AutoTherm: A Dataset and Benchmark for Thermal Comfort Estimation Indoors and in Vehicles

Overview

• This paper explores the gap between thermal comfort estimation in buildings and vehicles, which have different thermal state dynamics.
• The researchers present a new dataset for indoor thermal comfort prediction and compare deep neural network models for indoor and in-vehicle thermal comfort estimation.
• The key findings indicate that using time series data leads to superior performance compared to single-vector input, and modern machine learning architectures can effectively recognize and estimate human thermal comfort states.

Plain English Explanation

The paper examines the differences in thermal comfort estimation between indoor building environments and vehicle interiors. In buildings, thermal comfort is a well-studied topic, and researchers have collected human feedback on comfort levels. However, the thermal conditions in buildings are relatively static, so this data may not apply to more dynamic situations like inside a moving vehicle.

To explore this gap, the researchers created a new dataset for predicting thermal comfort indoors. They set up a climate-controlled chamber and had 18 people provide feedback on their comfort levels, along with 31 different sensor readings. The researchers also collected data from 20 people in a BMW 3 Series vehicle, gathering their judgments on thermal comfort.

By comparing deep learning models trained on the indoor and in-vehicle datasets, the researchers found that using time series data (a sequence of sensor readings over time) led to better performance than using a single snapshot of sensor data. The time series approach allowed the models to better capture the dynamic nature of thermal comfort.

The researchers also demonstrated that modern machine learning techniques, like recurrent neural networks, can effectively recognize and predict human thermal comfort states, both in the present and future. This could enable automatic systems to maintain optimal comfort levels in buildings or vehicles.

Overall, the study highlights the importance of considering the unique thermal characteristics of different environments when designing comfort estimation systems. The new datasets and modeling approaches provide a foundation for further research in this area.

Technical Explanation

The paper first establishes that while thermal comfort in indoor building scenarios is a well-studied topic, the static nature of these environments limits the applicability of the collected data to more dynamic situations, such as inside a vehicle.

To address this gap, the researchers introduce a new dataset for indoor thermal comfort prediction. They set up a self-built climatic chamber and had 18 subjects provide self-labeled thermal comfort ratings, along with 31 input signals, such as temperature, humidity, and air velocity. This temporal dataset aims to capture the dynamic nature of thermal comfort in an indoor setting.

For in-vehicle scenarios, the researchers acquired a second dataset featuring human judgments on thermal comfort from 20 subjects in a BMW 3 Series vehicle. This allows for a direct comparison between indoor and in-vehicle thermal comfort estimation.

The researchers then train and evaluate deep neural network classifiers on both datasets, comparing the performance of models using time series data (a sequence of sensor readings over time) versus single-vector input. Their experimental results indicate that the time series approach leads to superior performance, as it can better capture the dynamic nature of thermal comfort.

The paper also details the training of a recurrent network-based classifier and provides an initial performance benchmark for the proposed indoor thermal comfort dataset. Finally, the researchers compare their collected datasets to publicly available thermal comfort datasets, such as those mentioned in related work like Phone-based Distributed Ambient Temperature Measurement System, Unifying F1Tenth Autonomous Racing: Survey, Methods, and Benchmarks, Collaborative Perception Datasets for Autonomous Driving: A Survey, Data-driven Building Energy Efficiency Prediction Using Deep Learning, and Experimental Evaluation of Deep Reinforcement Learning Algorithms for HVAC Control.

Critical Analysis

The paper provides a valuable contribution by highlighting the differences between thermal comfort estimation in static indoor environments and more dynamic in-vehicle scenarios. The creation of the new indoor thermal comfort dataset is a significant step forward, as it allows for a more comprehensive evaluation of thermal comfort prediction models.

However, the paper does not address potential limitations of the dataset, such as the small number of participants or the specific settings of the climatic chamber. Additionally, the in-vehicle dataset, while useful for comparison, is limited to a single vehicle model, which may not represent the full diversity of in-vehicle thermal conditions.

Furthermore, the paper could have explored the potential impact of other factors, such as individual differences in thermal perception or the influence of clothing and activity levels, on the thermal comfort estimation task. Incorporating these additional variables could lead to more robust and generalizable models.

Finally, the paper could have delved deeper into the interpretation of the deep learning models' predictions, providing insights into the specific features or patterns that the models are using to recognize and estimate thermal comfort states. This could inform the design of more interpretable and explainable thermal comfort systems.

Conclusion

This paper presents a timely and relevant exploration of the gap between thermal comfort estimation in buildings and vehicles, which have different thermal dynamics. By introducing a new indoor thermal comfort dataset and comparing deep neural network models for both indoor and in-vehicle scenarios, the researchers provide valuable insights into the limitations of applying static building data to more dynamic environments.

The findings suggest that using time series data and modern machine learning architectures can lead to superior performance in thermal comfort estimation, paving the way for more accurate and adaptive comfort systems in buildings and vehicles. The datasets and modeling approaches introduced in this work offer a foundation for further research in this important area of thermal comfort analysis.