Machine Learning in Manufacturing

When businesses talk about AI, they are generally referring to two things: either Machine Learning (ML) or Generative AI, normally the Large Language Model (LLM). Also relevant to Manufacturing is Computer Vision. I wanted to outline these three and then give some reflections on ML, and I will write further on Generative AI and Computer Vision later.

It is worth noting that for Village Software and clients, we are integrators of existing technologies, possibly in innovative ways, but we are not doing the maths. This is what makes AI affordable and general Manufacturing systems integrators like us.

Machine Learning

Machine Learning is the application of clever maths to analyse sets of data, usually taking known results to train the system, then using this training to identify other results. I’ll talk shortly about our attempts to predict production speed on this basis as a use case. Machine learning and its statistical ancestors have been around for decades, but are now available and affordable to SME manufacturers. It is maths it is not based on information models like Generative AI.

Generative AI/LLM

Made famous by Chat GPT, this takes generally text inputs, compares these to its own ‘model’ of derived knowledge to generate text outputs. Outputs are subject to error, so good for drafting, marketing, searching documents for meaning, but not suitable for decisions requiring precision that must be correct.

In our practice, we have described in some technical detail an industrial scenario and asked it to generate database scripts to create exempla test data. For example, we have used it to get a spreadsheet of 10,000 imagined but plausible-sounding data points to simulate IOT recordings in the field. These speeds our work, but plausible sounding and true must not be mistaken. It allows us to work faster.

Computer Vision

Computer vision is looking at photos (or videos) to identify and classify. In a simple case, this is done by example. As an add-on, you might photograph 10,000 packages on a conveyor belt. A human marked up the 40 packages that need manual intervention. Then you can use this data to either maintain a quality record of the proportion of issues, or even to mark items for actual manual check. Manufacturing machinery specialists will build these things in. But it can be done in a more ad hoc manner too. We've used Microsoft cloud models but it can be used onsite or even on a phone as required.

Machine Learning and Production Scheduling.

A few years ago, we sat down to consider a problem. Our client was a print company; it produced large volumes of magazines. These went through several steps of production, and it was important to be able to schedule the high capital cost machinery to maximise output.

The problem was that although the characteristics of the product were known in advance. Thickness or paper, number of pages, additional elements (free cosmetic sample), type of finish, production quantity, it was proving difficult to schedule the right quantity of machine time or to set KPI expectations for speed in a way that people can be held to account.

This is an ideal problem for machine learning. This is what we did: we were heavily involved in marshalling the company's KPIs and data warehousing to provide activity-based costing and performance information required by the operations director, finance director, corporate HQ and others. So we were able to extract years of data on various machines, plotting the speed against product specification. For those unfamiliar with the concepts, speed is not as easy as it sounds. You have a speed curve, slower in make ready, faster speed means more stops, so potentially slower average throughput. We were mathematically on top of all that through the business intelligence systems.

The basic machine learning approach is.

1. Get the data into a single table of parameters and outcomes (item specification, production speed in our case).
2. Use 80% of the data to search for correlation between parameters and outcomes, a number of off-the-shelf regression models with names like Random Forrest and Decision Tree can be applied.
3. The chosen model is developed and tweaked, and then tested to see how often it gets the right answer on the remaining 20% of known data.
4. If the model is good, it can be deployed on new data and brought into your operation.

What happened for us, we did step one and extracted extensive data. We applied step 2 and had a team of our developers and an outside expert attempt to find a suitable model.  The models chosen by our expert panel were then compared to the remaining training data in step 3.

The result. We concluded that, against all our expectations, it was not possible to predict the speed of production. Or more accurately, the only significant factor was the name of the operator on duty at the time.

And here is a key thing: Machine Learning is unlikely to throw up huge surprises in an already well-managed situation.

I spoke to the operations team in the factory about our conclusions. They were not surprised at the conclusion that the operator was the main statistically significant factor. I asked them what made some operators ’better’ than others. The answer ’attention to detail in the production setup’, itself a lesson I am trying to apply.

For manufacturing, what can we learn?

If you have good, well-organised shop floor data (as in this case), you can relatively cheaply come to a conclusion. We took 2 days to prepare the data, and then gathered the technical war room shown (which included several trainees as well as experts) for less than a day to come to the conclusion.

If you don’t have well-organised data on production, well, this is probably the thing to do first, but this is really a business intelligence operation; you can later build your machine learning on top of this.

The number crunching of machine learning can check, validate, and show insight on what is happening, but the deep expertise of the operational staff must be included in the discussion.

It is worth saying what would have happened if the planned production speeds were predictable; this would be a mix of efficient scheduling and potential target setting, but this would still remain a matter of discussion between the operators and the operations managers and directors. Numbers don’t change things; they provide insight and a tool for change management.

The answer machine learning gives also has to make sense to the experts; the actual maths, although very computationally doable, is not transparent to even mathematically literate people on the shop floor, so it needs careful handling. The team at Village are experts in Business Intelligence across various industries. Get in touch to discuss your BI needs for your business today, this includes machine learning.

The team at Village are experts in Business Intelligence across various industries. Get in touch to discuss your BI needs for your business today, this includes machine learning.