LSTA-Net framework: Pioneering intelligent diagnostics for insulating bearings under real-world complex operational conditions and its interpretability.

Deep Learning has been attracting considerable attention as it can autonomously learn important signal features and has shown great potential for fault diagnosis. However, given the variability of high-power variable-frequency industrial systems, especially under unfavorable service conditions such as voltage and load fluctuations, the identification of insulated bearing faults in variable-frequency motors has become a key challenge to be addressed. Moreover, the time-varying nature of insulated bearing data poses common technical difficulties for existing models. To bridge these gaps, this research spearheaded the development of a lightweight spatial–temporal focusing framework equipped with a novel optimizer called LSTA-Net, which was designed to address the challenging problem of insulated bearing fault identification in real-world engineering scenarios. Specifically, this involves an innovative design of the Weight Reducing Recursive Unit (WDRU) strategy and an updated backpropagation formulation, which was skillfully applied to the feature extraction module of the LSTA-Net framework for the first time. Furthermore, a novel optimizer called Stochastic Newton Descent (SND) has been developed and skillfully integrated into the LSTA-Net framework with the newly derived weight update formula, providing unique insights to improve diagnostic performance. Finally, the diagnostic performance of the LSTA-Net framework was evaluated from multiple dimensions based on the same dataset, proving its superiority, generalizability, and robustness. t-SNE technique integration intuitively explains the framework's fault characterization mining process, improving its credibility, reliability, and accuracy. [ABSTRACT FROM AUTHOR]