Evaluating the Effect of Intrinsic Sensor Noise for Vibration Diagnostic in the Compressed Domain Using Convolutional Neural Networks

Machine learning allows designing intelligent sensing networks capable to perform automatic inferences about the integrity of technical facilities. Compression techniques decrease significantly energy requirements of the sensing networks proving essential when sensing nodes are not supported by constant power sources. Existing schemes pass through the reconstruction of the original time series data before moving to the diagnosis phase. However, this passage can be avoided, i.e., inference can be performed directly in the compressed domain, by exploiting the specific information retained in the compressed patterns. This paper fulfills the goal above in the context of vibration-based structural health monitoring by proving, from an empirical perspective, that Convolutional Neural Networks (CNNs) can be used to predict the structural health status directly in the compressed domain when properly combined with adapted Compressed Sensing mechanisms. Importantly, the study analyses the effect of the intrinsic noise that affects digital accelerometer sensors. Results confirm that CNNs can mine information in the compressed domain even in presence of strong noise components, i.e., accuracy remains above 94% even for ultra-low-cost solutions featuring a signal-to-noise-ratio below 20 dB.