Recurrent or non-recurrent temperature and humidity variations trigger various damages on the inside and outside surfaces of buildings, which eventually leads to poor insulation, additional energy consumption, and expensive repairing plan. Formal thermal inspection by professionals are expensive, often inconclusive and inconvenient for continuous or frequent monitoring. Thermal condition monitoring with sensors provides data-driven knowledge of thermal properties of built environments to residents and also helps in accelerating the process of thermal inspection by the professionals. In this work, we introduce a thermal condition monitoring framework which not only simultaneously learns temporal feature representations of thermal condition (i.e., temperature, humidity etc.) from inside and outside of built environment, but also cluster assignments on the unlabeled data using deep neural network. We installed thermo-hygrometers in three different homes for at least 40 days. Temporal clustering on the latent features of thermal data provides the pattern of indoor thermal conditions during different outside weather conditions. We demonstrate how indoor thermal variables respond to the outdoor thermal condition for each of the cluster patterns. Our proposed new algorithm for temporal clustering is evaluated on our thermal dataset from three buildings and compared with other temporal clustering algorithms, such as k-shape, Deep Temporal Clustering (DTC), Self-Organizing Map (SOM) from recent studies. For each of the considered buildings, our framework achieves better clustering metrics and provides a data-driven approach to monitor indoor thermal response which saves time and resources for visual inspection by professionals and keeps the homeowners informed about the thermal condition of home with no knowledge on structural properties of buildings.

经常性或非经常性的温度和湿度变化会在建筑物的内外表面上造成各种损坏，最终导致隔热效果差，能耗增加以及维修计划昂贵。由专业人员进行的正式热检查价格昂贵，对于连续或频繁的监控而言通常是不确定的，并且不便。带传感器的热状况监测为居民提供了有关建筑环境热属性的数据驱动知识，还有助于加速专业人员进行的热检查过程。在这项工作中，我们引入了一个热状态监测框架，该框架不仅可以从建筑环境的内部和外部同时学习热状态的时间特征表示（即温度，湿度等），而且还可以使用深度神经网络对未标记数据进行聚类分配。我们在三个不同的家庭中安装了温湿度计至少40天。基于热数据潜在特征的时间聚类提供了不同外部天气条件下室内热条件的模式。我们演示了室内热变量如何响应每个群集模式的室外热状况。我们在三座建筑物的热数据集上评估了我们提出的新的时间聚类算法，并将其与其他时间聚类算法进行了比较，例如最近研究的k形，深度时间聚类（DTC），自组织图（SOM）。对于每个考虑的建筑物，