Data-driven urban rail vehicle critical parts maintenance system based on cloud-edge collaboration

In view of the limitations of traditional maintenance methods for rail vehicles, this study proposes a data-driven maintenance system for critical parts of rail vehicles based on a cloud-side collaborative framework. Currently, there are several major issues with the maintenance of critical parts of rail vehicles, including long maintenance cycles, difficult troubleshooting, high maintenance costs, low maintenance efficiency, irregularities in data management, and a low level of informatization. To address these problems, a cloud-edge collaboration approach is adopted. The maintenance information of the vehicles is synchronized and exchanged in real-time with the cloud data center. Sensor and Internet of Things (IoT) technologies are used to collect real-time operational and status data of rail vehicles. The data is then analyzed and modelled in the cloud data center. The system achieves fault prediction and remaining useful life prediction for key components of the rail vehicles by applying machine learning algorithms. The experimental results demonstrate that compared to traditional maintenance methods, the system enables more effective troubleshooting and remaining useful life prediction of critical components of the vehicles. It improves maintenance efficiency and safety while also offering practical application value.


Introduction
In recent years, urban rail transit in China has experienced rapid progress, and the significance of maintaining urban rail vehicles has grown substantially.The bogie bearings play a crucial role in urban rail vehicles, and any failure in these bearings can severely impact operations.Therefore, ensuring the reliability and safety of the bearings is of utmost importance.However, the service life of bearings can vary significantly, even among bearings from the same batch operating under identical conditions.Consequently, the diagnosis of faults and prediction of bearing life have become urgent challenges that need to be addressed.
QIAN et al. [1] introduces the general maintenance methods for subway bogie bearings in China.Currently, the maintenance of bogie bearings on urban rail vehicles in China primarily relies on regular maintenance, which is divided into 5 levels: car repair in works, intermediate repair, periodical repair, monthly inspection, and train examination.During bogie repairs, which are conducted after the vehicles have traveled 600,000 kilometers, the bogies are disassembled, and the bearings are inspected.In recent

BiLSTM
The Bidirectional Long Short-Term Memory (BiLSTM) consists of a forward LSTM and a backward LSTM.This model leverages data information comprehensively, and its predictions often outperform those of LSTM models.The structure of BiLSTM is illustrated in figure 1.While the forward layer performs forward propagation, the reversed sequence is simultaneously computed in the backward LSTM using the same process.The final result is a concatenation of the outputs from both layers.Therefore, compared to the LSTM model, BiLSTM extracts more comprehensive features.This bidirectional propagation enables BiLSTM to have a stronger ability to capture sequential dependencies and extract more comprehensive features.

Remaining useful life
Remaining Useful Life (RUL) refers to the duration for which a system, component, or subsystem is expected to operate effectively before encountering a failure or reaching the end of its operational lifespan.Accurately predicting the RUL is crucial for minimizing losses caused by system failures and enhancing overall system reliability.It plays a vital role in proactive maintenance planning and resource allocation.The RUL prediction is projected as follows: Step 1. Data preprocessing.This involves analyzing and processing the raw signal using the Hilbert-Huang Transform (HHT) and calculating the Root Mean Square (RMS) value of the bearing.The HHT consists of two main components: Empirical Mode Decomposition (EMD) and Hilbert Spectrum Analysis (HSA).
Step 2: The running sequence of the bearing is divided into different bearing states using Singular Value Decomposition (SVD) with normalization.The main formula is shown below: U: an m×m unitary matrix (where the transpose of U is equal to the conjugate transpose of U), Σ is an m×n rectangular diagonal matrix, and VT is the transpose of an n×n unitary matrix.
Step 3. PCA downscaling analysis is performed using the Stacked Denoising Autoencoder (SDAE) model.The SDAE model consists of multiple Denoising Autoencoders (DAEs) stacked together to create a deep architecture.The structure of the SDAE model is illustrated in figure 5:

Cloud-edge collaboration architecture
The cloud-edge collaboration system is built with a cloud server as the upper computer, which serves as a platform for data visualization, storage management, and control in the monitoring system.It integrates edge devices for the collection of vibration signals, feature calculation, and remaining useful life prediction, forming the overall architecture of cloud-edge collaboration [5].

Edge
The data acquisition device at the edge is the WebDAQ504, which integrates data acquisition, edge computing, and model inference.In addition, it supports wireless transmission, enabling flexible device networking and model deployment.The WebDAQ504 performs real-time vibration data acquisition and edge computing for the key components of rail vehicles.It also performs high-speed protocol parsing via Ethernet and transfers the data to the cloud through the data acquisition system for analysis and diagnosis of the key component's status and health condition.

