Wear prediction method and system for subway door mechanical components under different working conditions

Abstract

The present invention discloses a method and system for predicting the degree of wear of subway door mechanical components under different operating conditions. The method includes the following steps: collecting motion data curves of door mechanical components in different wear states and performing preprocessing; obtaining motion data samples under different operating conditions and wear states of mechanical components to construct samples and label them according to the wear state; extracting features and performing normalization processing; reducing the feature dimension and performing clustering processing to determine the central points of feature data with different labels; constructing and filtering target domain censored data; and establishing a transfer learning prediction model to predict the degree of wear of mechanical components. The present invention can effectively reflect the actual situation of severe wear, provide an effective data source for subsequent model establishment and training, thereby improving the model prediction accuracy and solving the difficulty of model training caused by data loss.

Description

Technical Field

[0001] The present invention relates to the field of rail transit safety technology, and in particular to a method and system for predicting the wear degree of mechanical components of subway doors under different working conditions. Background Art

[0002] With the development of the rail transit industry, the safety and reliability of urban rail vehicle door operations have received widespread attention. Currently, the primary maintenance strategy for companies remains "traditional" time-based maintenance (TBM). TBM refers to preventive maintenance based entirely on the operating time of equipment / components, with continuous maintenance intervals generally set at fixed values. This strategy facilitates maintenance planning by simply formulating a maintenance plan based on the operating time of the equipment / components. However, if the equipment / components are still in good health during preventive maintenance, this can result in unnecessary waste of maintenance resources. Furthermore, if the machine deteriorates faster than expected, the equipment / components may fail and shut down before the next preventive maintenance is performed.

[0003] Thanks to the development of technologies such as monitoring, storage, and data analysis, condition-based maintenance (CBM) strategies have gradually become an important means for enterprises to optimize production management. CBM refers to maintenance implemented after equipment / components have shown significant deterioration, and the deterioration is reflected by changes in the monitored machine status parameters. The premise for implementing CBM strategies is to reliably predict the remaining useful life of the equipment / components at the current moment, thereby providing managers with valuable information to accurately formulate production and maintenance plans. The degree of wear of key mechanical wear parts in the door system can well reflect their performance and lifespan. Accurately predicting their diameter facilitates the implementation of CBM strategies.

[0004] The wear degree prediction of key mechanical wear components in vehicle door systems can be categorized into two main methods: model-based prediction methods and data-based prediction methods. Model-based prediction methods can be further divided into analytical model-based prediction methods and physical model-based prediction methods. Model-based prediction methods typically use empirical degradation models or physical analysis models as degradation models for the system. The kinematic coupling of multiple components and the complex wear mechanisms in vehicle door systems make constructing physical models difficult. Data-based prediction methods can be further categorized into three types: traditional machine learning-based prediction methods, deep learning-based prediction methods, and statistics-based prediction methods. The mainstream is machine learning-based methods, and deep learning-based methods are a special type of machine learning method with superior and more flexible performance.

[0005] Due to the unique application scenarios of rail vehicle door systems, their motor output curves are highly variable, operating conditions are complex, interference factors are numerous, and features are subtle and hidden. Consequently, research on predicting the wear of key mechanical components in these systems is limited. Current research focuses on fault diagnosis and sub-health status diagnosis. The idea is to extract relevant time-domain features from the motor output curves that monitor door movement, and then use machine learning methods such as clustering and classification to achieve the corresponding prediction objectives. However, during operation, the doors enter different operating conditions due to interactions between components and between components and the environment. The data collected from these conditions varies, and variations in operating conditions are often considered abnormal, while those that remain unchanged are considered normal. Given the limited amount of data from abnormal operating conditions, using only data from one operating condition to build a prediction model will result in inaccurate predictions for data from other operating conditions. In practice, it is rare for key mechanical components in door systems to reach the end of their lifespan without replacement. This means that the dataset used to build the model is incomplete, i.e., right-censored. Therefore, it is an urgent problem to predict different working conditions, especially severe wear conditions. Summary of the Invention

[0006] Purpose of the invention: The present invention aims to provide a method and system for predicting the degree of wear of subway door mechanical components under different working conditions, which can effectively reflect the actual situation of severe wear and improve the accuracy of model prediction.

