关键词: 神经网络, 反向传播算法, 误差复映, 参数优化

Abstract: The error reflection in machining parameters selection mainly depended on workers experience. To optimize the selection of parameters, the non-linear mapping ability of mended Back-Propagation (BP) network arithmetic was utilized to approximate the non-linear relationship between error reflection coefficiency and several factors such as hardness of work piece, feeding speed and so on. As machining conditions such as error of rough before machining, hardness of work piece were input to the network having been trained, times of processing and processing quantity of each time could be acquired as output. BP arithmetic was mended by self-adapting learning rate and appended momentum. Finally, an example was given to prove that the mended BP arithmetic could converge more quickly and reduce the chances of falling into a local minimum. The structure of BP network was determined by analyzing model of error reflection phenomenon and comparing training results with different network structures. The feasibility of optimizing machining parameters with BP network was validated by testing network with test collection in MATLAB.

Key words: neural network, back-propagation arithmetic, error reflection, parameters optimization