By Pat Dixon, PE, PMP
President of DPAS, (DPAS-INC.com)
If you take all the data in the world and use it to train a neural network, the likely result is a neural network that learns noise. We live in a noisy world. Noise is not information and doesn't help you optimize your process.
It is no surprise that your process produces mountains of data. Machine Learning is a way to find gold in those hills of noise. Before explaining how, it is important to define our terms. What is Machine Learning?
The term Machine Learning is often used interchangeably with Artificial Intelligence. They are distinctions between the two. Artificial Intelligence is a broad term that refers to the ability of non-humans to perform as humans. Machine Learning is a subset of Artificial Intelligence using data to learn patterns and make predictions. The following illustration (found at https://discuss.analyticsvidhya.com/t/difference-between-ai-ml-and-ds/74724/2) helps explain these distinctions:
How can Machine Learning help you? I will offer an example.
Back in the 1980's there were attempts to measure the strength of a sheet of paper online at the reel. This would allow real-time measurement, and possibly control, of a critical quality attribute. This would eliminate the chance that an entire reel is produced, followed by samples sent to a lab, with later discovery that the entire reel you turned up, and the one you are currently making, does not meet specifications. Unfortunately, online measurement of strength properties continues to elude us.
Although we lack the direct measurement, we have a mountain of data that could predict the strength of that sheet to a reasonable degree. We know there are many measurements we have that affect strength. Machine Learning, properly applied, could provide a prediction of previously unmeasured parameters, which can make for a much more efficient and profitable operation.
That is one example of the potential of Machine Learning. To turn that potential into reality requires being able to discern truth from the noise of buzzwords and hype. Below are some tips:
- While there may be some that claim they can take any and all data and turn it into gold, most of the work in Machine Learning is to take the noise out of the data to create a usable set of data for Machine Learning to train with. This is called pre-processing. Applications that lack the ability to help you pre-process data should be met with skepticism.
- Aligning data in time is part of the pre-processing. The moisture that is measured at the reel affects strength immediately, but the basis weight valve is many minutes before the reel, and the refiners are further back. Knowing what affects the prediction you are trying to make, and when, it imperative to create a usable model.
- If you are only trying to create an accurate prediction, you can be a little less picky about the dataset you create. If you are using Machine Learning to help you optimize or automatically control your process, you need to narrow down your dataset further. The inputs in your model should be independent of each other and have significant effect on the measurement. Failure to remove redundant or insignificant inputs can result in nonsensical predictions and behavior.
- A big advantage of a Machine Learning algorithm (such as a neural network) is that it can model a relationship of nearly any complexity. That is also a danger. A model used for optimization or control that has unreasonable relationships has essentially learned noise. While we have plenty on nonlinear steady state relationships in our processes, we know that most of them will have fewer than 3 inflection points. Although neural networks are a black box technique, it is possible to show gains (sensitivities) to reveal whether the model is realistic. This is a feature that I regard as essential in a Machine Learning product for an optimization or control application.
- Machine Learning efforts can benefit from a data scientist that knows how the algorithm works and what helps to create an accurate model. However, this alone would likely spell disaster. Without process knowledge, your prediction might be useless. Knowing how a process behaves and applying first principles cannot be replaced with data and algorithms. Combining a process expert with a data scientist can be an ideal team to produce successful Machine Learning results.
- Just like all of use, Machine Learning predictions will not be perfect. There will be mistakes. Setting expectations and knowing how to handle situations where the prediction leads the wrong way needs to be considered to maintain confidence in the application.
- To mitigate erroneous predictions, use your lab. It has not been rendered obsolete. You cannot measure everything. A furnish change (hardwood, softwood, recycle) may affect the strength of your sheet without a clear measurement in your system to account for the composition of fibers in your sheet. Using your lab as feedback to bias your prediction is a good practice.
This is a sampling of tips. There are many more, and experience is a great teacher. Creating your own datasets and playing with available products or some of the freeware that is available (such as Python and R) can help you get more familiar with the triumphs and pitfalls.
Machine Learning is being applied in our industry and producing gold. It can work. It can also fail if hype overshadows reality. Filtering the noise from the hype can help mitigate the noise in your Machine Learning investment.
