Fault diagnosis in rotors using adaptive neuro-fuzzy inference systems

Abstract

Condition monitoring of rotor dynamic systems is emerging research in recent years. The proposed research is a condition monitoring methodology based on Adaptive Neuro-Fuzzy Inference Systems (ANFIS) to detect the open cracks in the single disk rotor-bearing system. Condition monitoring systems generally requires a large amount of processed data for a specific output. The proposed methodology uses the same input data to train two different ANFIS without compromising on the accuracy of the results. The response of rotor-bearing system is generated by using finite element model analysis and harmonic balance method. All simulations are programed in MATLAB programing software. The effects of open groove or wedge cracks (notch crack) on natural frequency and resultant operational response (nodal deflections) of rotor-bearing systems are analyzed. Response orbit at 3× resonance of first natural frequency is analyzed to diagnose the crack in rotor shaft. Resultant operational response (Absolute response due to crack) is recorded for various crack locations and crack depths. Continuous Wavelet Transforms (CWT) are used, to extract the features from operational deflection shape (from operational response) to detect the crack location and their severity (depth). Location of maximum CWT coefficients provides the close vicinity of crack and their magnitude provides the severity of crack. Crack as small as 1% of crack depth to diameter ratio can be identified by CWT. ANFIS are used as a machine condition monitoring methodology to diagnose the crack and predict the crack parameters (crack location and depth). Two Parallel ANFIS are trained to predict the crack parameters. ANFIS-1 is trained for crack location and ANFIS-2 is trained for crack depth. CWT coefficients, maximum response amplitude at the vicinity of crack, and first three natural frequencies are provided as input to both ANFIS-1 and ANFIS-2 for training. The trained condition monitoring methodology accurately detects (predict) the crack location (ANFIS-1) and crack depth (ANFIS-2) with root mean squared error of 0.0833 and 0.137916 respectively.