Avoiding Pitfalls in Industrial ML Projects: 3 Key Mistakes to Skip

The use of artificial intelligence techniques, and more specifically machine learning, has been increasingly seen as a revolutionary tool. But what is machine learning?

An interesting description was proposed by Francois Chollet. In his book, "Deep Learning with Python," he defines machine learning (ML) as a new computational paradigm. In conventional computing, we provide the computer the rules and the data, and we expect correct results. Using machine learning, this order is changed. We provide the data and the results to the computer, and expect the rules as an answer.

This new computational paradigm significantly changes our way of solving everyday problems and opens a big a range of opportunities in all fields of study. In the past few years, machine learning has been used widely, including in the industrial sector. Despite this wide scope and having many great professionals working in this field, some common mistakes have been observed (and should be avoided) in industrial project development. Although there are many others, this article aims to discuss three of these possible mistakes.

1. Forgetting the Basics

Concepts like Industry 4.0, digitalization, artificial intelligence, machine learning, and IIoT are the top trends in the industrial world nowadays. Projects with this kind of approach frequently get special attention in a professional's portfolio, and some professionals tend to choose this kind of solution during a project design to improve the chances of getting their project approved. But the point is: is this tool (machine learning or any other) the right one to solve your problem? You need to ask yourself some questions before choosing any of these "modern" tools. I have listed a few here, although there are many others:

• Have I tried a simpler algorithm to solve this problem (like PLC logic)?
• Does a well-syntonized regulatory control lead to similar or better control results?
• Do we have an appropriate, well-trained team to keep these new applications working?
• Would a different set of quality tools (such as pareto charts or cause-and-effect diagrams) give enough data to get good insights?

Note that I am not saying that tools like machine learning can’t deliver great results, or that these tools are too complex. It’s important to keep in mind, however, that simple solutions frequently deliver good results and normally should be the first step for an Industry 4.0 journey.

2. Not Paying Attention to Data Quality

Machine learning tools for processing data are now widespread. Big companies like Google and many open source groups have developed great ML libraries. These libraries are available on the internet, some of them at low or no cost. When we look at the big picture, however, any machine learning project involves four major phases:

• Understanding and preparing data
• Processing this data
• Analyzing results
• Acting wisely based on data analysis

That said, when we look at the entire process of developing machine learning projects,  it is not uncommon for some professionals to go straight to the middle stages (processing data and analyzing results) without paying enough attention to the data collection and preparation.

The basis of any machine learning project is data. Like in cooking, where the starting point of any good dish is to use good ingredients, an essential factor when developing a good ML project is to obtain and use good data. Especially in industry, getting this data may be a challenging task. A non-exhaustive list of potential challenges is shown below:

• Lack of instrumentation (missing data)
• Less accurate data due to improper installation, maintenance, or configuration of instrumentation
• Wrong data labels due to misconceptions in analysis—supervised learning tasks

Even with these challenges, with a correct preparation of the data (sometimes with industrial adequation), it is possible to build a strong enough database to obtain great results.

So, take your time on this important task!

3. Ignoring Experts' Knowledge

This may be a controversial argument. Some professionals may say that if you want an expert opinion, you should opt for an expert system using fuzzy logic, for example. In an opposing point of view, when we investigate the great amount of knowledge that our companies possess, just ignoring it does not sound the best way to evolve our processes.

When we implement ML projects, the best way is to get a multidisciplinary team combining the ML developers (people who are knowledgeable about ML techniques—i.e., deep learning, natural language processing, ensemble methods, clustering, and so on) with the proper expert advisors (professionals who knows the process reality and the real problems to be solved). With this multidisciplinary team, good results are catalyzed. You will probably get more accurate solutions, with a higher likelihood that you will be able to implement them in the real world.

Going Further

Despite the great challenges and cautionary tales that exist, machine learning is increasingly showing itself as a powerful tool. ML, and the dozens of other tools aimed at modernizing and evolving the industrial world, are a trend and a natural (and necessary) evolutionary process. However—especially in critical processes in industrial, medical, and other fields—care must be taken. So, do not skip steps: keep your project as simple as possible, take good care of your data, and do not forget the experts. 

As we can see, a machine learning project is not a short run, but a long journey. Like any journey, this one is made from single steps, and the last step has the same importance as the first one.