Sensor Fault Recognition With Multioutput Regression Network

The detection of faulty sensor data is important to ensure the functional safety of electromechanical systems. This is especially relevant if additional data processing is done, which can disguise failures and lead to unrecognized measurement errors. Therefore, additional actions are required to detect such faults. On this account, this work combines two neural networks that process the data from industrial angle sensors. The first of these networks is used to correct a distorted measurement and estimate the measurement error. The results from the second network are then used to verify the results. This approach can be used to assess the quality of corrected measurements. Furthermore, faulty sensor data that indicate a hardware error can be recognize, which would otherwise be concealed by the network. This directly improves the acceptance of neural networks in a sensor context due to the improved functional safety.