Abstract:

Significance Human, clothing and external environment form an interactive system. As a barrier between the environment and human body, clothing directly affects the thermal comfort of people. It is indispensable to evaluate the thermal comfort or personal safety. However, individual differences would lead to discrepancy in subjective feelings and efficiency could be impaired if frequent subjective assessment is needed in working process. Therefore, effectively predicting the thermal comfort of individuals and returning timely suggestions to improve the micro-environment between clothing and body would be necessary. Conventional thermal comfort models including steady-state heat transfer models, thermal adaptative models and dynamic thermal physiology models were established based on physical equations or data from general population, without considering individual differences. Therefore, new methods should be introduced to study the personal thermal comfort. Researches have been carried out on the application of machine learning algorithms to establish personal thermal comfort models, predicting individual thermal comfort through data-driven methods. Compared with conventional models, the prediction of the personal thermal comfort models is significantly improved. The models overcome the defects of the conventional models which are complicated and inflexible, predict thermal comfort in real time, and are beneficial to improve of micro thermal environment more efficiently.

Progress The personal thermal comfort models established by machine learning could be regarded as a supervised learning process. Sample source, input features and output labels, machine learning algorithms and evaluation indicators are the main influencing factors encountered during the establishment. Sample source brings about the question of applicability. Models built upon laboratory data may not be fit for field studies, neither are models established with mild environments' data suitable for extreme conditions. The sample size for achieving stable prediction varied from models. Generally, input characteristic parameters included environmental parameters collected from the surrounding environment or meteorological platform, and individual parameters reflecting the state of the human, could be both considered when collecting input features. Subjective evaluation index as output labels depended on research purpose and the evaluation of human thermal comfort should consider at least two subjective indexes, including symmetric and asymmetric ones. When selecting machine learning algorithms, the sample size and applicability of the algorithm also should be taken into account, as well as the cost and interpretability. Evaluating the prediction performance helps to confirm the validity of models especially when conducting multi-index evaluation. Indexes such as accuracy, precision, recall are suitable for the binary-classification conditions, while Kappa coefficient could handle the multi-classification and imbalanced datasets. The models based on machine learning has a broad application prospect in clothing due to its personalization, flexibility and dynamic predictions. Developing intelligent temperature regulating clothing that could predict the thermal comfort of individuals in real time and change the control strategies accordingly has become a research hotspot. Personal thermal comfort models provide a feasible technical path by combining software and wearable hardware systems. Once achieved, the thermal security would be guaranteed and the work efficiency improved.

Conclusion and Prospect Personal thermal comfort models based on machine learning algorithm is a new method to achieve individual thermal comfort prediction, which has the advantages of user personalization, multi-dimensional input parameters and dynamic prediction. At present, some progress has been made in the research of this model, which is summarized as follows. 1) The data of the model usually come from simulated experiment environment or actual working environment, but the prediction model based on the two kinds of data is not universal. Therefore, the reasons and solutions for the differences can be further explored to expand the application scope of the model. 2) Personal thermal comfort is mainly affected by the environment and individual factors. The selection of feature parameters should adopt multi-parameter combination with different properties, and the number of feature parameters should be controlled according to different algorithms. When applied in the field of clothing research, attention should be paid to comprehensively consider the influence of clothing on human thermal regulation. 3) A variety of indicators are involved when evaluating models' prediction performance, and the evaluation objects and applicability of which should be considered. In order to overcome the limitation and incomparability of single index evaluation, future studies might focus on multi-index comprehensive evaluation to evaluate the model's prediction and generalization ability. 4) Personal thermal comfort models established by machine learning algorithm has high application value in the field of intelligent clothing. The modeling technique's improvement could provide key technical support for development of intelligent temperature regulating clothing. Accordingly, the real time thermal comfort requirements of operators would be met, while operational efficiency and thermal safety could be guaranteed.

Key words: machine learning, microclimate under clothing, thermal comfort, prediction model, intelligent clothing