Semiconductors and software make up the internals of the programmable logic controller (PLC). The line for computing power in the PLC has been blurred for a while, and most are more computers than PLCs, but we keep calling them PLCs. This evolution is possible because of advancements in semiconductors and software development. Add the onset of open protocols and smart edge devices for the field, and control systems have no boundaries.
The framework of sensors, actuators, controllers and communications interfaces is allowing control systems to expand in industrial and civic settings. Artificial intelligence adds another layer.
Artificial intelligence allows for adaptive control. Traditionally we do this with a proportional-integral-derivative (PID) loop. ABB has introduced diagnostics in its expanded product, Ability. Ability is an artificial intelligence system in its PLC. ABB allows for management of assets and transformers and gives feedback to a centralized location.
Genix Industrial IoT and AI Suite is used by ABB to look at data convergence and AI-driven insights, as well as scalability. These types of monitoring systems are getting more popular across many platforms, but people need to understand that the application of AI in a PLC may be slow. Why? It’s easier to deploy it with chemical changes and 24-hour type process engineering or energy monitoring than actual assembly. It is also not automatic. People still must interface and categorize so the system can learn upsets and normal operations.
The advantage of Genix APM Copilot is that there is a natural-language interface and real-time decision support. Can we take real-time decision support and adapt it to a PID interface? ABB says it can converse with the Genix system based on the natural-language interface and the real-time decision support, but does it matter if the system cannot be made when an upset occurs?
The response is to set up data-driven modeling based on inputs. Thus, the PID changes cannot just happen in an artificial intelligence environment. Why? The machine must learn, and the disturbances must be learned. This setup is called detection and classification. The next part is classification of faults based on sensor faults, process anomalies or external disruptions. Can you tell the system enough information to support it to classify sensor failure or wiring fault or to distinguish between a process disruption or an operator error? In summary, artificial intelligence may be used to identify problems, classify the disturbances and then predict upsets based on trend models.
If we build up this data, then the system may be able to predict future upsets based on trending. Humans still must classify the upsets and determine what data to use. Once the data is decided upon, then the machine could be set up to use adaptive PID tuning. Adaptive PID tuning means setting up trends to change the proportional, integral or derivative based on an input to the PLC, a comparison between current real-time data and trend models and then what outcome is expected. The system cannot do this in a vacuum. It requires reinforcement learning and model predictive control integration. What if there is data? What about timing? What if the computer adjusts and there is a major swing in a loop control? For those instances, the program still must be set to work within fault strategies and fallback based on anomalies. Why? Automatic changes could hurt equipment.
Get your subscription to Control Design’s daily newsletter.
The other is the option to produce a digital twin or set up test routines with the source data and see if you can predict output based on testing software and the process loop. What does this mean? This means coding the system that is in place and using shadow loops with the input data. Trend the operational data alongside the test data and then make small adjustments. This is labor-intensive and takes thought and planning, and probably a modern PLC. What would be next? Optimization.
Genetic algorithms, particle swarm optimization or other types of optimization might be used to create real-time PID adjustments. How does this differ from a current PID loop adjustment? It really does not, except maybe it gives more versatility to recognize upsets and set adjustments sooner. It still requires human forethought to program the instance so that the system might decide. How people are doing that on the PLC level is beginning with monitoring upsets and adaptive PID loops.
It’s confusing because traditional controls engineers may look at current PID loops as adaptive. The output of the PID is the setpoint of the control valve or the temperature change, correct? The output is changed based on the current value not meeting the setpoint, and the PID adapts.
Artificial-intelligence folks still classify that as a linear response and its based-on linearization of system feedback. Artificial intelligence will allow adapting to nonlinear real-time feedback and may propose different responses.
Traditional PID control is fixed, does not learn and is reactive. Case in point: a drive will try to keep speed if it loses reference and might run away. Most of the time, there are other parameters to eliminate runaway conditions. Artificial intelligence poses new ideas to respond to complex changes, and it does not always have to be linear.
How is it being used in reality? ABB is using ABB Ability OptimE and Marine Pilot Control to optimize propulsion efficiency and reduce emissions using vessel speed and toe angle. Genix APM Copilot and Snowflake AI Cloud are being used to allow manufacturers to ask asset health questions and unify data to optimize pricing, inventory and distribution.
