Hybrid Multi-model Feature Fusion-Based Vibration Monitoring for Rotating Machine Fault Diagnosis.

Purpose: Every Deep Learning (DL) model generates features, which basically consist of particular properties. This leads to the problem of loss of important features as the algorithms applied in the pre-trained DL models are tested on particular data, which is of a different classification problem. This paper addresses the data imbalance problem and the loss of important feature problems in DL models applied to condition monitoring of Machines. Methods: The acquired vibration signal for different conditions are divided into segments using the Average Frequency Band Decomposition (AFBD) approach. The two- dimensional image data set is developed using the time frequency approach. The four pre-reliable deep neural networks used in this approach for feature extraction are Alex Net, ResNet-18, Inception-ResNet-v2, and NASNet Mobile. The extracted features from the following networks are then fused using the Serial Based Feature Fusion (SBFF) method. Before providing fused features for classification, important features are selected using the Univariate Feature Selection (UFS) technique. The hyper-parameter tuning of the classifiers is done using the Grid Search (GS) Optimization Technique. The data imbalance problem has been solved using the Affine Image Transform (AFT), which generates augmented images based on rotations. Results: For validation of the proposed methodology, Case Western Reserve University (CWRU) benchmark data set is used. The experimental results and analysis shows, the highest achieved efficacy in the proposed technique is 99.7%. Conclusion: The developed technique is able to provide a solution for unplanned failures of Induction Motors. It is also able to detect multiple faults such as bearing and rotor faults with a high efficacy. [ABSTRACT FROM AUTHOR]