晨光
暗夜
晨光
极光
Bilingual Paper Reading · 中英对照精读

主动迁移学习用于结构健康监测

准大一 · 土木工程 × 结构健康监测 × AI —— 桥梁损伤分类精读材料
原文:arXiv:2510.27525 2025年10月31日发布 arXiv 预印本(cs.LG) 迁移学习 × 主动学习 × 群体式桥梁监测 附英文摘要朗读音频

一、论文档案

英文标题Active Transfer Learning for Structural Health Monitoring
中文标题主动迁移学习用于结构健康监测
作者J. 普尔, N. 德维利斯, K. 沃登, P. 加德纳, V. 吉廖尼, R. S. 米尔斯, A. J. 休斯(机构未在素材中标注)
发布时间2025年10月31日(v1)|分类:cs.LG(机器学习)
一句话概括迁移+主动学习组合拳,让少标签桥梁也能学出可靠的损伤分类。
💡 为什么选这篇给你:① 结构健康监测(SHM)是土木+AI 最热的交叉方向,直接连着桥梁群管养与运维成本;② 这篇把「迁移学习」和「主动学习」两个成熟工具第一次组合进 SHM,思路简单但工程价值大;③ 有完整的实验室缩尺桥实验(温度变化+伪损伤状态),故事闭环、方法可复现。

二、核心术语表(先扫一遍再读正文)

英文术语中文大白话解释
structural health monitoring (SHM)结构健康监测用传感器数据判断桥梁等结构有没有损伤、损伤在哪里的技术。
population-based SHM (PBSHM)群体式结构健康监测把「一群结构」的数据放在一起用——某一座桥数据不够,就用别的桥来补。
transfer learning (TL)迁移学习把在「数据充足的源结构」上学到的知识,搬到「数据稀缺的目标结构」上。
domain adaptation (DA)域适应迁移学习的一种:把源域和目标域的数据分布「对齐」,让模型能跨域使用。
unsupervised DA无监督域适应目标域完全没有标签(不知道数据对应什么状态)的域适应。
active learning主动学习让模型自己挑「最值得标注」的数据去问人,从而少标数据、省检测成本。
active transfer learning主动迁移学习把主动采样和迁移学习结合:先靠源域建模型,再在监控过程中挑最有效的数据标注。
negative transfer负迁移源域和目标域不够相关时,迁移不但没帮忙,反而把模型带坏。
class imbalance类别不平衡某些健康状态(如损伤)样本很少,模型容易「偏心」只学常见状态。
labelled data带标签数据标注了真实健康状态(正常/损伤等)的数据,获取成本高。
DA-RVM域适应相关向量机本文使用的贝叶斯分类器(RVM=相关向量机),配合域适应映射做损伤分类。
health-state健康状态结构所处的状态,如正常、轻微损伤、严重损伤等。
environmental chamber环境箱实验室里能精确控制温度等环境条件的箱子,用来模拟不同服役环境。
inspection检测(巡检)派人或仪器到结构现场检查并标注真实状态,成本高、周期长。

三、摘要中英对照(精读核心)

