Intelligent IoT Sensing System Based on Compressive Sensing with Adaptively Learned Dictionary

We propose a condition monitoring IoT system using compressive sensing with a learned dictionary that is adaptively changed, at the edge terminal, to track the target state. The learned dictionary is meant to be fitted to the features of the training signals; thus, its reconstruction performance degrades when the measurements acquired include features that are not embedded in the dictionary. This characteristic is exploited in our system to directly detect a change of state, even in the case in which the new state is an unknown one. We apply this technique to condition monitoring of rotating machinery and we evaluate the dictionary reconstruction performance for different working conditions. Then, we apply the technique to the online detection of fault states. The evaluation reveals that this method can be effective to reduce the burden on the data transmitted to the cloud and it can overcome the problems faced for feature extraction when the information on the target is unknown.