Adversarial training-based denoising autoencoder for robustness against industrial data noise

Background: The collection of industrial data will inevitably introduce various types of noise. Recently, deep learning has shown promising results in implementing various industrial data applications. However, existing deep learning-based model are still vulnerable to data noise, which could deteriorate the performance of the models. Methods: In this paper, we propose an adversarial training-based denoising autoencoder, called Adv-DAE, for industrial process modeling that is robust to noise interference. Due to the difficulty in obtaining datasets that cover all possible data noise characteristics, the proposed framework is developed to be zero-shot such that the model can be trained with only noise-free dataset without requiring any type of data noise, thus greatly saving the time and resources required to generate noise samples. In contrast, adversarial training is a gradient-based attack method, which can generate noisy corruption during the training processes, hence we serve the adversarial training method as the alternatives. Moreover, the adversarial examples generated by Adv-DAE are more difficult to identify than traditional noise types and can maximally corrupt clean sample data. We further provide theoretical convergence analysis for the proposed Adv-DAE to guarantee its successful practical application. Results: Case studies on benchmark image datasets, the benchmark Tennessee Eastman Process and the real-world aluminum electrolysis process are presented to verify the effectiveness of Adv-DAE.Experimental results show that our proposed Adv-DAE has more robust feature extraction ability than the compared methods. Conclusions: An adversarial training-based denoising autoencoder for robustness against industrial data noise is realized by our proposed algorithm.