🎧 音频在文末,可先听一遍原文再读;每个英文句都配了逐句翻译。

摘要 Abstract

EN · 原文
Data for training structural health monitoring (SHM) systems are often expensive and/or impractical to obtain, particularly for labelled data.
CN · 翻译
训练结构健康监测系统的数据往往昂贵甚至难以获得,尤其是带标签的数据。
EN · 原文
Population-based SHM (PBSHM) aims to address this limitation by leveraging data from multiple structures.
CN · 翻译
群体式结构健康监测(PBSHM)旨在通过利用多座结构的数据来弥补这一不足。
EN · 原文
However, data from different structures will follow distinct distributions, potentially leading to large generalisation errors for models learnt via conventional machine learning methods.
CN · 翻译
然而,不同结构的数据服从不同的分布,可能导致常规机器学习方法学到的模型产生很大的泛化误差。
EN · 原文
To address this issue, transfer learning -- in the form of domain adaptation (DA) -- can be used to align the data distributions.
CN · 翻译
为解决这一问题,可以使用迁移学习——具体形式为域适应(DA)——来对齐数据分布。
EN · 原文
Most previous approaches have only considered unsupervised DA, where no labelled target data are available; they do not consider how to incorporate these technologies in an online framework -- updating as labels are obtained throughout the monitoring campaign.
CN · 翻译
以往的方法大多只考虑无监督域适应(目标域没有任何标签数据),没有考虑如何把这些技术纳入在线框架——即在监测过程中随着标签不断获得而持续更新。
EN · 原文
This paper proposes a Bayesian framework for DA in PBSHM, that can improve unsupervised DA mappings using a limited quantity of labelled target data.
CN · 翻译
本文提出一个用于 PBSHM 的贝叶斯域适应框架,可以用少量带标签的目标数据改进无监督域适应的映射。
EN · 原文
In addition, this model is integrated into an active sampling strategy to guide inspections to select the most informative observations to label -- leading to further reductions in the required labelled data to learn a target classifier.
CN · 翻译
此外,该模型被整合进主动采样策略,用于指导检测工作去挑选信息量最大的观测进行标注——从而进一步减少学习目标分类器所需的标签数据。
EN · 原文
The effectiveness of this methodology is evaluated on a population of experimental bridges.
CN · 翻译
该方法的有效性在一组实验桥梁群体上得到评估。
EN · 原文
Specifically, this population includes data corresponding to several damage states, as well as, a comprehensive set of environmental conditions.
CN · 翻译
具体而言,该群体包含对应多种损伤状态的数据,以及一套全面的环境条件数据。
EN · 原文
It is found that combining transfer learning and active learning can improve data efficiency when learning classification models in label-scarce scenarios.
CN · 翻译
研究发现,在标签稀缺场景下学习分类模型时,把迁移学习和主动学习结合起来能够提高数据效率。
EN · 原文
This result has implications for data-informed operation and maintenance of structures, suggesting a reduction in inspections over the operational lifetime of a structure -- and therefore a reduction in operational costs -- can be achieved.
CN · 翻译
这一结果对数据驱动的结构运维具有重要意义:在结构全寿命期内可以减少检测次数——从而降低运维成本

主题词(摘自摘要):Structural Health Monitoring 结构健康监测 | Transfer Learning 迁移学习 | Domain Adaptation 域适应 | Active Learning 主动学习 | Population-Based SHM 群体式结构健康监测

四、引言精选(为什么这个问题重要)

① 为什么 SHM 数据又贵又少:同分布假设在真实结构上不成立

EN · 原文
Data for training structural health monitoring (SHM) systems are often expensive and/or difficult to obtain, particularly for labelled data relating to rare health-states, such as damage data. While unsupervised approaches have been shown to be capable of detecting damage [1, 2], data-based SHM systems typically cannot provide contextual information – often encoded as labels – without labelled data. Population-based SHM (PBSHM) is an emerging field that aims to address the issue of data scarcity by considering data from across a population of structures [3, 4, 5]. However, datasets derived from different structures will not follow the same underlying generative distribution, meaning the assumption that training and testing data were drawn from the same distribution is invalid; thus, supervised machine-learning techniques will likely have a large generalisation error [6].
CN · 翻译
训练 SHM 系统的数据往往昂贵且难以获得,尤其是与稀有健康状态(如损伤数据)相关的带标签数据。无监督方法已被证明能够检测损伤,但基于数据的 SHM 系统通常无法在没有标签数据的情况下提供情境信息(通常编码为标签)。群体式结构健康监测(PBSHM)是一个新兴领域,它通过利用整个结构群体的数据来应对数据稀缺问题。然而,来自不同结构的数据集并不服从同一个潜在生成分布,这意味着「训练与测试数据来自同一分布」的假设不成立;因此,有监督机器学习技术很可能产生较大的泛化误差。

