What do chatbots, self-driving cars, and predictive maintenance have in common? 

They each involve the application of artificial intelligence (AI). In recent years, the development of AI has become a hot topic. It seems that AI is being talked up at every conference, online event, and even in industry magazines.  

Vendors of equipment and software have also used the term to describe their offerings for some years now. Some might say perhaps a little too loosely. 

And why wouldn’t we be keen to adopt AI? 

The simulation of human thinking and the mimicking of actions in machines that can produce seemingly far more accurate results in a fraction of the time seems like an operational nirvana. There’s also the fact that they operate 24/7 without sleep, holidays or pay. 

But is it really everything that it is made out to be? Perhaps not yet. 

AI is often not much more than just a fancy algorithm. Sure, with more computing power it can handle more data in a shorter period and solve problems that would take a person much, much longer. But it nonetheless relies on the thinking and intelligence of the people who designed it. And that may be an algorithm developed decades ago. 

Does that mean it won’t be useful? Of course not. 

The connectivity created by the ‘internet of things’ (IoT) and Big Data enables a level of decision support analysis not imaginable early in my career. 

But before we get all excited about the extension of this to AI, we do need to recognize three main problems. 

First, how can we be sure about how to connect the dots?  

The successful use of AI requires an understanding of cause and effect before an event happens. A self-driving car needs to be programmed how to respond, based on the data collected from its sensors, just as people need to learn how to drive and respond to what they see and hear when in control of a vehicle.   

This means that what we call AI is limited by our existing understanding. It doesn’t (yet) have a conscious ability to learn. 

This might be fine for ensuring safety with the future application of self-driving cars. But with equipment maintenance, and by extension, spare parts decision-making, there are too many unknowns.  

The simulation of human thinking and the mimicking of actions in machines that can produce seemingly far more accurate results in a fraction of the time seems like an operational nirvana.

Therefore, AI won’t in the near term really improve the ability to predict parts failures. AI applications will be able to tell you to replace parts but won’t really be predicting failure. As most parts failures are not because of wear or age, there is a complexity here that we don’t yet understand. Without that understanding, we cannot program the algorithm. 

Second, with spare parts management, IoT, Big Data, and AI are not enough to ensure good outcomes because those outcomes rely on much more than data. The key links here are people and process.  

In practice, much spare parts decision-making can be described as emotional rather than rational. That is, people know they don’t know all that they need to know. So they decide to stock items or stock more of an item ‘just in case.’ They over-stock on spare parts because they want to avoid being the person who let the team down and caused excessive downtime. This happens even if the probability of the additional parts being required is vanishingly small to non-existent. 

Regarding process, the issue isn’t just is it defined, but is it followed? A simple example of this is when people override the stock level suggestions from their software. In my experience, this happens primarily because people don’t trust the results. 

Lastly, there is the old maxim of garbage in, garbage out. 

This expression was coined in the 1950s but is still applicable. With the vast amount of data being collected through the IoT it is perhaps even more relevant.  

In relation to spare parts management, consider the impact of the following: the direct procurement of parts that don’t go through the data collection required for AI; the hoarding of parts in workshops; requisitioning more than required; being slow to return those parts to the storeroom; poorly programmed operational data collection; not knowing what data to collect; how to weigh the inputs; and even the opportunistic replacement of a still-functioning machine part during a maintenance shutdown created for other work. These factors, and I am sure many more, lead to poor or inaccurate data. 

Ultimately, no matter how potentially useful a tool may be, the output can only be as accurate as the data entered into it. 

Through engineering and design development, equipment reliability continues to improve. This in theory reduces the demand for spare parts. Simultaneously, data connectivity and automation continue to improve the efficiency of the procurement and logistics associated with spare parts management. However, as the volume of stock held, and consequently the value of working capital tied up in spare parts, continues to rely on factors that are unreliable, we cannot reasonably expect AI to solve our spare parts management problems in the near term.