Adaptive fault components extraction based on physically interpretable optimized weight time–frequency matrix index and its applications in rotating machinery fault diagnosis

The adaptive fault component extraction (AFCE) process, crucial for diagnosing rotating machinery faults based on vibration signals, relies on statistical indices. Conventional indices like kurtosis and correlated kurtosis, sensitive to random noise, may produce imprecise results. To address this, this article proposes a new index—the physically interpretable optimized weight time–frequency matrix index (PIOWTFMI). Derived through large-margin convex optimization of time–frequency matrices from healthy and faulty samples, PIOWTFMI is not only interpretable but also enhances the separation of faulty components from interference. Additionally, an implementation method of AFCE based on PIOWTFMI is proposed, compatible with existing decomposition techniques like mode decomposition and requiring minimal parameter adjustments. Finally, the effectiveness of the method is verified by bearing public data from Intelligent Maintenance Systems (IMS) and industrial real truck transmission gear data. The results demonstrate the superior performance of proposed method compared to classical approaches such as fast kurtogram, feature mode decomposition, maximum correlation kurtosis deconvolution and multipoint optimized minimum entropy deconvolution adjusted.