solution: The method of fusing the extracted characteristic data to construct a high-dimensional feature vector is as follows: Concatenate the maximum value, minimum value, mean value, and root mean square of the tensile force data, the mean value and standard deviation of the displacement data, the change rate and maximum value of the temperature data, and the peak value and mean value of the vibration data according to the dimension to form a high-dimensional feature vector. Each element in the high-dimensional feature vector represents different characteristic data.

[0025] As a further description of the above technical solution: The training method of the coupler state prediction machine learning model includes:

[0026] Select a multi-layer perceptron MLP model;

[0027] Divide the collected multi-modal data according to the ratio of 70%, 15%, and 15%. The training set is used for the model to learn the complex relationship between the multi-modal data and the coupler health state. The validation set is used to evaluate the model performance in real time during the training process and adjust the hyperparameters of the model; the test set is used to finally evaluate the performance of the model on new data and determine whether the model can accurately predict the health state of the coupler, so as to verify the practicability and reliability of the model.

[0028] Use the training set to train the MLP model. The model passes through forward propagation, undergoes non-linear transformation in the hidden layer, and finally obtains the prediction result of the coupler health state at the output layer. Compare the prediction result with the true coupler health state label, calculate the loss function, and then calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm. Update the model parameters according to the gradient descent method, continuously adjust the weights and biases of the model, so that the prediction result of the model gradually approaches the true label. As the training progresses, the model gradually learns the complex mapping relationship between different combinations of multi-modal data and the coupler health state, and improves the prediction accuracy on the training data.

[0029] The performance of the model is evaluated using the accuracy metric on the validation set. The accuracy reflects the proportion of samples correctly predicted by the model. According to the evaluation results, the hyperparameters of the MLP are adjusted;

[0030] The tuned model is finally tested using the test set, and the accuracy of the model on the test set is calculated. If the test results meet the expectations, it indicates that the model has good generalization ability and can accurately predict the health status of the coupler in the actual scenario. Otherwise, retraining is required.

[0031] As a further description of the above technical solution: The method for obtaining the training set includes:

[0032] Collect the raw data through the deployed sensor group, process the raw data to obtain the corresponding multimodal data, organize a team consisting of railway vehicle maintenance experts, mechanical engineers, and data analysts to annotate the collected data, and establish the correspondence between the multimodal data and the health status of the coupler. The health status relationship of the coupler includes the healthy state, sub-healthy state, and fault state.

[0033] As a further description of the above technical solution: The sensor group includes a force sensor, a displacement sensor, a temperature sensor, and a vibration sensor.

[0034] Beneficial effects:

[0035] In the above technical solution, the fully automatic coupler multi-sensor status feedback and dynamic characteristic analysis system provided by the present invention configures a sensor group to collect the raw data of the fully automatic coupler in real time, adopts an advanced algorithm to fuse the data from different sensors to generate multimodal data, realizes multi-dimensional and high-precision collection of coupler status information, ensures the comprehensiveness and accuracy of the data, inputs the collected multimodal data into the pre-constructed machine learning model for coupler status prediction, and outputs the health status of the fully automatic coupler. That is, the invention realizes high-precision, real-time, and intelligent health management of the fully automatic coupler through multi-sensor fusion and machine learning prediction, improves safety, reliability, and maintenance efficiency, and provides a more efficient and safer technical guarantee for railway transportation.

