Although the data collected by modern industrial processes is increasing, it is quite difficult to obtain sufficient real fault labels. Therefore, semi-supervised fault diagnosis has become one of the problems of great practical significance in data-driven industrial process monitoring. To handle the semi-supervised fault diagnosis issue, this work proposes a novel general trainable pseudo-label generator (TPG) on the basis of deep reinforcement learning. In TPG, the generator of TPG is trainable, and the parameters of the generator are optimized through the feedback performance of the validation test. The iterative process of the pseudo-label generator can be formalized as a sequential decision process. The non-differentiable optimization issue of generating pseudo-labels based on feedback can be solved by reinforcement learning. As a data-level algorithm, the pseudo-label generator can be simply used in conjunction with the existing supervised diagnosis model. The practicability and superiority of TPG are demonstrated through one synthetic dataset and two industrial processes.

An imbalanced number of faulty and normal samples causes serious damage to the performance of the conventional diagnosis methods. To settle the data-imbalance fault diagnosis problem, this article presents a novel general imbalanced sample selection strategy (DiagSelect) based on deep reinforcement learning. In DiagSelect, the problem of imbalanced sample selection from the training set is formulated as a multiarmed bandit problem of deep reinforcement learning. The nondifferentiable optimization problem of imbalanced sample selection can be solved by the Markov decision process. The parameters of DiagSelect can be optimized by REINFORCE with the feedback of the validation set. DiagSelect performs intelligent imbalanced sample selection to obtain better diagnosis performance autonomously. As a data-level technique, DiagSelect can be easily used in conjunction with the conventional diagnosis models. DiagSelect is validated in a synthetic dataset and an industrial process dataset. The results have shown the effectiveness, stability, and transferability of DiagSelect.