Abnormal data detection for industrial processes using adversarial autoencoders support vector data description

Abstract

Abnormal data detection for industrial processes is essential in industrial process monitoring and is an important technology to ensure production safety. However, for most industrial processes, it is a challenge to establish an effective abnormal data detection model due to the following issues: (a) weak model performance due to the small amount of process data; (b) trade-offs between model sparsity and accuracy; and (c) weak generalization ability of abnormal data detection model. To address these issues, a method based on adversarial autoencoders support vector data description (AAESVDD) is presented in this work. First, a novel construction strategy is designed for a hybrid feature dataset based on the adversarial autoencoder (AAE). The hybrid feature dataset utilizes the latent feature and reconstruction residual extracted by the AAE to enhance the feature diversity of the process data. Then, combining the support vector data description (SVDD) and Bayesian optimization algorithm (BOA), an automatic detection model for abnormal data of the hybrid feature dataset is established. Meanwhile, a BOA objective function based on the criterion of the hybrid risk minimization is proposed to automatically optimize the model parameters, which further enhances the generalization ability of the SVDD-based model. Finally, the effectiveness of the proposed AAESVDD method is illustrated with the UCI benchmark datasets and an industrial penicillin fermentation process.