A prediction method for the remaining service life of bearings of urban rail trains

Abstract

The present invention discloses a method for predicting the remaining service life of urban rail train bearings, which includes the following steps: setting the rotational speed and sampling frequency of the bearings, collecting the original vibration signals of the bearings during the entire life cycle until failure, extracting the time-domain features of the original vibration signals of the bearings, and using a similarity measurement method to measure the similarity between the features; performing variational mode decomposition on the original vibration signals of the bearings to obtain modal component features, and using the energy entropy discrimination method based on the principle that the greater the energy entropy, the greater the information uncertainty; combining the time-domain features with the obtained modal component features and dividing them into a training set and a validation set; establishing an LSTM prediction network with optimized hyperparameters, inputting the feature matrix into the LSTM network, updating the weights through the backpropagation algorithm, updating the gradients through the Adam optimizer, and finally outputting the predicted value of the bearing RUL after calculation. The invention can scientifically, efficiently and comprehensively predict the RUL of urban rail train bearings, and has the advantages of quick and effective selection of network hyperparameters and good prediction performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault prediction of bearings of rotating components of rail transit trains, and particularly relates to a method for predicting the remaining service life of urban rail transit train bearings. Background Art

[0002] In recent years, urban rail transit has developed rapidly and become an important part of public transportation. The driving reliability and safety of urban rail transit trains have been widely concerned by all sectors of society. Bearings are indispensable components in the rotating mechanical equipment of trains. Research shows that about 50% of motor failures are caused by rolling bearing failures. Therefore, predicting the remaining service life of bearings is of great significance in the operation and maintenance of trains. At present, the industrial community mostly adopts periodic preventive maintenance for train components, combining major repairs, medium repairs, and minor repairs for maintenance and inspection, which can reduce the occurrence of failures as much as possible, but there are still phenomena of over-maintenance and under-maintenance. In order to ensure the reliable operation of trains, reduce the number and frequency of failures, and reduce the economic losses and even endangerment to personal safety caused by failure shutdowns, the academic and industrial communities are widely concerned about using the emerging technology of Prognostics and Health Management (PHM) to maintain trains. PHM refers to using sensors to monitor and collect system operation status data, perform data processing and feature extraction, and evaluate the health status of the monitored object and predict the occurrence of its failures through fault diagnosis and prediction models, so as to provide guidance and decision-making for the maintenance and repair of the monitored object. The remaining useful life (RUL) is an important basis for evaluating the health status of the object. PHM effectively shows its degradation trend by monitoring the bearing status and predicting its RUL, dynamically formulates and optimizes maintenance strategies according to the operation status, thereby avoiding failures of rotating components such as motors and improving the reliability and safety of train operation.

[0003] Bearing life prediction mainly includes two stages: data processing and feature extraction, and RUL prediction. At present, signal processing methods such as time-frequency decomposition and modal decomposition can be used to extract bearing features, but there is a problem of insufficient degradation state information caused by using single-domain features. In the RUL prediction method, the degradation state based on model fitting depends on expert prior knowledge, and it is difficult to achieve a high level of prediction accuracy and applicability. Data-driven methods mine degradation information in state data through machine learning, deep learning, etc., and map the relationship between monitoring signals and RUL values. However, existing RUL prediction methods such as support vector machines have great limitations in using the time series information of bearing vibration signals. Long short-term memory networks (LSTM) are suitable for dealing with long-term dependence problems and are widely used in bearing RUL prediction. The setting of LSTM hyperparameters will affect the prediction accuracy. Manually selecting hyperparameters is laborious and inaccurate. Some optimization algorithms for optimizing hyperparameters are prone to falling into local optima, with poor optimization ability and low accuracy, and it is difficult to comprehensively and accurately predict the remaining service life of bearings. Summary of the Invention

[0004] The object of the present invention is to provide a method for predicting the remaining service life of urban rail transit train bearings. The prediction method of the present invention retains the time series characteristics of the bearing vibration signal in the data processing and feature extraction stage, and adaptively optimizes and selects the hyperparameters of the neural network through the Harris Hawks Optimization (HHO) algorithm in the RUL prediction stage to improve the prediction accuracy; the present invention can scientifically, efficiently and comprehensively predict the RUL of urban rail transit train bearings, and has the advantages of diverse characteristic information, fast and effective selection of LSTM network hyperparameters, and good prediction performance. To achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] The present invention provides a method for predicting the remaining service life of urban rail transit train bearings, including the following steps:

