Research results from the US consulting firm Aberdeen Research show just how devastating downtime can be: one hour of machine downtime can cost companies up to 234,000 euros. Predictive maintenance, especially when it is based on artificial intelligence, is therefore increasingly finding its way into German production companies. However, there is still a lot of catching up to do when it comes to implementing AI and machine learning-controlled technologies, as a study conducted by Fraunhofer IAO in 2019 showed. For the survey "Artificial intelligence in business practice", 300 companies were questioned - it shows that 75% are already dealing with the topic of AI. However, the study also showed that only 16% of companies were actually using AI-based solutions in their operations. The fields of application were in the area of data and information extraction and the analyses and forecasts based on this (predictive analytics). Predictive maintenance (PdM) is used by 25 percent of the companies surveyed. one of the reasons why German industrial companies are lagging behind in terms of concrete AI implementation could be the fact that the world of data science is full of models and decision-makers from the production environment find it difficult to find their way around. They believe that they cannot directly see the concrete results and benefits of a predictive maintenance solution. They also associate implementation with high costs and a great deal of work, and are reluctant to call in outside expertise. It is important for operators to obtain sufficient information and find the right solution for their own operations. Although it is also possible to develop your own model and algorithms, it is not practical to implement them for various reasons.

Between theory and practice

In theory, a model and its implementation are the same - but not in practice. Academic papers on data science may include analyses that show how algorithms are superior to others by one or two percentage points, but in the real factory environment with millions of data and signals, it is difficult to recognize patterns at all.

However, this is only the first major obstacle that potential "do it yourself" (DIY) model developers have to overcome. Anyone who manages to develop a robust concept that works under real conditions immediately encounters the next big problem: useful models need to be implemented, not just developed.

On the one hand, implementation means running the models at the right scale. It also means providing a user interface that presents the results in a clear and user-friendly way and allows different groups to prioritize warnings or collect feedback, for example. If 20,000 robots are working in a plant, it is not trivial to provide interactive diagrams for all of them via a user interface. In fact, DIY developers usually find that they are actually developing their own apps. This ties up costs and resources.

For these reasons, it can be beneficial to look for a partner that can provide data science expertise and help with implementation. While companies may think that their own custom models can work better than generic algorithms, the difference is often marginal and the negatives of going it alone can far outweigh the positives.

For example, the models for PdM applications are comparable to custom, user-defined models, but work more efficiently. Ideally, machine learning algorithms are used here, which use data to accurately predict the Remaining Useful Life of assets - a technique known as prognostics.

When choosing a suitable PdM solution, care should be taken to ensure that the algorithms treat each machine as unique - even if they are the same make and model. The reason for this is that machines that start out the same will behave and wear differently over time as they are exposed to different loads in their immediate environment or due to the work they perform. Treating each asset as an individual with a unique "behavioral fingerprint" increases the accuracy of predictions and helps maximize uptime.

If an industrial company has already developed a model, an effective PdM solution can integrate it into the system via an API. Although the model itself is not integrated into the solution, the software can still use the results of the model as input. The advantage of generic algorithms from manufacturers is that they deal with the real world as it is - not as the ideal state would be. This means that the models are extremely robust, even in the noisiest data environments. If a user wants to implement prognostics and predictive maintenance in production, this approach is particularly important when collecting data from failures. In a relatively chaotic moment, meaningful information needs to be extracted from the 'noise' so that the system can identify a potential fault and trigger a warning before the machine breaks down.

Teamwork pays off

Even if the integration of external expertise is the most resource-efficient way to use models for condition-based monitoring and predictive maintenance: Users play an important role in realizing the full potential of generic data models. First of all, there is always a learning curve when deploying a generic model. For example, many solutions in the field initially take 14 days to deliver results - they create a "fingerprint" of each machine's unique behavior under normal operating conditions.

Thanks to the knowledge of interdisciplinary teams of condition monitoring specialists, mechanical engineers and technology experts, the system can be configured in advance so that data and events can be prioritized. This accelerates the initial learning process of the algorithms. In the longer term, a system of regular feedback allows the algorithms to build up a picture of events and trends that are important to the user. This is useful when providing generic models.

Potential through AI and machine learning

As part of the digital transformation, artificial intelligence and machine learning will play a major role in the future - especially in the area of predictive maintenance. There is still some catching up to do when it comes to implementing AI- and ML-based software. This is also due to the fact that those responsible are put off by the supposed complexity. External partners can help here, for example with well-founded, generic algorithms that continue to learn during operation, as well as with implementation. Support is also available in the areas of scaling, deployment, user-friendliness and security.

Some providers can reduce unplanned machine downtime by up to 50 percent, increase the productivity of maintenance personnel by up to 55 percent and increase the accuracy of downtime forecasting by 85 percent.

Peter Portner, Managing Director DACH at Senseye / ag