A Real-Time Polygonal Wheel-Rail Force Identification Method Based on Convolutional Neural Networks (CNN)

Urban Rail Transit https://doi.org/10.1007/s40864-024-00237-1 http://www.urt.cn/ ORIGINAL RESEARCH PAPERS A Real‑Time Polygonal Wheel‑Rail Force Identification Method Based on Convolutional Neural Networks (CNN) Zeteng Zhang1,2 · Jinhai Wang1,2 Dechen Yao1,2 · Jianwei Yang1,2 · Received: 11 April 2024 / Revised: 28 September 2024 / Accepted: 17 October 2024 © The Author(s) 2025 Abstract Wheel-rail force identification is one of the most challenging issues in the railway industry, which can provide real-time safety evaluation and fault diagnosis for railway vehciles in operation. A new real-time polygonal wheel-rail force identification method is proposed. Firstly, aiming at the characteristic of high-order polygon feature frequency of wheelset, multi-rigid dynamics model and flexibility-rigid dynamics model are established in SIMPACK to obtain data. Then, the data of rail force and vibration acceleration of vehicle components are normalized, graphically and discretized processed. Finally, the data are input into the designed real-time polygonal wheel-rail force identification network for learning. Simulation data are used for network learning and comparison. The experimental results demonstrate that the vibration acceleration of vehicle components along with the vertical displacement data of primary springs, exhibit optimal performance in the identification of wheel-rail forces when employed as inputs for the network. Interval usage polygonal data of different orders to fine-tuning the network yield the most accurate identification of polygonal wheel-rail forces. For the multi-rigid model, the average absolute error and determination coefficient of vertical force identification are 1039 N and 0.895, and the lateral force is 362 N and 0.833. For the flexibility-rigid model are 1529.2 N and 0.929 in vertical force identification, and 1734.5 N * Jinhai Wang 1 School of Mechanical‑Electronic and Vehicle Engineering, Engineering and Architecture, Beijing University of Civil Engineering and Architecture, Beijing 100044, China 2 Beijing Key Laboratory of Performance Guarantee on Urban Rail Transit Vehicles, Beijing 100044, China Communicated by Liang Gao. and 0.948 in lateral force identification. Furthermore, the wheel-rail identification can be real-time because the average calculation time is far less than the sampling time. Consequently, the proposed method can provide strong support for the safety evaluation of running railway vehicles based on monitoring data. Keywords Wheel-rail forces · Forces identification · Deep learning · Vehicle system dynamics · Wheel polygonization 1 Introduction Ensuring the safety of train operations is a fundamental requirement for the railway system. Therefore, wheel-rail force identification becomes a necessary engineering means for train operation, which can assist in analyzing derailment, structural durability, and reliability. Traditionally, wheel-rail force identification mainly adopts the direct method, which has been used in vehicle design and test phases. With condition monitoring and big data technology development, the indirect method has gradually become the focus of wheelrail force identification. It can make an essential contribution to ensuring the safety of trains at all times. Wheel-rail force identification can be divided into direct and indirect force identification methods. One of the main direct force identification methods is installing the measuring wheelset [1, 2]. The advantage of this method is that it can perform long-distance, high-precision, and real-time wheel-rail force identification. However, the manufacturing process of measuring wheelsets is complex, and sensor maintenance is challenging. Therefore, the application of measuring wheelsets comes with high costs. Another direct method is the ground-based wheel-rail force monitoring Vol.:(0123456789) Urban Rail Transit systems. This system utilizes strain gauges installed at the track to measure rail strains and obtains the wheel-rail forces based on the strain–force relationship [3, 4]. The drawback of this method is that it cannot accurately identify wheel-rail forces over long distances [5, 6]. Therefore, it can be seen that direct methods for wheel-rail force identification have many difficulties and are challenging to widely used. The indirect method mainly refers to wheel-rail force identification techniques, which utilize vehicle dynamics data to calculate wheel-rail forces [7, 8]. In vehicle dynamics experiments, measuring instruments commonly include accelerometers, displacement sensors, and gyroscopes. These measuring instruments have low usage costs and can achieve long-distance data monitoring. Therefore, indirect wheel-rail force identification technology is easier to promote and has significant research significance and engineering value. Traditionally, wheel-rail force identification techniques mainly include the frequency domain, time domain, and vehicle state estimation methods [9–13]. Uhl [14] studied the objective function by using the least squares error between the simulated system response and the measured system response. As this method belongs to the frequencydomain approach in force identification, one drawback is that it cannot provide real-time monitoring of the wheel-rail forces. Xia et al. [15, 16] developed a linear model for the freight train system. They computed the wheel-rail forces using the measured lateral and vertical accelerations, roll, pitch, and yaw angular velocities on the freight car body. This method can accurately predict the variation trend of the wheel-rail forces. However, predicting high-frequency wheel-rail force signals from low-frequency input data is challenging, which leads to lower numerical accuracy in the calculation of wheel-rail forces. It can be observed that early wheel-rail force identification techniques had issues such as low accuracy and the inability to identify wheel-rail forces in real-time. In recent years, with the significant advancement of computer technology, deep learning has shown vast prospects in multiple areas such as bearing life prediction, fault diagnosis, and image identification [17–21]. Deep learning has also been integrated into the research of wheel-rail force identification technology [22–24]. Pang et al. [25] used track unevenness as input and predicted the derailment coefficient based on the NARX neural network. Martin et al. [26] used the SAMSCAR multibody code to generate training data for the recurrent neural network. However, these methods require regular detection of the geometric shape of the track profile. This leads to limitations in identifying wheel-rail forces as it relies on the monitoring of steel rail profiles, resulting in additional costs. Rushabh et al. [27] studied the identification performance of MLP networks for lateral and vertical forces and track unevenness under different inputs. The drawback of this method is that it lacks realt (...truncated)