Research On Applications Of Improved Fuzzy System And ANFIS In Intelligent Choosing Of Machining Parameters

Developing tendency of current manufacturing industry is to be infomational, intelligent, networked and nimble. Computer Aided Process Planning (CAPP) is not only the bridge connecting the bottom layer and upper layers of manufacturing industry infomationiztion but also a link between machine design and maufacture. However development and application of CAPP is rather slow, far behind development of CAD or CAM.Because many problems to be solved in CAPP is complex, non-linear and indefinite. Research of CAPP can be classified into three gradations: macroscopic gradation, middle gradation and microscopic gradation. Most early domestic and oversears researches on CAPP belong to middle gradation, which concern products or parts and lay stress on determination of machining method and routine. Current researches lay emphasis on macroscopic gradation and middle gradation, which concern the whole enterprise and system model. Research on determination of machining parameters in microscopic gradation is still vacant. While the microscopic problem is just the essence of process planning and should be solved by combing machining, intelligence and information technologies. Many complex non-linear problems in CAPP can only be solved by simulating thinking ways and processes of experts. As a result intelligent CAPP system based on artificial intelligence is becoming a developing trend.At present neural network, fuzzy sytem, fuzzy neural network, etc. are methods which have rather good effects on solving non-linear problems. Compared to neural netwok, fuzzy system has the benefit of using experts'experiences. While Adaptive Neuro-Fuzzy Inference System (ANFIS) can model fuzzy system utilizing learning and adaptive ability of neural network. So it has special superiority over complex system wich is short of or hard to obtain definite knowledge.This dissertation combines science development plan project of Jilin province (20040333) named"Research on fuzzy-neural technoglogy in machining"to research on improvement of intelligent algorithms such as fuzzy system,ANFIS and their applications in solving the machining paprameters choosing problem in microscopic gradation.Inputs of current fuzzy inference systems are all accurate values or fuzzy aggregates. While in some applications accurate value and fuzzy linguistic value need to be input to the system simultaneously. Then current fuzzy inference systems should be improved. Now division of input/output space and determination of membership function parameters mainly depend on personal experiences in modeling of fuzzy system, which is rather subjective and indefinite. Many trials have to be made before fitful parameters are found. A new fuzzy system modeling method which can be used in more kinds of fuzzy system and easy to be realized should be put forward. Training speed of ANFIS is rather slow. Is there any method to improve it? Digital signal processor (DSP) has been the core part of many devices including CNC controller. If there is a method to transplant fuzzy system designed in MATLAB to DSP promptly. All these questions should try to be solved in research of this thesis. Main research work of this thesis can be summed up as following:1. An improved fuzzy clustering algorithm was advanced on the basis of analyzing traits of fuzzy c-means (FCM) custering algorithm and subtractive custering algorithm. In order to test performence of the improved fuzzy clustering algorithm standard FCM custering algorithm and the improved fuzzy clustering algorithm were respectively used to cluster IRIS data. Clustering results show that the improved fuzzy clustering algorithm can converge more quickly and steadily than standard FCM custering algorithm. Compared to FCM custering algorithm, initial objective function value of improved fuzzy clustering algorithm during training is much fewer than that of FCM custering algorithm. When the same minimum increment criteria of stop training was adopted standard FCM custering algorithm and the improved fuzzy clustering algorithm will have the same final objective function value, which indicates both algorithms have the same precision.2. Aimming at solving division of input/output space and determination of membership function parameters which influence application efficiency of fuzzy system, a new fuzzy system modeling method was put forward. In this method swatch data of each input or output are clustered by the improved fuzzy clustering algorithm. Then clustering results are used to fit gauss-type function and S-type function by trust-region method. This method can be used to fit mamdani-type, T-S type and other fuzzy system.It's easy to be realized too. Application of the new modeling method in controll of tank water level shows that it can save time spent on trials of parameters and create more precise system.3. On the basis of analyzing factors that influence maching precision a typical non-linear problem in CAPP—the error reflection phenomenon were analyzed in theory. The model of system to realize intelligent choosing of machining parameters was established. In order to set up fuzzy inference rules library and provide training data for ANFIS, experiment scheme and flow were designed.4. A new type of fuzzy inference system—Mixed-input-type fuzzy inference system was advanced. The structure and application range of Mixed-input-type fuzzy inference system were dwelled on. Two methods to realize the conventer were discussed. How to design GUI of Mixed-input-type fuzzy inference system with GUIDE and how to construct Mixed-input-type fuzzy inference system on the basis of Mamdani-type fuzzy inference system were enlarged on. Application in tip-counting shows that Mixed-input-type fuzzy inference system can increase system efficiency remarkbaly by sacrificing little precision. Mixed-input-type fuzzy inference system for intelligent choosing of machining parameters was designed. Test results of this system shows it works well and can satisfy machining demands basically.5. Peculiarity of ANFIS training algorithm and performances of several improved BP algorithms were analyzed. Two improved ANFIS algorithms were put forward by combing Fletcher-Reeves update (FRU) method and scaled conjugate gradient (SCG) method. These two improved ANFIS algorithms and standard ANFIS algorithm were used to forecast chaotic time series and approaching non-linear function so that their performances were compared. Results of applications indicate:①when BP method is adopted ANFIS improved by SCG method conveges fastest,ANFIS improved by FRU method takes second place,standard ANFIS conveges slowest. Training interation numbers and time of the former are much fewer than those of the latter two.②when hybrid method is adopted to approaching non-linear function, ANFIS improved by FRU method conveges fastest,ANFIS improved by SCG method takes second place, standard ANFIS conveges slowest.When hybrid method is adopted to forecast chaotic time series speed of NFIS improved by SCG method is rather close to that of ANFIS improved by FRU. Both of them are faster than standard ANFIS.③Hybrid method always converges faster than BP method no matter the algorithm is improved or not.ANFIS improved by FRU was used for intelligent choosing of machining parameters. Test results of the system indicate that Mixed-input-type fuzzy inference system is more robust than ANFIS.While ANFIS may get higher precision after it's trained with more data.6. A convenient method to realize fuzzy system in DSP quickly was advanced. Ways in which ANFIS can be realized on DSP were discussed on. According to traits of program and hardware many methods such as CCS optimizer, ragma directives, etc. are intergrated to optimize codes aimming at faster speed. Transplant and optimization of fuzzy system for tip-counting as an instance indicate that the method advanced in this thesis is valid.The execution time of codes reduce remarkably after optimization. Fuzzy sytem in DSP is more real-time than that in MATLAB. ANFIS for chaotic time series focasting was realized in DSP. Output results of the system indicate that forecast results in DSP coincide with those in MATLAB and the method of on-line inferencing offline training to realize ANFIS in DSP is feasible.In the same way ANFIS for intelligent choosing of machining parameters was realized in DSP, which laid the foundation for on-line intelligent choosing of machining parameters.