Addressing variance propagation with dynamic optimization

In continuous production processes, such as polymer extrusion, food processing or continuous drying, a multitude of factors are at play that influence the process behaviour, quality and efficiency. These factors include, but are not limited to variations in the input material, ambient conditions, machine wear, poorly designed control loops and natural process fluctuations.

As a consequence, variance is inevitable in any production process and a critical aspect for process quality and efficiency.

However, we need to distinguish between true random variance that manifests itself in process and measurement noise, systematic oscillatory patterns and the normal dynamic behaviour of complex processes. For instance, in any given moment 10-20 significant causes of variations are present in conventional extrusion lines, where examples of melt temperature, water temperature, haul-off speed and material properties are shown in Fig. 1.

Fig. 1: Examples of variation in a conventional extrusion line.

Variance propagates through the line

These sources of variations are particularly important in continuous processes where several process stages are chained together and the variance propagates through the line sometimes amplifying, sometimes dampening along the way. As a result, both the final product quality and process efficiency are severely impacted which reduces the competitive edge and leaves significant potential untapped. Moreover, when the normal process variance takes up a large portion of the specification range (see Fig. 2 for a pipe diameter as quality characteristic) then the safety margin is gone and no room is left for additional variations leading to longer startups, more production scrap and frequent downtimes.

Fig. 2:  Impact of process variance on the resulting pipe quality in a polymer extrusion line.

Therefore, minimizing process variance is of great importance at every stage, not just a the last mile (meaning the resulting product quality). Unfortunately, interactions along the line are complex, time delayed and difficult to comprehend, so the true source of quality deviations are often unclear lying behind the veil of complexity.

Prediction models to forecast the behaviour of entire production lines

At jora.tech, we are building dynamic data-driven models of entire production lines enriched with physics and process experience to address these challenges. That allows us to forecast the behaviour of essential process signals, such as melt temperature and pressure for a prolonged period of up to 60 minutes in advance. Together with our state propagation approach we can model the impact on the resulting quality to forecast the future behaviour and providing these insights to operators and process engineers in real-time (see Fig. 3). This approach represents the basis for our What-If analysis module which enables operators to glimpse into the future, conduct virtual experiments and receive targeted recommendations for specific goals (e.g. minimum overweight, maximum speed) based on live data from the line. As a result, the operating paradigm transformed from a reactive to a proactive and predictive approach.

Fig. 3: Dynamic predictions for process signals and resulting product quality.

Dynamic optimization to reduce process and quality variance

Nonetheless, while this human-in-the-loop approach proves effective and is usually the first step into the realm of predictive optimization, the full potential of predictive control is only realised when closed-loop control is enabled (see Solutions). This allows to continuously optimize the settings of the production line making sure the line runs at its true optimum at all times given the predicted future behaviour of key process and quality signals. Consequently, the oscillatory behaviour and normal process dynamics are addressed along the line reducing the variance and material consumption (e.g. meter weight of polymer pipes) significantly (see Fig. 4).

Fig. 4: Reduced variance and meter weight due to continuous predictive optimization.

The main ingredients for these improvements are both the dynamic model allowing to predict the future behaviour and the end-to-end perspective integrating the full process characteristics (not just individual components). As a result of predictive closed-loop control, the following process improvements can be expected:

• Reduce quality deviations by 80 %

• Reduce process variance by 60 %

• Increase output and process stability

• Detect anomalies faster to mitigate downtime and damages

To summarize, while other approaches, such as static models, allow you to optimize the average trajectory without considering the process variance along the line, only dynamic models in combination with predictive control allows you to optimize the entire production line from an end-to-end perspective making sure your line runs as optimal as possible.

Want to know more about our approach? Get in touch and let’s discuss on how jora.tech can help elevate your production to the next level!