Towards a Comprehensive Facąde Inspection Process: An NLP based Analysis of Historical Facąde Inspection Reports for Knowledge Discovery

Facąde condition assessment for buildings is essential to public safety in cities. Currently, twelve major cities across the U.S. ensure the building facąde safety with mandatory facąde inspection programs. Even in major cities with the facąde inspection programs, there have been seventeen falling debris accidents reported in 2019, three of which were fatal. These accidents indicate a need to improve the current facąde inspection practice. Shadowing work conducted by the research team with expert inspectors on three buildings, and analysis of facąde inspection programs and guidelines show that inspectors check facądes based on defect types or on facąde components, whereas existing documentation to guide inspectors are based on major material types. This mismatch results in inspectors checking facąde components based on their experience, which might not align well with the expectations of agencies. Systematic and detailed assessment guidance is necessary to get a comprehensive and consistent facąde inspection. Towards such systematic guidance and to understand the underlying reasons for continuing accidents, this paper provides the details of an approach to identify generic vocabularies and the relationships between major entities that play a role in the inspection domain for systematic inspection processes. To identify these, we developed a data-driven approach that analyzed around 100 facąde inspection reports that were filed to the NYC Department of Buildings (DOB) during the past inspection cycle (2014-2019). Among the twelve major cities, New York City (NYC) has the longest history of facąde inspection, and most buildings (14,000) enrolled in the facąde inspection program. We believe that study about NYC buildings can provide a general understanding of inspection requirements in other cities where similar problems exist. The developed mechanism is based on natural language processing and unsupervised machine learning techniques and is used to extract the vocabularies of facąde elements, defect types, associated defect attributes, and mapping between them. The results also provide the mapping relationship of facąde components and defect types for a specific facąde type (e.g., stone/limestone). This work provides the foundation for an ontology to be used to systematically guide facąde inspection for any given building.