Self-Supervised Anomaly Detection for Industrial Machines With Sensor Distance Classification

Abstract

The rise of low-power, affordable sensing technologies and machine learning algorithms has sparked a growing interest in using data-driven approaches to monitor industrial assets. In particular, applying machine learning to analyze the sounds produced by manufacturing tools is becoming an effective method for quickly detecting deviations from standard operating conditions. However, the implementation of these systems encounters a major challenge due to the difficulty in obtaining suitable training data. Self-supervised learning offers a promising solution to this issue. It enables the training of anomaly detection models exclusively with signals representing normal conditions, which are more accessible than anomalous signals. Despite its potential, achieving robust and consistent performance across machines of varying models and types remains critical. Existing methods generally follow one of two approaches. The first, single-machine training, involves training a dedicated model using local data from each machine. Although straightforward, this approach often yields suboptimal performance. The second, multi-machine training, aims to enhance detection capability by aggregating data from multiple machines to train a shared model. This strategy requires transmitting data from geographically dispersed manufacturing sites—often belonging to different clients—to centralized facilities, raising concerns about data transmission costs and risks of exposing sensitive production information. We address these issues by introducing a novel self-supervised strategy for effective single-machine training based on classifying the distance between the monitored machine (i.e., the sound source) and each microphone sensor of a multi-channel recording system. Our method bypasses the need for data aggregation, providing a cost-efficient and privacy-preserving alternative while delivering competitive detection performance. Experiments on the MIMII dataset highlight the effectiveness of our approach, with performance improvements up to 14.53% over conventional single-machine training strategies and 18.58% over multi-machine training strategies.