An Industrial Framework for Explainable Anomaly Detection: A Case Study for Pick-And-Place Machines

Abstract Machine learning (ML) techniques have become popular with their ability to learn and predict complex trends in data-rich sensing environments commonly found in the manufacturing industry. These techniques have improved cost reduction strategies, fault diagnosis, prognostics and health management, and maintenance decision-making. However, as ML models become more complex in the form of sophisticated deep learning architectures—particularly with the advent of generative artificial intelligence (AI)—their predictions have become more difficult to justify to industry practitioners, inhibiting trustworthy adoption. The research of explainable AI (XAI) aims to mitigate these issues and transition away from opaque “black box” predictions. Recent years have shown strides in XAI used for obtaining predictions of a trained model locally, globally, and with model-specific and model-agnostic techniques. However, much work remains on the integration of these techniques into established ML pipelines. This paper will discuss a new framework for adopting XAI in industrial settings under two realistic constraints: a) privacy-encoded inputs and b) weakly labeled or entirely unlabeled datasets. The framework consists of five steps: 1) handling raw process data; 2) applying nonlinear data transformations to mask the original features and labels of the dataset; 3) preprocessing and initial data exploration; 4) gathering explainable data-driven insights; and 5) producing actionable recourse to recommend process improvements. In particular, we will focus on explanations for anomaly detection models operating under heavy class imbalance. After explanations are obtained, a “cluster-and-describe” method is proposed to visualize global model behavior, allowing for operators to better trace the model’s decision pathways when predicting anomalous behavior. A case study in anomaly detection of pick-and-place (PNP) machines in semiconductor manufacturing is detailed to illustrate the capabilities of the approach. The case study objective is to investigate abnormal yield losses of multilayer ceramic capacitor chips in the PNP process, which separates the chips such that they can be placed individually in a ring-shaped tape. The lack of prior domain knowledge motivates the data-driven approach, which utilizes explainable AI to gain insights on the underlying chip loss phenomenon. In this high-dimensional case, privacy-protected time series data from the PNP process are discretized into periods, enabling statistical extraction of over 900 features. We demonstrate the effectiveness of combining XGBoost for anomaly detection alongside TreeSHAP explanations for swiftly obtaining predictions with corresponding feature importance rankings. Then, after clustering the anomaly explanations using hierarchical density-based clustering, we scope the clusters with 1-2 term rules based on the extracted and measurable statistical features. This allows operators to focus attention on about 1% of the extracted features, providing them with thresholds to improve real-time condition monitoring efforts. Importantly, the clusters and rules obtained in the real-world setting are similar to those obtained from the privacy-encoded case study, showcasing the utility of the approach for knowledge discovery in novel manufacturing systems.