Explainable fault diagnosis for smart cities

智能城市是基于无线传感器网络的。无线传感器网络中的故障和错误校准，如果没有被检测到，可能会降低为自主决策收集的“大数据”的质量，而这在智能城市应用中是必不可少的。对可靠故障诊断的需求在智能基础设施中尤为突出，智能基础设施是智慧城市的重要组成部分。智能基础设施配备了无线传感器网络，可以自动收集、分析和传递结构数据，称为“智能监控”。虽然故障诊断概念在相关研究领域并不新鲜，但这些概念没有跟上正在进行的“智能化”的步伐，无法适应智能基础设施。该项目旨在为部署在智能基础设施中的无线传感器网络开发一个故障诊断框架。与通常用于在分布式系统中实现容错的分析冗余方法不同，该项目提出了一种基于人工智能(AI)的新方法。基于人工智能的框架的新颖性在于提出了一种用于在无线传感器网络中分布式嵌入卷积神经网络的深度学习概念的强大数学公式。除了分散化本身，无线传感器节点有限的能量和计算资源也被考虑在内。此外，还介绍了一种普遍有效的基于分类的故障诊断问题的数学形式。基于分类的故障诊断问题公式的主要优点之一是将故障诊断问题转换为传感器数据中固有的数学特征，从而不需要分析冗余要求。在对本质上是黑箱的人工智能算法缺乏信任的推动下，基于人工智能的故障诊断框架得到了基于分类数学公式的解释界面的补充，从而增加了基于人工智能的故障诊断框架的透明度。最后，故障诊断框架通过双重验证和确认策略进行验证和验证，该策略建立在包豪斯大学魏玛分校DFG研究培训小组的结果基础上，使用实验室实验测试以及从实际运营的铁路桥记录的结构数据。通过故障诊断框架的人工智能算法，对工程师来说是可解释和透明的，预计智能基础设施将能够可靠地自我检测传感器故障和传感器错误校准-而不需要多个冗余传感器、第一原理模型(如有限元模型)或对智能基础设施的物理原理的先验知识。因此，智能基础设施自主决策的可靠性和准确性将得到增强，从而有助于降低智能城市的维护和运营成本。

Smart cities are based on wireless sensor networks. Faults and miscalibrations in wireless sensor networks, if undetected, may degrade the quality of "big data" collected for autonomous decision making, which is imperative in smart city applications. The need for reliable fault diagnosis is particularly prominent in smart infrastructure, which is an essential component of smart cities. Smart infrastructure is equipped with wireless sensor networks that autonomously collect, analyze, and communicate structural data, referred to as "smart monitoring". Although fault diagnosis concepts are not new in related research areas, these concepts have not kept pace with the ongoing "smartification" and cannot be adapted to smart infrastructure. This project aims at developing a fault diagnosis framework for wireless sensor networks deployed in smart infrastructure. Unlike analytical redundancy approaches that are usually used to achieve fault tolerance in distributed systems, this project proposes a new methodology based on artificial intelligence (AI). The novelty of the AI-based framework is a strong mathematical formulation of a deep learning concept proposed for distributedly embedding convolutional neural networks in wireless sensor networks. In addition to the decentralization itself, the limited energy and computing resources of wireless sensor nodes are also considered. Moreover, a generally valid classification-based mathematical formulation of the fault diagnosis problem is introduced. One of the key advantages of the classification-based fault diagnosis problem formulation is the absence of analytical redundancy requirements by shifting the fault diagnosis problem to the mathematical features inherent in sensor data. Propelled by the lack of trust in AI algorithms that are black-box by nature, the AI-based fault diagnosis framework is complemented by an explanation interface based on the classification-based mathematical formulation, thus adding transparency to the AI-based fault diagnosis framework. Finally, the fault diagnosis framework is verified and validated by means of a dual verification and validation strategy that builds upon the results of a DFG research training group at Bauhaus University Weimar, using experimental laboratory tests as well as structural data recorded from a real-world railway bridge in operation. Through the AI algorithms of the fault diagnosis framework, being explainable and transparent to engineers, it is expected that smart infrastructure will be enabled to reliably self-detect sensor faults and sensor miscalibrations – without the need for multiple redundant sensors, first-principle models (such as finite element models), or a priori knowledge on the physical principles of smart infrastructure. As a result, the dependability and the accuracy of autonomous decision making in smart infrastructure will be enhanced, thus facilitating reduced maintenance and operation costs in smart cities.