Research Of Key Technology In Prognostic And Health Management Based On Machine Learning

Under the background of "Made In China 2025”,it's the current research focus to utilize information technology to solve the pain points of traditional manufacturing.In order to solve the problem of equipment failure and maintenance,the industry and information fields put forward Prognostic and Health Management(PHM),of which the key technology is fault diagnosis and prediction.This paper studied the key technology of artificial intelligence applied in fault diagnosis and prediction,proposed an end-to-end model of fault diagnosis and prediction based on deep learning,and an improved method of unbalanced fault data sets.The main work is as follows:Firstly,feature engineering plays an important role in this task due to high dimensions and complexity of failure data sets.This paper studied the existing methods of this task,constructed a fault diagnosis and prediction model based on CNN to realize the automatic feature engineering,by making full use of its advantage in spatial feature extraction.The effectiveness of the model in fault data feature extraction was verified by comparison experiments.Secondly,according to the fault data's characteristic of time series,we combined CNN with Long Short-Term Memory(LSTM),which is suitable for time series analysis,to overcome the shortcomings of CNN's inability to extract time features fully.According to the research of LSTM in time series analysis,the stacked bidirectional LSTM structure that can learn context information and deep time features was selected.We proposed an end-to-end fault diagnosis and prediction model CNN-StBiLSTM based on CNN and stacked bidirectional LSTM,designed the appropriate method of data preprocessing through specific experiments.Then we verified the effectiveness of the combination of CNN and LSTM,the superiority of stacked bidirectional LSTM structure and the superiority of our proposed model compared with common ones through comparative experiments.Thirdly,since the fault data sets are always unbalanced,Focal Loss,which was proposed to deal with the unbalanced data sets problem in the field of image recognition,was introduced into the fault diagnosis and prediction task.The effectiveness of the method was verified by experiments.In addition,this paper proposed a new ensemble learning method that combined Focal Loss and random undersampling,which replaced the class weight coefficient in Focal Loss with random undersampling and combined with the bagging method of ensemble learning,to improve the model effect further by using its complementary advantages.Comparative experiments were carried out to verify the superiority of our proposed model and the improved method.