Abstract Despite extensive studies for the industrial applications of deep learning, its actual usage in manufacturing sites has been extremely restrained by the difficulty in obtaining sufficient industrial data, especially for production failure cases. In this study, we introduced a fault-detection module based on one-class deep learning for imbalanced industrial time-series data, which consists of three submodules, namely, time-series prediction based on deep learning, residual calculation, and one-class classification using one-class support vector machine and isolation forest. Four different networks were used for the time-series prediction: multilayer perception (MLP), residual network (ResNet), long–short-term memory (LSTM), and ResNet–LSTM, each trained with the one-class data having only the production success cases. We adopted the residuals of the deep-learning prediction as an elaborated feature for the construction of the one-class classification. We also tested the fault-detection module with the actual mass production data of a die-casting process. By adopting the features elaborated by the deep-learning time-series prediction, we showed that the total accuracy of the one-class classification significantly improved from 90.0% to 96.0%. Especially for its capability to detect production failures, the accuracy improved from 84.0% to 96.0%. The area under the receiver operating characteristics (AUROC) also improved from 87.56% to 98.96%. ResNet showed the best performance for detecting production failures, whereas ResNet–LSTM produced better results for ensuring the production success. Our results suggest that the one-class deep learning is a promising approach for extracting important features from time-series data to realize a one-class fault-detection module.

中文翻译：

---

基于一类深度学习的故障检测，适用于受限于不平衡数据库的制造应用

摘要 尽管对深度学习的工业应用进行了广泛的研究，但由于难以获得足够的工业数据，特别是在生产故障案例中，其在制造现场的实际应用受到极大限制。在本研究中，我们针对不平衡的工业时序数据引入了基于一类深度学习的故障检测模块，该模块由三个子模块组成，即基于深度学习的时序预测、残差计算和一类使用一类支持向量机和隔离森林进行分类。四种不同的网络用于时间序列预测：多层感知 (MLP)、残差网络 (ResNet)、长短期记忆 (LSTM) 和 ResNet-LSTM，每个网络都使用只有生产成功案例。我们采用深度学习预测的残差作为构建一类分类的详细特征。我们还使用压铸过程的实际批量生产数据测试了故障检测模块。通过采用深度学习时间序列预测所阐述的特征，我们表明一类分类的总准确率从 90.0% 显着提高到 96.0%。尤其是其检测生产故障的能力，准确率从 84.0% 提高到 96.0%。接收器操作特性 (AUROC) 下的面积也从 87.56% 提高到 98.96%。ResNet 在检测生产故障方面表现出最好的性能，而 ResNet-LSTM 在确保生产成功方面产生了更好的结果。