Is Machine Learning applicable for all industries? Where is the place of Machine Learning in the process industry? How safe and reliable are these algorithms to operate processes?
These answers are at the confluence of Control Systems theory, Machine Learning and the openness of the industry to new technologies.
The large majority of algorithms that are used are supervised learning: this means that these algorithms need training sets of data that are similar (have the same distribution) as the testing or normal operation data. Once trained, the Machine Learning model is used on new data in order to evaluate it.
Companies that developed their products and services integrating the use of data at their core and apply data collection backed by monetisation of this data are doing very well using the entire palette of algorithms that Machine Learning has to offer. The clearest example of this type of company is Facebook that generates revenues almost exclusively from digital advertisement. Another clue on how important it is to have the data integrated into the core of the business is the fact that all the data these companies have, it can only be monetised at their maximum value by the same company; taken out of their platform, let’s assume on a (very large) storage device, the same data does not generate the same value by itself; the value comes when the data is used in the context of the company’s platform: the products, the services, the interfaces with other third parties are the ones generating value out of data.
For companies that had a business before the age of Machine Learning, applying Machine Learning algorithms on the available data and modifying the operation to include the new technologies at their full potential is challenging – and this is what the digital transformation is about.
The process industry, as the name tells it, deals with operating various processes in order to produce goods, it is a "physical industry". The result of the operation is an end product. Weather is part of a pharmaceutical, petrochemical, energy or water treatment plant, the process needs to operate at controlled set-points, be robust to disturbances and failures, and have a guaranteed stability. This is all done by applying Control Systems theory, a branch of applied mathematics that in the last two centuries made possible the application of all the technological advances. From maintaining a constant pressure in a tank or a constant power output of an energy plant or water flow from a treatment plant, the Control Systems theory helps design the feedback loops that optimally maintain the processes operational.
The most important aspect when designing the control is assuring the stability of a process. Vivid remainder of stability failures are present in the popular culture as the memory of the Chernobyl plant disaster. The first IEEE Bode prize lecture, “Respect the unstable” (Stein, 2003) gives deep insights on the use of Control Systems tools in assessing the stability of a process.
Apart of stability, controlling the operation of a process implies robustness – assuring that the control actions are affected only in small part by noisy measurements, errors of design and assure the process operates inside design limits.
Another aspect is the deterministic nature of the process and the control: even though the processes are usually nonlinear and different types of Control Systems algorithms are applied, in the end a mathematical model is obtained and used; this model has a direct interpretation in the physical world and the entire system (process + control) can be understood in detail. This aspect is very different to the use of a Machine Learning neural network where the different coefficients of the neurons on all the different layers have no physical interpretation – this represents a major obstacle in adopting Machine Learning in this industry.
Closing the feedback through a neural network and letting a Machine Learning algorithm take full control of a process might be far away. However, Machine Learning algorithms, with their power of identifying patterns can be used in assessing the state of the process in an open loop and give support to the operators; in other words, to function as a decision support system.
The major benefit of using Machine Learning algorithms in such industries include: predicting failures, scheduling maintenance, reducing downtime and operation costs, improving efficiency by identifying bottlenecks and sub-optimal operation states.
The most common ML algorithms used for processes are anomaly detection, and classification. A Machine Learning algorithm can be trained with years of historical data in order to predict and alert when a component is staring to malfunction, when it deviates from the normal operation. The art is providing the correct data and correct labels to the algorithm: it will report an anomaly outside the region that was represented by the training set.
The reinforcement learning algorithm is bridging the gap between the Machine Learning and the Control Systems as it closes the feedback loop and interacts with the process in both ways: reading measurements and sending actions. This algorithm maps the environment into an internal model and at each step selects from a range of possible actions to interact with the real environment with the purpose of minimising an optimisation function. This type of algorithms was used by DeepMind to win at Go against the world champion (Silver, 2018) and also to win against human opponents playing PC games (Hodson, 2019). These algorithms also found their way in a more process critical application of optimising the energy consumption of a data centers (Judge, 2019). The challenge with this type of algorithm is the high sensitivity to the changes of the environment (as always happens in the industry): a change in size of the Go table, the change of possible actions in a PC game or a physical change of an element of a process makes the model obsolete and imposes an expensive and time-consuming re-train.
Even though it is still very optimistic to expect Machine Learning solving all the problems we have and optimising the operation of an entire plant, these algorithms can perform very well in some particular applications that can support decision making. As on the optimisation part, “optimal control” is a consecrated and mature branch of Control Systems with many years of study and successful implementations.