Fault diagnosis in a hydraulic directional valve using a two-stage multi-sensor information fusion

As it is usually operating in bad working conditions and subjected to the severe interference from diverse paths, internal faults of the hydraulic valve are difficult to be detected using conventional hydraulic testing technology (such as relying on pressure sensors or flow sensors). Moreover, the information collected from a single sensor may not provide accurate diagnostic evidence or a complete description on faults of hydraulic valves, even though employing intelligent fault diagnosis methods. Therefore, a two-stage multi-sensor information fusion method is proposed, including the fault feature fusion and the decision-making information fusion. The aim is to realize the diagnosis on internal faults of hydraulic directional valves using the vibration signal analysis method instead of conventional hydraulic testing ones. The method is mainly divided into three steps. First, the noise reduction of the vibration information collected by multiple acceleration sensors is done using ensemble empirical mode decomposition (EEMD) and Teager-Kaiser energy operator (TEO), so that fault features are more obvious. Then, multi-class fault features including the severity and the location of the wear are extracted from the preprocessing signals to form the original feature set. Second, combined with feature ranking and subset selection based on euclidean distance (FRSSED) and maximum relevance minimum redundancy (mRMR) feature selection method, the statistical features extracted from multiple sensor signals are optimized to form the optimal feature subset. This is the first-level information fusion concerned with fault feature information fusion. In the third step, based on Dempster-Shafer (DS) evidence theory and convolutional neural network (CNN), decisionmaking information is fused (called second-level information fusion) to obtain the final diagnosis results. A hydraulic test bench is built to test different failure valves. Experimental results indicate that this method is effective in extracting the fault features from multi-sensor signals and detecting the fault states including severity and location of internal wear of the hydraulic valve.