LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

by

LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

0 Introduction

The traction and braking forces of rail vehicles come from the adhesion between the wheel and rail, and the adhesion coefficient is influenced by various factors such as speed, axle load, and the condition of the wheel-rail surface. These factors are affected by the natural environment, leading to complex and variable adhesion characteristics. China has a large climate variation, a wide variety of rail vehicles, and frequent slip conditions, prompting research institutions to study anti-slip control. Sun Huanyang et al. proposed a slip detection and anti-slip control strategy for the Fuxing bullet train; Cao Hongfa proposed an anti-slip control method based on fuzzy control algorithms; Xu Chuanfang investigated train speed tracking control strategies and adhesion anti-slip control issues, proposing an active adhesion anti-slip control algorithm. When the braking system detects slipping, wear has already occurred between the wheel and rail. Anti-slip control can prevent further damage but relies on the effectiveness of the slip detection and anti-slip control algorithms. The earlier the slip state is detected, the more beneficial it is for anti-slip control.

LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

1 The Impact of Slip Detection Algorithms on Adhesion Control

This section discusses the impact of slip detection algorithms on adhesion control. Creep is the phenomenon of slight sliding at the contact surface between the wheel and the rail, which is crucial for generating adhesion. The adhesion coefficient is the ratio of adhesion force to axle load, and its maximum value determines the maximum allowable braking force between the wheel and rail. Adhesion utilization is the ratio of the instantaneous adhesion coefficient to the maximum adhesion coefficient. The braking system should improve adhesion utilization to achieve maximum braking force, but the closer the braking force is to the maximum, the higher the probability of slipping. Research on the adhesion coefficient mainly relies on experiments. The Oldrich Polach model is an approximate fit for the adhesion coefficient. The train braking system needs to collect axle speed in real-time to detect slipping states. Common slip detection criteria include speed difference, deceleration, and slip rate. The benchmark axle speed is a virtual axle speed used for slip detection. Existing slip detection methods have shortcomings, such as estimation error accumulation leading to simulated axle speeds deviating from actual vehicle speeds, affecting the validity of speed difference criteria. To address these issues, China National Railway Group Co., Ltd. has improved adhesion utilization and shortened braking distances by adjusting the sensitivity of slip detection conditions through software and other measures, thereby reducing total air consumption.

LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

2 Benchmark Axle Speed Prediction Based on LSTM

This section introduces a method for predicting the benchmark axle speed of rail vehicles based on the LSTM network. The LSTM network addresses long dependency issues and alleviates problems of gradient explosion and vanishing by incorporating cell state. The workflow for predicting the benchmark axle speed time series includes steps such as data collection and preprocessing, and model establishment. The real-time data recorded by the braking system EBCU supports the benchmark axle speed prediction, selecting 5 features as model inputs. The LSTM model is implemented based on TensorFlow 2.0 and Keras frameworks, using a multivariate, multi-step prediction framework, with 5-dimensional time series input of continuous window length n, and 5-dimensional time series prediction length m. The model employs the ReLU activation function, uses RMSE as the evaluation metric, and is optimized using the Adam optimizer, with Dropout set to 0.2 to mitigate overfitting.

LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

3 Model Training and Evaluation

This section describes the training and evaluation process of the benchmark axle speed prediction model for rail vehicles based on the LSTM network. First, a dataset covering normal braking data from 300 to 0 km/h was prepared, normalized, and then 80% was used as the training set and 20% as the validation set. The model input consists of normalized historical driving data from the braking system, while the output is a 5-dimensional time series, primarily predicting the benchmark axle speed vreft. The root mean square error (RMSE) is used as the evaluation metric. By adjusting key parameters such as batch size, sliding window length, prediction length, number of LSTM neurons, and training epochs, multiple training evaluations were conducted to find the optimal parameter combination. The experimental results indicate that the optimal parameters are batch_size=20, n=40, m=10, L=16, and epochs=30. The model converges quickly in the first 10 epochs and then reaches a stable state, with the validation loss convergence curve being smoother and no severe overfitting occurring. The impact of the prediction length m on the prediction performance was analyzed. The benchmark axle speed predicted by LSTM is closer to the actual speed than existing algorithms, even in full axle slip conditions, avoiding the defects of the original method. The LSTM predicted values can detect slipping earlier, allowing for proactive anti-slip control. However, the longer the prediction length, the worse the convergence; conversely, shorter lengths yield better convergence. A balance between prediction accuracy and duration is necessary based on control requirements. As long as the anti-slip control is timely, the next prediction input will be more accurate, improving the accuracy of the predicted values for the next time period.

LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

4 Conclusion

The benchmark axle speed predicted by the LSTM network is closer to the actual train speed curve, especially under full axle slip conditions, overcoming the shortcomings of existing algorithms and serving as a supplement. This study takes a single model as an example, and the model may have limitations that need to be retrained for different vehicle types and lines. The model establishment and training use Python, requiring high hardware configuration, with computation speed measured in seconds; it is recommended to configure a dedicated NPU processor for real-time online learning.

This content is automatically generated by AI for reference only. For details, see “China Railway” 2025 Issue 1.

Related Information

Authors:

Sun Weibing, Nanjing Railway Vocational and Technical College, Nanjing CRRC Puzhen Haite Brake Equipment Co., Ltd.

Yang Lei, Nanjing CRRC Puzhen Haite Brake Equipment Co., Ltd.

Fang Song, Nanjing Railway Vocational and Technical College.

Citation:Sun Weibing, Yang Lei, Fang Song. LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles [J]. China Railway, 2025(1): 92-99.

LSTM-Based Benchmark Axle Speed Prediction Method for Rail Vehicles

END

Leave a Comment