Adaptive multi-scale Laplace wavelet weighted fusion framework for heavy haul freight train fault diagnosis under limited sample with high-noise

As a crucial component, the condition monitoring of wheel-axle bearings of heavy haul freight trains is of significance. However, due to the influence of the number of real fault samples and complex operating conditions, the accuracy of the current algorithm still needs to be improved. To tackle the problem, a framework for diagnosing the bearing faults of heavy haul freight trains is proposed, based on an adaptive multi-scale Laplace wavelet module (MLWCL) and a multi-channel weighted fusion strategy. Firstly, MLWCL is devised to address the issue of indistinct features of real signals with noise. The input signal is divided into groups of different sizes, which are analyzed from different MLWCL analysis scales. Meanwhile, in order to solve the problem of the exponential component in Laplace wavelets affecting the initial value distribution of the model, a suppression function is proposed. The suppression function avoids edge perception decline by adaptively adjusting the extremum points to match the Laplacian median points. Finally, a multi-channel weighted fusion layer enhances impact signal perception by fusing global trend features and local detail features. Experiments on the dataset collected from real faulty heavy train bearings demonstrate that the model possesses excellent generalisation and anti-noise robustness.