Machine Learning in Manufacturing and Exploration

When we say the words Machine Learning and Manufacturing, probably the majority of the people will imagine futuristic robots working in closed environment to build silicon chips, cars or some other gadget we use. While that is part of the picture, in many ways it may not be the most important one. Not that industrial robots are not important – they have clearly revolutionized a number of fields, so much so that companies like Fanuc make their own robots in a lights-out factory now (no people, no lights, no heating). I say now, but this has been happening for almost 20 years, before the AI revolution hit mainstream technology. So what is the trick here, is it that they knew way in advance about the advantages of AI and machine learning and employed them such a long time ago? Well, probably not. A lot of industrial manufacturing works on predefined and well-defined tasks. A robot knows exactly when and where to make a weld for example. That does not mean there are no sensors involved or some algorithmic pre-defined decision. However if the task is well defined and all design steps are clear right from the beginning, you may argue that that task does not necessarily need machine learning involved. It requires deterministic step-by-step approach. So while industrial robotics has changed the industrial and manufacturing landscape dramatically (e.g. replacing production practices, replacing humans, improving quality and speed of operations), they do not have the ML capacity to teach the system to do something reactively (or proactively) to an environment where dynamic adaptation is required.

Today’s manufacturing unit performs effectively when people and machines work in collaboration. While machines typically carry out the actual operations, humans typically make decisions and manage the process based on the information available. These decisions often include complex judgemental, expert-oriented technical tasks such as what amount of chemicals to put in the flotation process to maximize performance of copper flotation factories, or what amount of that specific fuel to use under the specific conditions in the cement kiln so that better quality clinker is produced, or which part of the quarry to explore next, considering the material required and the observed chemical and physical characteristics of the fields. In all such cases expert opinions are often employed with sometimes mixed results.

The real question in manufacturing and exploration today is if we are ready to use machine learning to augment our human understanding and improve the processes beyond the expert-based judgements. A lot of trained experts may say their work is very sophisticated, combining a lot of experience, critical thinking, even gut-feeling, and that is likely true. However maybe they underestimate the advances in AI as well. I would suggest the future belongs more to either completely computerized systems or systems operating with a human-computer interaction to make the proper decisions and undertake the right actions. In some cases this can be achieved directly with PLC modules (last year Siemens launched a neural processing unit for their S7 line for example), but probably more often you would need much more powerful hardware to analyse vast amount of data and speed the decision process (e.g. for finding of new deposits, if you are in the mining industry).

While some of the machine learning research does require substantial investments and would likely be done by specialized large-scale equipment manufacturers, other ideas can be tested and executed on a much smaller scale. Even without engaging large investments or replacing entire lines of your equipment, performance can be still improved dramatically. Here are some examples:

  • Predictive maintenance. One of the areas where a lot of work has been done in the last years is how to do predictive maintenance. Typical maintenance involves doing regular checks at specific machine hours, mileage, etc., and changing oil and spare parts as prescribed by the manufacturer. However, often equipment works in very different environments. Some equipment may be used with highly abrasive rocks, increasing wear on tires, crushers, etc. Other equipment may be used in very dusty environments, potentially causing much faster issues with filters. And sometimes breakages happen with elements outside of the standard maintenance and spare part replacement programs. Can we try and improve repairs in these cases, aiming at lower downtime or repair/maintenance cost? To answer such questions, a field of machine learning called Predictive Maintenance is available. It focuses on using various statistical and ML techniques to predict and prescribe actions to improve operations (e.g. specific parts need replacement). There are many ML techniques that can be used in the process, but generally the method needs as much training data as possible, so if you have good amount of history for your equipment, predictive maintenance is a possible way to improve your equipment availability or lower costs.
  • Quality checks. Often quality checks are performed by trained human professionals. However machine learning models can be widely employed to improve quality management. That includes reading sensor information, analysing it based on predefined models and making decisions about potential quality control issues. By using machine learning techniques visual information for example such as product scans can be analysed to detect cracks and other defects. Training algorithms to perform such tasks decreases personnel costs and simplifies operations. We would also like to point out that machine learning techniques could be used simply for analysing large amounts of quality-related data to make certain conclusions about processes and quality practices. ML allows for complex analysis on the relationships between factors, including human factors, and quality performance.
  • The last and quite important element of machine learning that we would like to point out in regards to the topic is the opportunity to use vast amounts of PLC data to manage the equipment actions and setting. While this could sound broad, the point is that whenever large amounts of sensor data is gathered and we have a parameter to target (product quality, kiln temperature, installation output, etc.), machine learning algorithms can be employed to do so. ML allows analysing the target variable and assessing if the available data can be modelled to achieve that target. If such a model has good performance, substantial savings and improvements can be achieved. The aim here is to allow for humans to be at the helm and control the process direction, but the system to make operational decisions. Such algorithms lower work stress, can take over tedious work (e.g. night shift monitoring) and allow humans to focus on what needs to be done, rather than how specifically it should be done.