Mel frequency mapping for intelligent diagnosis of rolling element bearings across different working conditions

Abstract

The diagnosis of rolling bearing faults under constant operational conditions has garnered substantial scholarly interest, leading to the development of numerous robust and reliable diagnostic techniques. However, the dynamic and variable nature of industrial operating conditions complicates the acquisition of diagnostic data and renders fault prediction problematic, thus posing a formidable obstacle for fault diagnosis in unknown operational settings. Drawing an analogy from human communication, where listeners can comprehend speech irrespective of its speed or intensity, this paper aims to devise a methodology capable of distinguishing between various health states of machinery under fluctuating working conditions. To this end, we present a novel method, Mel Frequency Mapping Classification (MFMC), for nonlinear fault characteristic mapping and classification. This data-driven model, trained on datasets from known operational conditions, adapts well to new settings. Initially, vibration signals are converted from linear Hertz frequencies to nonlinear Mel frequencies through the Mel frequency mapping technique. Subsequently, the extracted fault signal features are transformed into Mel spectrograms, which serve as the input for the data-driven model. Finally, these Mel spectrograms are utilized as input variables for machine learning algorithms, specifically support vector machine (SVM) and multilayer perceptron (MLP), to accomplish automated, intelligent bearing diagnosis. To rigorously assess the efficacy of the proposed methodology in mechanical condition monitoring and intelligent diagnosis across diverse operational conditions, it was tested on two distinct datasets: CWRU and PU. These datasets encompass bearings with varying degrees and types of damage, situated in different components of the bearing, and operating under disparate conditions. Experimental results indicate that the proposed MFMC approach can obtain high cross-condition diagnosis accuracy on the training data under a single condition. This methodology offers a promising avenue for the intelligent diagnosis of mechanical equipment under cross-condition scenarios.