Intelligent risk-oriented predictive diagnostics system based on vibration signal analysis
Abstract
This paper presents a risk-based decision-making methodology for predictive diagnostics of rolling bearings based on vibration signal analysis. A formalization of classification threshold selection is proposed through Bayesian risk minimization, taking into account the asymmetry of error costs. A hybrid architecture is developed that combines convolutional neural networks for spectrogram analysis and tabular machine learning methods for statistical diagnostic features. An analytical optimality condition in ROC space is obtained, and the robustness of the risk functional to noise and to bounded distributional shift is analysed together with the sensitivity of the method to probability-calibration errors. Experimental studies on the Case Western Reserve University (CWRU) bearing benchmark confirm a reduction in the integral risk of decision making compared to baseline models.