The Research Of Neural Network In Solving Error Reflection Phenomenon In Machining

In the process of machining, Errors of rough in dimension, shape and location lead to the change in processing quantity. The material of work piece isn't equal. For all these reasons cutting force changes in machining which make the machining system deform. Consequently errors of work piece come into being. This is the so called error reflection phenomenon. Generally such errors are reduced through many times of processing and adopting appropriate processing quantity each time according the operator's experience.According to the theory of error reflection the error reflection coefficient indicates to which extent errors of rough influence errors of work piece. It is connected with several factors such as machining condition,hardness of work piece and so on. Such non-linear relation can not be worked out by formula,While BP network can simulate such non-linear relation for its random non-linear mapping ability. BP arithmetic is not limited by the input numbers and output numbers. In the actual research,you can modify the program freely as you need. What I have finished in the paper in research are as follows:1. Confirming the model of neural networkThrough analyzing the model of error reflection and comparing different kinds of neural networks the model of neural network are confirmed. The essential of solving error refection phenomenon in machining is to approximate a complex non-linear function. While perception can only solve sorting problem whose input can be linearly sorted, Adaptive linear element can only learn linear relation between input and output. BP network can approximate such complex non-linear function for its strong non-linear mapping ability. Through analysis of error refection's model, Error of rough (EB),error of work piece after machining(EE) , rigidity of machining system(KS), hardness of work piece(HBS) and feeding speed(f) are used as inputs of BP neural network. Proportions of each times' cutting depth to overall cutting depth(P1,P2) are outputs of BP neural network.2. Modification of BP arithmetic BP arithmetic used in my research is improved on aiming at training quicker and fitting my research. With the modified BP network, network can be trained much more quicker and avoid sliding to a local minimum .An example is given to prove the superiority of mended arithmetic over others with the use of neural network toolbox of MATLAB. 3. Determining the structure of BP networkThe number of layers and the number of neurons in each layer are determined by training network with different structures and comparing the results of training. The structure 5-15-12-2 is the best structure as in it network can be trained quicker and acquire less SSE as well as avoid excessive matching.4. Application in solving error reflection Test the network after it has been trained with testing collection and observe the difference between testing results and actual results. In this way feasibility of solving error reflection with BP network is proved. It can be concluded that the network can converge well. Although there are certain errors testing the network, when inputs are close to training collect, errors are smaller. It indicates that the training collect may be not self-contained. If only larger training collect is possessed, BP network can solve error refection phenomenon very well.5. Design of an eccentric clampA kind of eccentric clamp was devised according to the need of experiments. It works well in experiments.6. Design of GUIA GUI was designed with the use of GUI design tool GUIDE in MATLAB in which users could set training parameters, train BP network and test it.