How can machine learning be used to optimize industrial processes?

Machine Learning and manufacturing are still in the very early courting phase. Rather than trying to supplant current practices from the Six Sigma and Lean frameworks, ML should be incorporated as one of the tools in the analyze/improve step in the traditional DMAIC type processes. We must not ignore the limitations of machine learning. It isn't some sort of magical black box where data goes in, and answers comes out. Most analytics do not do well in a high mix, low volume environment. Sometimes, your production run is three parts, and you won't build that specific part ever again. Discrete Event Simulation, Operations Research, and Life Data/Survival Analysis are much more adept at handling tiny, messy, datasets with censoring thrown in. We must be sure of the objective and metrics to be reported. Do we only need the expected value/prediction? Do we need to speak to the largest decision variables/features? How does your model support the physics of failure? Does it line up with what the people on the floor see? As I've always told my direct reports, use the right tool for the job. Be creative and combine tools but know their limits.

The most difficult part of bringing ML into a manufacturing environment is not necessarily the data delivered, but rather training the decision makers to act on the collected data. Knowing the optimal time to do preventative maintenance is great, yet useless if the manager will not plan for the downtime to work on the machine. Oftentimes the leaders in production/manufacturing are nearly tech-illiterate, and will need an 'interpreter' to explain what the data result is explaining. I've had to explain it using a car analogy: take your car for regular oil changes or the engine will be ruined. We need to begin to train the decision makers and line operators on HOW to apply the ML software being pushed to production floors. AI/ML will do more improving in a few years that previously took decades. We need to bridge the disconnect from the programmer/data scientist that made the ML and the end user who is trying to use the tool. Otherwise ML is just a nice thought exercise for the 'office types'

Given a particular line of a factory which produces in batches. With enough data gathered from other production lines having unique technologies, instead of having some machines redundant until a machine is faulty, machine learning could be used to estimate the optimal time in the operation of each machine in a line. That way, it would not just help with a preventive maintenance but also extend the overall life span of the individual machines on that line. A similar approach is used by nutritionists whereby patients are asked to eat small chunks of food throughout the day instead of just 3 times a day. The duration for which the machines operate would be estimated partly by the?machine learning model, given the parameters of the system at any given time and, the needs of the line manager and his team put into consideration that they are specialists in the field.

Machine Learn (ML) can also be applied as an open box to discover new laws for processes! ML with symbolic regression can generated candidate equations for processes. With a human in the loop, you can evaluate the best candidates and potentially make fundamental discovers about the relationships in your process. The Machine Scientist!