Structure-based connectionist network for fault diagnosis of helicopter gearboxes

A diagnostic method is introduced for helicopter gearboxes that uses the gearbox structure and characteristics of the 'features' of vibration to define the influences of faults on features. The structural influences in this method are defined based on the root mean square value of vibration obtained from a simplified lumped-mass model of the gearbox. Featural influences characterize the frequency-specific information of the vibration features which correspond to the type of gearbox faults the features represent. These influences are defined as fuzzy variables to account for the approximate nature of the simplified model of the gearbox. The fuzzy structural and featural influences are then incorporated as the weights of a connectionist network for diagnosis, so as to avoid supervised training of the network. Diagnosis in this Structure-Based Connectionist Network (SBCN) is performed by propagating the abnormal features through the weights of SBCN to obtain fault possibility values for the components in the gearbox.; In the proposed diagnostic method, vibration features obtained from raw vibration are first utilized by an unsupervised Fault Detection Network (FDN) for identifying the presence of faults. Fault diagnosis is then performed by SBCN only if the presence of a fault is prompted by FDN. Since SBCN uses abnormal vibration features as inputs, an unsupervised pattern classifier is designed for abnormality-scaling of features. The abnormality-scaled features are then propagated through the weights of SBCN for isolating faulty components.; The proposed diagnostic method is experimentally evaluated in application to two helicopter gearboxes: OH-58A and S-61. Experimental vibration data for the OH-58A gearbox were collected at the NASA Lewis Research Center, and vibration data from three S-61 gearboxes rejected in field operation were collected at Sikorsky Aircraft. The proposed method is evaluated in diagnosis of the OH-58A gearbox faults as well as isolating the faults within the three S-61 gearboxes. The diagnostic results indicate that the SBCN is able to correctly diagnose about 80% of the OH-58A gearbox faults and all the faults in S-61 gearboxes. In addition to evaluation of the structural influences based on diagnostic results, they are validated by comparing them with influences obtained from experimental RMS values as well as the weights of a neural network structurally similar to SBCN, but trained through supervised learning. Moreover a sensitivity analysis is performed to study the effect of variations in structural influences on diagnostic results. The structural influences developed in this method can also be utilized for assessing the importance of various gearbox accelerometers in diagnosis. Three indices are defined based on the structural influences to quantify various aspects of accelerometer significance and are evaluated using the data from the OH-58A gearbox.