Beyond Monitoring: Data-driven Optimal Efficiency in Manufacturing

Modern-day technology is quickly turning manufacturing facilities into complex structures, as managers try to squeeze small efficiency improvements that translate into large end-of-year financial gains. Process insights are no longer solely derived by human know-how, but also through data-driven approaches, usually in the form of monitoring or planning. As more and more information is collected at the plant level, we must ask: are we leveraging this data to its full potential?

The current standard of data used in manufacturing is descriptive; it focuses on recording past information. This sensor data may, at a later time, be used to extract details of the nonstationary production process. Insights derived from these results are acted upon by updating control charts, improving operational training, and taking other actions. Although valuable, this paradigm can be suboptimal. It relies on subjective assessment and does not consider the restrictions imposed by the production process at the current time.

A better proposition is to use data prescriptively - to analytically propose future actions. In this case, sensors and analytical tools are combined to suggest, at each point in time, optimal bounds for process parameters. Maximizing efficiency in this new paradigm - whether in the form of faster productions, reduced costs, or increased product quality – is pursued at each point in time in an objective and data-driven manner, while considering the controllable and uncontrollable aspects of the production process. Prescriptive analytics can ensure that company objectives are always optimized with current resources.

‍

The many merits of an analytic prescriptive framework

Accurate prescriptive algorithms can impact plants in multitude of ways:

• In the short term we can expect to maximize efficiency, whether it is defined as rate, cost or quality. As long as it is a quantitative measure available in the already collected data – as is in most cases – and dependent on measured production parameters, it can be optimized. Based on historical data, analytics can minimize the probability of bad productions.
• Although insights are typically collected based on intuition and experience, the use of analytical tools can illuminate the blind spots of human judgement, guiding us in directions that would otherwise be missed to draw valuable conclusions. For example, these tools can unveil the critical aspects of the process, the most cost-effective solutions to an out-of-control period, or the correlations between productcharacteristics and the optimal production point.
• The proposed framework decreases subjectivity and improves stability. As such, more uniform productions can be expected, in terms of time, raw materials used, energy consumed, and machine wear. Less variation can benefit other operational aspects, including planning and long-term strategy.
• At the data collection and storage level, analytics can improve focus by selecting the most relevant variables, diminishing the operational costs of the current and future infrastructure, and simplifying data interpretation.

The competitive advantages from these tools ensure a more stable and efficient productionprocess which can translate into large profits for large-scale industries.

‍

The challenges of designing models for manufacturing

The design of prescriptive algorithms for the manufacturing industry introduces specific modelling challenges, that must often be addressed in a case-by-case manner. Process knowledge is crucial for many core tasks, including identifying artifacts, quantifying noise, making design choices, and validating predictions and conclusions.

Typically, historical production records fail to encompass the whole range of produced models or operating points. Extracting conclusions from this data is challenging and, in some cases, not appropriate. One must always quantify the error of their analysis and account for the everpresent possibility of a black swan. Inconsistency and noise in data can be another obstacle, as certain sensors may change, malfunction or be recalibrated during the considered sampleperiod. Detecting, quantifying, and correcting these effects are critical to ensure the validity of the analysis.

On the algorithmic front, a critical consideration is interpretability. In most cases, it is merely not enough to determine if the process is in- or out-of-control. Understanding the aspects of the process that contribute to this state are equally important. Likewise, recommending an action falls short if the manager cannot grasp the relationship between that action and the underlying problem. The design and use of interpretable algorithms is critical in the context of prescriptive analytics.

LTPlabs combines machine-learning methods and optimization techniques to highlight the key factors that influence the production process. With a focus on interpretability, we work closely with domain experts to extract actionable conclusions from the often-overwhelming landscape of process data.

‍

Final Remarks

As manufacturing collects ever-expanding repositories of historical records, the transformative capacity of this data grows exponentially. Today, a new paradigm is possible where algorithms are not merely descriptive but prescribe actions and maximize objectives. Employing a robust analytical approach is necessary to act on this potential and harness its competitive advantages. The future of manufacturing is more precise, stable, and efficient, and LTPlabs is solidifying its position at the forefront of this revolution

‍