Abstract: Aiming to address the difficulties in identifying hidden fatigue crack damage in steel bridge components and in the insufficient accuracy in quantitative assessments, this paper proposes a crack damage diagnosis method integrating multi-dimensional features of ultrasonic guided wave (UGW). Based on UGW tests of cracked steel plates and on a parametric high-fidelity numerical simulation model, the time-domain signals of UGWs under different crack damage conditions were obtained. Combining experimental and simulation data, nine types of damage-sensitive parameters were extracted from the multi-dimensional characteristics of UGWs, including time domain, frequency domain, energy and modes. A correlation mapping model between characteristic parameters of UGW and crack damage was established. A principal component analysis was used to reduce the dimensionality of the multidimensional characteristic parameters of UGWs. On this basis, the feedforward neural network (FNN) and extreme gradient boosting (XGBoost) were applied to achieve accurate diagnosis of different crack damage degrees and locations. The results show that the experimental and numerical results are highly consistent in terms of the first wave arrival time and of the waveform morphology. The numerical modeling method proposed for UGW is reliable. The multi-dimensional characteristic parameters of UGWs are highly sensitive to the geometric sizes and to the locations of cracks. The diagnostic model established based on the XGBoost achieved an accuracy of 93.8% and 99% in crack damage degree and location identification, respectively. The study provides a theoretical basis and a technical support for the precise detection and for the quantitative assessment of hidden cracks in steel bridges.