[0006] Step 1: Set the rotational speed and sampling frequency of the bearing, and collect the original vibration signal of the whole life cycle of the bearing running to failure;

[0007] Step 2: Extract the time domain features of the original vibration signal of the bearing, use the similarity measurement method to measure the similarity between the features, remove the features with low discrimination, and select the optimal feature parameters representing the bearing degradation ability;

[0008] Step 3: Perform variational mode decomposition on the original vibration signal of the bearing to obtain modal component features, use the energy entropy discrimination method, and select the modal components according to the principle that the greater the energy entropy, the greater the information uncertainty, with the highest energy and the smallest energy entropy as the criteria, expand the frequency domain information on the basis of the time domain features, and calculate the energy and energy entropy of the modal components respectively;

[0009] Step 4: Combine the time domain features obtained in Step 2 with the modal component features obtained in Step 3 to obtain a total of k feature vectors with a sequence length of n. The i-th feature vector is F i =[F1(i), F2(i),..., F n (i)], where F n (i) represents the eigenvalue corresponding to the n-th sampling time point. Then, combine the k feature vectors into a feature matrix F = [F 1 , F 2 ,..., F k ; After removing the outliers in each feature vector and complementing them by cubic spline interpolation, normalize the feature vectors to between 0 and 1, and then divide the feature matrix into a training set and a validation set;

[0010] Step 5: Use the LSTM network as the main network for RUL prediction. Set the hyperparameters of the LSTM network as the population position. Train the LSTM network with the training set data, use the mean square error of the prediction results of the validation set data as the fitness function, and use the Harris hawks optimization algorithm to optimize the LSTM hyperparameters. Iteratively update the optimal hyperparameter combination every time the LSTM network is trained; establish an LSTM prediction network with the optimized hyperparameters, input the feature matrix into the LSTM network, and output the bearing RUL prediction value after calculation.

[0011] In a further preferred embodiment of the above solution, in the energy entropy discrimination method of step 3, the energy satisfies the following expression:

[0012]

[0013] where u i is the i-th modal component sequence, t = {0, 1,..., n}, and n is the sequence length;

[0014] The energy entropy satisfies:

[0015] H i = -p i log 10 p i , (2);

[0016] where p i is the ratio of the energy of the i-th modal component to the total energy,

[0017]

[0018] where i = {0, 1,..., K}, and K is the total number of modal components.

[0019] In a further preferred embodiment of the above solution, in step 4, the feature matrix is constructed by extracting time-domain features and selecting variational mode decomposition modal components by the energy entropy discrimination method. The constructed feature matrix contains time-domain, frequency-domain, and entropy-domain feature information of bearing degradation.

[0020] In a further preferred embodiment of the above solution, in step 5, the specific process of using the Harris hawks optimization algorithm to optimize the LSTM network hyperparameters is as follows:

[0021] Step 5.1: Select the hyperparameters of the LSTM network, namely learning rate, batch size, number of epochs, number of units in the LSTM layer, and number of units in the Dense layer, as the population position, and set the population size and the maximum number of iterations;

[0022] Step 5.2: Initialize the Harris hawk population and the positions and energies of the prey. Train the LSTM network with the training set data, use the mean squared error of the predicted results with the validation set data as the fitness function, and compare and update the fitness and position of the optimal individual.

[0023] Step 5.3: Calculate the escape energy of the prey. When the escape energy is greater than or equal to 1, perform global search to update the position. When the escape energy is less than 1, perform local exploration to update the position. During local exploration, when the prey's escape energy is still greater than or equal to 0.5, the Harris hawk consumes its energy through soft encirclement. When the escape energy is less than 0.5, the Harris hawk directly captures the prey through hard encirclement. When the prey cannot escape, adopt the team rapid dive strategy to attack.

[0024] Step 5.4: Calculate the fitness of the individual after updating the position through the Harris hawk predation strategy, and compare and update the fitness and position of the optimal individual.

[0025] Step 5.5: Determine whether the termination condition of the optimization is met, that is, whether the maximum number of iterations set is reached or the fitness meets the requirements. If it is met, output the optimized fitness and the hyperparameters corresponding to the optimal population position. If not, return to Step 5.3 to continue the optimization.

[0026] In a further preferred embodiment of the above solution, during the training process of the LSTM network, the Harris hawk optimization algorithm is used to adaptively iteratively update its optimal hyperparameters, and the optimal LSTM network is directly constructed therefrom without the need for re-training.

