A Deep Neural Networks Based on Multi-task Learning and Its Application

Abstract

With the development of machine learning, the deep neural network has achieved excellent performance. However, the learning of deep neural network often needs a large number of data and the independent learning of a single task tends to ignore the information from other tasks, resulting in redundant training and the cost of learning resources. In order to solve these problems, we introduce the deep neural network based on multi-task learning model. We introduce the structure of the multi-tasking learning model in detail, and use demographic information to prove the effectiveness of the model. Finally deep neural network based on multi-task learning framework is adapted to builds a normal behavior model for wind turbine generator in which each subspace of operating condition are taken as a task. The operating data comes from Liaoning wind farm in northeast China are used to verify the proposed method. The results indicate the subspace division of different operating condition benefits to improve the accuracy of normal behavior model of wind turbine generator, and multi-task learning is suitable to deal with the integrate subspace model.