A percussion method with attention-based sequence modeling for bolt looseness detection and adaptive anomaly recognition

Ensuring the reliability of bolted connections is crucial for the safety of structures in engineering applications. Traditional detection methods rely on expensive equipment and complex sensors, limiting practicality. This study suggests an innovative percussion-based Deep Learning (DL) method for bolt looseness detection and adaptive anomaly recognition. In this method, Mel Frequency Cepstral Coefficients (MFCCs) sequences are extracted from audio signals, which are then processed by an improved BiGRU network to capture bidirectional dependencies and optimize information flow. The integrated attention mechanism further enhances feature extraction by dynamically focusing on key information, improving overall performance. To confirm that this approach is effective, actual steel plates were struck, and audio signals under different working conditions were collected using a smartphone for automatic recognition. The results indicate that the proposed model can accurately identify the loosening state of bolts while avoiding overfitting and demonstrating strong noise resistance. It shows significant advantages over current techniques. Furthermore, ablation and visualization analyses were conducted to validate the network design and assess its feature extraction performance. After the model's capability was validated, a new algorithm was proposed, enabling the model to detect anomalous states not present in the original training database without requiring preset thresholds or retraining. As a result, the model achieved robust generalization and adaptive anomaly detection. In conclusion, this study proposes a bolt loosening detection method that offers ease of use, cost-effectiveness, and high efficiency, demonstrating great potential for engineering applications.