A convolutional autoencoder-based method for learning and ranking personal daylighting preference
•Novel method for leaning personal binary daylighting preference.•Two-stage training method: feature extraction module and preference inference module.•CAE-based feature extractor designed to monitor unseen luminance map characteristics.•Inference module: classifying daylighting preference and estim...
Gespeichert in:
| Veröffentlicht in: | Building and environment Jg. 285; S. 113595 |
|---|---|
| Hauptverfasser: | , |
| Format: | Journal Article |
| Sprache: | Englisch |
| Veröffentlicht: |
Elsevier Ltd
01.11.2025
|
| Schlagworte: | |
| ISSN: | 0360-1323 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | •Novel method for leaning personal binary daylighting preference.•Two-stage training method: feature extraction module and preference inference module.•CAE-based feature extractor designed to monitor unseen luminance map characteristics.•Inference module: classifying daylighting preference and estimating ordinal ranking.
This paper presents a two-stage training method for inferring the relative daylight preference from pairs of luminance maps using convolutional autoencoder (CAE) and relative ranking concepts. It combines an updatable CAE-based feature extraction module and a binary daylighting preference inference module including relative ranking inference. A Python-based target area selection program was developed to enable the CAE model to compress the window and background areas separately from the input luminance map, ensuring that the compressed latent pixels represent these two areas distinctly. The developed CAE model was trained first using luminance maps collected from 11 individuals in private offices. Then, the trained CAE model's encoder was transferred to be a feature extractor of the personal visual preference learning models. Each model was trained to classify binary preference feedback between pairs of visual scenes and infer ordinal ranking scores for luminance distribution characteristics. The results showed that personal daylighting preference could be classified with over 85 % accuracy, and the model successfully identified the most preferred luminance scene by comparing ranking scores. In addition, the trained CAE model was able to recognize luminance maps containing significantly different characteristics by monitoring the reconstruction performance. Compared to a convolutional neural network (CNN)-based approach, the CAE-based model could leverage the condensed pixels and compare real-time luminance map characteristics with the selected conditions using L2 norm and Euclidean distance metrics, which can guide visual adjustments towards the most preferred daylighting conditions. Therefore, this study presents a significant step towards preference-based daylighting control. |
|---|---|
| ISSN: | 0360-1323 |
| DOI: | 10.1016/j.buildenv.2025.113595 |