Parameter optimization of rotor-bearing model using artificial neural network and improved cuckoo search algorithm

This research employs machine learning (ML) techniques to establish the crack parameters in any rotor-bearing system. The parameters are trained using artificial neural network (ANN) models and optimal crack parameters are identified by an improved cuckoo search (CS) algorithm. The response amplitude is reduced for the optimal crack parameters in comparison with reference parameters. To facilitate this process, a finite element (FE) rotor model containing two discs supported on two bearings is developed with/without crack. The parameters associated with cracks are determined through a complete analysis of the natural frequencies, time, and frequency responses at different operating speeds. The first two critical speeds were 2200 rpm and 2260 rpm for healthy model. For the cracked model, reduction in critical speeds like 1800 rpm and 2100 rpm were observed. The time spectrum shows a periodic pattern in both models. However, the dominant peak frequency of the healthy model at a rotor speed of 5000 rpm is obtained at 108.03 Hz and is reduced to 57.98 Hz after the crack is introduced in the model. Using the design of experiments (DOE), data sets with corresponding frequencies and amplitudes are obtained. An excellent input-output relationship is established after considering the crack parameters as input variables using back propagation neural network (BPNN), counter propagation neural network (CPNN), and radial basis function neural network (RBFNN) schemes. The results show that CPNN achieves faster convergence of mean squared error (MSE) with an error rate of less than 3% compared to other models. An improved CS algorithm combined with trained CPNN is used to train the optimal crack parameters. The corresponding frequency responses with the optimized parameters are also presented.