Step 4: Using Contextualized Data for Modeling and Predictive Analysis
In this section, we delve into using soft sensors, which are derived values based on several different inputs and how to predict future outcomes, based on ML.
Soft sensor values are the combination of data from multiple sources. The sources may be laboratory and equipment data, and soft sensors can be predictors of both laboratory and equipment data. To develop a soft sensor, the team combined the "running OK" data set with laboratory operational
Importing operations data to Excel filtered by status (in "running OK" mode). (Courtesy of O.A. Bascur, OSIsoft LLC.)
data to create a predictive model. They then incorporated the model into online asset-based template calculations. Using soft sensors increases the value of the control strategies, avoiding process constraints and moving into optimal operational states.
The operational variables are classified into (1) manipulated, (2) controlled, and (3) disturbances. "Selecting these variables lets us predict the results of the target variables (target tags). This is done based on the effect of the disturbances, which suggests how to move the manipulated variables to achieve the desired objectives. The target variable is what you model based on the operating variables. If it is a soft sensor, it can be a controlled variable," said Peter.
Having all process units available simplifies the selection of the manipulated variables to avoid constraints. This way, optimal yield can be achieved while reducing operational costs.
By extracting subsets of the "running OK" data into Excel spreadsheets, the team creates tabular data, which is easily imported into ML tools. Referring to the spreadsheet shown in Figure 4.14, it shows the algorithm that interpolates the requested data for further analysis with the ML tools. The ability to integrate contextualized real-time and event-based information and then pass the results to external modeling tools is an effective form of data-driven analytics.
Predictive model showing a soft sensor generated from the running OK mode data. (Courtesy of O. A. Bascur, OSIsoft LLC.)
Figure 4.15 shows the Microsoft Azure Machine Learning Studio objects, which are used to generate a predictive analytics model. Predictive analytics is the art of building and using models that make predictions based on patterns extracted from historical data. An example is sample laboratory data, such as percent assays, turbidity, temperature, and pressure distillation profiles.
Figure 4.15 shows the soft sensor results using EIDI analytics to calculate a predictive model. This model is generated with a dataset in Microsoft's Machine Learning Studio to calculate the coefficients for a multilinear regression algorithm. The real-time model shows the predictive soft sensor data for a particle size analyzer (Steyn et al. 2018). This particle size analyzer soft sensor is used in implementing an advanced control strategy along with the real-time sensor. It acts as a redundant sensor with the physical one, if the physical sensor is not working correctly, as shown in Figure 4.15.
The ability of time- or length-based data aggregation within the data infrastructure provides the necessary augmented data for analysis such as averages, standard deviations, totals, modes, as well as minimum and maximum for the many operating states. It has been found that by having the data subset contextualized for particular operational states, the development of predictive models is simplified. It is easy to understand that if everything is in the "running OK" mode, the models will be completely different than if there is a disturbance such as an equipment failure or a fire.
For further information on how to use Microsoft's Machine Learning Studio and other tools with these unit event templates, see the "Additional Reading and Template Implementation Materials" box at the end of this chapter.