Research On Remaining Useful Life Prediction Of Adaptive Tool Considering Feature Expansion And Iteration Speed

This paper proposes a LightGBM-based residual life prediction method that takes into account processing procedures and processing conditions.A multi-information fusion strategy was considered during the data set creation process.By extracting conventional features for verification and comparison,it was found that the combination of vibration,current,and load signals can achieve good results in the most basic feature extraction.In addition,the spatial information in the process of tool vibration is considered,and the vibration process is considered as a kind of spatial vibration.The vibration signals are collected in multiple dimensions and features are extracted to completely describe the tool vibration process.Further,the three-dimensional vibration signal of the tool is subjected to the dimensionality reduction and feature extraction of the three-dimensional vibration signal of the tool using a local linear embedding technique.By comparing two different spatial feature extraction methods,we verify the effectiveness of the spatial feature extraction method for mining vibration signals,and the use of local linear embedding technology to extract the spatial features of three-dimensional vibration has better advantages and can effectively improve the model The accuracy.And in order to solve the requirement of fastness,a reconstruction error index is proposed for the neighborhood size parameter k that affects the dimensionality reduction effect,and a parameter determination method that comprehensively considers accuracy and industrial timeliness is determined using this index,and verified Its effectiveness.In the traditional remaining life prediction research,a feature that can take into account the expansion of the number of features and cover a small time granularity is proposed.Using the sliding window-based method to extract small-time granularity sub-segment features and perform unsupervised cluster analysis on the sub-segments to reduce the amount of data in the sub-segments,on the one hand,it can effectively reduce the dimension of small samples,on the other hand,it also effectively reduces Introduced invalid features,or strong co-linear features and noise information.Through the above-mentioned methods,the dimensions of the features are effectively extended,and the accuracy and generalization ability of the model are effectively improved.In addition,in terms of models,due to the industry's actual requirements for speed and accuracy,a comparison between theoretical and practical perspectives has verified that the accuracy and speed of the use of the model are in line with requirements.When returning to the actual industrial problems,there is an unavoidable error in the prediction model in reality.The positive and negative values of the difference between the predicted value and the actual value are often not equivalent.Therefore,both ends of the predicted risk need to be corrected to solve the model Uneven measurement of risk at both ends.By proposing the importance coefficient to weigh the ratio of the two ends of the positive and negative risks,the process importance coefficient is quantitatively evaluated by three factors.And the triangular fuzzy number combined with the expert system can be used to modify the weight factors of process manufacturing influence factors.The obtained results effectively combine the influence of the relationship between this process and the previous process.It is introduced into the loss function,and the validity of this method is verified by experiments,and the residual distribution of the final predicted value is tested,and the results are consistent with expectations.