Concatenation contrastive adversarial learning for fault detection in chemical processes

While fault-relevant detection approaches achieve high sensitivity by learning fault-correlated features, they perform poorly when applied to new operating modes where only normal data are available, which is common in early deployment scenarios. This limitation makes it difficult to identify faults in a timely manner and ensure safe operation in chemical processes. To address this challenge, this work presents a domain adaptation (DA) strategy, where the source domain contains both fault and normal data, while the target domain contains only normal data. The aim is to leverage prior fault knowledge from historical modes to construct a reliable detection model for new modes. However, traditional DA methods often suffer from performance degradation due to the scarcity of fault data and the presence of previously unseen faults. To this end, a novel concatenation contrastive adversarial learning (CCAL) algorithm is proposed for fault detection. Specifically, a feature concatenation strategy is developed to generate feature pairs, which are used to train a contrastive adversarial adaptation network for robust fault modeling. Additionally, a concatenation reconstruction score is designed as the monitoring statistic to enhance the detection of unknown faults. Experiments conducted on the continuous stirred tank reactor, industrial three-phase flow process and Tennessee Eastman benchmarks demonstrate the superior performance of CCAL in both known and unknown fault scenarios.