Predicting the presence of hazardous materials in buildings using machine learning

Identifying in situ hazardous materials can improve demolition waste recyclability and reduce project uncertainties concerning cost overrun and delay. With the attempt to characterize their detection patterns in buildings, the study investigates the prediction potential of machine learning techniques with hazardous waste inventories and building registers as input data. By matching, validating, and assuring the quality of empirical data, a hazardous material dataset for training, testing, and validation was created. The objectives of the explorative study are to highlight the challenges in machine learning pipeline development and verify two prediction hypotheses. Our findings show an average of 74% and 83% accuracy rates in predicting asbestos pipe insulation in multifamily houses and PCB joints or sealants in school buildings in two major Swedish cities Gothenburg and Stockholm. Similarly, 78% and 83% of recall rates were obtained for imbalanced classification. By correlating the training sample size and cross-validation accuracy, the bias and variance issues were assessed in learning curves. In general, the models perform well on the limited dataset, yet collecting more training data can improve the model's generalizability to other building stocks, meanwhile decreasing the chance of over fitting. Furthermore, the average impact on the model output magnitude of each feature was illustrated. The proposed applied machine learning approach is promising for in situ hazardous material management and could support decision-making regarding risk evaluation in selective demolition work.

Automated summary

This study explores the detection patterns and prediction potential of hazardous materials in buildings using machine learning techniques, showing high accuracy rates for predicting asbestos pipe insulation and PCB joints or sealants. The research also evaluates bias and variance issues in learning curves and emphasizes the importance of collecting more training data to improve model generalizability.