Deep neural network and mixed-integer programming framework for scheduling textual construction workplace audits

Construction auditing is a vital step in designing new sites in fulfillment centers (FCs). The audits include inspection processes for newly launched buildings which ensure that the buildings meet workplace standards. In current practice the process of scheduling construction audits and prioritizing the checklists when launching new buildings is highly manual, and requires going over lengthy textual data to determine dates of implementation and sequence. The lack of visibility on actual equipment installation dates result in conducting more auditing visits than needed and significant traveling and labor costs. Natural Language Processing (NLP) has been recently proposed to automate the auditing processes in construction sites. In this research, a hybrid machine learning and mathematical optimization framework is proposed to schedule and sequence the auditing checks. The proposed model bundles the checklists and assigns visit months based on historical probability of occurrence using Mixed Integer Linear Programming (MILP). A deep neural network with fine-tuned BERT-TSDAE architecture is proposed to generate embeddings that capture semantic similarities between checklist and checklist types and predict the occurrence time of the checks. The proposed model was validated on a subset of 49 warehouses, contributing to a potential reduction of 51.3 % of visits per site due to optimal assignment of visits. The proposed Sentence BERT fined-tuned with TSDAE model over performed CNN-Glove architecture, proposed in the literature for construction auditing, and led to the highest accuracy and F1 score.