PID controllers can deliver more than just control. They are ready to talk continuously about process health if you are listening. Controllers produce data streams you can use to detect process health events.
This paper illustrates the basic steps to detect latent process events with streaming analytics against measures like PID controller strain. PID controller strain will be defined and techniques will be presented on how to stream simple pattern recognition logic. This paper digs deep to present best practices that automate detection and communication to the user. Techniques that leverage deeper historical looks are emphasized to trigger only meaningful events. Special focus is given to generating actionable alerts with detailed alert message content that both educates the user and defines troubleshooting steps.
Case examples from the power industry are presented. The analytic techniques are applied to historical data from a random set of sixty PID controllers. A restricted pipe, a process chemistry shift, a lack of control air pressure, a liquid-starved pump, a product viscosity shift, a worn impeller, a lab mistake, a misconfiguration of a motor, a failing transmitter, a sticky control valve are examples of process anomalies that are indicated by changes in PID controller strain.
Presented ISA Process Industry Conference 2019
Get ready to hear voices! Once you implement this strategy across operating units there will be many new voices talking. It will be like the folks who never say anything suddenly have a lot to say. PID controllers are all around us. We have them in nearly every industry, site, building and process unit.
Here are some physical problems in plants that are often indicated from PID controller data: machine wear/age, process fouling, flow restrictions, cavitation, contaminants, vapor in a liquid service, poor instrument calibration, poor machine repairs, shifts in material properties, chemistry, feed stock quality, upstream conditions, buffered energy, ambient conditions and many more. We don't have a sensor for every condition but we can combine sensor streams in smart ways to reveal more insight.
PID Controller strain is derived from simple math. Consider the PID controllers in your processes. They send a MANIPULATING VARIABLE (MV) to a machine and they measure the resulting response called the PROCESS VARIABLE (PV). The ratio of (MV / PV) indicates the controller's strain. That's it! Calculate this and store it in your process historian.
Consider that a pump's speed (MV) delivers a liquid flow rate (PV). During healthy operation a pump's strain is usually constant. Poor machine health and process disturbances show up right away in pump strain. For example, starving a pump of liquid pump shows up immediately as drop in flow rate. The PID controller does it job instantly by increasing the speed to raise the flow rate back on setpoint. BINGO... This damaging shift is not obvious to the plant operator and goes uncorrected. This results in major erosion damage to the pump and shortens the life of the pump.
Exploring strain data is exciting. You can discover rare events that are hidden. Some anomalies you find will be progressive over time (such as fouling, corrosion, blockage) and lend themselves to being forecasted. Forecasts then drive cost avoidance and cost optimization.
PID controller strain is a fantastic diagnostic for control valve health. Control valves are used throughout many processes so the impact of this technique at scale can be large. Again, all you need is PV and OUT data to monitor Strain. The insight you find can be significant. Shown below is a typical control valve installation.
A PI AF analytic can be used to analyze patterns in Controller Strain over time and detect when significant shifts have occurred. Industry-standard pattern recognition techniques for Statistical Quality Control (SQC) are useful in this application. PI Templates are a must to allow for mass admin across a large installed base. One new PI tag is required to store the Strain for each controller. Periods of unusual performance are captured as PI event frames and stored by asset. Let PI sends detailed email/Text messages with all of the pertinent stats for a shift in Strain to multiple subscribers. Keep everyone in informed immediately. Emails contain detailed instructions for investigating root causes and next steps. Users need not monitor dashboards but rather wait to be notified of the shifts. PI Vision, PI System Explorer, or PI Datalink for Excel can each be used to study event history. Here it is important to also let PI send eMail messages when the Strain event returns to normal. (Sample PI AF message formats and best practices are provided in the attached kit).
This technique can reveal dependencies and interactions between variables in your process. This can be a natural first step in understanding your interactions. This can form a baseline for describing multi-variable model predictive control strategies. The question of linearity is a good one here too. Many devices operate within the linear range of their output but some do not. Devices working over a larger range of their output often produce non-linear behavior in normal operation. These will require additional allowances for linearization in your PI templates beyond the samples provided here.
Feel free to post any questions here or contact me directly. Don't hesitate to reach out to OSIsoft for support in getting started. Our Success Services engineers can help you launch this and many other techniques to mine value from your PI data.
Best of luck,