Intelligent Informative Frequency Band Searching Assisted by a Dynamic Bandit Tree Method for Machine Fault Diagnosis

The fault informative frequency band searching is crucial to envelope analysis-based machine fault diagnosis. Its success often depends on effective filters. However, existing filters encounter three problems: 1) fixed filters are not adaptive; 2) the adaptive decomposition filters are affected by key parameters; and 3) popular swarm-intelligent filters lack a clear guidance of parameter settings. This article innovatively introduces a bandit optimization algorithm, the dynamic bandit tree (DBT), to help realize more adaptive filters with the lower parameter-tuning burden in frequency band searching. Particularly, we show that boundaries of Meyer wavelet filters can be optimized by the DBT with effortless parameter tunings. The DBT is constructed by refining its growth dynamically based on the proposed multitree space partition and reshaped Thompson sampling. Consequently, the filter boundaries are determined by the optimal trial of the DBT, enabling better identifications of demodulated fault frequencies. In verifications, we first benchmark the DBT against ten optimization algorithms via two multidimensional test functions. We then compare the proposed diagnosis method with seven existing fault diagnosis methods using bearing and gearbox fault data. Our methods can excel the benchmarks qualitatively and quantitatively. Additionally, a Python repository is provided to facilitate future studies.