Uncertainty-aware data pipeline of calibrated MEMS sensors used for machine learning

Abstract

Sensors are a key element of recent Industry 4.0 developments and currently further sophisticated functionality is embedded into them, leading to smart sensors. In a typical “Factory of the Future” (FoF) scenario, several smart sensors and different data acquisition units (DAQs) will be used to monitor the same process, e.g. the wear of a critical component, in this paper an electromechanical cylinder (EMC). If the use of machine learning (ML) applications is of interest, data of all sensors and DAQs need to be brought together in a consistent way. To enable quality information of the obtained ML results, decisions should also take the measurement uncertainty into account. This contribution shows an ML pipeline for time series data of calibrated Micro-Electro-Mechanical Systems (MEMS) sensors. Data from a lifetime test of an EMC from multiple DAQs is integrated by alignment, (different schemes of) interpolation and careful handling of data defects to feed an automated ML toolbox. In addition, uncertainty of the raw data is obtained from calibration information and is evaluated in all steps of the data processing pipeline. The results for the lifetime prognosis of the EMC are evaluated in the light of “fitness for purpose”.

Keywords

• Machine learning
• Dynamic measurement uncertainty
• Interpolation
• Time series
• Predictive maintenance
• Low cost sensor network