Convolutional Neural Networks in Civil Engineering: A Contribution towards Systematic Model Design addressing Limited Data, Costly Labeling, and Explainability

Convolutional Neural Networks (CNNs) have proven to be powerful Machine Learning (ML) tools for processing unstructured data, such as images and signals. However, their application in the construction industry is often particularly challenging due to limited data availability, high labeling costs, and the need for model explainability. To address these challenges, this thesis proposes a systematic approach for the model design of CNNs to improve efficiency, robustness and transparency enabling a more efficient use of CNNs in the construction industry. To incorporate engineering knowledge into the development of CNNs, this study proposes a Maximum Receptive Field (MRF) rule that complements the Nyquist criterion and provides a theoretical foundation for the design of CNNs with unstructured data. The MRF here describes the maximum signal length that can be processed (or perceived) by the respective CNN at once. The results show that individual hyperparameters are of minor importance and that the hyperparameter space can be significantly reduced by applying the MRF rule. This potentially eliminates the need for extensive tuning of hyperparameters for a wide range of applications, saving computing resources and energy. Conversely, it is also important that the requirements of civil engineers for the application of CNNs are met. Therefore, the study examined how Explainable Artificial Intelligence (XAI) methods can be used to gain additional insights, increase confidence in models, and optimize them. In addition to explainability, techniques for dealing with limited data availability and high labeling costs were also explored. In civil and construction engineering, real-time data analysis on-site is often constrained by limited computational resources. This research suggests that the use of raw data, such as acceleration signals, can be an effective alternative to more complex, resource-intensive representations, offering the potential for significant efficiency gains while maintaining or improving model performance. This approach is particularly advantageous for applications where computational power is limited, such as in monitoring systems. By combining these findings, a significant contribution is made to the more robust, appealing, and reliable use of CNNs in civil engineering. The MRF rule could also have an impact beyond the field of civil engineering by enabling more efficient optimization of CNN models, thereby achieving substantial energy savings.