A DAS Monitoring System for Underground Cables Using Multi-Scale Dynamic Mel Frequency Cepstral Features

In this article, multi-scale dynamic Mel frequency cepstral features are proposed to improve the recognition accuracy of the distributed acoustic sensing system. The vibration signal is segmented twice using different frame lengths. In one segmentation, improved Mel frequency cepstral coefficients are extracted, considering the characteristics of the frequency distribution. In the other segmentation, linear predictive cepstral coefficients are extracted. Dynamic features are then derived by the linear combination of static features from adjacent frames. All these features are weighted and fused based on mutual information, which measures the dependency between features and labels, to form the final features. These final features are used as the input to a neural network to identify the target events. Experiments are conducted by applying various vibrations to the cable buried in 3 different environments. The results indicate that the proposed features achieve a recognition accuracy of 99.38% for 8 types of vibration events. Moreover, these features maintain high recognition accuracy and performance stability, even with variations in input data, parameter changes, and noise interference.