信息工程与人工智能学院知识库

      铁路信号设备是包含电源屏、轨道电路、道岔、信号机以及各种控制设备在 内的重要铁路系统安全装置，但总会由于各种各样的诸如设备质量不良、工作人 员疏忽或误操作以及外部环境等内部或外部原因，出现各种类型的故障。目前的 铁路信号设备故障诊断大部分还是采用较为传统的人工排查等方式，需要工作人 员经验丰富且尽职尽责，存在较大的不确定性，稍有不慎设备故障就有可能导致 事故或损失等不良影响。因此，如何根据现有的科学技术和设备来准确及时地诊 断铁路信号设备故障是目前的研究热点，且要准确诊断出具有随机性、复杂性和 多样性的铁路信号设备故障更不容易。本文的主要研究内容如下：

    （1）建立铁路信号设备故障文本数据集。针对铁路信号设备故障文本类型 存在数据不均衡的问题，采用了朴素随机过采样的 SMOTE 算法对不平衡数据进行 少数类故障文本的重复随机采样，由此来生成了质量较好的少数类样本，使数据 集的故障类别数量整体上达到了较为均衡的状态。

    （2）提出了针对铁路信号设备故障的短文本分类模型（ERNIE_RCNN）。由于 故障文本具有长度短、多歧义、特征稀疏等特点，采用了适合处理中文短文本的 ERNIE 和 TextRCNN 复合模型来对铁路信号设备故障进行分类诊断。通过实验证 明，本文提出的 ERNIE_RCNN 模型在分类精确度上要明显高于其他深度学习模型。

    （3）构建起基于短文本分类结果和专家评价的数据融合技术的故障诊断模 型。根据 ERNIE_RCNN 模型的精度输出和专家评价结果分别构造 D-S 证据理论的 基本概率分配值，然后将短文本分类模型和专家评价后的故障诊断结果通过 D-S 证据理论在决策级进行数据融合，最后通过实例证明了融合结果的可靠性。 本文通过将预处理后各类别数量均衡的铁路信号设备故障文本，在故障短文 本分类模型和根据 D-S 证据理论的数据融合技术来对铁路信号设备故障诊断。 这种将历史经验和现场专家评价相结合的方法很适合处理具有很大不确定性的 故障短文本，经过实验验证，两种方法都取得了很好的诊断效果，为铁路信号设 备维护人员提供了决策的参考。

      Railway signal equipment is an important safety device of railway system, including power screen, track circuit, switch, signal machine and various control equipment. However, there are always various kinds of failures due to various internal or external reasons, such as poor quality of equipment, negligence or misoperation of staff and external environment.At present, the fault diagnosis of railway signal equipment mostly adopts the traditional manual troubleshooting and other methods, which requires experienced and responsible staff, and there is great uncertainty. A slight careless equipment failure may lead to accidents or losses and other adverse effects.Therefore, how to diagnose the fault of railway signal equipment accurately and timely according to the existing science and technology and equipment is a research hotspot now, and it is also a difficulty to diagnose the fault of railway signal equipment accurately with randomness, complexity and diversity.The main research contents of this paper are as follows:

      (1) Establish railway signal equipment fault text data set.For railway signal equipment failure problems data imbalance of text type, the simple random sampling of SMOTE algorithm to a few classes of unbalanced data fault text repeated random sampling, a better quality of afterlife became a few samples from this, the number of fault category data set as a whole achieves a relatively balanced state.

      (2) A short text classification model for railway signal equipment faults (ERNIE_RCNN) is proposed. Because the fault text has the characteristics of short length, multiple ambiguity and sparse features, ERNIE and TextRCNN composite models which are suitable for Chinese short texts are adopted to classify and diagnose faults of railway signal equipment.Experiments show that ERNIE_RCNN model proposed in this paper is better than other deep learning models in classification accuracy.         (3) Build a fault diagnosis model based on short text classification results and expert evaluation.The accuracy output of ERNIE_RCNN model and expert evaluation results were used to construct the basic reliability allocation of D-S evidence theory, and then the short text classification model and expert evaluation fault diagnosis results were fused by D-S evidence theory at the decision level. Finally, an example was given to prove the high precision and reliability of fusion results.

       In this paper, the railway signal equipment fault diagnosis is carried out on the basis of the fault short text classification model and the data fusion method based on D-S evidence theory.The method of combining historical experience and field expert evaluation is very suitable to deal with the fault text with great uncertainty. Through experimental verification, both methods have achieved good diagnosis effect, which provides a reference for decision making for auxiliary railway signal equipment maintenance personnel.