The Ongoing AI Revolution: Part 1: Predicting Machine Failure with Machine Learning

Machine learning has made significant progress in recent years, with the emergence of ChatGPT and other generative AI showcasing the potential of techniques that have been in development for decades. While real-time facial recognition and orientation were once theoretical concepts, they are now commonplace. These advancements have had a significant impact on various industries, mainly internet companies. But some traditional industries, such as heavy industry, have been improved, for example where AI has been inspecting the end of assembly lines for defects. However, one area that has yet to realise the benefits of the ongoing AI revolution is that of maintenance and reliability.

An ML Algorithm materialising the answer we want from the data….

In this blog post, we will explore how current machine-learning techniques can be applied to maintenance and reliability problems to predict the lifespan of machines. Using real-time sensor data, imagery, and configuration information, predicting when a machine will fail is an enormous challenge in the field. At present, there are two main ways that existing machine learning methods can be used to make these predictions: to predict the time to failure or to estimate the probability of failure within a given timeframe.

Let’s consider an electric motor that drives conveyor belts in a fulfilment centre as an example. In this case, the motor is vital for the continued operation of the centre, and the operator wants to know when the motor will fail. Using the first approach, a conventional approach would be to use a regression algorithm to predict precisely how many hours of operation are left until the machine fails. Predicting a single number might be useful for ordering parts and preparing for the failure in a rough way, but it does not reveal anything about the risk of failure from now until that predicted time.

To overcome the problem of not knowing the risk of failure up to the predicted time, the second suggested approach could be used. That is, we can estimate the probability of failure directly. In this type of model, the machine learning algorithm outputs the probability of an event occurring in some period. This prediction has the advantage of localising the failure with a measure that can be used. However, the window-based probability model cannot tell us where in the window the failure is likely to occur. Also, if we use a window that is too wide, say from now until 1,000 hours, it tells us almost nothing. Conversely, if the window is too narrow it is not very useful for long-term planning.

Machine learning models can therefore be supplemented with traditional reliability analysis. Reliability analysis can provide insights into the risk of failure which can support decisions about maintenance plans, tasking priorities, and spares analysis. It is to these methods that we will turn to next…

Make sure you keep an eye out for part 2 of our blog series which will be covering survival analysis.