[0036] Secondly, through linear discriminant analysis, the high-dimensional feature vector is dimensionally reduced. The original data of the coupler status involves multiple sensors such as tensile force, displacement, temperature, and vibration, which may form a high-dimensional feature vector. By mapping the high-dimensional data to a lower-dimensional space (such as two-dimensional or three-dimensional), the amount of calculation is reduced, and the processing speed is increased. After dimensional reduction, the training and inference speeds of the machine learning model for coupler status prediction are greatly improved, and the data after dimensional reduction retains the most important features for classification, while the noise components are weakened or removed, further improving the accuracy of its prediction. Description of the drawings

[0037] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments:

[0038] Figure 1 It is a module connection diagram of the full - automatic coupler multi - sensor status feedback and dynamic characteristic analysis system provided by the embodiment of the present invention;

[0039] Figure 2 It is a flowchart of the method for pre - processing the original data provided by the embodiment of the present invention;

[0040] Figure 3 It is a flowchart of the method for reducing the dimension of high - dimensional feature vectors through linear discriminant analysis to generate multi - modal data provided by the embodiment of the present invention. Specific embodiments

[0041] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below in conjunction with specific illustrations. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0042] Embodiment 1

[0043] Please refer to Figure 1 - Figure 2 , the embodiment of the present invention provides a technical solution: a full - automatic coupler multi - sensor status feedback and dynamic characteristic analysis system, including:

[0044] A data acquisition module, configured with a sensor group to collect the original data of the full - automatic coupler in real time. The original data of the full - automatic coupler includes tensile force data, displacement data, temperature data and vibration data;

[0045] The sensor group includes a force sensor, a displacement sensor, a temperature sensor and a vibration sensor;

[0046] It should be noted that the force sensor monitors the tensile force received by the coupler; the displacement sensor monitors the relative displacement change of the coupler to judge whether there is looseness or abnormal movement of the coupler; the temperature sensor monitors the temperature of the coupler to prevent structural problems caused by temperature difference; the vibration sensor monitors the vibration state of the coupler during high - speed operation to judge whether there is wear or looseness.

[0047] A data pre - processing module, which pre - processes the original data. The pre - processing includes time synchronization, noise filtering and normalization processing.

[0048] Specifically, the time intervals and noise levels of data collected by different sensors vary. A unified time benchmark and normalization processing can ensure that data can be compared on the same numerical scale, reducing errors caused by sampling inconsistency and noise, thereby laying a solid foundation for subsequent feature extraction and fusion. Time synchronization is to unify all data to the same time benchmark, eliminate the timing deviation caused by different sampling frequencies, and use methods such as low-pass filtering or wavelet transform to remove high-frequency noise to obtain smoother and more reliable data. Through normalization processing, the data of each sensor is converted to the same scale for subsequent fusion.

[0049] The methods for preprocessing the original data include:

[0050] S01: Through the time interpolation method for the obtained original data, unify all data to the same time benchmark;

[0051] Specifically, to ensure the precise alignment of the data of each sensor in time, since the sampling frequencies of different sensors may be different, for this reason, the time interpolation method is adopted to resample all data to the unified time benchmark. For example: for the original data at any moment t, the interpolation formula to ensure that each data point corresponds to the same moment is:

[0052]

[0053] In the formula, t0 is the first time of the known data point, that is, the starting time of the interpolation interval, t1 is the second time of the known data point, that is, the end time of the interpolation interval, t is the target time to be interpolated, that is, the interpolation at this moment is desired to be calculated, x(t0) is the value of the original data observed or recorded at time t0, x(t1) is the value of the original data observed or recorded at time t1, and x(t) is the interpolation value obtained by using the interpolation formula at the target time t.

[0054] S02: Use low-pass filtering or wavelet transform to remove high-frequency noise;

[0055] Optionally, the method for removing high-frequency noise through a low-pass filter screen includes:

[0056] In the formula, F(t) represents the value of the filter output signal at time t, F(τ) represents the value of the original input signal at time τ, and h(t - τ) represents the value of the impulse response of the filter at the time difference t - τ.

[0057] S03: Use the Z-score normalization formula to convert the data of each sensor to the same scale.

[0058] The Z-score normalization formula: Wherein, xnorm is the result after normalization, x is the original data value, μ is the mean of the data set, and σ is the standard deviation of the data set, where the original data refers to a data in the data set.

