塑料注射成型工艺参数优化的模糊规则网络模型

注射成型是塑料产品成型的最主要工艺，工艺参数是影响成型产品外观、尺寸与性能的关键因素之一。工艺参数的设置与优化属于弱理论、强经验的问题，迫切需要发展科学化、系统化的方法。针对产品缺陷修正中人工经验依赖性强的问题，构建知识的统一模糊化规则形式，建立工艺优化知识表示和推理于一体的Takagi-Sugeno-Kang(TSK)模糊规则网络模型。进一步，提出从工艺数据集自动发现工艺参数优化规则的学习方法，采用Dropout策略与Bagging集成学习策略缓解高维工艺数据下工艺知识库增长出现的规则数量爆炸等问题。分析了模糊规则网络参数、结构对知识表示和推理的影响，建立模型的参数学习与结构优化的双重进化方法。提出基于经验回放的工艺数据增量学习方法，建立数据的增量学习策略。在注射成型工艺数据集上的结果表明，模型的规则数量和长度降低了50%，具有高可解释性以及增量学习稳定性。

A Fuzzy Rule-based Network Model for Optimization of Process Parameters in Plastic Injection Molding

Injection molding is the most crucial process for forming plastic products, and process parameters are one of the critical factors affecting the appearance, size, and performance of products. However, the optimization of the process parameters is weakly theoretical and strongly empirical problems, and there is an urgent need to develop a scientific and systematic method. In response to the problem of the strong dependence of manual experience in product defect correction, a unified knowledge form using fuzzy rules was constructed and a Takagi-Sugeno-Kang (TSK) fuzzy rule network model integrating knowledge representation and inference of process parameters optimization was established. Furthermore, a learning method was proposed to automatically discover optimization rules of process parameters from process datasets. The Dropout strategy and Bagging ensemble learning strategy were adopted to alleviate the problem of rule explosion caused by the growth of process knowledge bases in high-dimensional process data. Then, the influences of the fuzzy rule network parameters and structure on knowledge representation and inference were analyzed. Based on these analyses, two methods, parameters learning and structure learning of the model were developed respectively. The learning method of process data based on experience pool replay was proposed, establishing the incremental learning strategy of process data. The test results on the injection molding dataset showed that the number and length of rules were reduced by 50% in the proposed fuzzy rule-base network model, realizing high interpretability as well as incremental learning stability.