Predictive maintenance: Intelligent maintenance for maximum efficiency
How can machine downtimes due to breakdowns be reduced? A simple solution is regular maintenance and the precautionary replacement of components, for example after a visual inspection. This procedure is known as preventive maintenance. It can significantly reduce downtime, but there is a risk that maintenance will be carried out too early and components will be replaced unnecessarily early.
Preventive vs. Predictive Maintenance
Instead of relying on preventive maintenance, predictive maintenance can be a better alternative. Here, machine and process data is analyzed in order to predict impending failures. This means that maintenance can be carried out as required when problems with a machine become apparent. Maintenance can still be planned so that personnel and spare parts can be provided in good time.
Remaining Useful Life
But how does predictive maintenance work? The remaining useful life of a machine before it needs to be serviced or replaced is referred to as “Remaining Useful Life” (RUL). There are two methods for calculating the RUL: model-based and data-driven approaches.
- Model-driven method
The model-driven method uses mathematical models based on current conditions. This requires extensive specialist knowledge, and complex physical relationships often have to be simplified in order to make the model usable. - Data-driven method
In contrast, the data-driven method does not require explicit physical models, but works with actual machine data. This data includes measured values such as vibrations, flow rates or noise levels of the machine, but also environmental influences such as temperature and humidity. This data is collected, digitized and analysed. How exactly the RUL is then calculated depends on the available data: How long has it taken for similar machines to break down (Lifetime Data)? What does their history look like (run-to-failure data)? What are the threshold values at which a failure is detected (Threshold Data)?
Automation
Processing large volumes of data requires in-depth knowledge of data science and machine learning. Microsoft Azure already offers AutoML solutions, and an AutoRUL pipeline has been developed specifically for RUL, which makes it possible to take advantage of predictive maintenance even without in-depth knowledge of data science.
Another challenge is data gaps, i.e. incorrect or missing measured values in the collected time series. Time series refers to the chronological sequence of data. A simple approach is to calculate a mean value from neighboring data points. Large language models (LLMs) are also used in current research. These models, which are normally used for text completion, can also be fed with numbers. The model then provides possible values that fit the existing data well and fill in the gaps.
The accuracy of the results depends heavily on the language model used. Initial results show that a suitable language model often performs better than conventional methods such as section-by-section linear interpolation, even without specific training on the corresponding data set. Modern research can therefore significantly advance the application and implementation of predictive maintenance.
If predictive maintenance sounds interesting for your company, our experts will be happy to help you. Find out how this technology can be used in your specific case and how you can benefit from the advantages.
