.png){kind=icon}
.png){kind=icon}
Return
Transformer-Based Explainable Framework for Human Activity Recognition in IoT-Enabled Smart Homes
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2026.3661415.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The growing integration of the Internet of Things (IoT) devices in smart homes has enabled applications in comfort, energy efficiency, healthcare, and security. A key enabler of these applications is Human Activity Recognition (HAR), which enables smart systems to understand residents’ activities and deliver personalised services. While single-resident HAR has been well studied, multi-resident scenarios are more complex due to collaborative and parallel activities, requiring effective modelling of temporal dependencies. Transformer-based models have shown promising results in this context as their self-attention can capture long-range temporal dependencies and allows parallel computation. However, standard attention uses both past and future information, causing information leakage in real-time HAR. It also treats all time steps equally, ignoring the fact that event correlations in smart homes naturally decay over time and require locality bias. Since these systems involve real residents, explainability is also crucial to ensure trust and transparency in activity classification. This paper proposes a Powerformer-based multi-output explainable framework incorporating a Weighted Causal Multi-Head Attention (WCMHA) mechanism with power-law decay to emphasise recent events while maintaining causal order, improving accuracy. A Multi-Output Temporal Local Interpretable Model-agnostic Explanations (MO-TLIME) technique is proposed, which identifies key sensors and time segments influencing each resident’s activity. Evaluated on publicly available multi-resident datasets, the proposed model outperforms existing deep learning baselines in both accuracy and efficiency. Statistical validation and ablation studies confirm its effectiveness, while explainability analyses demonstrate that the model focuses on contextually relevant sensor data and temporal segments.
Keywords:
Explainable AI (XAI)
Internet of Things (IoT)
local interpretable model-agnostic explanations (LIME)
multi-resident human activity recognition
smart homes
transformer