[0007] Technical solution: The method for predicting the wear degree of mechanical components of subway doors under different working conditions described in the present invention includes the following steps:

[0008] (1) Collecting motion data curves of door mechanical components in different wear states and performing preprocessing;

[0009] (2) Obtain motion data samples under different working conditions and wear states of mechanical parts, and label them according to the wear state;

[0010] (3) Feature extraction and standardization;

[0011] (4) Feature dimensionality reduction and clustering to determine the feature data centers of different labels;

[0012] (5) Construction and screening of censored data in the target domain;

[0013] (6) Establish a transfer learning prediction model to predict the degree of wear of mechanical parts.

[0014] Preferably, the different wear states in step (1) include healthy, lightly worn, moderately worn and severely worn states.

[0015] Preferably, the different working conditions in step (2) include normal working conditions and abnormal working conditions, the samples under normal working conditions are the source domain, the samples under abnormal working conditions are the target domain, and the target domain sample data are right-censored data.

[0016] Preferably, the feature extraction in step (3) adopts a wavelet packet decomposition method, performs wavelet packet decomposition on the time domain data, and extracts the energy proportion and energy entropy characteristics of the decomposed sub-band signal.

[0017] Preferably, the feature dimensionality reduction in step (4) adopts a linear discriminant analysis method.

[0018] Preferably, in step (4), a density-based clustering method is used to determine the feature data centers of different tags.

[0019] Preferably, the step (5) comprises:

[0020] (5.1) Based on the existing target domain feature data center points with different labels, determine the possible center points of the right-censored feature data, construct low-dimensional censored data around the right-censored feature data center points, and map them to high dimensions;

[0021] (5.2) Use the existing target domain feature data with different labels to train a linear regression model to predict the constructed censored data, and filter out the censored data whose difference between the predicted result and the preset value is no greater than the threshold;

[0022] (5.3) Based on the center point of the data obtained in step (5.2), continue to perform step (5.1) around it; after constructing the data, continue to perform step (5.2), select the data whose prediction result is closest to the required label, and add it to the target domain feature dataset.

[0023] Preferably, the step (5.1) comprises:

[0024] First, based on the center point locations and trends of the feature data of healthy, lightly worn, and moderately worn states in the target domain feature dataset, the possible locations of the center point of the feature data of the heavy wear state are determined.

[0025] Secondly, define the characteristic distance between the characteristic data of the light wear state and the medium wear state as R', then the center point of the characteristic data of the heavy wear state may be located in the following areas: the sector ring area with the characteristic data center point of the light wear state as the center and R' and 2R' as the radius, and the sector area with the characteristic data center point of the medium wear state as the center and 2R' as the radius;

[0026] Finally, several points are evenly selected in the area where the central points of the characteristic data under the heavy wear state determined in the above steps may exist. These points are used as the center of the circle and the average radius R of the existing characteristic data of the healthy, lightly worn and moderately worn states is used as the radius to construct low-dimensional missing data and map them to high dimensions.

[0027] Preferably, establishing the transfer learning prediction model in step (6) includes establishing a source domain model and a target domain model, wherein the source domain model is a convolutional neural network model, and the target domain model is the same as the source domain model. During training, the parameters of the convolution layer and the pooling layer of the target domain model are frozen, and no parameter update is performed.

[0028] Preferably, the prediction of the degree of wear of mechanical parts in step (6) includes: inputting the features and labels in the source domain feature data set into the source domain model to train the structural parameters of the source domain model; calling the trained source domain model and freezing the parameters of all convolutional layers and pooling layers to form a target domain model; dividing the target domain feature set into a training set and a test set, and then inputting the training set in the target domain feature set into the target domain model to perform secondary training on the parameters of the fully connected layer; finally, inputting the test set in the target domain feature set into the trained target domain model to obtain a prediction result.

