Bolt looseness fault diagnosis method based on multi domain feature fusion

Abstract Aiming at the problem of bolt loosening detection in multi-bolt connection structures under complex operating conditions, a data-fusion bolt loosening detection model is proposed based on multi-domain feature extraction and attention mechanisms. The method first applies fast Fourier transform (FFT) and a convolutional network to perform time–frequency joint denoising on the raw vibration signals, followed by inverse transformation to reconstruct the denoised time-domain signal. The denoised signal is then fed in parallel into three branches—time domain, frequency domain, and wavelet domain—for multi-domain feature extraction. By integrating domain features via a multi-domain semantic fusion attention mechanism, complementary cross-domain fusion features are obtained. To further capture the long-term dependencies and dynamic patterns of bolt vibration signals in temporal sequences, this study introduces a bidirectional long short-term memory network for time-series modeling of the fused features. Finally, a multilayer perceptron (MLP)-based detection head is employed to detect the bolt connection states. A bolt loosening test rig is built to simulate 16 distinct bolt loosening conditions. Ablation studies using the experimental data demonstrate the effectiveness and necessity of each module for vibration signal identification. On the self-constructed dataset, the proposed model achieves an average accuracy of 95.66% and an average F1-score of 95.65, outperforming traditional approaches by +3.91 percentage points in accuracy and +5.02 in F1-score.