%0 Journal Article %A Gowid, Samer S.A.A. %A Dixon, Roger %A Ghani, Saud %D 2017 %T Performance comparison between FFT-based segmentation, feature selection and fault identification algorithm and neural network for the condition monitoring of centrifugal equipment %U https://repository.lboro.ac.uk/articles/journal_contribution/Performance_comparison_between_FFT-based_segmentation_feature_selection_and_fault_identification_algorithm_and_neural_network_for_the_condition_monitoring_of_centrifugal_equipment/9546362 %K Condition based monitoring %K Condition monitoring %K Feature selection %K Neural network %K Centrifugal equipment %K Fault detection %K FFT %K Mechanical Engineering not elsewhere classified %X This paper compares and evaluates the performance of two major feature selection and fault identification methods utilized for the Condition Monitoring (CM) of centrifugal equipment, namely FFT-based Segmentation, Feature Selection, and Fault Identification (FS2FI) algorithm and Neural Network (NN). Mutli-Layer Perceptron is the most commonly used NN model for fault pattern recognition. Feature-selection and Trending play an important role in pattern recognition, and hence, affect the performance of CM systems. The technical and developmental challenges of both methods were investigated experimentally on a Paxton industrial centrifugal air blower system with a rotational speed of 15,650 RPM‟s. Five different machine conditions were experimentally emulated in the laboratory. A low training-to-testing ratio of 50% was utilized to evaluate the performance of both methods. In order to maximize fault identification accuracy and minimize computing time and cost, a near-optimal NN configuration was identified. The results showed that both techniques operated with a fault identification accuracy of 100%. However, the FS2FI algorithm showed a number of advantages over NN. These advantages include the ease of implementation and a reduction of cost and time in development and computing, as it processed the data from the first trial in less than 6.2% of the time taken by the neural network. %I Loughborough University