Time series forecasting of train axle fatigue crack acoustic emission signals by integrating multi-head attention mechanism into DLinear model
利明 若林, Xiaonan Shang
Abstract
Among the many factors affecting train safety, the health of train axles is particularly critical. As an important part of the train, cracks in the axle can lead to serious safety accidents if not detected and treated in a timely manner. Acoustic emission technology can detect cracks at an early stage and, as an online real-time detection method, is essential to ensure the reliability of axles. This technology not only improves the timeliness of fault detection but also provides strong support for train maintenance and management. However, obtaining complete and continuous crack extension data is challenging due to environmental and equipment limitations. Therefore, real-time prediction of crack development during train operation has become particularly important. The real-time prediction of the sequence of acoustic emission signals enables the early detection of potential faults, thus effectively preventing the occurrence of major accidents. Therefore, we propose an improved model based on DLinear, designed for real-time prediction of acoustic emission signal time series. This model innovatively incorporates a multi-head attention mechanism into both the trend and seasonal branches. This unique architectural design enables the trend branch to more accurately capture nonlinear variation features while significantly enhancing the seasonal branch’s ability to analyze high-frequency fluctuating signals. Experimental results demonstrate that our proposed algorithm can effectively predict the time series of acoustic emission signals from fatigue cracks in axles.