Research Of Key Technologies Of A Tool Wear Monitoring System Based On Multi-Model Decision Fusion

Tool condition monitoring is an important part of the advanced manufacturing technology. It is an emerging technology which has been developing with the modern sensor technology, computer technology, signal processing technology and artificial intelligence. In recent years, some research institutes have conducted thorough research on tool condition monitoring and have achieved many results. However, there are still a lot of problems needed to be solved. To solve the existing problems, a series of research works have been carried out. Firstly, this paper proposes an improved wavelet threshold noise canceling method to solve the question about the pollution of noises in tool monitoring data. Secondly, the feature extraction and optimization techniques of tool condition monitoring were investigated. The PPCA (Parts Principle Component Analysis) which is an improved PCA algorithm has been discussed. At last, to solve the big size tool wear state space recognition problem, the Support Vector Machine and3-level Integrated Neural Network were proposed. An improved D-S evidence theory method was used to fuse the results. These methods further improve the performance of the tool condition monitoring system. The primary contents of this paper are as follow.(1) Common tool wear phenomenons during manufacturing have been discussed. By comparing many monitoring methodes, the wear condition of back surface of cutters has been used to represent the tool wear state. The hardware platform of monitoring system is composed of three-dimensional dynamometer and three-dimensional accelerometer. This multi-sensor data fusion monitoring method has been proved to be effective by experimental study.(2) In the signal processing process, treatments of tendency elimination, zero-mean and wavelet denoising have been carried out. To overcome the shortcomings of classic threshold denoising algorithm, of wavelet filtering algorithms and to increase the nonlinearity of threshold function, an improved wavelet threshold noise reduction method called single parameter logarithm wavelet packet filtering algorithm has been proposed by introducing nonlinear component in the process of constructing the thresholding function. Experimental results have proven the reliability and effectiveness of the method.(3) Signal features are extracted form time domain, frequency domain and the time-frequency domain. The fractal geometry signal feature has been extracted too. It is not applied widely in tool condition monitoring processes right now. Further investigation on this technique has been made and the fractal box dimension feature of the signal is obtained and proved to be effective. Research work for applying the wavelet analysis methode in the tool monitoring process was also carried out. And wavelet sub-band power features based on statistical theory were proposed. All features extracted from the tool monitoring signal are as follow:Sum of absolute value, Maximum value, Distance between extremum, Standard deviation, Mean of absolute value, Variance, RMS, Kurtosis, Skewness, Sum of amplitude of frequency, Maximum frequency, Fractal box-dimension, Wavelet sub-band power, Sub-band power rate, Distance of sub-band power rate.(4) To avoid the curse of dimensionality and reduce the cost of computing which may be caused by the large feature space, the PPCA (Parts principle component analysis) methode is proposed by improving the PCA algorithm. Feature optimization was successfully achieved. Compared with the PCA, this method avoids the interference of features between different classes. The convergence speed is increased by using of PPCA approach and the precision of monitoring system was improved.(5) This paper proposes a fusion model of the Support Vector Machine and3-level Integrated Neural Network for tool condtion monitoring, for solving the problem of high dimensional input/output mapping. Traditional neural networks can’t converge well when identifying large pattern spaces. So a3-level Integrated Neural Network was proposed and it was successfully applied in tool condition recogniztion. Support Vector Machine recognition model was also established. This paper proposes the fusion model of SVM and ANNs that have different mathematical characteristics. The results of each model and twin-model have been compared in chapter6and chapter7.In the process of decision fusion, the problem of geting the accuracy of SVM model must be solved. The commonly method is cross validation. But it is not applicable when treating the data in this programme, because the cross validation is nolonger meaningful for small size sample data with multi-patterns. A methode of calculating the mean distance between forecast classification and the target classification is proposed to solve the above problem.D-S evidence theory is applied in decision-making level. Two intelligent models with different mathematical characteristics (ANNs and SVM) are applied separately to recognize tool wear states and the results are fused to make the final decision. The problem of geting the confidence function value of D-S theory is a key point. The root mean value of the output error of the intelligent recognition model is used as the confidence function value. According to the requirements of D-S theory, the confidence function space of SVM and ANN is normalized. Experimental results showed that the improved D-S evidence theory fusion model for SVM and ANNs successfully solved the tool wear recognition problem in the full life span of the cutting tool.