For pneumatically operated train doors, continuous monitoring and earlier detection of air leakages would greatly reduce train downtime and ensure both safety and high availability of the rail system. Deep learning, a powerful tool to extract useful information from a large amount of sensor data is widely used in fault detection. A large collection of known fault samples is necessary to generate reliable deep learning models. However, in real-world applications, it is very difficult to precollect all possible types of fault or anomaly conditions to train the model. When an unknown sample is fed into a traditional deep learning model, the model may make misclassification and recognize this sample as one of the known classes. To address this issue, the concept of OSR was proposed, whose goal is not only to classify known classes but also to detect unknown samples. Existing OSR methods are commonly based on complex generative models, and the testing phase is divided into multiple stages, which makes it difficult to apply these OSR methods to real-time industrial tasks. In this article, we proposed an end-to-end OSR method with less computational cost based on a streamlined and lightweight convolutional neural network. The proposed method is applied to detect different levels of air leakage on the pneumatic train door subsystem. Compared with other existing OSR methods, the proposed method dramatically shortens the running time and reduces the use of parameters in the model with negligible loss of OSR accuracy.

中文翻译：

---

使用开集识别对气动火车门子系统进行漏气检测


对于气动列车门，持续监控和早期检测漏气将大大减少列车停机时间，确保铁路系统的安全性和高可用性。深度学习是一种从大量传感器数据中提取有用信息的强大工具，广泛应用于故障检测。要生成可靠的深度学习模型，需要大量已知故障样本。然而，在实际应用中，很难预先收集所有可能类型的故障或异常情况来训练模型。当将未知样本输入传统深度学习模型时，该模型可能会错误分类并将该样本识别为已知类别之一。为了解决这个问题，提出了OSR的概念，其目标不仅是对已知类别进行分类，而且还可以检测未知样本。现有的OSR方法通常基于复杂的生成模型，并且测试阶段分为多个阶段，这使得这些OSR方法很难应用于实时工业任务。在本文中，我们提出了一种基于精简且轻量级卷积神经网络的计算成本较低的端到端 OSR 方法。该方法适用于检测气动火车门子系统上不同程度的漏气情况。与其他现有的OSR方法相比，该方法极大地缩短了运行时间并减少了模型中参数的使用，并且OSR精度的损失可以忽略不计。