Blade fault diagnosis using artificial intelligence technique

Abstract

Blade fault diagnosis is conventionally based on interpretation of vibration spectrum and wavelet map. These methods are however found to be difficult and subjective as it requires visual interpretation of chart and wavelet color map. To overcome this problem, important features for blade fault diagnosis in a multi row of rotor blade system was selected to develop a novel blade fault diagnosis method based on artificial intelligence techniques to reduce subjective interpretation. Three artificial neural network models were developed to detect blade fault, classify the type of blade fault, and locate the blade fault location. An experimental study was conducted to simulate different types of blade faults involving blade rubbing, loss of blade part, and twisted blade. Vibration signals for all blade fault conditions were measured with a sampling rate of 5 kHz under steady-state conditions at a constant rotating speed. Continuous wavelet transform was used to analyse the vibration signals and its results were used subsequently for feature extraction. Statistical features were extracted from the continuous wavelet coefficients of the rotor operating frequency and its corresponding blade passing frequencies. The extracted statistical features were grouped into three different feature sets. In addition, two new feature sets were proposed: blade statistical curve area and blade statistical summation. The effectiveness of the five different feature sets for blade fault detection, classification, and localisation was investigated. Classification results showed that the statistical features extracted from the operating frequency to be more effective for blade fault detection, classification, and localisation than the statistical features from blade passing frequencies. Feature sets of blade statistical curve area was found to be more effective for blade fault classification, while feature sets of blade statistical summation were more effective for blade fault localisation. The application of feature selection using genetic algorithm showed good accuracy performance with fewer features achieved. The neural network developed for blade fault detection, classification, and localisation achieved accuracy of 100%, 98.15% and 83.47% respectively. With the developed blade fault diagnosis methods, manual interpretation solely dependent on knowledge and the experience of individuals can be reduced. The novel methods can therefore be used as an alternative method for blade fault diagnosis.