② 迁移学习有用,但有两个坑:负迁移与类别不平衡

EN · 原文
This issue motivates the application of transfer learning (TL), a field of machine learning that aims to improve the performance in a target domain (structure) by leveraging related information from a more data-rich source domain (structure) [7]. Unsupervised-transfer learning approaches, such as unsupervised domain adaptation (DA), have emerged as a promising method and have been demonstrated in a variety of SHM applications [8, 9, 10, 11, 12, 13, 14]. These methods do not require labelled data in the target, meaning that they have the potential to facilitate damage classification when label information is missing in the target, by allowing a predictive function learnt using source data to generalise to the target. Nevertheless, these approaches are not infallible and they may result in performance degradation - so-called negative transfer [7] – if the datasets are not sufficiently related. They are also particularly sensitive to issues related to class imbalance [15] – where certain classes are under-represented in the training data – which is already a pertinent issue in SHM datasets, as some health-states are naturally more common.
CN · 翻译
这促使我们应用迁移学习(TL)——它通过利用数据更丰富的源域(结构)的相关信息,来提高目标域(结构)的性能。无监督迁移学习方法(如无监督域适应)已被证明很有前景,并在多种 SHM 应用中得到验证。这些方法不需要目标域标签,因此当目标域缺少标签信息时,仍有望让源数据学到的预测函数泛化到目标域,从而完成损伤分类。但这类方法并非万无一失:如果数据集不够相关,就可能出现性能下降——即所谓的负迁移;它们还对类别不平衡问题特别敏感——训练数据中某些类别代表性不足,而这在 SHM 数据集中本就普遍存在,因为某些健康状态天然更常见。

③ 标签那么贵,不如让模型「主动点名」最值得检测的数据

EN · 原文
Given the cost of acquiring labels, it would be beneficial to schedule inspections to coincide with the most informative data. Active learning has been demonstrated to significantly reduce the label requirement in SHM by using a predictive model to infer which unlabelled data would provide the largest improvement if they were labelled [17, 18, 19, 20, 21]. In the transfer-learning literature, guided sampling strategies have been proposed to leverage source data to improve the initial model [22, 23, 24, 25], and have been demonstrated to mitigate the class-imbalance issue in DA [25]; these methods will be referred to as active transfer learning methods. However, to the authors knowledge active transfer learning has not been investigated in SHM.
CN · 翻译
考虑到获取标签的成本,把检测安排在最有信息量的数据上是有利的。主动学习已被证明能显著降低 SHM 的标签需求:用预测模型推断「哪些未标注数据一旦标注能带来最大改进」。在迁移学习文献中,也有人提出引导式采样策略来利用源数据改善初始模型,并证明能缓解 DA 中的类别不平衡问题;这些方法被称为主动迁移学习方法。然而,据作者所知,主动迁移学习尚未在 SHM 中被研究过

④ 本文方案:贝叶斯域适应 + 主动学习,三座缩尺桥验证

EN · 原文
This paper proposes the first online transfer-learning strategy for PBSHM by incorporating a novel Bayesian DA method into an active-learning framework. This online framework is able to update mappings estimated via unsupervised DA using limited labels; the general proposed approach is illustrated in Figure 1. This approach is validated using an experimental dataset consisting of three laboratory-scale bridges with varied support locations; these structures were subjected to testing over a range of damage-states and environmental conditions using an environmental chamber.
CN · 翻译
本文提出首个面向 PBSHM 的在线迁移学习策略:把一种新颖的贝叶斯域适应方法嵌入主动学习框架。该在线框架能用有限的标签更新无监督域适应估计出的映射。方法在三座不同支座位置的实验室缩尺桥实验数据集上验证:这些结构在环境箱中经历了一系列损伤状态和环境条件的测试。
💡 这是全文最有味道的一句“datasets derived from different structures will not follow the same underlying generative distribution”——机器学习里「训练/测试同分布」的默认假设,在真实桥梁面前根本不成立。看懂这一句,就懂了为什么要迁移学习。

