Case Study: The challenges of sensor based data analytics to support predictive maintenance

Frederic Kihm, Product Manager, HBM Prenscia


Digital transformation and the Industrial Internet of Things (IIot) can provide insights to the way products are actually used in service. IIoT uses sensors to gather information for the understanding of usage, maintenance and operations. IIot can be used in various domains, from manufacturing to energy production, from automotive to aerospace industries.

Reliability issues in operation are often related to the way the equipment is actually used. A Health and Usage Monitoring System (HUMS) monitors the health state of the asset, it also monitors the usage of the asset and targets the improvement of its availability, reliability and performance. The detection of early signs of failure or loss of performance helps to anticipate and schedule the maintenance down times and associated tasks such as the provision of spare parts, etc.
This presentation will describe the practical use of both descriptive and predictive analytics in an industrial case study, where an automotive component is monitored. This case study required the collection of usage data off the vehicle such as digital bus data, vibration, temperature, strain, etc. It is however not trivial to turn this sensor data into actionable knowledge about the health of the equipment. Similarly, predictive maintenance based on the processing of sensor data is challenging. Some of the challenges encountered in the case study will be discussed in this presentation. Challenges include the need for pre-processing the data, the adoption of more or less sophisticated data reduction techniques, dealing with false positives, the need for high performance data processing, the limits of using purely statistics for the prediction, etc. This case study will also illustrate the importance of capturing the scatter in the applied loads not only for scheduling the maintenance accordingly but also for enabling design improvements or for generating more realistic qualification tests.


  1.  Sensor data require a post-acquisition preprocessing pass, including cleaning, smoothing, resampling, etc.
  2. To limit the number of false positives, apply changes to the criterion and do retrospective analyses on the historical data
  3.  Physics of failure can dramatically enhance predictions



Frédéric is responsible for the signal processing related software products including GlyphWorks and VibeSys. Frédéric is managing the various aspects of product release including understanding customer requirements, technical specification, documentation and software quality. Frédéric previously worked as an Engineering Consultant for HBM Prenscia, where he provided training and engineering services activities, supporting FEA, signal processing, fatigue/durability topics and vibration analysis in the automotive, aerospace, and defense industries. Frédéric holds a MS in Mechanical Engineering from IFMA University in France and a PhD with the Institute of Sound & Vibration Research (ISVR) in Southampton.