Agentic Retrieval-Augmented Generation for Industrial Anomaly Detection

Ensuring operational integrity is critical in industrial environments. However, traditional anomaly detection models often require large amounts of labeled data and involve extensive preprocessing and training. This study presents a novel approach that leverages vector-based embeddings stored in a dedicated vector database, combined with an Agentic retrieval-augmented generation (RAG), to detect anomalies in real-time pipeline sensor data. Using an open-source dataset from Kaggle, the process begins by converting sensor readings into textual descriptions and embedding them into high-dimensional vectors. These embeddings are indexed in a vector database, which enables rapid similarity searches to retrieve the most relevant historical cases. The Agentic RAG, powered by a large language model (LLM), integrates these retrieved cases to make context-aware decisions about the presence of anomalies—for example, by identifying a significant drop in RMS volumetric flow rate as a strong indicator of abnormal behavior. Experimental results demonstrate that this Agentic RAG approach achieves highly accurate anomaly predictions on unseen data—reaching 100\% accuracy in the tests—while significantly reducing the training effort and time compared to traditional supervised models such as XGBoost or SVM. This method minimizes dependency on large labeled datasets and offers a scalable, adaptable, and interpretable solution for real-time industrial monitoring.