With sustainability and cost pressures in the spotlight, chemical manufacturers are increasingly seeking to boost efficiency through data-driven decision making. Advanced analytics and explainable machine learning offer complementary ways to mine new insights from existing data. In this post, we'll explore how these solutions deal with the all-too-common problem of asset degradation, specifically focusing on heat exchanger fouling.
Heat exchangers inevitably foul over time, although the specific rate of fouling differs depending on various factors including temperature and chemical properties. Once their efficiency drops below a certain threshold, they are taken offline for maintenance, slowing down production. Thus, at a global-scale continuous production facility, even minor heat exchanger fouling can have a significant impact on P&L.
How does advanced analytics help?
Given the deluge of data that a typical process plant produces each day, it's not always easy to see the whole process, especially using an old-school tool like Excel, which many plants continue to rely on. With advanced analytics—a set of solutions that help visualize data and diagnose issues, such as dashboards and traditional Six Sigma models—engineers can easily and quickly flag important issues and trends.
When it comes to heat exchanger fouling, the immediate application of advanced analytics offers valuable benefits. From examining the device's historical fouling instances, its behavioral pattern can be plotted, producing visualizations that can easily be shared with others on the team. These recent trends can then be extrapolated to the future to give an idea of how the heat exchanger's efficiency will evolve over time. Using this pattern, engineers can understand roughly how often the heat exchanger needs to be cleaned and plan the maintenance schedule accordingly—assuming this pattern continues to hold true.
What about explainable machine learning?
Once the heat exchanger's fouling behavior has been visualized and plotted using advanced analytics techniques, you can use explainable machine learning to dig deeper into the data. This is particularly valuable in one particular type of use case that we often see in process plants, where a heat exchanger starts rapidly degrading for unknown reasons, rather than following an understandable, linear pattern.
Unlike advanced analytics, which relies on the user's domain expertise when it comes to root cause identification, machine learning models can take in historical data and automatically learn all the factors affecting production KPIs as well as relationships between them, in a matter of hours. In the heat exchanger use case, for example, the models learn both what factors are affecting the device's performance and degradation and what those effects are.
Sometimes the factors are fairly obvious, i.e. bulk temperature. In other instances, however, a machine learning algorithm may be able to dig deeper than conventional solutions and uncover complex root causes occurring upstream of the heat exchanger. For example, a process medium may contain certain components which lead to crystallization or solids forming on the device's surface. The specific composition may vary depending on various upstream factors.
Traditional machine learning models can also pinpoint root causes. However, in a traditional black-box machine learning algorithm, the results aren't visualized in a way that the user can understand, limiting their usability in a plant where engineers need to actively analyze and interpret data. Explainable solutions, in contrast, quantify the potential influence of each parameter and provide confidence measures, so that engineers can understand them the same as they might an advanced analytics dashboard.
Along with helping track down the root cause of fouling, explainable machine learning solutions can make predictions about the heat exchanger's future behavior and come up with recommendations that extend its life and minimize costly repairs. Earlier we noted that advanced analytics has a similar prediction-generating capability; however, this capability is limited to linear relationships. In cases where the device deteriorates rapidly for mysterious reasons, a traditional plot would simply be inaccurate. Using explainable machine learning solutions can help supply more accurate predictions about degradation, limiting maintenance downtime and ultimately lowering operational costs.
Another benefit is that engineers might also be more proactive about changing their workflow based on the models' discoveries about optimization. Once certain upstream factors are pinpointed as affecting the heat exchanger's behavior, they can adjust their behavior accordingly to minimize these factors.
In short, advanced analytics and explainable machine learning can serve as complementary solutions to solve the costly problem of heat exchanger fouling. Indeed, many software tools include both functionalities. Whether you're looking for quick insights via advanced analytics, or more in-depth problem-solving and recommendations via explainable machine learning, these solutions can work together to take your process to the next level.