ABSTRACT This manuscript presents an effective diagnosis algorithm for permanent magnet synchronous motors running with an array of faults of varying severity over a wide speed range. The fault diagnosis is based on a current signature analysis. The complete fault motor diagnosis system requires the extraction of data based on the proposed method, and a subsequent method for adding classifications. In this paper, we propose a feature extraction method using a stacked autoencoder and a classification method using a softmax layer. The results show that the proposed methods can effectively diagnose five different motor states, including two different demagnetization fault states and two bearing fault states. INTRODUCTION Several studies have employed time-domain analysis to diagnose faulty motors. In [1], a fast and accurate motor condition monitoring and early fault detection system was developed using one-dimensional convolutional neural networks. The experimental results obtained using real motor data demonstrated the effectiveness of the proposed method for real-time motor condition monitoring. In this study, we present a feature extraction method using a stacked autoencoder with a softmax layer to classify permanent magnet synchronous motor (PMSMs) into five states. The experimental results show that our proposed methods can effectively diagnose five different motor states running at speeds of 150–3,000 RPM. PMSM FAULT SIGNAL A. Normal PMSM We measured the current time-domain signal of a normal motor. The signal is composed of sine wave combinations, and the period of each sine wave is the period for which the driver operates the motor. Furthermore, by observing the changes in the size of each peak, we can see that the magnetic distribution is not perfect. Connecting all the peaks reveals that the peak connection is a sinusoidal waveform, and not a horizontal line. However, there is an acceptable tolerance for the amount of magnetization present. B. Demagnetization Faults in PMSM In our experiments, we utilized two types of demagnetization fault, each with a different degree of demagnetization. One demagnetization fault was created by replacing one of the permanent magnets in the rotor with an aluminum magnet of equal weight (aluminum has no magnetic properties). This method ensures that the balance of the motor dynamics is not affected, preventing eccentricity failures. A second demagnetization fault was created by reducing the magnetic flux of one of the permanent magnets by half. We refer to this as a semi-demagnetization fault for the remainder of this paper. By measuring the demagnetization fault, we find that the demagnetization phenomenon leads to a non-sinusoidal component in the current signal. Therefore, nearby waveforms are also affected. The current signal for the motor with a demagnetization fault is significantly different from the current signal generated by the normal motor. Using high- and low- frequency harmonics that are close to the main frequency, the frequency of the fault feature can be determined as in [2]:  ௠ = ௦ ሺͳ±  ⁄ ሻ, k = ͳ,ʹ,͵, …, (1) where  ௠ is the fault feature frequency,  ௦ is the electrical angular frequency, and  is the number of pole pairs in the rotor. By observing the motor with a semi-demagnetization fault, we determined that the peaks of this motor are relatively small. C. Bearing Fault in PMSM According to [3] and [4], the relationship between the bearing eccentricity and electrical angular frequency is defined by  ௡ = | ௦ ±  ,௢ |, (2) where  ,௢ is related to the specification of the bearing by  ,௢ = ௡ 2  ௥ [ͳ ±  ௣ cos ], (3) where m is a positive integer, n is the number of balls in the bearing,  ,௢ is the feature vibration frequency,  ௥ is the mechanical angular frequency, db is the diameter of the ball, and pd is the diameter of the raceway. In the experiments, motors with two types of bearing faults were examined. The faults were located at different parts of the bearing. One of the bearing faults was designed to simulate a bearing inner ring that has been damaged by metal fatigue in the form of raceway cracks. The processing method used a drill on the inner ring to create four small holes. These holes are symmetrical, in order to maintain the rotational balance. The second bearing fault was designed to simulate a bearing raceway produced in a factory environment, where it was subjected to a large amount of fine iron powder. The processing method involves removing the bearing and adding soft aluminum powder to the raceway. The motor was subjected to higher friction during rotation, but no sound or vibration can be directly observed. When observing the motor with the first type of bearing fault (damaged by drilling the inner ring), as the ball struck the borehole, the disturbance current caused the waveform to turn slightly at the crest. However, it is not easy to observe the difference between this type of motor and the normal motor. When observing the other type of bearing fault (bearing raceway filled with aluminum powder), the increase in rotor friction caused the amplitude of the stator current to slightly Implementation of Permanent Magnet Synchronous Motor Fault Diagnosis by a Stacked Autoencoder I-His Kao, Wei-Jen Wang, I-Chieh Chiang, and Jau-Woei Perng, Member, IEEE