Reliability-coupled evidence fusion strengthens construction behavior recognition across shifting domains

A study published in Scientific Reports on 24 October 2026 reports a machine-learning method designed to keep construction behavior recognition reliable when acquisition conditions or participants differ from the training data. The method, reliability-coupled domain-adaptive evidence fusion (RDAEF), developed by Changbo Liu, Cheng Zhong, Liangxian Li, Jing Lian and colleagues, achieved target-domain macro-F1 scores of 0.798 on the Construction Meta Action (CMA) benchmark and 0.783 on the Scaffolding Worker IMU Time-series (SWIT) benchmark, exceeding the strongest unsupervised baselines by 3.4 and 5.9 percentage points respectively.

The authors identify two weaknesses in existing approaches. Domain adaptation can admit target samples that the model cannot actually support, meaning the system accepts inputs it has no reliable basis to classify. Multimodal fusion, which combines information from several sensor streams, can become overconfident precisely when those streams are degraded, missing, or temporally misaligned, so a fused decision may look authoritative even when one channel has gone bad.

RDAEF tackles both weaknesses by deriving a shared diagnostic state from multiple reliability signals. These include signal quality, the distance between a sample and learned class prototypes, evidential uncertainty, and open-set risk, which flags inputs that may not belong to any known class. When multiple sensor streams are available, the diagnostic state also incorporates temporal lag between streams and opinion conflict, a measure of how much the different modalities disagree with one another.

This shared state then plays two distinct roles. During training, it is mapped to a target-admission weight, controlling which target-domain samples the adaptation process is allowed to rely on. During inference, it modulates evidence discounts that are specific to each modality, so a degraded or conflicting sensor stream contributes less to the final decision rather than dragging the fused result astray. Evaluation relied on two complementary configurations: the CMA benchmark supplies RGB clips for testing adaptation across acquisition conditions, while the SWIT benchmark supplies synchronized pose and inertial sequences for testing adaptation across participants and for multimodal fusion. The reported macro-F1 figures are model-estimated performance on benchmark data, not field-deployment measurements.

An ablation reported in the study addresses the role of conflict handling. Removing conflict discounting increased the risk–coverage area (RCA) from 0.118 to 0.198, and a higher RCA indicates poorer selective decisions, meaning the system could no longer judge as accurately when to answer and when to abstain. Within this ablation, the reported change indicates that conflict discounting contributed to the method’s selective reliability rather than being an incidental design choice. By coupling admission of target samples during training with per-modality evidence discounts at inference, RDAEF keeps unreliable evidence from masquerading as confident classification, a property the authors position as essential when acquisition conditions or participants differ from the training data.

Subject of Research: Dynamic construction behavior recognition through domain-adaptive multimodal evidence fusion

Article Title: Dynamic construction behavior recognition through domain-adaptive multimodal evidence fusion

Article References: Liu, C., Zhong, C., Li, L., & Lian, J. (2026). Dynamic construction behavior recognition through domain-adaptive multimodal evidence fusion. Scientific Reports. https://doi.org/10.1038/s41598-026-74823-7

Image Credits: AI Generated

DOI: 10.1038/s41598-026-74823-7

Keywords: Dynamic, construction, behavior, recognition, domain-adaptive, multimodal, evidence, fusion, scientific research