By Jennifer Shine, Principal Solution Engineer, Imubit
The quality of a plant's history sets the quality of the model learned from it, and closed-loop optimization and control raises that bar further. That does not mean the data has to be perfect before you start. Send the historian export as it is: gaps, outages and noisy tags are normal, and our engineers screen and clean them in the platform. What follows, drawn from Imubit's closed-loop deployments, is what helps most over time, and what to set up now so the next model learns from more of your plant's history.
1. Data volume
"The first thing I learned from my academic mentor is that the only thing better than data is more data," said Imubit CTO Prof. Nadav Cohen in Imubit's Demystifying Industrial AI webinar series. The more data a model learns from, the wider the range of operating experience it gains and the better it performs. Neural networks depend on large, high-quality and varied datasets, because their accuracy rests on how rich and reliable that data is.
For robust modeling, we recommend six months or more of unit process data, and at least 200 lab samples to support inferentials. That much history usually covers a wide range of operating scenarios. The most common setup we see is data stored at 1-minute intervals on a historian on the business network.
2. Data compression
Historian compression settings affect how accurately a model can learn from historical data. When instrumentation is added or modified, the historization and compression settings are often overlooked. For example, re-ranging an instrument without adjusting its default compression settings can hide much of the real movement in the data.
Validate historian compression settings as part of the site's management of change process. We recommend setting compression to record a data point every minute for all process PVs and control loop data. The earlier those settings change, the more usable history the site will have when a project starts.
3. Data extraction
High-frequency historized data is often hard to extract and export for use in cloud computing environments. It is a common early hurdle in adopting new AI technology, and it has two causes. Legacy historians and computing hardware often lack easy export capabilities, and outdated data backup strategies make the problem worse.
To address this, invest in basic database retrieval tools and modernize data backup strategies so that this process data can be retrieved when it is needed. Putting a process in place now to access the data prepares the site for future projects.
4. Data quality
Data quality is usually the first concern a site raises, and it rarely holds up a start. You do not need to filter or clean the data before uploading it. Our engineers screen it in the platform for missing data, compression, frozen or flatlined tags and outliers, remove the periods that should not train the model, such as turnarounds and outages, and agree with your engineers which tags matter. Because the model learns from change, a fixed offset or a calibration error matters less than whether the change itself is measured accurately.
What can limit a use case is data that carries too little information: heavily compressed history, setpoints that never moved, or lab samples without timestamps. If the screening finds that, we say so early, and show what would fix it, such as one-minute logging going forward or a planned set of moves.
5. Instrumentation
When designing models for closed-loop optimization and control, include the data that shows how each instrument reached its process value. Check that the historized dataset is complete, with every component of the primary control loops recorded:
- process value (.PV)
- setpoint (.SPT or .SP)
- output (.OUT or .OP)
- mode (.Mode or .MD)
If advanced automation systems are in use, capture their key parameters too, including error and status codes, upper limits, targets and ON/OFF statuses. Confirm that product prices and lab samples are historized as well. Lab results are most useful when they are backdated to the timestamp when the sample was pulled.
A proactive instrumentation maintenance and repair program keeps instruments correctly ranged and calibrated.
6. Existing controls infrastructure
Assess the readiness of the time-series data together with the readiness of the base-layer control system.
- Evaluate control loop operation. Determine whether the primary control loops run in automatic or cascade, or in manual.
- Assess manual field manipulations to find the control loops that need manual intervention.
- Review regulatory control performance to confirm that the base layer performs acceptably.
- Analyze data movement to confirm there is enough variability in the key variables and that tolerances are set appropriately.
- Identify delays in data collection or processing, and address latency and synchronization issues across the business IT and OT networks. Make sure every server is synchronized to a master time source.
Building on the data foundation
None of these has to be settled before the first upload. Working through them over time widens what the models can learn and sets the foundation for closed-loop optimization in the plant. In Deep Learning Process Control® (DLPC®), a Deep Learning Predictor learns a nonlinear dynamic model of the unit from its own history, guided by first principles, so the quality of that history sets the quality of the model. The controller then writes to the existing APC or DCS, inside limits set by plant engineers, which is why the base layer needs to perform well. Imubit's engineers and tools help plants work through these data challenges in refining, chemicals and polymers, cement, and mining and minerals.
