Novel computing method for uncertainty-aware fault diagnosis of rotating machinery based on image processing and interval theory

Traditional image processing-based fault diagnosis methods for rotating machinery often exhibit limited accuracy in the presence of uncertainty factors such as sensor errors. To address this limitation, this study proposes a novel interval theory-based fault diagnosis method integrating a weighting distribution strategy. Based on high-dimensional images transformed from vibration signals, a time-frequency imaging-based feature extraction method integrating histogram of oriented gradients (HOG) and principal component analysis (PCA) is developed to extract representative features. To characterize the dispersion in fault features induced by uncertainty factors, an interval-formed feature is further established and a Euclidean distance-based pattern matching strategy is proposed to quantitatively measure the severity of unknown modes. Considering the indeterminate classification outcomes caused by inaccurate features, an interval Fisher feature weight distribution (IFFWD) method is developed to prioritize discriminative features by optimizing between-class scatter against within-class scatter. Three case studies have validated the competitive diagnostic performance and robustness of the proposed methods in dispersion-intensive environments compared to conventional methods.