Manifold Learning And Its Application In The Noise Source Detection Of Hydroelectric Generating Unit

Small hydropower is a kind of clear and reproducible power and is the key strategic direction of the national energy development. But at present most hydroelectric generating units detect the abnormal noise and fault by artificial judgement. This way has low speed and will influence the accuracy of judgement. So that it is necessary to consider the machine learning method. Manifold learning has the ability of finding the inner structure of the dataset. It is valuable to research the way to apply the manifold learning algorithm to the hydroelectric generating unit noise source detection.This article has analysed the characteristics of the hydroelectric generating unit noise signal and the problems when applying the manifold learning to the hydroelectric generating unit noise source detection, and then designed the corresponding manifold learning algorithm. The main contributions of the work are as follow:1) Analyze the reasons of abnormal vibration of the hydroelectric generating unit and the characteristics of the noise signal and introduce the vibration signal denoising method in details.2) Design an incremental within-class locality preserving dimension reduction algorithm, IWDA. This algorithm is designed to solve the linear problem and has the ability of the incremental computation and can preserve the local space structure. This algorithm is effective on the data which is multimode or overlapping.3) Design an incremental kernel discrimination analysis method via QR Decomposition, IKDR/QR. This algorithm is suitable for complex nolinear device like hydroelectric generating unit by combining IWDA with kernel method. Because the abnormal vibration noise is hard to get in the early stage of the noise detection, the online learning is needed. IKDR/QR can decrease the time and space of creating the kernel matrix during the process of online learning.4) Apply IKDR/QR to hydroelectric generating unit noise source online detection. The noise source online detection experiment proved the effectiveness of IKDR/QR.