Automatic anomaly detection in engineering diagrams using machine learning

This study implements a method of automating anomaly detection in engineering diagrams by extracting patterns within graphs after recognizing graphs from a piping and instrumentation diagram (P&ID). The framework consists of three parts: graph generation, subgraph extraction, and graph classification. Graphs are generated through symbol recognition and line recognition, and subgraphs are extracted using the frequent subgraph mining algorithm. The graph classification targets are divided into two categories according to the frequency of the main equipment of the extracted subgraph. If the frequency is low, it is classified through whether to include a user-defined subgraph, and if it is high, it is trained in a support vector machine (SVM) algorithm after vector embedding to generate a classification model. K-fold cross-validation is also applied to increase classification accuracy. The proposed framework shows 85% accuracy for a given test drawing through cross-validation. These outcomes contribute to the field of engineering diagram analysis and have potential applications in plant industries.