五、方法要点(4 个要点)

EN · 原文
1. Bayesian DA framework. This paper proposes a Bayesian framework for DA in PBSHM, that can improve unsupervised DA mappings using a limited quantity of labelled target data.
CN · 翻译
1. 贝叶斯域适应框架。提出用于 PBSHM 的贝叶斯域适应框架:用少量目标标签即可改进无监督域适应映射。
EN · 原文
2. Active sampling strategy. In addition, this model is integrated into an active sampling strategy to guide inspections to select the most informative observations to label -- leading to further reductions in the required labelled data to learn a target classifier.
CN · 翻译
2. 主动采样策略。模型被整合进主动采样策略,指导检测挑选最有信息量的观测进行标注,进一步降低学习目标分类器所需的标签量。
EN · 原文
3. First online TL strategy for PBSHM. This paper proposes the first online transfer-learning strategy for PBSHM by incorporating a novel Bayesian DA method into an active-learning framework.
CN · 翻译
3. 首个 PBSHM 在线迁移学习策略。把新颖的贝叶斯域适应方法融入主动学习框架,支持监控过程中在线更新
EN · 原文
4. Data efficiency result. It is found that combining transfer learning and active learning can improve data efficiency when learning classification models in label-scarce scenarios.
CN · 翻译
4. 数据效率结论。在标签稀缺场景下,迁移学习+主动学习的组合能显著提升分类模型学习的数据效率。

六、结论中英对照

EN · 原文
A critical limitation of conventional data-driven approaches to SHM is that supervised machine learning methods require a fully-labelled dataset with examples representing each health state of interest, which is often costly and/or unfeasible. Two technologies for reducing the label requirement of supervised classifiers are transfer learning and active learning. Previous studies have considered these technologies independently; however, this paper proposes an active transfer-learning strategy to address several key challenges of considering either approach independently, resulting in a practical framework for online learning in PBSHM.
CN · 翻译
常规数据驱动 SHM 方法的一个关键局限是:有监督机器学习需要覆盖每个目标健康状态的完整标签数据集,这往往代价高昂甚至不可行。降低有监督分类器标签需求的两种技术是迁移学习和主动学习。以往研究把它们分开考虑;本文则提出主动迁移学习策略,解决单独使用任一方法时的几个关键挑战,最终形成一套实用的 PBSHM 在线学习框架
EN · 原文
Four transfer tasks were used to demonstrate the proposed framework by transferring a damage classifier between lab-scale bridge structures subject to various temperatures and pseudo-damage states. In all cases, leveraging labelled source data enabled the DA-RVM to classify health states that had not been observed in the target, even when the target dataset only contained data corresponding to a subset of the classes in the source dataset – showing robustness to class imbalance. On the other hand, conventional active-learning approaches can only classify data from previously-observed health-states. The ability to classify health-states prior to their observation in the target domain has significant implications in SHM, as predictions about health-states critical to decision making could be achieved before these health-states are observed in the target structure and without repeating labelling efforts. Furthermore, the active transfer-learning approach resulted in fewer overall queries compared to conventional active learning, which in practice would result in a reduction in inspections; hence, lower operational costs throughout a monitoring campaign.
CN · 翻译
研究用四个迁移任务验证框架:在不同温度与伪损伤状态下的缩尺桥之间迁移损伤分类器。在所有情况下,利用带标签的源数据使 DA-RVM 能够分类目标域尚未观测到的健康状态——即使目标数据集只包含源数据集类别的一个子集,展现出对类别不平衡的鲁棒性。相比之下,常规主动学习只能分类之前观测过的健康状态。能够在目标域亲眼观测到某健康状态之前就对其分类,对 SHM 意义重大:可以在目标结构出现关键健康状态之前就做出决策所需的预测,且无需重复标注工作。此外,主动迁移学习相比常规主动学习需要的总查询次数更少,实际中意味着检测次数减少、整个监测期间运维成本降低
EN · 原文
There are several interesting potential directions for future work. One of the main limitations of the current approach is that the number of observations labelled is not directly related to a labelling budget. In practice, an operator would have a limited budget for inspections; thus, the current approach may exhaust the budget prior to observing all data. Ensuring the labelling budget is not depleted early in the sampling process is a common challenge in stream-based active learning [35].
CN · 翻译
未来有几个有趣的潜在方向。当前方法的一个主要局限是:标注的观测数量与标注预算没有直接关联。实际中检测预算有限,当前方法可能在观测完所有数据之前就耗尽预算。在流式主动学习中,确保采样过程早期不耗尽标注预算是一个常见挑战。
EN · 原文
This paper assumes that labelled data are available for all classes in the source domain. In practice, obtaining such comprehensive source datasets could be challenging, and a more feasible approach might assume data are distributed across multiple source domains. Previous DA methods are often prone to negative transfer when aligning datasets with only a subset of shared classes, whereas the presented framework can effectively align data using a limited number of shared classes.
CN · 翻译
本文假设源域中所有类别都有标签数据。实际中获取这样全面的源数据集可能很困难,更可行的做法是假设数据分布在多个源域中。以往的 DA 方法在只共享部分类别的数据集对齐时容易发生负迁移,而本文框架用少量共享类别就能有效对齐数据——这为扩展到多源场景铺平了道路。

