Evaluation of Machine Learning Techniques for Electro-Mechanical System Diagnosis M. Delgado, A. García, J. C. Urresty, J.-R. Riba, J. A. Ortega MCIA Center, Electronic Dept., Technical University of Catalonia Rbla. San Nebridi s/n – TR14 GAIA - 0822 Terrassa, Spain Tel.: +34 / (93) – 739 85 18 Fax: +34 / (93) – 739 89 72 E-Mail: miguel.delgado@upc.edu Acknowledgments This work was supported in part by the Spanish Ministry of Science and Technology under the TRA2010-21598-C02-01 Research Project. Keywords << Diagnostics, Maintenance, Actuator, Permanent magnet motor >> Abstract The application of intelligent algorithms, in electro-mechanical diagnosis systems, is increasing in order to reach high reliability and performance ratios in critical and complex scenarios. In this context, different multidimensional intelligent diagnosis systems, based on different machine learning techniques, are presented and evaluated in an electro-mechanical actuator diagnosis scheme. The used diagnosis methodology includes the acquisition of different physical magnitudes from the system, such as machine vibrations and stator currents, to enhance the monitoring capabilities. The features calculation process is based on statistical time and frequency domains features, as well as time- frequency fault indicators. A features reduction stage is, additionally, included to compress the descriptive fault information in a reduced feature set. After, different classification algorithms such as Support Vector Machines, Neural Network, k-Nearest Neighbors and Classification Trees are implemented. Classification ratios over inputs corresponding to previously learnt classes, and generalization capabilities with inputs corresponding to learnt classes slightly modified are evaluated in an experimental test bench to analyze the suitability of each algorithm for this kind of application. Introduction It has been increased the use of new electrical machines, as Permanent Magnet Synchronous Motors (PMSM), during the last years. Although it is difficult to replace the classical Induction Motors in the most of industrial applications, the use of PMSM is growing in critical sectors such as automotive or aeronautical [1]. High performance, light weight and small size, are important characteristics which make the PMSM a good option for electrical traction or drive actuation tasks. In this context, PMSMs are implemented in these critical applications with a mandatory monitoring system. For that reason it has been necessary during the last years, the analysis of electro-mechanical systems PMSM based and fault diagnosis systems to cover any kind of condition. A great deal of studies has been performed about PMSM behavior under different fault conditions [2]. However, in order to characterize the PMSM behavior, most of the works are focused on single fault detection, and even, under stationary speed and torque conditions. Due to the aforesaid critical applications, there is a demand of diagnosis methods able to detect different kinds of faults in an electro-mechanical actuator PMSM based. Moreover, it is necessary the PMSM analysis under a complete set of varying conditions, even if different faults appear at the same time combined in the system.