Leveraging Machine Learning and AI for Predictive Maintenance: Insights from a Senior Software Engineer & Data Scientist

Hi, André. You work as Senior Software Engineer & Data Scientist at Brunel Car Synergies in Hildesheim. Could you tell us a little bit about what exactly Brunel Car Synergies does? 

Certainly! Brunel Car Synergies (BCS) is a subsidiary of Brunel GmbH that serves as a testing and development centre for high-tech clients. With two locations in Hildesheim and Dortmund, BCS caters to national and international clients. At the Hildesheim location, where I work, a team of 80 professionals specialise in researching and developing embedded systems – products that integrate hardware and software – for companies in the agricultural technology, railway technology, medical and renewable energy sectors, providing tailored solutions to meet their specific needs.

After over fifteen years’ experience working with machine learning methods, what is your main focus at Brunel?

My team and I focus on exploring how artificial intelligence can be optimally integrated and utilised across various industries. One common characteristic of AI models is that they often assign probabilities to predictions and decisions, rather than providing definite answers. This is due to the inherent uncertainty of probabilistic models used by AI algorithms. However, relying on probabilities presents a challenge when working with industrial machines, where smooth operation and reliability are critical factors. Ensuring seamless integration between AI and industrial machines thus remains a key concern that we strive to address in our work.

Can you give us an example of how you currently use AI in an industrial setting? 

In areas where the use of AI does not conflict with the core requirements of machines, it can provide data-driven benefits. For example, proactive maintenance involves performing regular maintenance tasks on equipment before any signs of failure are detected, and this is an area where AI can be particularly useful. Most industrial machines have many sensors, which is how AI can collect the necessary data. In the first step, AI continuously measures key indicators such as oil levels in bearings. The data collected can cover all conceivable aspects, including resonance vibrations, electric currents, temperatures, pressure levels and actual failures. In the second step, AI continuously compares this data with learned limits and combinations where something broke, for example. We collect as much data as possible, interpret it and then accurately estimate the remaining service life of machines and individual components.

What advantages can be gained by estimating the remaining service life of machines?

Predictive maintenance offers the advantage of recognising the need for maintenance interventions before they become critical. This is especially beneficial for machines like offshore wind turbines that are exposed to harsh environments. For instance, a sealing ring that can no longer withstand the salty air at sea can be detected early through unexpected changes in the gearbox's frictional resistance. Using predictive maintenance, it becomes possible to estimate when the damaged gearbox will reach a critical range and potentially fail. Companies can then schedule maintenance at an ideal time and bundle maintenance appointments. As a result, technicians can carry out all the necessary maintenance appointments during a single trip instead of making multiple trips for different systems, significantly reducing maintenance and personnel costs per system. That said, it's important to note that this is not about saving personnel, but about helping technical experts to provide better service and maintain a larger pool of equipment.

That sounds like a very efficient way of saving time and money. However, I’m sure that AI is not perfect yet. As the technology continues to evolve, what do you see as one of the next big challenges in AI research? 

AI is indeed still in its infancy. As I mentioned earlier, data collection is essential to enable the algorithm to work. But this in turn requires data that accurately defines the limits of a component or machine – which essentially means breaking it. To do that without actually having to destroy expensive machines, we use a different approach. For example, I can drive the machine to certain limits through the AI and then tell it: “Stop, the component will soon break." This creates a multidimensional data sphere that consists of green and red data points representing good and bad states, respectively. To accurately estimate when something is breaking, it's important to create trends from time series data. Ideally, the data set will be continually expanded by the machines in the field of operation. However, machine learning technology is still highly application-specific and has its limitations. While our AI models are extremely helpful, they lack the ability to transfer their learning from one machine to another, even if it's the same task on a different machine. Changing the type of sensor data means that the learning process must start anew. So, this transfer learning is what I see as the next big challenge in AI research.