七、编者解读:这篇论文到底讲了什么(大白话版)

  1. 问题:训练「桥梁损伤识别」需要大量标注了真实状态的传感器数据,但损伤数据又贵又少;每座桥的数据分布还不一样,拿 A 桥训练好的模型直接用到 B 桥,误差会很大。
  2. 做法:两步走——先用无监督域适应把「源桥」和「目标桥」的数据分布对齐,让模型初步能跨桥用;然后在监测过程中,每做一次人工检测(拿到少量标签),就用贝叶斯方法在线更新映射,越用越准。
  3. 聪明的地方:检测那么贵,不能盲目做——模型用主动学习「点名」那些最不确定、最值得检查的数据,优先去标注它们,把每次检测的收益最大化。
  4. 结果:在三座缩尺桥(不同支座、多种温度、多种伪损伤)的四个迁移任务上,DA-RVM 能分类出目标桥从未见过的健康状态(常规主动学习做不到),而且需要的查询次数更少。
  5. 最值钱的观点:健康状态可以「先预测、后看见」——在目标结构还没出现某类损伤前,就能借助别的结构的经验提前给出判断;这意味着检测次数下降、运维成本下降。
  6. 边界:论文坦承目前标签数量与预算脱钩、假设源域各类别都有标签,并给出了多源扩展方向——不吹牛、留了路。
🎯 对保研的启示:这篇论文是「组合创新」的教科书——两个成熟工具(迁移学习、主动学习)都没发明,但把它们的组合用到新领域(SHM)并解决真实痛点(标签贵、检测贵),就是一篇好论文。复试时讲「别人各用各的,我把它俩合起来并说明为什么合起来更好」,比堆砌新名词更有说服力。

八、给准大一的阅读路线图 & 延伸方向

📖 怎么读这篇论文(三遍法)

  1. 第一遍(10 分钟):只读摘要和术语表,回答三个问题——问题是什么?方法是什么?结果是什么?
  2. 第二遍(20 分钟):读引言 + 结论,重点体会「为什么要迁移 + 为什么要主动」以及「结论的边界(预算、多源假设)」。
  3. 第三遍(30 分钟):读方法部分文字描述(贝叶斯 DA、主动采样、DA-RVM),跳过公式和编号,遇到不懂的术语回查术语表。

🚀 这个方向你能延伸做什么

九、英文摘要朗读(练听力用)

先盲听一遍→再看对照稿→再听一遍。目标是听出数字(three laboratory-scale bridges、four transfer tasks)和术语(domain adaptation、active learning、class imbalance、negative transfer)。