Cloud database
The cloud database is a database that is optimized or deployed in a virtual computing environment.Compared to traditional databases, cloud databases overcome limitations such as geographical and device constraints.The design of cloud-based databases includes user information databases, vibration signal storage databases, eigenvalue storage databases, and so on.

User interface
The user interface serves as the central display platform for the cloud-edge collaborative system's datasets.Users can access the website at any time and from anywhere to view real-time monitoring of bearing health status and remaining useful life.

Presentation of experimental data
5.1.1.Introduction to the XJTU-SY dataset.The XJTU-SY dataset comprises experimental bearings, specifically the LDK UER204 rolling bearings, which are exposed to three different operating conditions as outlined in Table 1 [6].

Experimental parameter settings
The RMS of the vibration signal is calculated to construct the health indicator (HI) feature vector.The RMS curve of the bearing is shown in figure 6.The number of iterations is set to 500, and the size of the small batch for each training iteration is 2. The learning rate is set to 0.02, and the gradient threshold is set to the default value of 1.

Results of analysis
To assess the effectiveness of the proposed model, the author performed validation using the corresponding data.Moreover, to demonstrate the superiority of the proposed model, a comparative analysis was conducted against the LSTM and TCN models.
The prediction platform is connected to the BiLSTM, LSTM, and TCN models.Upon reading the bearing data, the backend system automatically conducts the prediction of remaining useful life and generates a graph illustrating the predicted remaining useful life of the bearing.Based on the prediction results, a comparison is made.The specific operational details are depicted in the figure below.

Conclusion
This study proposes a BiLSTM-based algorithm for predicting the RUL of bearings.The algorithm utilizes the HHT method for data preprocessing, the normalized SVD for segmenting bearing operating sequences, and the SDAE model for PCA dimensionality reduction analysis.Experimental results using measured rolling bearing data demonstrate the effectiveness of the BiLSTM model in accurately predicting RUL.This approach provides a valuable tool for improving maintenance planning and minimizing losses caused by bearing failures.

Figure 1 .
Figure 1.BiLSTM network structure.While the forward layer performs forward propagation, the reversed sequence is simultaneously computed in the backward LSTM using the same process.The final result is a concatenation of the outputs from both layers.Therefore, compared to the LSTM model, BiLSTM extracts more comprehensive features.This bidirectional propagation enables BiLSTM to have a stronger ability to capture sequential dependencies and extract more comprehensive features.
The Temporal Convolutional Network (TCN) combines modeling capabilities in the temporal domain with the feature extraction power of convolutional operations.TCN effectively utilizes the inherent strengths of convolutions to extract features across time steps.It offers faster training compared to LSTM-based recurrent neural network models and excels at capturing dependencies in time sequences.The TCN structure is shown in figure 2.

Figure 3 .
Figure 3. RUL prediction schematic diagram.By extracting vibration signal data throughout the entire life cycle of the bearings, including both horizontal and vertical vibration signals, we can calculate the acceleration-time relationship.This relationship provides valuable insights into the behaviour of the bearings over their entire lifespan, as illustrated in figure 4.

Figure 4 .
Figure 4. Extraction of full-life data for bearings.The RUL prediction is projected as follows:Step 1.Data preprocessing.This involves analyzing and processing the raw signal using the Hilbert-Huang Transform (HHT) and calculating the Root Mean Square (RMS) value of the bearing.The HHT

Figure 5 .
Figure 5. SDAE structure.Step 4. Bearing RUL forecasting involves inputting the test set into the model and performing RUL predictions for the bearings.

Figure 6 .
Figure 6.Full life cycle RMS curve of bearings.The BiLSTM model utilizes the ADAM algorithm with four layers of bidirectional LSTM layers.The number of iterations is set to 500, and the size of the small batch for each training iteration is 2. The learning rate is set to 0.02, and the gradient threshold is set to the default value of 1.

Figure 8 .
Figure 8. Prediction results of TCN.It is evident from this dataset that the BiLSTM model outperforms the TCN model in terms of prediction accuracy.

Table 1 .
Operating conditions of XJTU-SY dataset.Introduction to the PHM2012 dataset.The PHM2012 dataset, provided by the FEMTO-ST research institute, consists of 6 training sets and 11 testing sets for estimating the RUL.The dataset is divided into three operating conditions, as shown in Table2.

Table 2 .
Operating conditions of PHM2012 dataset.