计算机科学
人工智能
加权
机器学习
控制图
卷积神经网络
图表
人工神经网络
方案(数学)
故障检测与隔离
控制(管理)
数据挖掘
模式识别(心理学)
放射科
数学分析
操作系统
执行机构
统计
过程(计算)
医学
数学
作者
Donovan Fuqua,Talayeh Razzaghi
标识
DOI:10.1016/j.eswa.2020.113275
摘要
Abnormal control chart patterns are naturally infrequent in an industrial setting. However, such patterns may indicate manufacturing faults that, if not treated in a timely manner, can lead to significant internal and external failure costs, ultimately threatening the product reputation. Therefore, the detection of abnormalities, which is sought in the well-known control chart pattern recognition (CCPR) problem, is of utmost importance. Standard machine learning algorithms have been extensively applied to this problem. However, they often produce biased classifiers unless the inherent data imbalancedness, which originates from the scarcity of abnormal patterns, are carefully addressed. In this paper, we develop a cost-sensitive classification scheme within a deep convolutional neural network (CSCNN) for the imbalanced CCPR problem. We further investigate the performance of our algorithm on both simulated and real-world datasets to determine separable and non-separable common fault patterns in a manufacturing setting. As the contribution of this work, we particularly demonstrate that the cost weighting strategy is both robust and efficient for moderately- and severely-imbalanced cases. We further show that our method can either be fine-tuned to specific faults or trained to detect multiple faults while remaining efficient for large datasets. To the best of our knowledge, this is the first deep CSCNN designed for imbalanced CCPR problems, which presents great promise for other manufacturing applications in the presence of imbalanced datasets.
科研通智能强力驱动
Strongly Powered by AbleSci AI