Research On Data-Driven Health Status Assessment Method For Chain-Type Tool Magazine In Machining Center

Chain-type tool magazine is a complex mechatronic equipment,which has the functions of tool storage and automatic exchange.It is one of the key functional components of machining center.However,the domestic chain-type tool magazine is also the functional component with the highest failure rate of the machining center.Compared with the imported products,the domestic chain-type tool magazine has frequent failures,low reliability and short service life,resulting in large shutdown losses and high maintenance costs for the user enterprises.At present,user enterprises generally adopt the maintenance strategy of breakdown maintenance and regular maintenance.Breakdown maintenance will increase unplanned downtime,low maintenance efficiency and high maintenance cost;Regular maintenance will cause waste of maintenance resources and excessive maintenance.Compared with breakdown maintenance and regular maintenance,the condition-based maintenance develops corresponding repair and maintenance strategies according to the health status of the product,which can significantly reduce the unplanned downtime and maintenance costs and improve the reliability of the product.Therefore,accurate assessment of the health status of the whole chain-type tool magazine is helpful to formulate the optimal conditionbased maintenance strategy,thus reducing the unplanned downtime and maintenance cost of products and improving the reliability of domestic machining centers.It has important theoretical significance and engineering application value.In order to accurately evaluate the health status of the whole chain-type tool magazine,with the support of major national science and technology projects,this paper takes the chaintype tool magazine of the machining center as the research object,builds a chain-type tool magazine health status monitoring system,and obtains multi-source sensor signals such as vibration,voltage,current,air pressure and temperature through automatic tool change experiments.Then,a data-driven chain-type tool magazine health status assessment is carried out.The main contributions of this thesis are as follows:(1)Based on the analysis of the structure and working principle of the chain-type tool magazine,the fault tree analysis method is used to analyze the overall fault of the chain-type tool magazine.Aiming at the two typical failure modes(tool dropping of manipulator and tool dropping of tool-case)determined based on the on-site tracking test data of machine tool users,the failure cause analysis and failure signal analysis are carried out by constructing statics model and dynamic simulation.Based on the results of fault tree analysis and typical fault analysis,two key components(manipulator and tool-case)with high failure rate and the overall health status monitoring indicators are determined.(2)Aiming at the problem of identifying pulse signal segments containing state information from monitoring signals of key parts of chain-type tool magazine,a frequency domain dynamic time warping(DTW)algorithm is proposed.The experimental results based on simulation and measured signals indicate that compared to time-domain DTW,the proposed frequency-domain DTW can more accurately identify specified pulse signal segments.A dual tree complex wavelet domain overlapping group sparse denoising algorithm is proposed for transient impact signals containing only Gaussian noise;A denoising algorithm based on digital filtering and overlapping group sparsity is proposed for transient impulse signals containing both narrowband interference noise and Gaussian noise.The experimental results based on simulation and measured signals indicate that the proposed method has better denoising performance compared to several classic sparse denoising algorithms.(3)Health status assessment based on shallow neural networks usually includes key technologies such as feature extraction,feature selection,construction and training of shallow neural networks.The selection of wavelet basis functions for extracting features in the wavelet domain has significant subjectivity,the accuracy of feature subsets selected using commonly used feature selection algorithms is low for health state assessment,and the use of commonly used gradient descent methods to train Hybrid Class RBM(H-Class RBM)is prone to falling into local optima,A key component health status assessment method based on H-Class RBM is proposed.This method extracts the time-domain,frequency-domain,and wavelet domain features of the monitoring signal,and uses the proposed XGBDT-ISFS(extreme gradient boosting decision tree and improved sequential floating selection)to select the extracted feature set.The health status of key components is evaluated using the proposed H-Class RBM based on ADAM-SSD(Adaptive motion estimation and Stochastic Spectral Descent)optimization.When extracting features in the wavelet domain,this method adopts the principle of maximizing the improved energy aroma entropy ratio as the selection basis for the wavelet basis function.The experimental results based on the health status assessment of key components indicate that using the principle of maximizing the improved energy aroma entropy ratio can accurately select the optimal wavelet basis function;Compared to several commonly used feature selection methods,the proposed method can select higher quality feature subsets from the feature set;Compared to several commonly used optimization algorithms,ADAM-SSD optimized H-Class RBM has higher training set accuracy;Compared to several evaluation models based on commonly used machine learning algorithms,the proposed evaluation model achieves lower training time and higher evaluation accuracy.(4)In order to solve the problem that the common convolutional recurrent neural network has a weak ability to capture the long-distance spatio-temporal dependent information contained in the input when dealing with long-term series data,an attention convolution mogrifier long term short-term memory network(ACMLSTM)was proposed and used to assess the health status of key components.The proposed attention convolution module is used to learn local spatial features with rich state information from each segment of the input signal,and then the mogrifier long term short-term memory network is used to fuse these local spatial features.The experimental results based on the health status assessment of key components indicate that using attention convolution modules can improve the accuracy of the health status assessment model and reduce its hinge loss;Compared to several commonly used evaluation models,the proposed model has higher classification accuracy.(5)A comprehensive health status assessment method based on probability interval number grey clustering and barrel theory is proposed to address the uncertainty of measurement values of chain-type tool magazine status monitoring sensors and the fusion of status information from different subsystems.This method first uses Fisher Rao registration,probability interval number grey clustering,and a health status evaluation model based on neural networks to evaluate the status of each subsystem.Then,based on the barrel theory,the overall status of the chain-type tool magazine is evaluated.The experimental results based on the health status evaluation of chain-type tool magazine show that for the health status evaluation of subsystems,compared to interval number grey clustering,the evaluation results of probability interval number grey clustering are more reasonable and accurate;Compared to several evaluation methods based on evidence theory,the evaluation results of the proposed method are more reasonable and accurate for overall health status assessment.