[0059] The feature extraction module extracts the maximum value, minimum value, mean value and root mean square of the tensile force data, obtains the mean value and standard deviation of the displacement data, obtains the change rate of the temperature data, and obtains the peak value and mean value of the vibration data.

[0060] It should be noted that through feature extraction, a large amount of original data is transformed into a small amount of information describing key dynamic behaviors, thereby reducing the computational complexity. The features extracted by each sensor can highlight its most representative physical attributes, enabling the subsequent fusion to give full play to their respective advantages and achieve the effect of "learning from each other's strengths and compensating for each other's weaknesses".

[0061] The feature fusion module fuses the extracted feature data, constructs a high-dimensional feature vector X, and reduces the dimension of the high-dimensional feature vector through linear discriminant analysis to generate multi-modal data.

[0062] The feature data are the maximum value, minimum value, mean value and root mean square of the tensile force data, the mean value and standard deviation of the displacement data, the change rate and maximum value of the temperature data, and the peak value and mean value of the vibration data.

[0063] The method of fusing the extracted feature data to construct a high-dimensional feature vector is to splice the maximum value, minimum value, mean value and root mean square of the tensile force data, the mean value and standard deviation of the displacement data, the change rate and maximum value of the temperature data, and the peak value and mean value of the vibration data according to the dimension to form a high-dimensional feature vector. Each element in the high-dimensional feature vector represents different feature data.

[0064] The state prediction module inputs the collected multi-modal data into a pre-constructed coupler state prediction machine learning model and outputs the health state of the full automatic coupler;

[0065] The training method of the coupler state prediction machine learning model includes:

[0066] Select a multi-layer perceptron (MLP) model. MLP is a feedforward neural network that can handle complex non-linear relationships between input data. The feature combinations of different types of data such as coupler tensile force, displacement, temperature, and vibration are not simply linearly related to the coupler health state. MLP performs non-linear transformation and combination on these numerical features through multiple hidden layers, and can effectively learn the mapping relationship between the input data and the coupler health state. Compared with some simple linear models, the powerful fitting ability of MLP makes it more suitable for processing such complex multi-modal numerical data prediction tasks;

[0067] Divide the collected multimodal data according to the ratio of 70%, 15%, and 15%. The training set is used for the model to learn the complex relationship between multimodal data and the health status of the coupler. A large amount of data enables the model to fully capture the patterns in the data. The validation set is used to evaluate the model performance in real time during the training process, adjust the hyperparameters of the model (such as the number of nodes in the MLP hidden layer, learning rate, number of iterations, etc.), and prevent the model from overfitting. The test set is used to finally evaluate the performance of the model on new data and determine whether the model can accurately predict the health status of the coupler, thereby verifying the applicability and reliability of the model. This division method is commonly used and effective in machine learning and can make full use of data resources for model training and optimization;

[0068] Use the training set to train the MLP model. Take the preprocessed coupler tension, displacement, temperature, and vibration data as inputs and input them into the MLP model. Through forward propagation, the model undergoes non-linear transformation in the hidden layer and finally obtains the prediction result of the coupler health status at the output layer. Compare the prediction result with the true coupler health status label and calculate the loss function (such as the cross-entropy loss function). Then, calculate the gradient of the loss function with respect to the model parameters (weights and biases) through the backpropagation algorithm, update the model parameters according to the gradient descent method, and continuously adjust the weights and biases of the model to make the prediction result of the model gradually approach the true label. As the training progresses, the model gradually learns the complex mapping relationship between different multimodal data combinations and the coupler health status, improving the prediction accuracy on the training data;

[0069] Use the accuracy metric to evaluate the model performance on the validation set. The accuracy reflects the proportion of samples correctly predicted by the model; according to the evaluation results, adjust the hyperparameters of the MLP;

[0070] Use the test set to conduct the final test on the tuned model and calculate the accuracy of the model on the test set. If the test results meet the expectations, it indicates that the model has good generalization ability and can accurately predict the health status of the coupler in the actual scenario. Otherwise, retrain the model.

