An AI-based Automatic Risks Detection Solution of Plant Owner’s Requirements in Equipment Purchasing-Order

Abstract

Maintainability and project investments for the plant owner are exposed to various uncertain risks from the bidding to execution stages. Technical specifications as the owner’s requirements that are not sufficiently reviewed at the bidding stage may directly or indirectly affect future performance and cause disputes with the contractors and equipment suppliers. Therefore, it is necessary to review the key risk clauses in the technical specification at the bidding stage. In this study, an AI-based automatic risks detection model was developed to detect the owner’s requirements from the technical specification during the bidding stage and purchasing order (PO) thereafter. B. Methodology The technology for AI-based automatic risks detection for technical specifications is: first, to develop a table recognition and conversion technique as well as a comparison model on the technical specifications between the owner and the supplier in bidding for equipment purchasing contracts. An automatic table recognition technology in a document was developed using Optical Character Recognition (OCR) technology. To this end, the tables in technical specifications were categorized by table format patterns, and the recognition accuracy was inversely proportional to the complexity of the tables. After table recognition, the results were converted into a database. To compare the tables provided by the owner and the supplier, the artificial neural network Ditto model was applied to match the entities in the table. Then, the table comparison analysis rules were applied to compare and analyze the contents of the tables between the owner and supplier. Analysis results were automatically extracted with a summary table. Second, for the text recognition and comparison model for the contextual and semantic analysis of the equipment POs, the key risk clauses such as the performance test requirements and site delivery schedule were selected as the proof-of-concept targets for the contextual analysis with NLP (Natural Language Processing). This model applied specific pattern-based extraction rules to extract risk clauses. After filtering the numeric data from the extracted clauses by the phrasing matching rule, the key values were compared in a rule-based algorithm. The contextual analysis results of these two models were automatically extracted into a summary table (Fig 1). This model has been developed using the Python programming language. All documents of this study consist of the English language due to the English NLP engines’ applicability but will be expanded to Korean in the near future. The pilot tests were conducted with the developed model for the motor drive PO with a fairly good match with manual comparison.