Wasserstein Distance Based Domain Adversarial Autoencoder for Industrial Few-sample Fault Detection

In the era of industry 4.0, it is particularly essential to ensure the safe and stable operation of industrial processes. Data-driven fault detection methods are favored for their exceptional model fitting capabilities. However, few-sample scenarios are prevalent in real industrial processes, particularly during the initial operation phase of a new mode and in the situations where data collection is expensive. To address this challenge, we propose a model termed Wasserstein distance-based domain adversarial autoencoder (WDAA). Based on the domain adversarial framework, WDAA integrates a Wasserstein adaptor to optimize domain alignment between the source and target domain. By employing Wasserstein distance as a metric, WDAA facilitates more efficient alignment, thereby fostering the synthesis of a shared feature space conductive to effective knowledge transfer between disparate domains. The efficacy of our approach is demonstrated on an industrial model, the continuous stirred tank reactor (CSTR). Experiments show the proposed method has promising fault detection results than other advanced methods.