Pouring Sound-Based Cross-Container Water Temperature Recognition Using Single-Container Training
Open Access
- 25 February 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 32350-32364
Abstract
Acoustic-based temperature recognition (ATR) for flowing water offers a non-contact sensing alternative for various applications. However, practical deployment faces a critical challenge: training data is often available from only a single reference container, yet the system must operate across diverse containers in real-world use. Container-specific acoustic characteristics, such as resonance modes, material properties, and geometric effects, can dominate recorded pouring sounds and prevent direct transfer of models trained on one container to others. We tackle this single-container training scenario using the Audio Spectrogram Transformer (AST) as the base model, combining data augmentation (DA) with cross-entropy (CE) objective and supervised contrastive learning (SupCon) to learn container-invariant temperature representations. To contextualize our approach, we first evaluate several unsupervised source separation (USS) methods for extracting temperature-related sources as potential solutions; however, all separation approaches underperform the original, unseparated data, indicating that explicit source decomposition is ineffective for this task. Our proposed DA+CE+SupCon framework does not require explicit source decomposition or ground truth separated signals; instead, it leverages data augmentation and contrastive loss to enforce similarity between same-temperature samples regardless of container identity. Evaluating this approach on a dataset of four containers (two ceramic, two glass) across three temperature ranges (cold 20-25°C, warm 50-55°C, hot 80-95°C), we demonstrate significant improvements in cross-container transferability within the evaluated material set. When training on a single container with full data and testing on others, our approach achieves 4-10% higher cross-container accuracy than models trained with standard cross-entropy loss. In extreme data-scarcity scenarios (5-shot learning), the proposed approach further yields 13-23% gains, demonstrating exceptional robustness even with severely limited training data. These results enable practical ATR deployment with minimal training data collection requirements across the evaluated container types; generalization to other materials remains unexplored and is discussed as future work.Keywords
Funding Information
- National Science and Technology Council, Taiwan (NSTC 114-2221-E-027-039)
This publication has 30 references indexed in Scilit:
- Acoustic scene classification based on three-dimensional multi-channel feature-correlated deep learning networksScientific Reports, 2022
- Multi-representation knowledge distillation for audio classificationMultimedia Tools and Applications, 2022
- Statistics of natural reverberation enable perceptual separation of sound and spaceProceedings of the National Academy of Sciences, 2016
- Deep Residual Learning for Image Recognition2016
- Review of the present status of optical fiber sensorsOptical Fiber Technology, 2003
- Traceable Temperatures2001
- Independent component analysis: algorithms and applicationsNeural Networks, 2000
- Learning the parts of objects by non-negative matrix factorizationNature, 1999
- Theory and Practice of Radiation Thermometry1988
- XVI.On musical air-bubbles and the sounds of running waterJournal of Computers in Education, 1933