The Discriminative Elastic Embedding Algorithm For Classification And Its Application In The Anomaly Detection Of Hydroelectric Generating Unit

Renewable energy development strategy is an important issue of the12th five-year plan. Small hydropower is a kind of renewable energy with widely distributed resources, large development potential, small impact on environment and extensible utilization. So it has important significance in the state energy development strategy. At present, because of the complexity of the hydroelectric generating unit and the punitive of small hydropower station’ location, it usually needs specialist staff on duty for equipment maintenance and abnormal monitoring. This method is inefficient and depends much on experience knowledge. And the probability of misjudging will be high. So it is necessary to research on machine learning theory, statistical theory, and implement high-performance recognition algorithm. At last, an unattended monitoring system for small hydropower stations should be constructed taking advantage of these theories and algorithms. Neighbor embedding algorithm is effective for data clustering visualization operation. How to improve the identification performance of neighbor embedding algorithm and apply it to hydropower noise source identification has important value in research.This paper focuses on the problem of hydroelectric generating units’anomaly detection which depends on the noise source, analyzes the expansibility with supervision and linear projection technology in neighbor embedded analysis algorithm, designs the corresponding diagnostic neighborhood embedded classification algorithm. The main work is as follows:(1) A discriminative neighbor embedding classification algorithm with Laplacian direction, named as DEE, is presented. In anomaly detection task, existing neighbor embedding analysis algorithms have low recognition rate due to the lack of class labels. DEE algorithm not only can cluster data intuitively and visually, but also enhance the efficiency of the model building by Laplacian direction, so that it can be capable of performing the identification task with large-scale data. With three categories of public available data sets, the experimental results verified the clustering capabilities and recognition performance of the proposed algorithm.(2) Two different types of kernel discriminative neighbor embedding classification algorithm, KDEE1and KDEE2, are proposed. Taking into account the non-linear nature of the noise and vibration characteristics when hydroelectric generating equipment operates abnormally, this paper propose the nonlinear expansion of DEE by kernel trick, while retaining its linear projection characteristics. In the model construction process, based on the different differential object, the kernel DEE version is called KDEE1and KDEE2, both of which can be applied to non-linear input data, and the clustering capability and identify efficiency are further enhanced compared to DEE.(3) We analyze the noise characteristics collected from hydroelectric generating abnormal vibrations and pretreatment methods, applied these algorithms mentioned in this article, KDEE1, KDEE2and DEE to an actual hydroelectric machine anomaly detection. Through the two experimental results, noise subspace clustering and recognition rate, it shows that the kernel discriminant neighbor embedded analysis algorithm has a high practical value.