跳转至

DomusFM：用于智能家居基于事件行为监测的基础模型

文章背景与核心概要

随着医疗保健、独立生活以及对功能或认知变化早期检测的需求日益增长，智能家居环境中的传感器行为监测展现出了巨大的应用潜力。然而，现有的监督学习方法需要海量且不切实际的标签数据，而现有的基础模型大多针对连续的惯性或生理传感器数据设计，难以适应智能家居中稀疏、离散且语义丰富的事件流。

为了弥补这一技术空白，本文介绍了 DomusFM——一个专门针对智能家居环境设计的领域特定基础模型。该模型采用自监督的双重对比学习范式，通过集成轻量级语言模型的语义嵌入以及专门用于二进制状态和时间模式的编码器，能够同时捕捉事件级的语义属性和序列级的时序依赖关系。经过七个公共数据集的验证，DomusFM在日常活动（ADL）识别、未来 \(k\) 个事件预测以及无监督聚类等下游任务中均表现出卓越性能，同时保持了极低的计算占用，支持在边缘设备上直接部署。

DomusFM：用于智能家居基于事件行为监测的基础模型

DomusFM: A Foundation Model for Event-Based Behavioral Monitoring in Smart-Homes

摘要

DomusFM 是一个专门针对智能家居环境中基于传感器行为监测的领域特定基础模型，它充分利用了稀疏、离散且语义丰富的事件流。为了解决当前监督方法（需要海量不切实际的标签数据）以及现有连续传感器基础模型的局限性，DomusFM 采用了一种自监督的双重对比学习范式。通过将来自轻量级语言模型的语义嵌入与针对二进制状态和时间模式的专用编码器相结合，该方法能够同时捕捉事件级的语义属性和序列级的时序依赖关系。经过七个公共数据集的验证，DomusFM 在各项下游任务（如日常活动（ADL）识别、未来 \(k\) 个事件预测和无监督聚类）中展现出优异的性能，同时保留了足够小的计算占用，适合在边缘设备上部署。

Summary

DomusFM is a domain-specific foundation model designed for sensor-based behavioral monitoring within smart-home environments, leveraging sparse, discrete, and semantically rich event streams. Addressing the limitations of current supervised approaches (which require impractical amounts of labeled data) and existing continuous-sensor foundation models, DomusFM utilizes a self-supervised dual contrastive learning paradigm. This approach captures both event-level semantic attributes and sequence-level temporal dependencies by integrating semantic embeddings from a lightweight language model with specialized encoders for binary states and temporal patterns. Validated across seven public datasets, DomusFM demonstrates superior performance in downstream tasks—such as Activities of Daily Living (ADL) recognition, next-\(k\) event prediction, and unsupervised clustering—while retaining a small enough footprint for edge-device deployment.

元数据与出版详情

Metadata & Publication Details

字段 详情
arXiv ID arXiv:2602.01910 [cs.AI]
主学科 Computer Science > Artificial Intelligence (cs.AI)
作者 Michele Fiori, Gabriele Civitarese, Flora D. Salim, Claudio Bettini
时间线 2026年2月2日提交；2026年8月13日最后修订
许可协议 知识共享署名-相同方式共享 4.0 国际版
DOI 10.48550/arXiv.2602.01910
Field Details
arXiv ID arXiv:2602.01910 [cs.AI]
Primary Subject Computer Science > Artificial Intelligence (cs.AI)
Authors Michele Fiori, Gabriele Civitarese, Flora D. Salim, Claudio Bettini
Timeline Submitted on 2 Feb 2026; Last revised 13 Aug 2026
License Creative Commons Attribution-ShareAlike 4.0 International
DOI 10.48550/arXiv.2602.01910

摘要正文

Abstract

智能家居基于传感器的行为监测在医疗保健、独立生活以及功能或认知变化的早期检测方面具有巨大潜力。在这种背景下，活动识别、预测和模式发现等任务为日常生活提供了互补的视角，支持个人日常规律与习惯及其长期变化的建模。

Smart-home sensor-based behavioral monitoring holds significant potential for healthcare, independent living, and early detection of functional or cognitive changes. In this setting, tasks like activity recognition, prediction, and pattern discovery provide complementary views of daily life, supporting the modeling of personal routines and habits, and their long-term changes.

然而，现有方法面临着严峻的局限性： * 监督学习方法： 需要海量且不切实际的标记活动数据，才能捕捉不同居民和环境中日常行为的变异性。 * 当前的基础模型： 大多专为连续的惯性或生理传感器数据设计，无法有效处理智能家居产生的稀疏、离散且语义丰富的事件流。

Existing approaches, however, face critical limitations: * Supervised Methods: Require impractical amounts of labeled activity data to capture the variability of daily behavior across diverse residents and environments. * Current Foundation Models: Mostly designed for continuous inertial or physiological sensor data, failing to address the sparse, discrete, and semantically rich event streams produced by smart homes.

为了填补这一空白，作者推出了 DomusFM，该模型采用自监督的双重对比学习范式来捕捉： 1. 事件级语义属性（通过轻量级语言模型） 2. 序列级时序依赖关系（通过针对时间模式和二进制状态的专用编码器）

To bridge this gap, the authors introduce DomusFM, which employs a self-supervised dual contrastive learning paradigm to capture: 1. Event-level semantic attributes (via a lightweight language model) 2. Sequence-level temporal dependencies (via specialized encoders for temporal patterns and binary states)

核心成就与性能

  • 可迁移表示： 能够适应异构的智能家居环境以及各种事件分析任务。
  • 评估： 通过跨 七个公共智能家居数据集 的“留一数据集法”（leave-one-dataset-out）评估进行测试。
  • 下游任务： 在以下任务中持续优于基线模型：
  • 日常活动（ADL）识别
  • 未来 \(k\) 个事件预测
  • 无监督聚类
  • 边缘计算兼容性： 具备较小的计算占用，允许直接在边缘设备上部署。

Key Achievements & Performance

  • Transferable Representations: Adaptable across heterogeneous smart-home environments and various event-analysis tasks.
  • Evaluation: Tested via leave-one-dataset-out evaluations across seven public smart-home datasets.
  • Downstream Tasks: Consistently outperforms baselines on:
  • ADL (Activities of Daily Living) recognition
  • Next-\(k\) event prediction
  • Unsupervised clustering
  • Edge Compatibility: Features a small computational footprint allowing direct deployment on edge devices.

全文与资源访问

Full-Text & Resource Access

外部参考与引用

External References & Citations

(注：本文关联的许可证图标保留自原始文档：license icon)

(Note: License icon associated with this work is preserved from the original document: license icon)