AI Anomaly Detection for Smart Manufacturing

Abstract

AI-based anomaly detection represents a significant research field with myriad application possibilities in smart manufacturing that involves using machine learning (ML) algorithms to identify abnormal patterns in the data collected from sensors and other sources. While substantial progress has been made in this field in recent years, some gaps in the literature still need to be addressed. One key challenge is the lack of labeled data for training ML algorithms, especially for rare anomalies. Moreover, unsupervised anomaly detection is also an obstacle to smart manufacturing implementation, as it can be prone to false positives or missing some anomalies. Additionally, the imbalanced dataset, where the collected data samples are predominantly from one specific class, results in no specific information about the anomalies that must be detected. Addressing these gaps, the present research provides a comparative analysis of three unsupervised ML techniques, namely: One-class Support Vector Machine – OCSVM, Isolation Forest – IF, and Local and Outlier Factor – LOF for product anomaly detection. The ML models were developed using an imbalanced dataset collected in the manufacturing company from the process industry. Finally, the results show the OCSVM technique provides the best performance regarding accuracy, precision, recall, and F1 score.