A Novel Approach for Unsupervised Sensor Error Classification

Abstract

Listen

Translate article

The precise detection and classification of sensor errors is crucial for the optimal operation of smart manufacturing systems. Sensor errors, which can manifest in various forms, such as drifts or constant values, significantly challenge data integrity. This work introduces a novel approach that combines unsupervised learning and system simulations to detect and classify relevant sensor errors. Our approach avoids extensive manual labeling and uses only error-free simulated time series data to create initial embeddings of sensor data. This establishes a baseline for normal sensor behavior. To facilitate a precise identification of error types through a classifier, simulated sensor errors are introduced to create context-aware embeddings. We further analyze and compare conventional to advanced embedding methods such as TS2Vec. We apply our approach to a synthetic dataset obtained from a vacuum handling system and accurately identify and classify relevant sensor error types. Our work contributes to advancing operational efficiency and reliability in automated manufacturing processes within the Industry 4.0 framework by improving the detection capabilities for sensor errors.