[0029] The wear degree prediction system of subway door mechanical components under different working conditions according to the present invention includes:

[0030] The sample and label creation module is used to construct samples using the pre-processed motion data of different mechanical parts in different working conditions and different wear states, and label them according to the wear state;

[0031] The feature processing module is used to extract features from the sample data to form a source domain feature dataset and a target domain feature dataset; reduce the dimension of the target domain feature data and determine the feature data center point and radius of different labels;

[0032] The missing data construction and screening module is used to construct missing data based on the feature data center points and radii of different labels in the existing target domain and screen them through the linear regression model;

[0033] The modeling and prediction module is used to establish a transfer learning prediction model to predict the degree of wear of mechanical parts.

[0034] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. In order to solve the problem that most monitoring data are right-censored data, the linear discriminant analysis dimensionality reduction method is adopted. Based on the right-censored data, censored data is constructed in a low dimension and screened by a linear regression model. The censored feature data construction method is more directional and targeted. The severe wear position is inferred according to the two-dimensional spatial position of the label wear degree through the reduced feature data. The area of ​​the constructed data center can be preliminarily constrained and the data aggregation range can be constrained by the data aggregation radius, so that the constructed censored data can effectively reflect the actual situation of severe wear data, provide an effective data source for training for the establishment of subsequent models, thereby improving the model prediction accuracy and solving the difficulty of model training caused by data missing. 2. In order to solve the problem that different door states have an impact on the prediction of the wear degree of mechanical parts, and the amount of abnormal working condition data is small, the transfer learning method is adopted. The normal working condition data with complete data is used as the source domain, and the abnormal working condition data with missing data is used as the target domain for model training. The source domain model is adjusted using the target domain data. This method can train corresponding models for different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method of the present invention;

[0036] Figure 2 Schematic diagram of censored data constructed for the present invention

[0037] Figure 3 This is the one-dimensional convolutional neural network model structure in an embodiment of the present invention;

[0038] Figure 4 Schematic diagram of the wear state prediction result of mechanical parts after constructing censored data in an embodiment of the present invention;

[0039] Figure 5 Schematic diagram of prediction results of wear status of mechanical parts when censored data is not constructed in an embodiment of the present invention;

[0040] Figure 6 This is a schematic diagram of the wear state prediction results of mechanical components when verifying the rationality of missing data in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0042] like Figure 1 As shown, the method for predicting the wear degree of mechanical components of subway doors under different working conditions according to the present invention includes the following steps:

[0043] (1) Collecting motion data curves of door mechanical components in different wear states. Taking the motor monitoring signals collected by the door system as an example, we obtain motion data curves of door motors in different wear states. The different wear states include healthy, light wear, moderate wear, and severe wear. Simulating two working conditions, we obtain eight sets of door transmission component motor monitoring signal data, as shown in Table 1. For each combination, we collect several speed data, each of which includes door opening and closing curves. Preprocessing is performed, including data separation and deletion of bad points, to obtain the corrected data collected by the rail vehicle door system.

[0044] Table 1 Wear status combination on both sides of the upper door

[0045]

[0046] (2) Sample and label production. Determine the corresponding number of sample points based on the actual door opening and closing time and the motor data sampling frequency, including p sampling points for door opening and q sampling points for door closing. Concatenate the door opening and closing curves to form a sample, and the sample label is the smaller value of the wear state in the combination. Produce n samples and corresponding labels to form a sample data set, which contains n samples, and the dimension of each sample is p+q. Divide the samples into source domain samples and target domain samples. The source domain samples are samples under normal working conditions, and the target domain samples are samples under abnormal working conditions. The target domain sample data is right-censored data.

