A novel lightweight classification model based on dynamic separation representation and adaptive mapping decision for mechanical bearing fault diagnosis under strong noise interference

Abstract

Mechanical bearing fault diagnosis is crucial for ensuring the reliability and safety of industrial equipment. However, in the presence of strong noise interference, existing methods face challenges such as high resource consumption, limited generalization across different operating modes of mechanical bearings (time-varying or constant speed), and insufficient capabilities in feature learning and decision-making. To address these issues, this paper proposes a lightweight classification model based on dynamic separation representation and adaptive mapping decision for mechanical bearing fault diagnosis under strong noise interference (LCMFMB). The core components include dynamic asymmetric calibration convolution (DACConv) and adaptive decision classification block (ADCB). First, DACConv effectively identifies and extracts the intrinsic feature information most relevant to bearing health from noisy signals under different bearing operating modes, improving the precision and diversity of feature learning. Then, the feature extraction layer, centered on DACConv, further refines fault features with high discriminability while reducing computational overhead and progressively represents fault information in a stacked manner to form high-dimensional features. Finally, during the classification decision stage, ADCB’s adaptive coordination ability reveals the complex underlying relationships between high-dimensional features and fault types, achieving high-accuracy fault diagnosis. We conducted comparative experiments on publicly available bearing fault diagnosis datasets under variable and constant-speed conditions (compared with seven recent methods), and the competitive experimental results fully validate the effectiveness, generalization, and lightweight nature of LCMFMB under noise interference. Ablation experiments also demonstrate the effectiveness of its core components.