Robust-GSR driven lightweight CNN framework for intelligent fault diagnosis under variable-speed conditions

Rolling bearings are critical to the reliability and safety of rotating machinery. Under variable-speed conditions, vibration signals exhibit strong non-stationarity and weak fault features, posing significant challenges for accurate and efficient fault diagnosis. To address these issues, this study proposes a fast and intelligent diagnostic framework, which integrates a robust generalized stochastic resonance (robust-GSR) system with a lightweight convolutional neural network (CNN), referred to as RGSRNN. Specifically, the robust-GSR system is designed based on first- and second-order stochastic asymptotic stability analysis in the angular domain and incorporates fault feature amplification (FFA)-based multi-parameter optimization strategy, enabling consistent feature enhancement. To improve feature expressiveness, the enhanced signals are converted into Markov transition field (MTF) images in the fault order spectrum, facilitating structured and discriminative two-dimensional (2D) representation. The lightweight CNN, composed of only two convolutional and two fully connected layers, performs fast and accurate classification with significantly reduced computational cost. Experimental results under complex variable-speed scenarios also demonstrate that RGSRNN outperforms existing methods in both accuracy and stability, making it a practical and efficient solution for intelligent fault diagnosis in real-world industrial environments.