Fault Diagnosis in a Permanent Magnet Synchronous Motor using Deep Learning

Publication date
2024-03-14
Document type
PhD thesis (dissertation)
Author
Quseiri Darbandeh, Pedram
Advisor
Referee
Granting institution
Helmut-Schmidt-Universität / Universität der Bundeswehr Hamburg
Exam date
2023-02-03
Organisational unit
DDC Class
620 Ingenieurwissenschaften
Keyword
Demagnetization
Variational autoencoder
Permanent magnet synchronous machine
Linear discriminant analysis
Principal component analysis
Abstract
Condition Monitoring (CM) of electrical equipment, especially electrical machines, has an important role in enhancing reliability and preventing disturbance in production centers such as power plants and factories. One of the important steps of CM's process is diagnosing faults at the proper time to prevent the defect from spreading to other components of the system. Nowadays, the Permanent Magnet Synchronous Machine (PMSM) which uses permanent magnets instead of winding in its structure is widely used among other types of electrical machines. This thesis proposes a new algorithm based on Variational Autoencoder (VAE) as an unsupervised deep learning method to detect different faults with different severities. In this research work, current, vibration and sound signals are collected from a 1 kW prototyped PMSM. This motor is simulated in Finite Element Analysis (FEA) software and compared with the prototyped one. Demagnetization, bearing, and eccentricity faults are selected to be implemented on a prototyped PMSM. The performance of VAE is evaluated for different operating conditions consisting of different loads and speeds. Also, the VAE is investigated for classifying data with different complexity. The feature extraction methods which are known as the pre-processing methods can improve the performance of VAE by reducing dimensionality and extracting important data features. Besides, these methods improve the network's training time by reducing input data size. In this thesis, Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are used as feature extraction methods. The performance of VAE-PCA and VAE-LDA is compared, and it is shown that VAE-PCA has a slightly better performance than VAE-LDA in most operating conditions, especially in the high-class number in which the complexity is higher compared to other cases. Also, this method is developed for analyzing different combinations of sensors. Results show that the combination of current and sound signals can be a successful way to diagnose faults with and without pre-processing methods. In addition, the performance of VAE is compared with other machine learning methods such as Naïve Bayes (NB) and K-Nearest Neighbourhood (KNN) with and without pre-processing methods from accuracy and training time aspects. It is demonstrated that the performance of VAE for classifying the low and high number of classes of data is proper from accuracy and training time aspects. However, the performance of machine learning methods in the high number of classes is not proper for training time.
Version
Not applicable (or unknown)
Access right on openHSU
Open access