Predictive Maintenance In Tomorrow’s Industries

By Olaf Enge-Rosenblatt and Steven Brandt

Tomorrow’s production plants must be efficient and adaptive, which is the key to survival in modern, highly competitive markets brought about by digitalization and automation. Future-oriented companies are increasingly focusing on a tight combination of automation and computer technology as promised by the Industry 4.0 paradigm. More and more globally distributed technical systems facilitate exchange of data and remote analysis, e.g. in superordinate IT infrastructures like in a cloud. Local and cloud-based data analysis can be used to realize comprehensive predictive maintenance solutions in order to optimize maintenance intervals, maximize machine lifetime and reduce expensive machine downtime.

In smart factories, massive amounts of data are collected from the complete production cycle in order to control the production process and to monitor product quality. Additional data is often used for condition-monitoring systems yielding insights into behavior patterns of technical equipment in terms of abrasion and anomalies during operation. All this data may be useful for companies through meaningful analysis methods for predictive maintenance purposes. Because of the steadily growing amount of data, it may be necessary to incorporate big data aspects at a certain stage of complexity.

Well-prepared, successful predictive maintenance solutions not only require knowledge about mathematical statistics and classification methods but also about the complete production process. Only the right combination of these different kinds of expertise makes it possible to reliably predict machine failures in advance and prevent their occurrence through planned maintenance ahead of those predicted events.

A variety of approaches for predictive maintenance purposes are known, e.g. analyzing vibration and sound data, oil particles, temperatures and many more. But in principle, any kind of physical quantity showing a certain relationship to abrasion processes or anomalies and the like are potentially appropriate candidates for monitoring and analysis. Usually, the collected data is analyzed by complex machine learning algorithms that can learn data patterns and then predict future events related to abrasion and anomalies within production processes. Choosing suitable techniques and incorporating expert knowledge of both machine and process details as context information is key to gaining usable insights into the analyzed system and making reliable predictions of future events. Hence, cooperation of manufacturers and data analysis experts is highly effective in developing reasonable predictive maintenance solutions.

Before raw sensor data can be analyzed, it usually needs to be pre-processed. Commonly used methods range from error correction and noise reduction to transformation of the signal into another domain, e.g. the frequency domain. Once the signal is prepared, data scientists have to extract problem-related features from it. At this stage, working together with process experts and incorporating their special domain expertise is key to finding suitable features. These features can then be used with classification and regression machine learning algorithms in order to learn from and predict future machine behavior.

Successfully trained models are then applied in a variety of ways, ranging from small independent devices close to the observed technology to complex combinations of sensors, local pre-processing and cloud-based analysis of the whole production process with state-of-the-art artificial intelligence.

The use of artificial intelligence in sustainable product development requires new ways of thinking, fundamental integration of modern information technology in product experience and a leading CTO with deep understanding of state-of-the-art technologies like in traditional software companies. There is still a long way to go for today’s industry, but these days research provides both tools and know-how to pave the way towards a highly digitalized, connected and smart production world.

For more, check out https://www.eas.iis.fraunhofer.de/en/research_topics/industrial_data_analysis.html.

Steven Brandt is member of the Computational Analytics group at Fraunhofer EAS.