Hydraulic failure diagnosis of tractor hydro-mechanical continuously variable transmission in shifting process

Abstract: Unlike conventional mechanical transmission with manual shift, the hydro-mechanical CVT (continuously variable transmission) shifts from one range to another automatically under the control of hydraulic system, however, some faults of this hydraulic system can affect the establishment of clutch engagement pressure and affect the shift quality further. In order to improve the reliability of the hydro-mechanical CVT in shifting process and avoid safety accident, the problem of fault diagnosis on this hydraulic system was studied in this paper. First, a measurement and control system for the hydro-mechanical CVT was built, including the transmission, clutch hydraulic control system, sensors, transmission control unit and IPC (industrial personal computer). Software developed under Labview can build the communication between the IPC and the sensors, and record the feedback data from sensors to the table file with a frequency of 62.5 Hz. Then, 5 fault modes were defined in this paper, i.e. normal state, piston seizure, seal ring damage, oil passages clog and pipeline sealing failure. Through experiments, 60 groups of feedback data from pressure gauges and flowmeter with different fault modes were acquired. Thirty groups of the testing data were selected as training samples and the other 30 groups were selected as testing samples. For each sample, 6 kinds of statistic features such as root mean square value of sample data in a second were calculated. Because any combination of the 6 statistic features could not classify the fault sample directly in a low-dimensional space, it was necessary to map the sample to a higher dimensional space and to find the hyperplane of demarcation. So, a projection vector was calculated using the method of kernel based on the rule of Fisher. According to this projection vector, the sample data which was difficult to classify in original space could be transferred to feature space, in which the sample data could be classified with a simple linear function. After the kernel calculation, the methods to reduce the computation complexity were studied. Based on the rough set theory, the kernel features were found and the FAI (fisher based attribute importance function) value was calculated. The kernel features and the non-kernel features with a high FAI value were selected as the training features. Based on the heuristic algorithm, the training sample data corresponding to the minimum component in projection vector could be deleted in turn. Finally, in order to test the robustness of the algorithm, 10 groups of training set and testing set were rebuilt through random sample selection. The results revealed that the training and testing sample data could be classified correctly using the method of kernel, and proper reduction on features and training samples could reduce the computation complexity of fault diagnosis without accuracy loss. In addition, The average fault diagnostic accuracies of the random samples under the 5 modes were 99.67%, 100%, 99%, 98% and 99%, respectively, and the variation range for each group of sample was from 96.7% to 100%. The results indicated a high robustness of the algorithm used in this paper. The study of this paper provides a method to improve the reliability of the hydro-mechanical CVT.