Anomaly detection is crucial in identifying unexpected behaviors or faults within complex industrial systems, where early detection can prevent costly failures and downtime. With the increased data generation by today’s smart factories, developing efficient methods for detecting anomalies has become a requirement to ensure reliable operations. This study introduces an approach for detecting anomalies in time series data, leveraging the strengths of Case-Based Reasoning (CBR) and time series representation learning. By employing Transformer-based models, we transform raw time series from complex industrial systems to fixed-length embeddings. By pretraining the Transformer on healthy/normal data, the system generates reference embeddings, which are subsequently used to identify anomalies through similarity comparisons with new observations. This method does not rely on faulty data, making it particularly well-suited for industrial applications where labeled fault data is often limited. The proposed approach is validated through two case studies: a public gearbox fault dataset and a private industrial system dataset. In the public dataset, our method achieved an accuracy of 99.17%, outperforming other techniques across varying load conditions. The private case study further demonstrated the model’s ability to detect anomalies and identify early fault indicators several days before failures were reported. These results underscore the potential of our approach in predictive maintenance, offering a powerful tool for reducing unplanned downtime and optimizing industrial operations without the need for faulty samples. • Anomaly detection using Transformer representation learning and case-based reasoning. • Adapts to varying loads via load-aware matching and to component changes without retraining. • Representation model only requires healthy reference data for training. • 99.17% accuracy on public data and early fault detection on private data.
Naqvi et al. (Thu,) studied this question.