Getting Important Quality Measurements in Real-Time

Most process plants have at least one key quality parameter that is not measured in real-time. Traditionally, process manufacturers have relied upon manual lab samples to verify that the process was producing products within the required specifications. The issue with this approach is the time delay between when the sample is drawn and when the lab analysis is complete. If the results are not within the required specifications the product during this period of time must either be reworked or in some processes, disposed of. This reduces the efficiency and profitability of the process.
Over the past decade, neural network technology has been introduced as a way to create models to act as soft sensors for properties where no physical sensors exist. Soft sensors are also used in conjunction with analyzers to fill in gaps between sample points or for validation/backup of expensive analyzers. Some examples include: Kappa analysis in the pulp and paper digestion process, end (cut) points of products in the CDU/VDU columns of a refining process, food properties, end of fermentation process prediction, and emissions analysis.
In spite of the indicated potential and the improved tools, the acceptability of neural net soft sensors has been fairly limited. I spoke with Ashish Mehta, a lead developer in the DeltaV APC technology organization, who feels the complexity (actual and perceived) has been a major factor. He presented a paper, Successfully developing a property estimator with DeltaV Neural (7.8Mb), at the last Emerson Exchange to help alleviate such concerns.
Ashish mentioned the real benefits of using soft sensors:

  • They provide real-time online predictions of important quality variables (as fast as 1s)
  • They reduce process variability as predictions can be used in feedback
  • They improve control as quality parameters can be incorporated into APC/optimization strategies

Like other process optimization projects, it’s important to make sure your instruments providing the data for the model are sound. Ashish recommends auditing the sensors and valves providing the measurement data to make sure they are reliable and reasonably well tuned. Important process equipment should not be out of service or bypassed causing it to be operating in a non-standard way. Finally, it is critical that the lab measurement and analysis process is streamlined to ensure that measurement values and delays are accurate.
When it comes to selecting the variables as input to the neural network, make sure you capture the ones causing the dominant effects. More inputs are better, although avoid those offering redundant (highly correlated information.) Use calculated inputs like ratios/first principles-based inputs.
Ashish stressed that neural networks are empirical models where the underlying model knows, and is therefore only as good as the data it is trained on. As a result it’s important to collect the data over a wide operating range, and make sure outliers (say due to shutdown) are removed from the training data set. Generally, you should also maintain an additional data set, different from the test set, to verify the model by comparing its prediction with the actual data.
If the soft sensor has been created using DeltaV Neural, it is commissioned by a simple download to an NN function block. The function block approach greatly simplifies the online operations and increases the soft sensor lifetime. For example, it will monitor for any of the inputs being outside its trained range and mark the soft sensor output as uncertain, so that the control strategy can take this uncertain information into account. It continues to use the lab analysis results in a feedback fashion to automatically adapt the prediction value to any process changes after training.
In addition, the NN function block can be used in a closed loop control strategy as the PV of an MPC (or PID) block. According to Ashish, the greatest opportunity is the ability to use the key quality measures, that were available only infrequently, in full closed loop (advanced) control and optimization strategies thereby resulting in significant variability reduction (and likely increasing the operation’s profitability.) You should always look ahead to closing the loop on the quality parameters that you want to develop a soft sensor for.

Posted Monday, May 15th, 2006 under Process Optimization.

Leave a Reply