The data scientist conundrum – model selection and process visualization through dashboards (Part 3/3)

Guiding data scientists to create the right model to optimize the manufacturing process.

By Ilias Iliopoulos PhD, Yoram Barak PhD

In Part 1 and Part 2 of the data scientist conundrum series, we briefly discussed the tasks of data aggregation and preprocessing and data transformation to identify the key features and patterns in the data. In this article, part 3 of our series, we will briefly examine how SmartFactory Rx® helps data scientists choose and create the right model for the process and helps them visualize that process (Figure 1).

Figure 1: Typical steps in the modeling process
Manufacturers spend millions of dollars on Fault Detection & Classification (FDC) and Statistical Process Control (SPC) to:
  • Reduce process variability and scrap
  • Improve productivity
  • Reduce cost and unscheduled equipment downtime
  • Make real-time decisions on the manufacturing floor and react to process changes

The overwhelming majority of manufacturing sites have rudimentary SPC that follows Univariate Analysis (UVA) to monitor key properties of the product. SmartFactory Rx effortlessly provides the capability to monitor any real or virtual variables. Algebraic expressions can be written based on observed parameters and then treated as variables themselves (virtual sensors). There is flexibility in treating brand new or mature processes: limits around the variables can be adaptive based on user defined rules e.g., Exponentially Weighted Moving Average (EWMA), based on the standard deviation of the parameter, or simply expressed as % deviation around a target value.

Multivariate Analysis (MVA) adds another layer of knowledge and analytics power. The user can define the data to be modeled and the application will take care of the training and model creation, as well as the precision of the created model versus incoming subsequent production data. The next key component is supervised modeling. Batch sampling or metrology results can be ingested as they become available and lead to an improvement in model accuracy.

As discussed in previous blogs, the models can be hybrid – marrying the convenience of statistical modeling with the accuracy of first principles based modeling – by allowing the inclusion of virtual sensors tied to the process being monitored. In addition to the UVA and MVA modeling capability, the code-savvy user can integrate their own python script in the SmartFactory Rx modeling arsenal. The models are able to calculate process health metrics based on the process parameters.

The subsequent health metrics are plotted against time for the duration of the process and the operator on the manufacturing floor has a clear understanding of whether the process is deviating, or not, from the expected optimal behavior (Figure 2).

Figure 2: High level overview for equipment condition
If the operator identifies an issue, they can select any of the colored points and drill down to try to understand the cause of the process drift – that is, select any of the available green, amber or red dots on the equipment diagram that signify the process parameters (Figure 3).
Figure 3: Drill down of equipment health to understand variability in recent runs
(temperature for this example)


SmartFactory Rx enables the scientist or operator to gather data from dissimilar sources into a single database. Through its analytics toolbox that can “slice and dice” the data, it aggregates sensor information into equipment health metrics, streamlining the analysis of the hardware and process health status. It simply provides an easy-to-understand index on a single screen. This is used both on R&D and manufacturing floors to identify any potential issues either with the equipment (by operators) or the process (by scientists). This enables the user to take proactive action to resolve such issues before they manifest and cause an unscheduled event. The implementation of this methodology can help minimize the loss of valuable material as scrap and/or expensive unscheduled equipment downtime.

Related Posts

Predict, don’t react

Predict Dont React
Drive adaptive maintenance and avoid down-time using your manufacturing data

Beyond the MES: Fly. (Part 2/2)

Beyond the MES
Advanced capabilities and MES provide the intelligence that is key to the ‘lights out’ factory

Minding the Gap with AI and ML Processes

Minding the Gap with AI and ML Processes
Experience a smarter way of controlling manufacturing using the SmartFactory Productivity AI/ML platform.