[0047] (3) Feature extraction and standardization. The time domain motor signal data is decomposed into 6 layers of wavelet packets to extract the energy proportion of the decomposed sub-band signal and the energy entropy characteristics of the total frequency band. Each diameter combination has 15-16 samples, and each sample corresponds to three channels with a total of (1+9)*3=30 features. The feature set of each sample under normal working conditions (source domain) is The source domain feature dataset is D1 = [x n1 ,x n2 ,…,x nn ] T ; The abnormal working condition (target domain) feature data set is D2 = [x a1 ,x a2 ,…,x an ] T , the target domain does not use a dataset with severe wear to simulate right-censored data. The feature data is normalized to a mean of 0 and a variance of 1: each element x in the dataset D1 aij Transformed to y aij , forming a new data set D1'; the transformation formula is as follows: in

[0048] (4) Feature dimensionality reduction and clustering processing are performed. Linear discriminant analysis, i.e., LDA method, is used to reduce the 30-dimensional feature data of the target domain health, light wear, and moderate wear states to an appropriate dimension (the C2 working condition is reduced to 2 dimensions) to determine the center point of the feature data under different wear conditions; and the density-based clustering method (DBSCAN) is used to determine the center point of the feature set of different wear degrees and the cluster radius.

[0049] (5) Construction and screening of censored data in the target domain. Specifically including:

[0050] (5.1) First, based on the center point positions and trend trends of the feature data of healthy, lightly worn, and moderately worn states in the target domain feature dataset, determine the area where the center point of the feature data of the heavy wear state may exist, such as Figure 2 Middle A and B areas;

[0051] Secondly, define the characteristic distance between the characteristic data of the light wear state and the medium wear state as R', then the center point of the characteristic data of the heavy wear state may be located in the following areas: the sector ring area with the characteristic data center point of the light wear state as the center and R' and 2R' as the radius, and the sector area with the characteristic data center point of the medium wear state as the center and 2R' as the radius;

[0052] Finally, 10 points are evenly selected in the area where the characteristic data center points in the heavy wear state determined in the above steps may exist. These points are used as the center of the circle and the average radius R of the existing characteristic data of healthy, light wear and moderate wear states is used as the radius to construct the 2D missing data and map it to 30 dimensions, as shown in the following example: Figure 2 shown.

[0053] (5.2) Using the existing feature data for healthy, lightly worn, and moderately worn states in the target domain feature dataset, a linear regression model is trained. Predictions are then made for the constructed censored data, retaining any censored data whose predicted results differ from a preset value by no more than a threshold. The preset value is set to a value close to the severe wear state. This step provides a preliminary constraint on the region where the data center is constructed. By setting the preset value for perturbation, the workload of subsequent iterations is reduced, improving iteration efficiency.

[0054] (5.3) Determine a new region based on the approximate location of the center point of the data obtained in step (5.2), and repeat steps (5.1)-(5.2) within the region, selecting the data whose prediction result is closest to the required label, that is, the severe wear state, until the same amount of feature data as the healthy, lightly worn, and moderately worn states is constructed, then stop, and add the constructed data and its labels to the target domain training set.

[0055] (6) Establish a transfer learning prediction model to predict the wear degree of mechanical parts. Specifically including:

[0056] (6.1) Establish a transfer learning prediction model, including establishing a source domain model and a target domain model. The source domain model is a one-dimensional convolutional neural network model consisting of three convolutional layers, three pooling layers, and two fully connected layers, such as Figure 3 As shown. The convolution layer performs convolution operation on the input layer signal, the formula is: in represents the jth value of the kth layer, M j Indicates input, Represents the value of each input element, represents the weight matrix of the convolution kernel, Denotes bias, and f(·) denotes activation function. The activation function used in this model is the Relu activation function, which can speed up the model convergence and avoid gradient saturation. The maximum pooling layer downsamples the data of the convolution layer and uses the maximum value within the operation area as the output. This can effectively simplify the complexity of network calculations and avoid overfitting. The calculation formula is: X k =f(down(X k-1 )+b), where X k is the output, X k-1 is the input, down(·) is the downsampling function, f(·) represents the activation function, and b represents the bias. The model can predict the source domain input data. The target domain model is the same as the source domain model, except that the parameters of the convolutional layer and pooling layer are frozen during training and no parameter updates are performed. The parameters of the one-dimensional convolutional neural network model in this embodiment are shown in Table 2-3. This model can predict the source domain input data and can adapt to different working conditions and train different target domain models.

