Interpretable data-knowledge integrated AI models with application in civil engineering

Abstract

Current artificial intelligence (AI) learning paradigms are predominantly data-driven, with deep neural networks representing the mainstream of machine learning algorithms. While these methods have achieved remarkable progress, the lack of interpretability remains one of the primary challenges in practical applications of machine learning. Research on the interpretability of machine learning is still in its early stages. Their support and willingness to help me with any issue that arose meant a lot to me. Translating theory into practice is essential to narrow the gap between machine learning research and real-world applications.

This thesis proposes a multi-level framework for interpretable AI models, comprising three distinct types: knowledge-centric, knowledge-enhanced, and data-knowledge integrated models. These models are applied in various engineering domains, including project evaluation, construction scheduling optimization, and fault diagnosis, to demonstrate their effectiveness and practical value. The main contributions of this research are summarized as follows:

1. Construction and Application of Knowledge-Centric Interpretable Evaluation Models
The knowledge-centric evaluation model emphasizes the dominant role of domain knowledge in system assessment, where knowledge is derived from expert experience, scientific principles, and logical reasoning. In this study, an interpretable decision analysis model is developed by integrating uncertainty-aware multi-criteria decision-making methods with domain knowledge. Specifically, the model is constructed based on Spherical Fuzzy Set theory and the Analytic Hierarchy Process (SF-AHP), enabling effective representation and handling of uncertainty inexpert judgments. By incorporating spherical fuzzy numbers into the hierarchical evaluation structure, the model derives attribute weights in a consistent and transparent manner, thereby enhancing interpretability and decision-making reliability. The proposed approach is applied to carbon emission evaluation in prefabricated construction, demonstrating the effectiveness of knowledge-dominant modeling in complex engineering assessment tasks.

2. Construction and Application of Knowledge-Enhanced Interpretable Optimization Models
In this research, knowledge-guided mechanisms are introduced into optimization modeling. A novel state space is designed, incorporating a three-stage heuristic scheduling rule set as the action space, along with a newly developed reward discrimination mechanism. This framework enables adaptive optimization of scheduling strategies, accelerating the search for optimal solutions and reducing the number of iterations required. The model ultimately identifies Pareto-optimal solutions and is applied to the construction scheduling of prefabricated buildings to verify its effectiveness and applicability.

3. Construction and Application of Knowledge-Integrated Interpretable Classification (Prediction) Models
This model integrates both domain knowledge and data-driven insights. On one hand, domain knowledge is transformed into effective rules; on the other hand, the rich information embedded in the data is leveraged to generate additional rules. During the rule extraction phase, both accuracy enhancement and interpretability improvement are considered as joint optimization objectives. In the reasoning phase, the model demonstrates a rigorous logical structure that integrates rule generation, extraction, and inference, emphasizing causal reasoning. The model is applied to fault diagnosis of pumping units in oil extraction, enabling accurate classification and interpretable results.

In summary, this thesis investigates the construction of interpretable decision models driven by varying levels of knowledge involvement. The proposed theoretical framework and methodological system serve as a valuable reference for research in interpretable machine learning. Corresponding rule generation, extraction, and reasoning models have been developed and empirically validated. The findings demonstrate strong applicability across various civil engineering scenarios. These contributions not only deepen the understanding of interpretable models but also advance the practical application of machine learning in engineering domains.

Date of Award	Jun 2026
Original language	English
Supervisor	Aftab Ali (Supervisor) & Jun Liu (Supervisor)