Onboard monitoring plays an important role in real-time condition assessment of rail systems. However, the data amount is typically tremendous due to the high sampling rate needed and long traveling distance, especially for vibration data collected from high-speed trains (HSTs). As for fault diagnosis of mechanical systems, compressive sensing (CS) has been increasingly adopted to reduce the data amount. In comparison to rotary bearings and bolted joints in machinery that operate in relatively steady working environments, HSTs run in an open and varying environment throughout the traveling mileage, and the data amount is normally much larger, making it tricky to directly apply the classical CS methods. This study aims to bridge the gap by investigating the sparsity of HST vibration signals and CS approaches. Considering the lack of sparsity and long reconstruction time, we propose an efficient adaptive CS approach for dynamic responses of HSTs. More specifically, we unroll the iterative soft thresholding algorithm (ISTA) in a deep learning (DL) framework and configure it into a data reconstruction machine. Compared to the conventional CS methods, our approach exhibits two advantages: (i) The dictionary learning and signal reconstruction are integrated into one neural network and can be conducted in an end-to-end manner; (ii) the process is highly efficient since encapsulating ISTA in a DL framework can naturally leverage the capability of GPU. The proposed approach is validated using data collected from an in-service HST, and results show that our approach achieves superior reconstruction performance over fixed bases and redundant dictionaries.adaptive compressive sensing, deep learning, high-speed train, onboard monitoring, sparse coding
| INTRODUCTIONCondition monitoring is vital for the safety and punctuality of railway systems, and various onboard monitoring systems have been implemented on high-speed trains (HSTs) collecting real-time data of key components. Specifically,
Ultrasonic guided waves (UGWs) have been extensively utilized in nondestructive testing and structural health monitoring (SHM) for detection and real-time monitoring of structural defects. By implementing multiple piezoelectric sensors onto a plane of the target structure to form a sensor network, damages within the sensing range can be detected or even visualized through a pitch-catch configuration. On the other hand, deep learning (DL) techniques have recently been widely used to aid UGW-based SHM when the waveform is over complicated to extract a specific mode of interest due to irregular structure or boundary reflections. However, not too much research work has been conducted to thoroughly combine sensor networks with DL. Existing research using DL approaches is mainly used to train and interpret waveforms from isolated sensor pairs. The topological structure of sensor layout and sensor-damage relative positions are hardly considered in the data-driven process. Motivated by these concerns, this study offers a first-of-its-kind perspective to interpret UGW data collected from a sensor network by mapping the physical sensor-damage layout into a graph, in which sensors and potential damages serve as graph vertices bearing heterogenous properties upon coming to UGWs and the process of UGW transmission between sensors are encapsulated as wavelike message passing between the vertices. A novel physics-informed end-to-end graph neural network model, named as WaveNet, was exquisitely and meticulously developed. By utilizing wave information and topological structure, WaveNet enables inference of multiple damages in terms of severity and location with satisfactory accuracy, even when the waveforms are chaotic, and the sensor arrangement is different at the training and testing stages. More importantly, beyond the SHM scenario, the present study is expected to enlighten new thinking on interconnecting physical wave propagation with virtual messaging passing in neural networks.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.