The progress of AI technology in the industrial field
Linear dynamics, nonlinear static parts modeled using neural networks. These industrial applications explicitly address extrapolation issues outside of their training base.
In the same time frame, most soft sensor development takes a different modeling approach.
In the 1990s, process systems made important academic contributions to neural network applications. These include hybrid modeling using neural networks, where unknown relationships and/or parameters are fitted to the neural network model. Another noteworthy approach incorporates PLS-type functionality into the network, but allows for non-linear terms rather than linear terms like PLS. Other contributions involve the use of neural networks in classification methods to detect abnormal operations (which can be regarded as nonlinear PCA).
The subsequent development of AI and machine learning (ML) is largely done by large tech companies and is therefore not driven by the applications or needs of the process industry. Therefore, the application of these methods may not be applied 100% in our field. Of course, it's great where they do it. Image processing is an example. Newer networks now offer dynamic modeling capabilities that are an improvement over the cyclic networks used in the past. One example is ChatGPT, which was developed for large language models but has proven equally successful in modeling time series data. We've seen promising results with this technology in soft sensors and hybrid modeling, but so far we've seen few real industrial applications.
We are still in the early stages of the journey to figure out what new developments in AI and ML mean for the process industry. There is a lot of hype, but I believe there is a lot of hope. I think the biggest impact will be in leveraging these AI and ML tools or combining them with existing methods, rather than assuming they will replace them completely.
Comparison of different process control methods
PID (Proportional-Integral-differential Control) : PID control acts as an error regulator, focusing on driving the error to zero. It is often applied in systems with variable or nonlinear models, so it is essential to carefully select the adjustment parameters for stable performance. PID operates in a single input, single output (SISO) manner, but combining multiple PID controllers can introduce complexity to the control scheme.
MPC(Model Predictive Control) : In contrast to PID, MPC utilizes a process model to optimize multiple variables simultaneously to achieve predefined goals. A key challenge with MPC is the need for a known process model. Unlike PID, variations in the model can lead to poor performance, and a model matrix is often required for effective control in complex processes.
FLC(Fuzzy Logic Controller) : Alternatively, FLC intervenes when dealing with different or unknown models by simulating a skilled operator. Instead of modeling processes directly (such as MPC) or focusing on reducing errors (such as PID), FLC simulates ideal operator behavior in different scenarios.
AI Control: Using historical and real-time data, AI controllers strive to achieve goals without prior knowledge of the process. Unlike FLC, AI systems operate like a black box, providing data-based adaptation without explicit knowledge of processes or operations.
Each control method has its own characteristics: with PID, tuning involves using process knowledge to quickly set the appropriate controller parameters based on the desired relationship between these parameters and the process response. For example, a flow loop typically requires a low proportional gain (<0.1), while a level loop requires a higher value, depending on the application. In MPC, complex modeling replaces educated guesswork and emphasizes the importance of well-defined process models. FLC relies on understanding operational success rather than a detailed process model, making it a valuable option for processes that are not clearly characterized. For AI control, large amounts of data and clear goals are essential to guide the system to effectively achieve its goals.
Ultimately, effective process control goes beyond the complexity of the controller itself. Just as in racing, skilled drivers (controllers) need high-performance vehicles (well-designed processes and equipment) to be successful, achieving optimal performance requires a holistic approach, not just the adoption of "smart controllers."
The challenges of AI and ML in the process domain
AI, ML, or deep learning (DL) are all equivalent to large statistical regressions. To get useful models from these applications requires a lot of "high frequency" data, which contains a lot of movement, and a lot of offsets beyond the desired performance boundaries. All of this is required so that the model "knows" the nominal location of the "cliff edge." Much long-term historical data is over-compressed in the name of saving disk space. Therefore, the saying "garbage in, garbage out" is very applicable.
As with any other statistical model, ML does a fairly good job of interpolating, but overfitting at best has the well-known effect of making extrapolations sneaky. As has already been pointed out, closed-loop data often skews model results in strange ways. And, as with all ML applications, "domain expertise" is still required to ensure that the model nominally reflects reality.
One area we haven't seen effectively addressed for process control applications is understanding the physical limitations of control valves, instrument ranges, and so on. This is a problem that early model predictive control (MPC) developers recognized: applications are built to recognize that they have no direct control over processes. Therefore, understanding when the PID controller's motion is restricted or limited in one or both directions is fundamental. ML applications don't seem to grasp this concept at the moment.
Finally, "learning" with historical data depends on ensuring that the underlying processes and control structures for learning data and current operations are the same (except for the compression issues mentioned above). Therefore, changing control valve capacity, heat exchangers, and/or pumps, etc., can skew the model and give unreliable/unpredictable results.
Research progress on the application of AI in process control
In recent years, a number of recent studies from industry experts and researchers have shown that increasing the use of AI technologies could bring efficiency gains to enhance and support process control, as well as those working in the field of process automation.
AI can both be a threat and enhance our work in threat search and intelligence. Our younger colleagues currently working in the expanding field of industrial process automation and control will benefit from gaining AI knowledge; Basic principles, theories, methods, differences between them and their applications.
As many in the industry agree, our future jobs will not be taken away by AI, but by other engineers who know how to use AI and gain a competitive advantage in the field.
AI is being used to directly control factories
An unattended facility (NUF) is a facility that operates fully automated or remotely, usually with no personnel on site. Wider adoption of the NUF approach in the industry faces several challenges (technical, logistical, financial, and regulatory). There are a number of industry-led initiatives aimed at moving in this direction, while encouraging technology development initiatives that enable this new operational philosophy and ultimately position NUF as a safe, cost-effective and widely accepted approach to the design and operation of oil and gas facilities.
AI combined with advanced model predictive control and advanced regulatory control strategies may help achieve this goal.
Compared to previous manual operations, AI systems exhibit greater stability and efficiency, successfully controlling stability even in the face of external interference by keeping critical operational values close to target values. This is the first example of reinforcement learning AI being formally used to directly control a factory.