[0057] Table 2 Structural parameters of the source domain model in the present invention

[0058]

[0059] Table 3 Target domain model structure parameters in the present invention

[0060]

[0061]

[0062] (6.2) Prediction of wear degree of mechanical parts

[0063] First, the features and labels in the source domain feature dataset are input into the source domain model to train the source domain model structure parameters;

[0064] Secondly, call the trained source domain model and freeze the parameters of all convolutional layers and pooling layers to form the target domain model, as shown in Table 3.

[0065] Then, the target domain feature set is divided into training set and test set in a ratio of 7:3, and the training set is input into the target domain model for secondary training of the fully connected layer parameters.

[0066] Finally, the test set is input into the trained target domain model to obtain the prediction results. The test results of the model are as follows Figure 4 As shown, A, B, C, and D represent healthy, lightly worn, moderately worn, and severely worn states, respectively, with a mean absolute error (MAE) of 0.208 mm and a mean square error (MSE) of 0.061 mm, indicating that the wear degree prediction model of the present invention can more accurately predict data under abnormal working conditions of centering change (C2).

[0067] The present invention also provides a system for predicting the degree of wear of mechanical components of subway doors under different working conditions, comprising:

[0068] The sample and label creation module is used to construct samples using the pre-processed motion data of different mechanical parts in different working conditions and different wear states, and label them according to the wear state;

[0069] The feature processing module is used to extract features from the sample data to form a source domain feature dataset and a target domain feature dataset; reduce the dimension of the target domain feature data and determine the feature data center point and radius of different labels;

[0070] The missing data construction and screening module is used to construct missing data based on the feature data center points and radii of different labels in the existing target domain and screen them through the linear regression model;

[0071] The modeling and prediction module is used to establish a transfer learning prediction model to predict the degree of wear of mechanical parts.

[0072] In order to further verify the effect of the present invention, when the model is trained without using the constructed deleted data, the prediction results of the test samples are as follows: Figure 5 As shown, the mean absolute error (MAE) is 0.242mm and the mean square error (MSE) is 0.142mm. Figure 4 and Figure 5 As can be seen, the model trained using the constructed severe wear data can effectively predict actual severe wear data, reducing prediction errors and meeting the needs of practical applications. This method can be applied to predict the wear degree of other mechanical wear components in rail door systems and can also be extended to other abnormal operating conditions.

[0073] In order to verify whether the constructed deleted data is consistent with the real deleted data, the constructed severe wear state data is used as a training set to train the target domain model, and the real deleted data (severe wear state) is used to test the model. The test results are as follows: Figure 6 As shown in FIG, D represents a severe wear state. The mean absolute error (MAE) is 0.218 mm, and the mean square error (MSE) is 0.067 mm, indicating that the data construction method used in the present invention can be applied to the construction of right-censored data of this abnormal working condition.

Claims

1. A method for predicting the wear degree of mechanical components of subway doors under different working conditions, characterized by: The following steps are involved: (1) Collecting motion data curves of door mechanical components in different wear states and performing preprocessing; (2) Obtain motion data samples under different working conditions and wear states of mechanical parts, and label them according to the wear state; (3) Feature extraction and standardization; (4) Feature dimensionality reduction and clustering to determine the feature data centers of different labels; (5) Construction and screening of censored data in the target domain; The step (5) comprises: (5.1) Based on the existing target domain feature data center points with different labels, determine the possible center points of the right-censored feature data, construct low-dimensional censored data around the right-censored feature data center points, and map them to high dimensions; (5.2) Use the existing target domain feature data with different labels to train a linear regression model to predict the constructed censored data, and filter out the censored data whose difference between the predicted result and the preset value is no greater than the threshold; (5.3) Based on the center point of the data obtained in step (5.2), continue to perform step (5.1) around it; after constructing the data, continue to perform step (5.2), select the data whose prediction result is closest to the required label, and add it to the target domain feature dataset; (6) Establish a transfer learning prediction model to predict the degree of wear of mechanical parts.

