A Unified Framework for Replicability of Anomaly Detection in Multiple Time Series: Enhancing GANF and RanSynCoders Pipelines with MFCC Feature Extraction for Acoustic Data

Anomaly detection, a ubiquitous pattern recognition methodology for identifying atypical data, has been applied across a myriad of domains. In the industry, it has been used to identify malfunctions in industrial machinery in advance and suggest predictive maintenance based on multiple time series of sensor data, including temperature and pressure. With the availability of high-quality acoustic data, our understanding of acoustic scenes involving industrial machines has enhanced. However, the state-of-the-art approaches in anomaly detection, involve increasingly complex data processing pipelines, making replication with acoustic data not fully understood. Our proposal are three-fold: (i) we developed MTSA (Multiple Time Series Analysis), a unified framework for replicability of state-of-the-art anomaly detection approaches; (ii) we implemented Hitachi, RANSynCoders, and GANF on MTSA—three state-of-the-art anomaly detection approaches— improving RANSyn-Coders and GANF pipelines with a feature extraction based on Mel-Frequency Cepstral Coefficients (MFCC); and (iii) we conducted a comparative analysis of these approaches on MTSA for the anomaly detection with acoustic data from valves, pumps, fans, and slide rails using the open dataset for malfunctioning industrial machine investigation and inspection (MIMII). Overall, our results, measured by AUC-ROC, suggest RANSynCoders and GANF can be improved to handle acoustic data effectively, since these models achieved solid results such as 0.94 (95% CI, 0.92-0.97) and 0.79 (95%, 0.75-0.82) respectively. As a unified framework, MTSA can decompose complex pipelines into simple data processing building blocks, facilitating the replication of existing approaches and the development of novel models.