Towards more autonomous and intelligent industrial AC drives for Mechatronics

The modern industrial world calls for efficient, reliable and safe systems. A contribution to the solution to all these problems is the predictive maintenance. According to this trend and tailoring the analysis to the electric drives field, this thesis performs a step forward for the realization of more reliable drives through their condition monitoring. Different AC motors have been considered in the dissertation: Permanent Magnet Synchronous Motors, Induction Motors and Synchronous Reluctance Motors, covering the actual and also the next future industrial drives scenario. Some of the more relevant faults that can occur on these machines have been taken into account: interturn short circuits, demagnetization and damage at the rotor bars. The development of adequate observation indexes for the recognition of these failures has been researched deeply in the past. Nowadays, the prompt recognition of the incoming failure condition is the issue that modern research in this field has to face. In the following chapters, some innovative Artificial Intelligence-based tools will be applied for the condition monitoring of electric motors. Artificial Neural Networks and Convolutional Neural Networks are used here in different ways: for an effective modelling of the machine behaviour and for the knowledge-based recognition of the motor state of health. The main bottleneck in developing Neural Networks is the availability of a proper training dataset for the efficient tuning of their weights. In case of electric motors, the problem is even more relevant. A huge amount of healthy and damaged motors are needed, an unaffordable condition for this industry-oriented context. As an innovative and never-used-before approach, very precise models of the motors have been used in the thesis to generate artificially the training dataset. When these models were not available, the Data Augmentation theory was used instead as a keen and innovative approach for the artificial enhancing of the available training datasets. Very relevant results have been obtained and the principal and more significant ones are reported in this dissertation. Three different Convolutional Neural Network designs are reported. The first, trained only on a simulative accurate model of the motor, was able to efficiently recognize demagnetization and the interturn fault on Permanent Magnet Synchronous Motors. The second one, oriented to Induction Motors, the model was not available and so Data Augmentation was used to train the network that recognizes broken bars in the rotor. The third network was used again for Induction Motors but it made use of a model which definition was still based on a built-in Neural Network. Finally, this same Artificial Intelligence-based modelling methodology was used for the effective implementation of the Extended Kalman Filter for the sensorless control of Synchronous Reluctance Motors, thus further enhancing the reliability of the drive, since the position sensor is avoided. The motor non linearities have been managed through a custom Artificial Neural Network and new approaches to the original Extended Kalman Filter implementations have been studied. As common root of all the treated topics, Neural Networks applicability on the electric drives field has been investigated. In the development of these tools, a special and careful eye was taken to maintain the solution feasible and attractive from an industrial point of view. Therefore, each of the arguments was fully validated through an intensive simulation and experimental stages, as reported in the thesis.