Optimizing manufacturing with machine learning

The importance of manufacturing to modern society is hard to overstate. To put it in commercial terms, global manufacturing output was more than $35 trillion in 2017.

To make this value more tangible, let’s consider a small sub sector, namely disposable glove manufacturing. Sounds a bit eclectic, perhaps?

In fact, the manufacture of disposable gloves is a large global industry that produces in the region of 150 billion pairs of gloves per year, with a market value of over USD $5 billion. Thus, even a tiny improvement in the glove production process has the potential to create enormous added value.

By leveraging machine learning, we can do just that. 

The process

We’ll stick with glove manufacturing as an example to illustrate how you can optimize manufacturing with AI. Specifically latex gloves, which are one of the three main types of disposable gloves (see [1].)

First, here’s a simplified view of a production process for latex gloves:

And here is a nice “how it’s made” (don’t you just love those?) video for glove manufacturing.

Even this simplified process has nine steps (see [1] for a more detailed description). Each step has one or more controllable parameters: temperature, ingredient mix, speed/time, and so on. Just the source of the latex itself is complex, as it is produced by 20,000 species from over 40 families of trees.

The exact formulation of the latex drip has a particularly strong impact on glove properties – a point we will return to.

What makes a glove “good”?

Before we can optimize the glove manufacturing process, we must first decide what the goal is. 

Glove quality is not a single number. A latex glove has several important properties such as elasticity, strength, and cost. In addition some parameters can change throughput. For example, if we increase drying time fewer gloves will be produced per unit of time.  

So the first step is to design a scoring function. We might give 0-10 points for elasticity, going from very bad to very good on whichever elasticity scale we are using. We could do the same for the other parameters, such as strength. 

Then we could add these together to get a single quality metric. Perhaps we would weigh strength more than elasticity. This is easily achieved by simply doing a weighted summation of the scores. 

Make sure quality measurement is consistent

A key assumption is that the tools and methods used to measure quality don’t change. This can be a dangerous assumption. What if a new employee uses the strength testing tool in a slightly incorrect way which overestimates strength, relative to the more experienced test staff? 

I won’t go into detail about this issue here. Suffice it to say that regular reviews of quality test procedures and tools are extremely important for optimal manufacturing. 

Optimize the process

Now let’s focus on the important ingredient mix for the latex itself. Which ingredients should be added to the latex in which ratios so we obtain the best end product?

For example, zinc oxide and activated  zinc oxide may be used in the vulcanization process, the  general  range of usage being 0.5-2.5 phr (ratio to parts by weight of rubber). What is the optimal value?

The challenge here is an example of optimizing an unknown function – see the figure. At the outset, we know that there is some relation between the concentration of ZnO, but we don’t know what that relation looks like. 

Imagine we do an experiment with a middle value of zinc oxide. This produces gloves with medium quality (top figure). We have only a vague idea (gray shaded area) what higher and lower values would do to quality, but we know that we now need to explore either higher or lower values. In this case, our algorithm – taking uncertainty into account – suggests we try a high value. 

As the bottom figure shows this higher value resulted in low quality. After these two experiments, our algorithm intelligently suggests trying a low value. After a few iterations of this, we will find the best relative concentration of Zinc oxide. 

By going about the optimization in this principled way, using something like Bayesian optimization, we obtain the best possible quality with as few experiments as possible. This is crucial as each experiment is very expensive – it requires running a whole production batch and testing its quality.

This same approach can be used to optimize across many dimensions at the same time, so you can optimize all the manufacturing process parameters together for the best and efficient result. 

Optimization in practice

To optimize the process like this you must first monitor it, collecting reliable data for each parameter. Ideally, measure what really happened rather than what was prescribed. 

This may require adding sensors to production line machines and investing in data collection hardware and software. Even very old machines can often be retrofitted with sensors with a very good ROI, because the increase in quality creates profits that dwarf the data collection costs. 

If you would like to get started with leveraging machine learning to optimize your production processes, I’m here to help – schedule a free consult today. 


[1] Production method & market trend of rubber gloves,  T. Akabane, International Polymer Science and Technology, 43, No. 1, 2016.

Leave a Reply

Your email address will not be published.