Intelligent IoT sensing and diagnosis method for rotating machinery based on low-dimensional compressed measurements

Condition monitoring and predictive maintenance of industrial machinery require to continuously measure vibrations or wave data with a high sampling rate for accurate diagnosis. However, such measuring strategies cause high time and energy consumption for transferring and handling a large amount of data. For this reason, a compression sensing approach is investigated. Compression sensing is a technique that allows to completely reconstruct a signal from its samples, while sampling it below the Nyquist limit. It relies on the hypothesis of sparsity of the signal in some domain. In particular, the higher the sparsity level, the lower will be the dimension of the compressed measurements required to recover the original high-dimensional signal. The sparsifying basis (or dictionary) can be fixed and used for every signal (f.e. wavelet transform) or learned from a set of training signals and adapted to a specific class of signals only (Dictionary learning technique). The latter is the approach used in the thesis in order to improve the representation performance. In fact, this basis can decompose the signal very sparsely since it is specifically generated to represent the characteristics of the training signals. In particular, the test performed shows that this algorithm allows to the reconstruct the signal in presence of noise with low dimensional measurements and a simple sensing mechanics. On the other hand, the possibility to use this technique to detect a fault is investigated. The idea is that the basis specifically represents the features of the training class of signals. This means that only signals in the same state as the training once will be well represented, while signals in other states will not be decomposed sparsely on this dictionary. In the presence of a fault condition, the vibration signal generated will show different features compared to a signal generated in a normal condition. Using the dictionary trained during a normal working condition and evaluating the reconstruction error on the compressed measurement, it is possible to recognize a change in the state of the machine. Using this idea, an online fault detection can be achieved. Moreover, the possibility to use different dictionaries to discriminate between different fault conditions is investigated. In the case in which signals from different fault conditions are available, it is possible to train offline a basis for each condition and use the representation error on all the basis to recognize the type of fault. The first part of the thesis regards the simulation of the reconstruction algorithm and the evaluation of the reconstruction quality and the achievable dimensionality. In the second part, the fault diagnosis method is simulated and the achievable recognition rate is confirmed. Finally, an online testing in a real factory environment is performed to ensure the feasibility.