Enhancing Φ-OTDR vibration event classification by stacking MFCCs and differential phase features using CNNs

Abstract

In the realm of vibration event classification, using the Phi-Optical Time-Domain Reflectometer (Φ-OTDR) and deep learning techniques like Convolutional Neural Networks (CNNs) requires a substantial amount of training data, which can be expensive to collect and annotate. Yet, maximizing the utility of data features from a limited set of samples could enhance training efficacy and classification precision. This study introduces an innovative approach that utilizes a combination of Mel-Frequency Cepstral Coefficients (MFCC) and Differential Phase (DP) features, referred to as MFCC-DP. The MFCCs are extracted from the Rayleigh Backscattered (RBS) signal intensities, while DP features are extracted from the analytic signals of RBS. The MFCC-DP features are used to train a CNN model for event classification. Experimental findings demonstrate a noteworthy enhancement in accuracy, reaching 98.2% with MFCC-DP compared to 92.1% and 94% when using DP and MFCCs, respectively. Furthermore, the results indicate that the use of MFCC-DP reduces the number of events that are difficult to classify due to overlapping features.