Comparison of support vector machine methods and neural network in the classification of air fan defects and empirical vibration analysis

Comparison of support vector machine methods and neural network in the classification of air fan defects and empirical vibration analysis


Comparison of support vector machine methods and neural network in the classification of air fan defects and empirical vibration analysis

نوع: Type: thesis

مقطع: Segment: masters

عنوان: Title: Comparison of support vector machine methods and neural network in the classification of air fan defects and empirical vibration analysis

ارائه دهنده: Provider: Sajad Khalili Rad

اساتید راهنما: Supervisors: Mehdi Karimi (Ph.D)

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees:

زمان و تاریخ ارائه: Time and date of presentation: 2020-9-28

مکان ارائه: Place of presentation: Salon amfi

چکیده: Abstract: In this dissertation, the condition of the cooling fan system in unit 109A of phase 13 of Rajab South Pars has been investigated by vibration analysis method. The proposed methods include 7 steps: data collection, signal processing, feature calculation, feature extraction, feature selection, classification using support vector machine and classification using neural network. First, using the accelerometer sensor, vibration signals related to healthy states, unbalance, bearing failure, belt and pulley misalignment, bearing failure and unbalance combination, bearing failure combination and belt and pulley misalignment were removed and Stored in SPECTRPRO software for data analysis. Divide the time signal taken from each state into 10 subsets, so that from each state 10 to FFT is obtained. Now from 12 statistical properties (including: root mean square, mean, geometric mean, harmonic mean, standard deviation, skewness Corticosis, peak, sum, peak coefficient, shape coefficient, impact coefficient) in time dimension and 12 statistical properties in frequency dimension and 2 entropy properties (permutation entropy, approximate entropy) in time dimension were used to extract the feature. A total of 26 features obtained using statistical methods were limited to 6 features by the ICA dimension reduction method and 6 features by the LDA dimension reduction method. And 2 of the best LDA features have been used to form the smart matrix input matrix. Finally, a 4x60 matrix was obtained, of which 60% of this data was used for training and the other 40% for testing calibration algorithms, and the performance accuracy of both algorithms became very close. In the MLP neural network performance test, the performance of the backup vector machine with radial base kernel function was 95.84. The proposed intelligent algorithm for the laboratory data of Case Western Reserve University of America was also examined for comparison with industrial results. For this purpose, each time signal related to normal states, defective external cannons, defective internal cans and defective bullets was divided into 50 parts and 50 FFTs were taken and expressed using 26 statistical features and application of ICA and LDA as input to intelligent algorithms, neural network testing accuracy and support vector machine was 100%, which indicates the acceptability of the method used. According to the results of the studied algorithms, the combination of LDA and ICA with backup vector machine is proposed as a reliable alternative in future intelligent troubleshooting. Key Words: Vibration analysis, cooling fan, frequency and time domain signal analysis, ICA feature extraction, LDA feature extraction, neural network, support vector machine

فایل: ّFile: