Study On The Method Of Tool Wear Condition Monitoring Based On Multi-sensor Information Fusion

In metal cutting process,the real-time monitoring of tool wear condition is an important link for realizing automated production,ensuring product quality,improving production efficiency and reducing equipment faults.With the rapid development of sensor technology,data processing technology and computer software and hardware technology,the application of multi-sensor information fusion technology on tool wear condition monitoring will become an important research direction in this field.The core problem of multi-sensor information fusion is to select sensor types reasonably and achieve complementary information effectively.Links including signal aquisition,signal filtering,signal feature extraction and optimization factors which affects the effect of tool wear condition monitoring.Base on the viewpoints above,the main study contents are as below:1.The traditional data acquisition method in experiments was improved.Two sets of experiments with the same cutting conditions were used for realizing the experimental data acquisition in the continuous cutting state.It need not to stop cutting for tool wear quantity measurement,and can make the cutting state in experiments closer to that in actual production.Taking account of the cost,installation and fusion effect,the AE,microphone and current sensors were taken as the sensors for tool wear condition monitoring.The orthogonal experiment method was adopted for the acquisition of AE,cutting sound and spindle motor current signals.2.According to signal characteristics of each sensor respectively,different methods were selected for signal filtering and feature extraction.For the AE signal,the wavelet packet transformation method was used for signal filtering and the generalized fractal dimension method was used for feature extraction.For the cutting sound signal,the empirical mode decomposition method was used for signal filtering and the Hilbert-Huang transformation method was used for feature extraction.For the spindle motor current signal,the wavelet packet transformation method was used for getting the reconstruction coefficient matrix.Features of wavelet packet energy and singular value were extracted from the matrix.3.Some improvement was made on the base of the traditional box-counting method for extracting the generalized fractal dimension features of the AE signal.The new method is a step-scan method,it makes the grid division more in line with the distribution of sampling points.The generalized fractal dimension features of AE signal acquired were extracted by the improved and traditional box-counting methods.Through comparison,it can be found that the improved method can make the extracting precision of the generalized fractal dimension features improved and the time for calculation reduced correspondingly.4.Through comparing the support vector machine(SVM)with BP neural network and fuzzy neural network based on experimental data analysis,the SVM model was selected as the model for the feature level fusion and the sub-model for the decision level fusion of multi-sensor information because of its better performance.Because some noise also exists in the signals filtered and the correlation degrees between tool wear quantity and each signal feature are different,not all the features extracted are useful and the existence of interference features will reduce the accuracy of prediction results.The penalizing factor C and the parameter y of kernel function also affect the performance of the SVM model.Therefore,based on multiple population genetic algorithm(MPGA),a method for the optimization of monitoring signal features and SVM model parameters was designed by combining the training and prediction of SVM model and MPGA.Through comparing monitoring effects before and after optimination,the effectiveness of the optimization method was verified.5.Taking the SVM model as the fusion model,the feature level fusion method of multi-sensor information for tool wear condition monitoring was studied.The signal features and model parameters were optimized and the feature level fusion effects for different types of signal combination was analysed.Then the monitoring effects of the feature level fusion method and single signal monitoring methods were compared.6.The decision level fusion method of multi-sensor information for tool wear condition monitoring was studied and it was realized by constructing integrated model and designing effective fusion method.The base layer of the integrated model is composed of SVR models,which correspond to the different types of signal respectively.Their outputs are the predicted values of tool wear quantities,which are used as sub-decisions.The decision layer of the integrated model is used for the combined decision of classification or prediction according to sub-decisions by designing effective fusion method.For tool wear condition classification,the decision layer is a SVC model.For tool wear quantity prediction,it can be found through study that taking the FNN as the model in the decision layer can guarantee the predictive accuracy,reduce the maximum predictive error and enhance predictive stability and reliability.The decision level fusion effects for different types of signal combination was analysed,and the monitoring effects of the feature level fusion and single signal monitoring methods were compared.It can be proved that the decision level fusion method can make information from different types of monitoring signal complemented,and its monitoring effect is better than that of single signal monitoring method.The accuray and reliability of the monitoring system can be improved by the decision level fusion method of multi-sensor information.