What are the pros and cons of machine learning in control systems design?

During my time spent babysitting critical systems, a lot of mental energy was spent wondering where the sweet spot was for setpoints for given conditions. Even if I found a more efficient mode of operation, I had to consider if it was reliable as conditions changed, and eventually my shift was over and I had to either pass on a change or set it back. A lot of missed efficiency happened whereas an ML system can be constantly hunting for those gains and operators can focus more on alarm triage and things that require eyes and ears.

**Pros:** - **Handling complexity:** ML excels at handling complex, nonlinear systems that are challenging for traditional methods. - **Adaptability:** models can adapt to changing environments, - **Data-driven insights:** can uncover hidden patterns and relationships in data, leading to better control strategies. **Cons:** - **Explainability:** ML models can be opaque, making it difficult to understand their decision-making processes. - **Data dependency: ** relies on high-quality, representative data for training and performance. - **Robustness and safety:** ML models are not inherently robust or safe, requiring careful validation and testing. - **Computational resources:** especially for real-time control applications.

A drawback of ML is often that it isn't trusted by operators, who many times are opposed to any sort of change. Some ML optimizations are wholistic and if one set of actions or settings is ignored all of the gains may be lost. ML systems are also only as good as the data you feed them; and control systems are rife with bad data (chiller status of 65535 for example.) A lot of control systems are not given the time to organize a good standard when it comes to naming because it's hard to quantify the cost, so in addition to data quality issues there are data organization issues which turn into an headwind when you are trying to integrate with a system.

EXAMPLE BELLOW?? Machine Learning could be an intimidating subject for many but the best way to be introduced to it is with a couple of real world examples: ?? Amazon Monitron: this device is used on electrical motor to detect abnormal condition to enable predictive maintenance, There are others brands that do the same for example SEW, I will have the opportunity to be introduced to this product soon. ?? KEYENCE MP-F series: the ultimate pressure regulator that can be set to learn optimal conditions of the pneumatic network. In case there any deviation you can analyse the data or let the device alert you, also it has Leak detection, and energy saving features.

I'm not a fan of control systems adjusting parameters on their own. There are issues of liability and safety and such if the ML "learns" incorrectly. I also have issues with the transparency of the learning process. That said, I believe ML is a very valid method for monitoring a process and generating alerts or recommendations. I think you go as far as to say digital twins is an outgrowth of ML processes.