本文为荷兰特文特大学（作者：Jan Thiemen Postema）的硕士论文，共151页。

雷达系统是一种大型、复杂、技术先进的机器，具有很长的使用寿命。这就意味着有很多部件和子系统可能会发生故障。荷兰泰雷兹公司开发了一整套雷达系统，其中包括SMART-L MM雷达系统，这是世界上最先进的雷达系统之一，能够探测距离2000公里的目标。为了辅助雷达的维护和维修，它配备了多种传感器，总共产生1100种传感器信号。传感器信号目前由两个程序处理，一个是内置的测试系统，它根据一组规则和一个离群点检测算法发出警报。在异常检测算法的情况下，主要的缺点是缺乏解释。即使可能检测到异常值，仍然没有为其指定任何解释或标签。为了解决这一缺陷，泰雷兹公司希望建立一个系统，能够识别和分组以前发现的行为。这就产生了以下研究问题：1. 在多大程度上可以自动诊断系统状态？（a）哪些技术可用于诊断异常值，哪些技术在第1.1节所述情况和可用数据下最适合？（b）如何评估用于提供诊断方法的质量？（c）在RQ 1.a中选择的方法如何根据RQ 1.b中的度量、训练和诊断速度相互叠加？在广泛的文献回顾的基础上，本报告提出使用一种聚类算法来提供基于注释的解释。为了找出哪种算法最有效，总共测试了七种组合。为了研究半监督学习是否比无监督学习具有更大的优势，本文还提出了一种新的基于半监督约束的自组织映射（SOM）变体，称为基于约束的半监督自组织映射（CB-SSSOM）。

测试这些算法的方法包括四个步骤：（1）预处理，（2）降维，（3）聚类和（4）评估。这是在三个合成数据集和一个真实数据集上完成的。后者由领域专家手动注释，以便于评估。最重要结果的快速概述见表1，大多数算法都是通过深度信任网络（DBN）进行降维和不降维的尝试。该报告的结论是，无监督聚类很可能不是一个可行的选择，尽管在形式上的子空间聚类仍有一些希望。然而，半监督聚类确实提供了一些有希望的结果，并且可能是一个可行的解决方案，特别是当与主动学习相结合的情况下。

Radar systems are large, complex, andtechnologically advanced machines that have a very long life-span. Thisinherently means that there are a lot of parts and subsystems that can break.Thales Nederland develops a whole range of radar systems, including the SMART-LMM, one of the world’s most advances radar systems, capable of detectingtargets at a distance of up to 2.000 kilometers. In order to aid themaintenance and repair of the radar it is equipped with a wide range ofsensors, which results in a total of 1.100 sensor signals. The sensor signalare currently processed by two programs, a Built-In Test system, which givesalarms based on a set of rules and an outlier detection algorithm. In the caseof the anomaly detection algorithm the main shortcoming is the lack ofexplanations. Even though an outlier might be detected, there is still noexplanation or label assigned to it. In order to resolve this shortcomingThales wants to create a system which is capable of recognizing and groupingpreviously seen behaviour. This results in the following research questions: 1.To what extent can the system state be diagnosed automatically? (a) Whichtechniques are available to diagnose the outliers and which are most suitablegiven the case described in Section 1.1 and the available data? (b) How toassess the quality of the methods used to provide a diagnosis? (c) How do themethods selected in RQ 1.a stack up against each other based on the metricfound in RQ 1.b and training and diagnostic speed? Based on an extensiveliterature review this report proposes to use a clustering algorithm to providethe explanations based on annotations. To find out which algorithm works best,a total of seven combinations are tested. To find out if semi-supervisedlearning provides a substantial benefit over unsupervised learning for the caseof Thales, this report also proposes a novel, semi-supervised constraint-basedvariant of the Self-Organizing Map (SOM) called the Constraint-BasedSemi-Supervised Self-Organizing Map (CB-SSSOM).

The methodology with which these algorithmsare tested consists of four steps, (1) pre-processing, (2) dimensionalityreduction, (3) clustering and (4) evaluation. This is done on three syntheticdata sets and one real data set. The latter is annotated manually by a domainexpert to ease the evaluation. A quick overview of the most important resultscan be found in Table 1. Most algorithms were tried both with and withoutdimensionality reduction performed by a Deep Belief Network (DBN). Theconclusion of the report is that unsupervised clustering is most likely not aviable option, although there is still some hope in the form of subspaceclustering. However semi-supervised clustering did offer some promising resultsand could be a viable solution, especially when combined with Active Learning.

1. 引言

1.1 研究动机

1.2 数据

1.3 研究的问题

1.4 研究方法

1.5 报告组织方式

2. 项目背景

2.1 聚类方法

2.2 降维

3. 文献回顾

3.1 故障检测

3.2 性能衡量准则

3.3 相关技术

3.4 小结

4. 研究方法

4.1 数据

4.2 预处理

4.3 降维

4.4 聚类

4.5 评估

4.6 超参数估计

5. 结果

5.1 合成数据集

5.2 真实数据集

5.3 实时运行

6. 讨论

6.1 含义

6.2 问题

7. 结论

附录A 故障诊断

附录B 分类

附录C 数据

附录D 基于约束的半监督自组织映射

附录E 软件包

更多精彩文章请关注公众号：