[0071] The method for obtaining the training set includes:

[0072] Collect the original data through the deployed sensor group, process the original data to obtain the corresponding multimodal data, organize a team consisting of railway vehicle maintenance experts, mechanical engineers, and data analysts to annotate the collected data, and establish the corresponding relationship between the multimodal data and the coupler health status. The coupler health status relationship includes healthy status, sub-healthy status, and faulty status;

[0073] It should be noted that railway vehicle maintenance experts have rich experience in coupler maintenance and can judge the health status of the coupler at the data acquisition moment based on the actual maintenance situation and long-term observation of the coupler. Mechanical engineers assist in judging the health status of the coupler from professional perspectives such as mechanics and mechanical principles, in combination with the design parameters and operating conditions of the coupler. Data analysts are responsible for ensuring the format standardization, consistency, and accuracy of the labeled data to meet the requirements of model training. The collaboration of multiple experts can ensure that the labeling results are based on actual engineering experience and can meet the data quality requirements of model training, providing reliable labeled data for the model to learn the mapping relationship between accurate multi-modal data and the health status of the coupler.

[0074] Embodiment 2

[0075] Please refer to Figure 3 , the embodiment of the present invention provides a technical solution:

[0076] A method for dimensionality reduction of high-dimensional feature vectors and generating multi-modal data through linear discriminant analysis includes:

[0077] Calculate the within-class scatter matrix \(S_W\). The calculation method of the within-class scatter matrix is to calculate the center point, that is, the average value, of each data category in turn, calculate the distance from each sample to the center point of this category, denoted as the within-class difference distance, and add up the within-class difference distances of all categories to obtain the within-class scatter matrix \(S_W\);

[0078] Calculate the between-class scatter matrix \(S_B\). The calculation method of the between-class scatter matrix is to calculate the overall average value of all data, and then calculate the distance between the average value of each category and the overall average value, denoted as the between-class difference distance, and add up the between-class difference distances of all categories to obtain the between-class scatter matrix \(S_B\);

[0079] By maximizing the between-class scatter matrix and minimizing the within-class scatter matrix, find the optimal projection matrix \(W\). The calculation formula of the optimal projection matrix is: By solving the eigenvalues and eigenvectors of, obtain the projection matrix \(W\), where \(W^T\) is the transpose of \(W\). For example, if \(W\) is an \(n\times m\) matrix, then \(W^T\) is an \(m\times n\) matrix;

[0080] Generate a dimensionality reduction matrix \(Y\) based on the optimal projection matrix \(W\), and denote the dimensionality reduction matrix \(Y\) as multi-modal data. The generation method of the dimensionality reduction matrix \(Y\) is: \(Y = W^TX\).

[0081] The basic principles, main features and advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the descriptions in the above embodiments and the specification are only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. Full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system, characterized in that Including: A data acquisition module, configured with a sensor group to collect raw data of the fully automatic coupler in real time. The raw data of the fully automatic coupler includes tensile force data, displacement data, temperature data, and vibration data. A data preprocessing module for preprocessing the raw data. The preprocessing includes time synchronization, noise filtering, and normalization. A feature extraction module that extracts the maximum value, minimum value, mean value, and root mean square of the tensile force data, obtains the mean value and standard deviation of the displacement data, obtains the change rate and maximum value of the temperature data, and obtains the peak value and mean value of the vibration data. A feature fusion module that fuses the extracted feature data, constructs a high-dimensional feature vector, and performs dimensionality reduction on the high-dimensional feature vector through linear discriminant analysis to generate multimodal data. A state prediction module that inputs the collected multimodal data into a pre-constructed machine learning model of the coupler state and outputs the health state of the fully automatic coupler.

2. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 1, characterized in that, The method for preprocessing the raw data includes: S01: Unify all data to the same time base through time interpolation method for the obtained raw data. S02: Use low-pass filtering or wavelet transform to remove high-frequency noise. S03: Convert the data of each sensor to the same scale using the Z-score normalization formula.

3. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 1, characterized in that The method for performing dimensionality reduction on the high-dimensional feature vector through linear discriminant analysis to generate multimodal data includes: Calculating the within-class scatter matrix SW. Calculating the between-class scatter matrix SB. Finding the optimal projection matrix W by maximizing the between-class scatter matrix SB and minimizing the within-class scatter matrix SW. Generating a dimensionality reduction matrix Y based on the optimal projection matrix W and denoting the dimensionality reduction matrix Y as multimodal data.

4. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 3, characterized in that The calculation method of the within-class scatter matrix is to calculate the center point, that is, the average value, of each data category in turn, calculate the distance from each sample to the center point of the category, denoted as the within-class difference distance, and add up the within-class difference distances of all categories to obtain the within-class scatter matrix SW.

5. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 3, characterized in that The calculation method of the between-class scatter matrix is: calculate the overall average value of all data, then calculate the distance between the average value of each category and the overall average value, denoted as the between-class difference distance, and add up the between-class difference distances of all categories to obtain the between-class scatter matrix SB.

6. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 1, characterized in that The feature data is the maximum value, minimum value, mean value, and root mean square of the tensile force data, the mean value and standard deviation of the displacement data, the change rate and maximum value of the temperature data, and the peak value and mean value of the vibration data.

7. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 6, wherein The method for fusing the extracted feature data to construct a high-dimensional feature vector is to splice the maximum value, minimum value, mean value, and root mean square of the tensile force data, the mean value and standard deviation of the displacement data, the change rate and maximum value of the temperature data, and the peak value and mean value of the vibration data according to dimensions to form a high-dimensional feature vector. Each element in the high-dimensional feature vector represents different feature data.

8. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 1, characterized in that The training method of the machine learning model for predicting the coupler state includes: Selecting a multi-layer perceptron MLP model. Divide the collected multi-modal data according to the ratio of 70%, 15%, and 15%. The training set is used for the model to learn the complex relationship between multi-modal data and the health status of the coupler. The validation set is used to evaluate the model performance in real time during the training process and adjust the hyperparameters of the model. The test set is used to finally evaluate the performance of the model on new data and determine whether the model can accurately predict the health status of the coupler, so as to verify the practicability and reliability of the model; Use the training set to train the MLP model. Through forward propagation, the model undergoes non-linear transformation in the hidden layer and finally obtains the prediction result of the coupler health status at the output layer. Compare the prediction result with the true coupler health status label, calculate the loss function, and then calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm. Update the model parameters according to the gradient descent method, continuously adjust the weights and biases of the model, and make the prediction result of the model gradually approach the true label. As the training progresses, the model gradually learns the complex mapping relationship between different multi-modal data combinations and the coupler health status, improving the prediction accuracy on the training data; Use the accuracy metric to evaluate the model performance on the validation set. The accuracy reflects the proportion of samples correctly predicted by the model. Adjust the hyperparameters of the MLP according to the evaluation results; Use the test set to conduct the final test on the optimized model and calculate the accuracy of the model on the test set. If the test results meet the expectations, it indicates that the model has good generalization ability and can accurately predict the health status of the coupler in the actual scenario. Otherwise, retrain the model.

9. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 8, characterized in that, The method for obtaining the training set includes: Collect the original data through the arranged sensor group, process the original data to obtain the corresponding multi-modal data, organize a team composed of railway vehicle maintenance experts, mechanical engineers, and data analysts to label the collected data, and establish the corresponding relationship between the multi-modal data and the coupler health status. The coupler health status relationship includes healthy status, sub-healthy status, and faulty status.

10. The full-automatic coupler multi-sensor status feedback and dynamic characteristic analysis system according to claim 1, wherein The sensor group includes force sensors, displacement sensors, temperature sensors, and vibration sensors.

Images