Large margin-learning methodology from time-frequency maps and its physically interpretable weights for simultaneous machine health monitoring and fault diagnosis

Abstract

Artificial intelligence (AI) enables machine health monitoring including fault detection and diagnosis that are becoming prevalent and have been actively investigated during recent years. Nonetheless, poor interpretation and black-box characteristics hinder their performance in machine condition monitoring. Specifically, abstract features extracted from AI models are difficult to physically interpret and correlate with fault mechanisms. In this paper, to enrich interpretation of learned features, a large margin-learning methodology from time-frequency maps is newly proposed. The original idea is to construct a health indicator (HI) for degradation process tracking in a feature extraction module and it is defined as the sum of element-wise products between a time–frequency map and a weight matrix. Here, the weight matrix can indicate fundamental components that always exist in healthy and faulty signals and newly generated components that are frequency components caused by machine faults. Moreover, the significance of the two-dimensional weight matrix can clearly show how weights are varied with time and indicate a cyclo-stationarity behavior of repetitive transients existing in fault vibration signals, which results in a new perspective for revisiting time–frequency maps for intelligent large margin-learning. Following, thanks to the physically interpretable feature matrix, a single-hidden layer network in a decision-making module is seamlessly designed for immediate machine fault detection, in which weights between an input layer and a hidden layer are tactfully set to the physically interpretable feature. At last, since the proposed methodology is based on convex optimization, it only needs a small amount of samples for training and realizing the physically interpretable weight matrix to understand a significant difference between different time–frequency maps for machine health monitoring and diagnosis. Experiments verify the effectiveness and superiority of the proposed large margin-learning methodology for simultaneous machine health monitoring and fault diagnosis.