Multi-Stage Auditory Anomaly Detection and Classification for Natural Gas Compressors

Monitoring natural gas compressors is essential because compression systems connect upstream gas production with downstream usage by maintaining the pressure needed for smooth gas flow through pipelines. This study introduces a multi-stage audio-based anomaly detection and classification system for natural gas compressors. The system aims to identify anomalous audio inputs and categorize them into four classes: normal, benign anomalies (e.g., normal audio with human voices), known anomalies, and unknown anomalies. In the first stage, a variational autoencoder (VAE) is employed to separate anomalies using Mel Frequency Cepstral Coefficients (MFCC) from normal operational audio. In the second stage, the reconstruction error and latent features from the VAE are used to distinguish known anomaly classes from unknown anomalies via Out-of-distribution (OOD) analysis. In the third stage, known anomalies are classified by a neural network. The proposed approach was evaluated using an industrial compressor dataset, and the experimental results demonstrated its effectiveness in accurate anomaly detection and classification.