Research On Rotational Error Prediction Method Of Spindle Through Bi-LSTM Network

CNC is used as the manufacturing industry "working machine",and the spindle is the core function and implementation for CNC.Rotation error is an important measurement index of spindle’s precision,which is directly related to the quality of workpiece.Expensive measuring equipment,harsh installation requirements,and the existing measuring technology will affect the normal processing tasks,resulting in the spindle rotation error is inconvenient monitored directly.Vibration signal is regarded as an important indicator of equipment state,because it contains rich state information,and has been widely used in state monitoring of rotating equipment.Therefore,this paper studies the mapping relationship between the vibration signal,which is easy to acquire,and the spindle rotation error.Investigate and summarize the vibration signal processing method,equipment state prediction method based on vibration signal and the equipment prediction method under the small samples.According to the current problems,the research thought begin improving ability from vibration signal about the characteristics of wear and speed,to gradually improving the prediction model,and then to the research ideas of model application,and the specific research contents are as follows:Focusing on the LMD,which is adaptive time-frequency processing method,and its endeffects,this paper analyzes the reasons why the end-effects lead to the decomposition waveform distortion and interference feature extraction,and determines the basic idea of extending to find new extreme points.After that,an slef-adaptive waveform point extended method based on Bi-LSTM regression network is proposed,and the extended method divides the trainable samples into two steps to optimize learning parameters and determine initialization matrix: First-step: finds out the optimal training parameters beyond effective samples adaptively;Second-step: learns final network parameters with all trainable samples.Additionally,the extended method is combined with conventional LMD to construct an ILMD with eimination of end-effects.The results of the final simulational and the actual vibration experiments show that the ILMD can extract accurate characteristic frequencies with higher characteristic amplitudes from the simulation signals,and more accurate wear and speed characteristic frequencies can be extracted from the vibration signals under no load,constant load and variable load conditions,while higher amplitude can be ensured at the characteristic frequencies.In order to solve the problem of vibration signal interference and the time-correlation characteristic of spindle rotation error,as well as combined with the prediction requirements of practical engineering to the rotation error.Therefore,a prediction method based on ILMD filter fusion and BI-LSTM classification network is proposed in this paper.In this method,the vibration signals are preprocessed by ILMDFF and rotation error at 1um intervals,and the preprocessing results are finally input into the Bi-LSTM classification network for training and prediction.A spindle simulation loading experiment platform was designed and built to complete the wear loading experiment more than 1700 hours,and 170 sets of vibration signals and the corresponding rotation error values were collected under the varied speed.Finally,this method is used to predict the rotation speed of 1000,2000,3000 and 4000 respectively,and these results are analyzed.By comparing the four types of input signals,ILMDFF is better than other input signals,which indicates that the prediction method has the characteristics of filtering fusion,and the prediction accuracy is better than the traditional data compression method.In view of the problem that the extreme spindle rotation error cannot be accurately predicted,and the two-dimensional image convolution method can be used to predict with higher sensitivity.Furthermore,a new prediction method based on 2D-ILMDFF and convolutional BI-LSTM classification network is proposed.Firstly,this method converts the one-dimensional temporal sequence into the two-dimensional graph,which not only preserves the temporal memory characteristics of the original sequence,but also incorporates the learning ability of high-dimensional convolution features.Additionally,optimization training parameters and image size are added in the training process of the new method.Finally,this method is applied to the training and prediction of the collected experimental data,and the prediction results at multiple rotational speeds are also analyzed: this method improves the overall prediction accuracy even under the conditions of different input modes,and twodimensional convolution is introduced to effectively solve the problem that extreme rotation error cannot be predicted.Considering the popularization and application of rotation error prediction for different spindles,the data sets,which are collected by different spindles,are easy to make few samples or zero samples,and these can be resulting in the lack trained for convolutional Bi-LSTM model with supervised learning and use the prior model directly.This paper studies the application of transfer learning in monitoring equipment under few sample condition,and proposes an improved method based on inhibited transfer learning and convolutional BiLSTM model.This improved method not only proposes the adjustment strategy under the missing samples condition,and the under-learning problem can be solved by pre-training small samples and retraining by adding inhibitory factors.Finally,this improved method is applied to predict a new spindle rotation error,which only have 20 groups date sets under no load,constant load and variable load conditions,to verify the effectiveness in the case of few samples.