Research On Life Prediction Based On Neural Network And Exponential Smoothing

Effective remaining useful life prediction have important research value in the field of the machining system which requires high safety and reliability.In recent years,with the improvement of storage capacity,the monitoring data used for analysis and processing is characterized by massive,high-dimensional and multi-measurement.The prediction and the model convergence will be affected by a large amount of input data.Although traditional methods such as support vector regression,have achieved good results in monitoring,it still has a great impact on the results when the input characteristics lack sufficient information,which poses a huge challenge.What’s more,the remaining useful life prediction based on the exponential smoothing method has the problem that the smoothing coefficient cannot be adaptively adjusted.The prediction will be biased when the degradation trend changes,and reduces the accuracy of the remaining useful life prediction.The paper proposes a condition evaluation and remaining useful life prediction method based on neural network and exponential smoothing method.The method mainly includes three steps.In the step of feature extraction,the high-dimensional data sets are constructed by the original cutting force signal of the tool and the timefrequency characteristics of the vibration signal of the harmonic reducer respectively.The sparse auto-encoder is used to fuse the high-dimensional data features to eliminate the redundancy and obtain a concise feature representation.In order to improve the robustness,the dropout method is introduced to reduce the over-fitting problem.The tool wear experiment and the harmonic reducer vibration experiment are designed to prove the proposed feature is more correlated with the performance degradation than the traditional features.In the step of condition evaluation,the BP neural network model is trained to predict performance degradation parameters.The results show that the prediction accuracy based on new features is higher than the traditional features;In the step of remaining useful life prediction,the traditional exponential smoothing method cannot adaptively adjust the smoothing coefficient.The introduction of upper and lower limit threshold parameters to update the model parameters in real time solves this problem and improves the life prediction accuracy.The research results show that the BP neural network model based on the features of sparse auto-encoder can effectively predict the tool wear value and the harmonic reducer transmission error,and the proposed exponential smoothing algorithm can accurately predict the remaining useful life when the trend changes.