[0027] In summary, due to the adoption of the above technical solution in the present invention, the present invention has the following remarkable effects:

[0028] The present invention comprehensively considers the problems existing in signal processing and the RUL prediction network. Signal processing methods such as time-domain analysis and variational mode decomposition are used to extract features, integrating time-domain and frequency-domain information to avoid the problem of incomplete effective degradation information of a single feature. The LSTM network is used as the main RUL prediction network, and the Harris hawk optimization algorithm is used to adaptively optimize its hyperparameters during the network training process, effectively solving the problems of too much manual selection work for hyperparameters and difficulty in finding the optimal hyperparameters for prediction performance. The method of the present invention is scientific and comprehensive for predicting the RUL of the bearings of urban rail trains, and is an effective and practical prediction method that can be widely applied to predicting the remaining service life of the bearings of various rotating components of trains. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of a method for predicting the remaining service life of the bearings of urban rail trains;

[0030] Figure 2 is the fitness curve of the HHO optimized LSTM hyperparameters;

[0031] Figure 3 It is a comparison graph of the prediction curves of the HHO-optimized LSTM and the original network. Detailed implementation manners

[0032] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the following provides preferred embodiments with reference to the accompanying drawings and further elaborates on the present invention in detail. However, it should be noted that many details listed in the specification are only for enabling the reader to have a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be implemented even without these specific details.

[0033] As Figures 1 to 3 shown, a method for predicting the remaining service life of a bearing of an urban rail transit train according to the present invention includes the following steps:

[0034] Step 1: Set the rotational speed and sampling frequency of the bearing, and collect the original vibration signals of the entire life cycle of the bearing running until failure.

[0035] Step 2: Extract the time-domain features of the original vibration signals of the bearing, use the similarity measurement method to measure the similarity degree between the features, remove the features with low discrimination, and select the feature parameters that best represent the degradation ability of the bearing; the time-domain features are shown in Table 1;

[0036] Table 1 Time-domain features

[0037]

[0038] Step 3: Perform variational mode decomposition on the original vibration signals of the bearing to obtain the features of the modal components (IMF components), use the energy entropy discrimination method, and select the modal components according to the principle that the greater the energy entropy, the greater the information uncertainty, with the criteria of the highest energy and the smallest energy entropy, expand the frequency-domain information on the basis of the time-domain features, and calculate the energy and energy entropy of the modal components respectively;

[0039] The energy satisfies the following expression:

[0040]

[0041] where u i is the sequence of the i-th modal component, t = {0, 1,..., n}, and n is the sequence length;

[0042] The energy entropy satisfies:

[0043] H i = -p i log 10 p i , (2);

[0044] where p i is the ratio of the energy of the i-th modal component to the total energy,

[0045]

[0046] where i = {0, 1, ..., K}, and K is the total number of modal components.

[0047] Step 4: Combine the time-domain features obtained in Step 2 with the modal-component features obtained in Step 3 to obtain k feature vectors of sequence length n. The i-th feature vector is F i = [F1(i), F2(i), ..., F n (i)], where F n (i) represents the eigenvalue corresponding to the n-th sampling time point. Then, combine the k feature vectors into a feature matrix F = [F 1 , F 2 , ..., F k . After removing the outliers within each feature vector and filling them in by cubic spline interpolation, normalize the feature vectors to between 0 and 1, and then divide the feature matrix into a training set and a validation set. In the present invention, by combining time-domain feature extraction and energy-entropy discrimination method to select variational mode decomposition modal components to construct a feature matrix, the constructed feature matrix contains time-domain, frequency-domain, and entropy-domain feature information of bearing degradation;

[0048] Step 5: Use the LSTM network as the main network for RUL prediction. Set the hyperparameters of the LSTM network as the population position, train the LSTM network with the training set data, use the mean square error of the prediction results of the validation set data as the fitness function, and use the Harris hawk optimization algorithm to optimize the LSTM hyperparameters. Iteratively update the optimal hyperparameter combination every time the LSTM network is trained. Establish an LSTM prediction network with the optimized hyperparameters, input the feature matrix into the LSTM network, update the weights through the backpropagation algorithm, update the gradients through the Adam optimizer, and output the bearing RUL prediction value after calculation. In the embodiment of the present invention, the specific process of using the Harris hawk optimization algorithm to optimize the LSTM network hyperparameters is as follows:

[0049] Step 5.1: Select the hyperparameters of the LSTM network, namely learning rate, batch size, number of iterations, number of units in the LSTM layer, and number of units in the Dense layer, as the population position, and set the population size and the maximum number of iterations;

