AI and Machine Learning for Control Applications

Abstract

The rapid advancement of artificial intelligence (AI) and machine learning technologies has fundamentally changed the traditional paradigm of control engineering. The focus of this special issue was to inspire people to discuss how AI and machine learning techniques can be used to enhance control applications in a wide range of fields, such as industrial process monitoring and fault diagnosis, optimal process design and control, deep generative model-based target recognition, and so forth. The varieties of methodologies and application studies within this special issue fully revealed the potential and necessity to further promote control-oriented AI and machine learning techniques. It is believed that this subject will continue to flourish and become one of the centerpieces of control research communities.

Among the papers accepted in the special issue, the first element to emerge is the development of AI and machine learning techniques for industrial process monitoring and anomaly localization [1-3]. Modern industrial processes often exhibit complicated characteristics of time-varying, multi-unit collaboration, multi-rate measurements, and significant process noises. There is an urgent need to understand and handle these characteristics. People within the study by Wu et al. [3] developed an adaptive spatiotemporal decouple graph convolution network to deal with the time-varying characteristics of large-scale process. The adaptive spatiotemporal graph is capable of incorporating prior knowledge and better reflecting the dynamic relationships among process variables. The proposed feature redundancy reduction scheme can simplify the graph structure and results in a more interpretable model. The enhanced fault detection performance revealed the potential of the adaptive graph neural network in industrial process monitoring. A further research issue is the multi-unit collaboration and multi-rate measurements in industrial processes. The work of Dong et al. [1] introduced a subsystem decomposition method and the multi-rate partial least squares, which showed promising performance in identifying process faults. In handling process noises, Jia et al. [2] introduced a slow feature-constrained decomposition autoencoder for anomaly detection isolation in industrial processes, which reduced the high-frequency noise and translated into better fault detection performance and isolation accuracy.

The second element discussed by the papers within this special issue is fault diagnosis and performance degradation prediction of rotating machinery and fuel cell stack [4-10]. Despite the numerous research progress made in fault diagnosis of rotating machinery in recent years, there is still a lack of effective solution to address issues like domain drift and unknown faults, data imbalance, strong noise, and so forth. Lin et al. [4] introduced a few-shot learning-based unknown recognition and classification method to deal with domain drift and unknown faults. The domain drift problem is handled by incorporating data scaling using Min-Max scaling, so that the drift in the vibration data can be dealt with without changing the source data distribution. Other issues like irregular sampling intervals are also considered. The work by Lu [5] focuses on the imbalanced data problem, which involves the multi-scale convolution neural networks and transformer. Wei et al. [6] developed a graph convolution network-based framework to deal with strongly noisy environments. The work of Zhang et al. [7-9] develops a belief-rule-based (BRB) technique for machinery fault diagnosis. The BRB method involves a two-stage feature extraction procedure using complex network and principal component analysis, which improved the separability of fault features. Another important research issue for machinery products is degradation prediction. Zhou et al. [10] developed a remaining useful life predictive method based on the adaptive continuous deep belief networks and improved kernel extreme learning machine. The work of Zhou et al. [10] involves two-stage prediction procedures, with feature extraction using deep belief networks being the first stage and prediction using kernel extreme learning machine being the second stage. On the other hand, the work of Zhang et al. [7-9] focuses on the multi-step performance degradation prediction problem of the proton-exchange membrane fuel cell stack. By incorporating the 1D convolution layer and the interactive learning mechanism of CatBoost, multi-step prediction can be achieved.

The third element of the special issue involves the incorporation of AI and machine learning methods with control problems [7-9, 11-13], covering control problems like robot control, iterative learning control, and disturbance compensation control. The work by Zhang et al. [7-9] introduces a conditional adversarial motion priors method based on reinforcement learning for humanoid robot control, which can be used to control straight-legged walking. The work by Aarnoudse and Oomen [11] proposed a data-driven MIMO iterative learning control method, which uses random learning in the form of unbiased gradient estimates. The convergence speed of the random learning-based method is further verified in an industrial printing process. Finally, the disturbance compensation control problem for discrete-time systems using reinforcement learning is discussed in Li et al. [12, 13], which used a new off-policy Q-learning algorithm to update the state feedback controller and compensator parameters.

The fourth element in this special issue covers the problems of system identification, neural operator approximation of partial differential equations (PDE) and pump scheduling [12-15]. Parameter identification of the Hammerstein system is an important problem in system identification. The work by Li et al. [12, 13] applies the neural fuzzy model and ARMAX model to decouple the Hammerstein system and uses combined signals to identify parameters in the system. In Lv et al. [14], a neural operator learning method is applied to accelerate the control design of cascaded parabolic PDEs, with the nonlinear operators approximated by the deep neural network of DeepONet. In Shao et al. [15], a deep reinforcement learning scheme is designed for pump scheduling of large-scale multiproduct pipelines, which is solved using the enhanced proximal policy optimization algorithm.

It should be noted that this special issue only covers a small fraction of the potential applications of artificial and machine learning in control engineering. We firmly believe that more and more promising control applications of AI and machine learning will be made in the future.

The authors declare no conflicts of interest.