400586 Embedding Continuous Data Analytic in Industry Big Data Applications
This paper/presentation details an approach for embedding data analytics in process control systems. The analytics are executed as function blocks inside control system controllers and are configured and trained using techniques familiar to control system users.
Embedded data analytics is a core component of big data applications in the process industries and can be used as the initial approach for big data applications and to test data-driven techniques before moving to the much larger big data approach.
An integrated approach toward analytic functionality design provides an easy-to-use set of tools for analytic model development, verification and subsequent download to the DCS controller for on-line operation. The model development and download for on-line operation is streamlined into clearly defined and easily executed steps:
- Defining analytic model configuration, i.e. identifying potential process parameters for PCA/PLS model
- Creating analytic module based on the defined model configuration, downloading module to the DCS controller.
- The module is configured to collect history data, including lab analysis, that is defined by the analytic model.
- Developing analytic models, including PCA, PLS, NN and MLR from collected historical data as defined by the analytic module or from an external historical data file created prior to the analytic module download
- Validating and downloading model for on-line operation
On-line analytic modeling monitors faults in the process operations and predicts product quality. The results are presented in a web based interface and can also be used to enhance an existing alarming system. The predicted property quality can be used in designing the control.
The continuous data analytic design has been tested on simulated data and in the field. An analytic application’s PLS model was used to predict heavy components in a distillation column. When the results were compared with an online analyzer, the analytic modeling demonstrated several advantages: more reliable operation, less demanding maintenance, and integrated fault monitoring that prevented the use of the prediction results when a significant fault was detected in the process operation. Process engineers greatly appreciated the ability to identify abnormal distillation column operation and the potential for including this functionality into an existing alarming system.
The field trial facilitated the development of an iterative procedure for analytic model improvement. Similar results were obtained from a polibutene unit analytic model development and an on-line test. Specific to this process is the product quality (viscosity) which can be set on two distinct levels, depending on the manufacturing needs. Two alternative approaches were explored: one using a state parameter associated with viscosity and the other using two models for Low and High viscosity values. In the majority of tests two separate analytic models performed better. The primary problem with the multistate analytic model was identifying a process parameter that was defined in the analytic model as a state parameter that correlates well with quality level.
We greatly appreciate the professional and extremely dedicated support in the field testing we received from Ashok Dasgupta and Tom Vernor from Huntsman and from Efren Hernandez and Bob Wojewodka from Lubrizol. James Beall from Emerson provided a wealth of control expertise and helped us to develop and install analytic models.
See more of this Group/Topical: Big Data Analytics