Knowledge-integrated transfer learning with multi-scale CNN and graph attention for multi-condition anomaly detection in process industry

Abstract In process industries, production conditions are influenced by a multitude of factors, including manual operations, variations in raw material composition, and component degradation, which result in diverse working states, termed multi-conditions. Consequently, detecting anomalies across these multi-conditions presents a significant challenge in industrial process monitoring. However, the inherent complexity and nonlinearity of production processes, characterized by interdependent variables and data biases, pose limitations to the application of purely data-driven models. Furthermore, the process industry has domain knowledge, and the integration of data and knowledge during the modeling phase could be considerable. This paper proposes a knowledge-integrated transfer learning (TL) framework for multi-condition anomaly detection. Our method combines domain expertise with temporal patterns through a multi-scale convolutional neural network architecture augmented by graph attention layers, enhancing temporal feature extraction. An industrial knowledge TL framework is established based on the pre-trained model of a one-condition experiment through implicit knowledge transfer with model fine-tuning. TL experiments on four working condition monitoring datasets of a vertical roller mill, the proposed method achieves optimal performance (with an average F1 = 0.9226, and AUC = 0.9353 on all transferred dataset), its F1 average increases of 25.14%, 25.09%, 26.41%, 4.17%, 2.94%, and 1.58%, while corresponding curve area under the curve average increases of 21.88%, 22.04%, 23.12%, 3.47%, 2.74% and 1.41% compared with other methods, respectively.