Publication
Combining informed data-driven anomaly detection with knowledge graphs for root cause analysis in predictive maintenance
Patrick Klein; Lukas Malburg; Ralph Bergmann
In: Engineering Applications of Artificial Intelligence, Vol. 145, Page 110152, Elsevier, 2025.
Abstract
Industry 4.0 has facilitated the access to sensor and actuator data from manufacturing systems, leading to studies on data-driven anomaly detection, but limited attention has been paid to finding root causes and automating this process using formalized expert knowledge. This is crucial due to the scarcity of qualified engineers and the time-consuming nature of diagnosing issues in large production systems. To address this gap, we present a framework that combines data-driven anomaly detection with a knowledge graph that provides domain knowledge by leveraging typical explanations of such models (i.e., data streams potentially caused the detection) for further diagnosis. The framework's usefulness to infer affected components or data set labels has been evaluated using two deep anomaly detection approaches. For knowledge-based diagnosis, three query strategies that utilize various knowledge graph relationships are implemented through three Artificial Intelligence (AI) techniques. The proposed anomaly detection approach, informed by integrating expert knowledge via the graph structure of the knowledge graph and node embeddings for encoding time series, outperforms baselines and a deep autoencoder in detecting anomalies and in identifying anomalous data streams. In subsequent diagnosis, it achieves the best performance on a complete knowledge graph in combination with a graph pattern matching query by identifying the label or affected component in 60% of detected anomalies by providing 4.1 labels or 2.3 components until the correct one is identified. In case of a corrupted one, Symbolic-Driven Neural Reasoning (SDNR) and Case-Based Reasoning (CBR) with knowledge graph embeddings demonstrate advantages by halving the number of incorrect labels and unaffected components.