Abstract: The core of quality control lies in the full utilization and analysis of quality-related data.Aiming at the problems of the loss of abnormal instances and the lag of quality anomaly discovery in the quality database,a method of recognizing quality anomaly from the quality control chart data by combining genetic algorithm with Probabilistic Neural Network (PNN) was proposed,which made up deficiencies of widely using SPC control chart in practical applications.The abnormal pattern of control chart was derived by analyzing the insufficiency of judgment criterion in the abnormality judgment of control graph.To reduce the training time of the model,the Principal Component Analysis (PCA) method was adopted for reducing the dimension and extracting feature for the original data of control map.PNN network had the advantages of simple structure and good recognition effect,which could realize the recognition of single mode and mixed mode of control chart.To eliminate the lack of experience value,the improved SGA single target optimization genetic algorithm was used to optimize the key parameters of PNN network.The proposed method was validated with simulation experiment and proved to be effective by comparing to traditional BP neural network,single PNN network,PCA-PNN model without parameters optimization and SVM model optimized by PSO particle swarm optimization algorithm.

Key words: quality control, pattern recognition, genetic algorithms, probabilistic neural network

摘要： 质量控制的核心在于对质量相关数据的充分利用和分析,如何有效组织和利用质量数据已成为企业和学者们广泛研究和关注的问题。针对质量数据库中异常实例缺失以及质量异常发现滞后的问题,提出利用遗传算法结合概率神经网络从质量控制图中挖掘质量异常现象的方法,弥补了当前广泛使用的统计过程控制(SPC)控制图在实际应用中存在的不足。首先通过分析判异准则在控制图异常判定方面的不足,引出控制图的异常模式;然后使用主成分分析法(PCA)对控制图原始数据进行降维和特征提取,以减少模型的训练时间;利用概率神经网络(PNN)结构简单、识别效果好的特点,实现控制图单一模式和混合模式的识别;通过改进的单目标优化遗传算法(SGA)对PNN的关键参数进行寻优,以消除经验取值的不足;最后通过仿真实验对所提方法进行了验证,并与传统的BP神经网络、单一的PNN、未进行参数优化的PCA-PNN模型,以及PSO优化的SVM模型进行了对比,证明了所提方法的有效性。

关键词: 质量控制, 模式识别, 遗传算法, 概率神经网络