Vibration Event Classification in Φ-OTDR Systems Using MFCC Features and ResNet50-CBAM

To address the challenge of event recognition in phase-sensitive optical time-domain reflectometer (Φ-OTDR) systems, this paper proposes a method that combines Mel Frequency Cepstrum Coefficient (MFCC) features with an attention-enhanced convolutional neural network. The method first extracts the MFCC feature maps from the phase signals acquired by the Φ-OTDR, and then utilizes an improved network with the Convolutional Block Attention Module (CBAM) embedded in the ResNet50 residual block for classification. Optimisation strategies such as Mixup data augmentation and Focal Loss are also used during training in order to improve performance. Experimental results on a self-constructed dataset containing six typical event types show that the proposed ResNet50-CBAM model achieves an overall recognition accuracy of 82.6%. Compared with baseline CNN models trained on the same MFCC features, the proposed method achieves significantly higher accuracy.