Search

Based on human perception and machine learning methods, this study proposes a measurable method for evaluating visual comfort in underground spaces. First, a comfort evaluation index based on the characteristics of human visual perception is proposed, and color features and segmentation extraction methods for intelligent methods are given. Then, using probability statistics and machine learning methods, a multi-class intelligent sorting and classification algorithm for ranking visual comfort levels is constructed and a comparison is made of the suitability of different intelligent methods for evaluating visual comfort. The random forest algorithm is then selected as the most effective measurable intelligent evaluation algorithm for underground spaces. Finally, the proposed method is compared to intelligent methods employed by previous research, and a case study, the Wujiaochang underground space in Shanghai, China, is applied as the background. Results show that the proposed method can effectively improve the quantification and refinement of human perception and evaluation of comfort in underground spaces; this method will also be useful in computer-aided generative design in the future.

With the rapid development of underground infrastructure and the uncertainty of its location, the possibility of damage due to nearby construction has increased. Thus, for the early warning of dangerous construction behaviors around underground facilities, this paper proposes a novel real-time distributed monitoring method with three levels, comprised of the terminal node, relay node, and server. Corresponding to these three monitoring levels, a vibration-based intelligent solution for recognizing the construction source is presented and compared with the traditional method. First, the blind source separation method was used to separate collected signals into a limited number of monitoring object sources; this helped to minimize the number of required classification categories and reduce the recognition uncertainty caused by signal mixing. Then, the mutual information (MI) method was used to select suitable vibration features, which were used as the input matrix for the resulting intelligent recognition. Finally, the construction behaviors were identified at the server based on returned features. Guided by this method, a sample dataset including pile-driving, train-operation, and environment-vibration signals was constructed and combined with a multi-layer perceptron (MLP) and a long short-term memory (LSTM) network. The effects of blind source separation and the MI method are discussed in depth in this paper.