2. The wear degree prediction method of subway door mechanical components under different working conditions according to claim 1 is characterized in that: The different wear states in step (1) include healthy, lightly worn, moderately worn and severely worn states.

3. The wear degree prediction method of subway door mechanical components under different working conditions according to claim 2 is characterized in that: The different working conditions in step (2) include normal working conditions and abnormal working conditions. The samples under normal working conditions are the source domain, and the samples under abnormal working conditions are the target domain. The sample data of the target domain are right-censored data.

4. The wear degree prediction method of subway door mechanical components under different working conditions according to claim 3 is characterized in that: In step (3), the feature extraction adopts the wavelet packet decomposition method, the time domain data is decomposed by wavelet packets, and the energy proportion and energy entropy characteristics of the decomposed sub-band signal are extracted.

5. The method for predicting the wear degree of subway door mechanical components under different working conditions according to claim 1 is characterized in that: In the step (4), the feature dimension reduction adopts the linear discriminant analysis method.

6. The method for predicting the wear degree of subway door mechanical components under different working conditions according to claim 1 is characterized in that: In the step (4), a density-based clustering method is used to determine the feature data center points of different labels.

7. The method for predicting the wear degree of subway door mechanical components under different working conditions according to claim 1 is characterized in that: The step (5.1) comprises: First, based on the center point locations and trends of the feature data of healthy, lightly worn, and moderately worn states in the target domain feature dataset, the possible locations of the center point of the feature data of the heavy wear state are determined. Secondly, define the characteristic distance between the characteristic data of the light wear state and the medium wear state as R', then the center point of the characteristic data of the heavy wear state may be located in the following areas: the sector ring area with the characteristic data center point of the light wear state as the center and R' and 2R' as the radius, and the sector area with the characteristic data center point of the medium wear state as the center and 2R' as the radius; Finally, several points are evenly selected in the area where the central points of the characteristic data under the heavy wear state determined in the above steps may exist. These points are used as the center of the circle and the average radius R of the existing characteristic data of the healthy, lightly worn and moderately worn states is used as the radius to construct low-dimensional missing data and map them to high dimensions.

8. The wear degree prediction method of subway door mechanical components under different working conditions according to claim 1 is characterized in that: Establishing the transfer learning prediction model in step (6) includes establishing a source domain model and a target domain model, wherein the source domain model is a convolutional neural network model, and the target domain model is the same as the source domain model. During training, the parameters of the convolution layer and the pooling layer of the target domain model are frozen and no parameter update is performed.

9. The method for predicting the wear degree of mechanical components of subway doors under different working conditions according to claim 1 is characterized in that: The prediction of the degree of wear of mechanical parts in step (6) includes: inputting the features and labels in the source domain feature data set into the source domain model to train the structural parameters of the source domain model; calling the trained source domain model and freezing the parameters of all convolutional layers and pooling layers to form a target domain model; dividing the target domain feature set into a training set and a test set, and then inputting the training set in the target domain feature set into the target domain model to perform secondary training on the parameters of the fully connected layer; finally, inputting the test set in the target domain feature set into the trained target domain model to obtain a prediction result.

10. A system for predicting the degree of wear of mechanical components of subway doors under different working conditions, used to implement the method for predicting the degree of wear of mechanical components of subway doors under different working conditions as claimed in any one of claims 1 to 9, characterized in that: include: The sample and label creation module is used to construct samples using the pre-processed motion data of different mechanical parts in different working conditions and different wear states, and label them according to the wear state; The feature processing module is used to extract features from the sample data to form a source domain feature dataset and a target domain feature dataset; reduce the dimension of the target domain feature data and determine the feature data center point and radius of different labels; The missing data construction and screening module is used to construct missing data based on the feature data center points and radii of different labels in the existing target domain and screen them through the linear regression model; The modeling and prediction module is used to establish a transfer learning prediction model to predict the degree of wear of mechanical parts.

Images