Machine Teaching: "AI as a means to boost collective intelligence"
Machine Learning, AI, Lean 4.0 and other concepts promise to enhance industrial performance thanks to big data. But what about human know-how? What about the instinct of the experienced field technician? Founder of InUse, Laurent Couillard, describes his ethos “Making collective intelligence digital”
You have said that you prefer machine teaching to machine learning. What is the difference?
At InUse, we consider that the machine must be taught before it learns by itself and especially before it performs maintenance duties. And this is where AI comes in to complement machine teaching in specific areas. The distinction of these two complementary approaches seems important to us. Machine learning requires large amounts of data (that are not always available) that can feature the largest amount of maximum combinations. Fortunately, there are very few machine breakdowns in industry. Typically, we are prone to preventive maintenance overkill which prevents algorithms from learning to recognise a malfunction in its initial stages. This is due to the fact that there are too few events recorded in the historical data for mathematical models to correlate. What Machine teaching is really about is the human knowledge that enriches these statistical models so that the technical know-how is complete. Basically, this gives us a two-tier structure. On the one hand what we already know, machine teaching, and then on the other hand what we will learn thanks to the data (machine learning). Generally speaking, 80% of the expected value is reached with Machine Teaching alone.
“Manufacturers, operators, maintenance technicians…each player has its own specific insight related to how the machines behave”
Where does the data used to teach the machines originate from?
The initial data actually comes from the skills and experiences of all the players involved with the machine from the manufacturers, operators, to the maintenance technicians and others. Each player has its own specific insight related to how the machines behave, ranging from how the machines behave at certain times to a situational instinct like when preliminary symptoms occur before the machine breaks down and these are detected. These individual data are already valuable, but they take on a new cooperative dimension when they are crossed with those of the other actors. A collective intelligence is thus truly built and its accuracy is as good as any mathematical model! When several operators of the same type of machine on different sites propose similar and / or consistent rules, the information becomes very reliable.
How can this vast industrial memory be digitalised?
Simply by asking questions. We like to use what we call the “5 Whys” rule. This approach is not only the simplest, it is the most efficient in operational terms because when an issue occurs, it is mostly described by the operator using words like “it vibrates …” or “it stops and starts”. Then we go back to the root cause of the problem and the objective data from sensors can model these descriptions and sequence of events from the data and thus create patterns. It is then important to restore this information operationally in written or verbal format directly on the shopfloor.
“It is important that the human remains at the centre of the system and that he keeps control over the interactions with his machines”
Isn’t the information system taking over from human expertise? How do you ensure that the model does not become inferior over time?
In fact, the opposite is happening. It is important that the human remains at the centre of the system and that he keeps control over the interactions with his machines, it is not up to software to act in his place. Certainly, professions evolve a little, but in the right direction, so operators gain autonomy because they can perform a level I maintenance alone, maintenance technicians have the experience from other sites and managers reduce their maintenance costs and improve availability. Finally, continuous improvement initiatives are intrinsically embedded into the system since each action on a machine is listed and then its consequences analysed. Basically, everyone improves their performance by becoming an “augmented worker”.