VRFST: Voiceprint-Based Robust Wear Detection for Diamond Abrasive Tools via Fuzzy-Enhanced Spatiotemporal Features

Accurate and robust wear detection of diamond abrasive tools is crucial in Industry 5.0, yet challenging due to inherent fuzziness and uncertainties in multi-condition, high-noise industrial environments. This paper presents a novel voiceprint-based robust method via fuzzy-enhanced spatiotemporal features (VRFST) to address this challenge by exploiting the power of fuzzy neural networks in handling uncertainty. First, to theoretically guide model design, the acoustic-based wear detection problem of diamond abrasive tools is formally described as a multi-objective optimization task. Second, to obtain robust representations, acoustic signals are preprocessed, including time series encoding and frame-level feature aggregation. Critically, the core innovation lies in integrating three synergistic feature extractors: a temporal feature extractor using stacked bidirectional long short-term memory (SBiLSTM) and multi-head attention, a spatial feature extractor with convolutional block and spatial pyramid pooling (SPP) block, and a fuzzy feature extractor utilizing overlapping Gaussian membership functions to suppress multi-source uncertainties adaptively. Third, voiceprint representations are derived by multi-scale convolutional fusion of fuzzy-spatial-temporal features and classified using a theory-guided loss function to enhance accuracy and generalization. Experiments show that incorporating the fuzzy feature extractor significantly enhances robustness, with the extracted temporal, spatial, and fuzzy features offering inherently complementary representations. VRFST achieves exceptional accuracy, ultra-low inference time, and strong generalization under diverse operating conditions and high-noise scenarios across three datasets, outperforming state-of-the-art methods.