Introducing data-based techniques into the domain of traditional industrial process modeling often raises the question, Are these approaches friends of foes? Our Chief Data Scientist Emeli Dral explores the differences and practical implications for performing process optimization tasks.
The Two Worlds
In the industrial world, when we talk about modeling, we usually mean first-principle models. These are the classical models that are based on known physical laws, such as the laws of mechanics or thermodynamics, and their dependencies. The starting point for building such models is the ability to describe the relationships between key parameters based on domain knowledge of the processes involved.
With machine learning, the models are instead built on the basis of the actual behavior of the system and observed relationships between input data and results. Such models do not require the ability to explain the key relationships. Instead, the system is fed with historical examples so that the algorithm can "grasp" the patterns by itself.
If we look at how these models are created and used for process-optimization tasks, we can notice a few important differences.
Classical models are easily interpreted, which allows the expert to explain each and every model response. However, classical models cannot use all the available data about the process; they rely on the creator's ability to establish relationships between the variables.
These models cannot accommodate infinite complexity, and often have to presuppose some "ideal" system state.
On the other hand, machine-learning models are much less interpretable: there is no straightforward way to explain each individual prediction they make. But on the upside, they are often much more accurate due to their ability to account for many weak factors and hidden dependencies.
Although classical models require much fewer examples for model tuning, the data must be complete and accurate. Classical models cannot deal with measurement errors or gaps; all the input data has to be "cleaned" and verified for the model to work.
As a result, one might create a great model with a limited "golden set" of examples describing a certain process but be unable to apply it in a real production environment, where data never comes in a complete and perfect state.
In comparison, machine learning is based on statistics and probability rather than process knowledge, so it can work even when the data is noisy or partly missing, and sensors are imperfectly calibrated. However, it also requires many more examples to learn from.
Classical models require subject-matter expertise. They are so powerful because they are based on a deep understanding of industrial processes and underlying physical, chemical, and mechanical transformations.
As a result, they depend on the creator's ability to understand and study processes in-depth, as well as to maintain and calibrate them when the system changes.
On the other hand, building machine learning models requires expertise in data analysis because you need to collect, process, and combine data as well as selecting the appropriate technique for model training. Due to its ability to indirectly account for unmeasured variables, the technology can be applied even for complex, less-understood processes, or when classical models might be hard to create and maintain.
Interested to know how AI can make your production processes more efficient? Come and talk to us!