Data-driven diagnosis of railway wheel flat based on multi-channel vibration data fusion

Rail vehicle wheel flats are a major source of damage to both railway infrastructure and rolling stock, contributing to the increased noise, vibration, and even safety–critical incidents such as derailments. This study proposes a data-driven diagnostic approach for detecting wheel flats using axle box acceleration signals—a solution suitable for deployment on on-board edge computing devices with low wireless data-transmission requirements. A multi-channel vibration signal feature fusion framework is introduced, which consistently outperforms models based on single-channel data. Experimental validation is carried out on a full-scale high-speed test rig in laboratory, operating at speeds up to 300 km/h, with up to 12000 km of mileage data collected. Results show that triaxial sensor signals capture complementary and distinctive fault characteristics. The proposed method exclusively utilizes acceleration data as input and does not require an encoder or any additional instrumentation for speed measurement. The results also indicate that vertical-direction signals are not necessarily the best data source for data-driven wheel flats diagnosis or monitoring tasks. The proposed diagnostic method achieved a classification accuracy of 98 %, demonstrating the effectiveness of both the measurement scheme and the diagnostic framework.