A deep learning approach for anomaly detection with industrial time series data: a refrigerators manufacturing case study

Abstract

In refrigerators production, vacuum creation is fundamental to guarantee the correct manufacturing of the product. Before inserting the refrigerant in the refrigerator cabinet, the vacuum is tested through a Pirani gauge that assesses the pressure within the cabinet. Such readings are used to evaluate the vacuum creation process and to verify if leakings are present. In this work, we employ a Deep Learning-based Anomaly Detection approach to associate an Anomaly Score to each pressure profile; this score can be exploited to optimize actions performed by human operators like more detailed inspections or unit exclusion from the downstream production stages. We propose a native time series-based approach based on Deep Learning and compare it with classic ones based on hand-craft features. The proposed approach is designed to be deployed in a Decision Support System for assisting human operators in the following testing operations, helping them in reducing evaluation bias and attention losses that are inevitable in production line environment. Moreover, costs associated with false positives (normally operating units detected as anomalous) and false negatives (undetected anomalies) are considered here to optimize decision making in a cost-reduction perspective. We also describe promising results obtained on real industrial data spanning on a 5-month period and consisting of thousands of tested household units.