[0050] Step 5.2: Initialize the Harris hawk population and the positions and energies of the prey. Train the LSTM network with the training set data, use the mean square error of the prediction results of the validation set data as the fitness function, and compare and update the fitness and position of the optimal individual;

[0051] Step 5.3: Calculate the escape energy of the prey. When the escape energy is greater than or equal to 1, perform global search to update the position. When the escape energy is less than 1, perform local exploration to update the position. During local exploration, when the prey's escape energy is still greater than or equal to 0.5, the Harris hawk consumes its energy through soft encirclement. When the escape energy is less than 0.5, the Harris hawk directly captures the prey through hard encirclement and adopts the team rapid dive strategy to attack when the prey cannot escape.

[0052] Step 5.4: Calculate the fitness of the individual after updating the position through the Harris hawk predation strategy, compare and update the fitness and position of the optimal individual.

[0053] Step 5.5: Determine whether the termination condition of the optimization is satisfied, that is, whether the set maximum number of iterations is reached or the fitness meets the requirements. If satisfied, output the optimized fitness and the hyperparameters corresponding to the optimal population position. If not satisfied, return to Step 5.3 to continue the optimization.

[0054] In the present invention, the bearing vibration signal is monitored and collected, the time-domain features of the vibration signal are extracted and selected through similarity measurement, then variational mode decomposition is adopted and combined with the energy entropy discrimination method to select the modal component features, and multiple feature vectors F i =[F1(i), F2(i),..., F n (i)] are combined to construct the feature matrix F = [F 1 , F 2 ,..., F k . The training set and the validation set are divided. Subsequently, the training set is input into the Harris hawk optimized LSTM network for model training. The mean square error of the prediction results of the validation set data is used as the fitness function to adaptively optimize the hyperparameters, and the LSTM prediction model is established with the iteratively updated optimal hyperparameters to predict the bearing RUL. Taking the remaining service life of a rolling bearing as an example, the load of the test bench for collecting the bearing vibration signal in the experiment is 12 kN, the rotational speed is 2100 rpm, the sampling frequency is 25.6 kHz, the sampling interval is 1 min, and the sampling time for each time is 1.28 s; and the prediction is carried out according to the following steps:

[0055] 1). Real-time monitor and collect the bearing vibration signal, sample once every 1 min, and sample 1.28 s at 25.6 kHz each time to obtain 32768 data points. A total of 123 groups of samples are sampled until the bearing fails.

[0056] 2). Perform time-domain analysis on the collected bearing vibration signal to obtain 14 time-domain features, and select the feature types with smaller similarity according to the similarity between the features to characterize the degradation state.

[0057] 3) Variationally decompose the collected bearing vibration signals to obtain 5 modal components. According to the energy entropy discrimination method, select the modal component with as large energy as possible and as small energy entropy as possible to characterize the degradation state;

[0058] 4) Normalize, remove outliers and interpolate and complete each obtained eigenvector respectively, construct an eigenmatrix, and divide the first 90 samples into the training set and the 91st to 120th samples into the validation set;

[0059] 5) Use the training set data as the input of the LSTM network, use the mean square error of the validation set data prediction as the fitness function of the Harris hawk optimization. Set the population size to 10 and the maximum number of optimization times to 10. The population positions correspond to hyperparameters such as the learning rate, batch size, and number of epochs. While training the LSTM network, adaptively optimize the best hyperparameter combination, and establish a prediction network with the hyperparameters after the optimization iteration is completed. The fitness curve during the hyperparameter optimization process of the Harris hawk optimization algorithm is as Figure 2 shown. It can be seen that it has strong global search ability, fast convergence speed, and the fitness value reaches convergence after four optimizations.

[0060] Table 2 Hyperparameter range

[0061]

[0062] 6) For the characteristic sample data, use the long short-term memory network optimized by the adaptive HHO to predict the remaining service life of the bearing. Compare the prediction accuracy with the LSTM optimized by multiple other algorithms and the original LSTM network model. The mean square error (MSE) index is reduced by 27.3% - 62.8%, and the mean absolute error (MAE) index is reduced by 19.0% - 40.9%. In comparison, the prediction performance of the method of the present invention has been significantly improved. The comparison curve of the prediction results of the LSTM network optimized by HHO and the original network is as Figure 3 shown.

