Mechanical Fault Diagnostic And Prognostic Based On Multi-task Learning

In production practice,in order to avoid the huge losses caused by mechanical equipment downtime due to failure,it is necessary to obtain the failure information and life degradation information of mechanical equipment and make maintenance plan in time.In recent years,diagnosis and prognostic based on deep learning has been extensively researched.However,there are still many challenges in the practical application of the existing methods: in practice,the samples in normal state are far more than those in abnormal state,and the imbalance of categories easily leads to the difficulty in training the diagnosis model and low generalization performance;the fault prediction methods based on traditional convolutional neural network can’t extract multi-scale features,and can’t achieve the fusion of multi-source sensor timing information;In order to obtain the fault information and degradation information of mechanical equipment,the existing methods need separate training model to realize fault diagnosis or prognostic,which not only greatly increases the development and deployment costs,but also reduces the practicability of intelligent operation and maintenance.In order to solve such problems,this paper mainly completes the following work:(1)Firstly,in order to solve the problem that the intelligent model is difficult to train and the generalization performance is poor due to the class imbalance of negative samples far more than positive samples in production practice,a class imbalance fault diagnosis method based on improved focal loss is proposed,which introduces class weight,sample weight and L2 regularization,to avoid the model overfit to the majority class.Through the bearing fault experiment,it is confirmed that the proposed method still has better accuracy and generalization under various unbalance conditions.(2)A prognostic model to fuse multi-sensor’s temporal information established.The original signal of multi-sensor is taken as input,the deep multi-scale features are extracted by multi-scale convolution feature extractor,and the multi-sensor’s temporal information is fused in the meantime.Then,the extracted features have been inputted into Long Short-Term Memory Network(LSTM)to realize prediction.Finally,Through the tool wear prediction experiments,it is verified that the proposed method has the ability to extract multi-scale features and multisensor’s temporal information,and it has better performance than the traditional methods.(3)An intelligent diagnosis and prognostic method based on Multi-task learning is proposed.Because of the knowledge sharing of multi task learning,this method can extract the features of multi task sharing,and realize fault diagnosis and prognostic of mechanical equipment through end-to-end supervised joint training.Compared with the traditional single task method,the proposed method only needs to train and deploy one model to get fault information and life information at the same time.The proposed method can satisfy the demands of industrial application better.Finally,the bearing life degradation test results show that the proposed method can accurately diagnose the fault category and predict the Remaining Useful Life(RUL),and has a good adaptability to class imbalance.The proposed method can realize the fault diagnosis and prognostic of mechanical equipment at the same time,and can provide important decision-making basis for operation and maintenance,which has strong practical significance.