Abstract: Ultrasonic guided wave-based nondestructive testing is an effective technique for assessing the structural health of rock bolts in complex environments. However, challenges such as wave packet overlap and echo signal loss often compromise the accuracy of crack detection. To overcome these issues, this study proposes a WFE multi-source data fusion-based convolutional neural network (CNN). The method integrates the original waveform (W), fast Fourier transform spectrum (F), and instantaneous spectral entropy (E) of guided wave signals to construct multidimensional feature inputs. The optimal low-frequency range and dispersion characteristics for bolt inspection are determined through theoretical analysis. An experimental ultrasonic guided wave platform is established to collect echo signals from rock bolt cracks under varying locations, excitation frequencies, and damage severities. An improved AlexNet architecture is developed to perform both crack severity classification and defect location regression. The results indicate that, under a single-signal representation method, the WFE fusion method achieves a classification accuracy of 99.48%, which increases to 99.65% under multi-signal representation incorporating STFT and CWT time-frequency analysis. For defect location prediction method, the WFE method yields root mean square errors of 0.145 8 m and 0.118 0 m, and coefficients of determination of 0.747 5 and 0.850 9 under single and multi-signal representations, respectively, demonstrating higher predictive accuracy and robustness. This study demonstrates that the proposed method can effectively integrate multi-source time-frequency features to enhance the accuracy and reliability of crack identification and localization in rock bolts.