[0063] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of a bearing of an urban rail train, characterized in that: It includes the following steps: Step 1: Set the rotational speed and sampling frequency of the bearing, and collect the original vibration signals of the bearing during its entire life cycle until failure. Step 2: Extract the time-domain features of the original vibration signals of the bearing, use the similarity measurement method to measure the similarity between features, remove the features with low discrimination, and select the feature parameters that best characterize the degradation ability of the bearing. Step 3: Perform variational mode decomposition on the original vibration signals of the bearing to obtain modal component features. Use the energy entropy discrimination method. According to the principle that the greater the energy entropy, the greater the information uncertainty, select the modal component with the highest energy and the smallest energy entropy. Expand the frequency-domain information based on the time-domain features, and calculate the energy and energy entropy of the modal components respectively. Step 4: Combine the time-domain features obtained in Step 2 with the modal component features obtained in Step 3 to obtain a total of k feature vectors with a sequence length of n. The i-th feature vector is F i =[F1(i), F2(i),..., F n (i)], where F n (i) represents the eigenvalue corresponding to the n-th sampling time point. Then, combine the k feature vectors into a feature matrix F = [F 1 , F 2 ,..., F k ; After removing the outliers within each feature vector and completing the interpolation through cubic splines, normalize the feature vectors to between 0 and 1, and then divide the feature matrix into a training set and a validation set; Step 5: Use the LSTM network as the main network for RUL prediction. Set the hyperparameters of the LSTM network as the population position. Train the LSTM network with the training set data, use the mean square error of the prediction results of the validation set data as the fitness function, and use the Harris hawk optimization algorithm to optimize the LSTM hyperparameters. Iteratively update the optimal hyperparameter combination every time the LSTM network is trained. Build an LSTM prediction network with the optimized hyperparameters, input the feature matrix into the LSTM network, update the weights through the backpropagation algorithm, update the gradients through the Adam optimizer, and output the predicted value of the bearing RUL after calculation.

2. The prediction method for the remaining service life of the bearing of an urban rail train according to claim 1, wherein: In the energy entropy discrimination method of Step 3, the energy satisfies the following expression: where u i is the i-th modal component sequence, t = {0, 1, ..., n}, and n is the sequence length; The energy entropy satisfies: H i = -p i log 10 p i , (2); where p i is the ratio of the energy of the i-th modal component to the total energy, where i = {0, 1,..., K}, and K is the total number of modal components.

3. A method for predicting the remaining service life of an urban rail transit train bearing according to claim 1, characterized in that: In Step 4, the feature matrix is constructed by extracting time-domain features and selecting variational mode decomposition modal components using the energy entropy discrimination method. The constructed feature matrix contains time-domain, frequency-domain, and entropy-domain feature information of bearing degradation.

4. A prediction method for the remaining service life of a bearing of an urban rail train according to claim 1, characterized in that: In Step 5, the specific process of using the Harris hawk optimization algorithm to optimize the LSTM network hyperparameters is as follows: Step 5.1: Select the hyperparameters of the LSTM network, namely learning rate, batch size, number of iterations, number of units in the LSTM layer, and number of units in the Dense layer, as the population position, and set the population size and the maximum number of iterations. Step 5.2: Initialize the Harris hawk population and the position and energy of the prey. Train the LSTM network with the training set data, use the mean square error of the prediction results of the validation set data as the fitness function, and compare and update the fitness and position of the optimal individual. Step 5.3: Calculate the escape energy of the prey. When the escape energy is greater than or equal to 1, perform global search to update the position. When the escape energy is less than 1, perform local exploration to update the position. During the local exploration process, when the escape energy of the prey is still greater than or equal to 0.5, the Harris hawk consumes its energy through soft siege. When the escape energy is less than 0.5, the Harris hawk directly captures the prey through hard siege and uses the team rapid dive strategy to attack when the prey cannot escape. Step 5.4: Calculate the fitness of the individual after updating the position using the Harris hawk predation strategy, and compare and update the fitness and position of the optimal individual. Step 5.5: Determine whether the termination condition of optimization is satisfied, that is, whether the maximum number of iterations set is reached or the fitness meets the requirements. If satisfied, output the optimized fitness and the hyperparameters corresponding to the optimal population position. If not satisfied, return to Step 5.3 to continue the optimization search.

5. A prediction method for the remaining service life of a bearing of an urban rail train according to claim 4, characterized in that: During the training process, the LSTM network adaptively iteratively updates its optimal hyperparameters through the Harris hawks optimization algorithm, and directly constructs the optimal LSTM network with these hyperparameters, without the need for retraining.

Images