Abstract:Aiming at the problem that full modification of sensor’s evidence has infirm pertinence and fails to distinguish conflict evidence, a fault diagnosis approach based on improved evidence theory is proposed. Similar evidence and conflict evidence are recognized by means of conflict evidence criterion. Similar evidence is reserved, while conflict evidence is modified in virtue of similitude. This method can reduce the uncertainty of evidence information of sensors and decrease the effect of conflict towards combination results. Then neural network builds up the mapping from characteristic space to evidence space, and the results are used to construct original evidence by way of information entropy. Finally, all the evidence is combined in improved D-S formula. The gear pump’s early fault test proves that new approach has higher diagnosis precision by contrast with neural network and other combination methods.