The operational condition of high-speed train axle box bearings has a direct impact on both train safety and dynamic performance. However, under complex working environments, bearing fault signals are often contaminated by strong noise interference and random impacts, making it challenging to effectively extract fault impulses and leading to reduced diagnostic accuracy. To address this challenge, this paper proposes an improved adaptive frequency band optimization strategy based on a variable iteration mechanism, aimed at enhancing fault diagnosis accuracy and robustness. The method first leverages the cyclostationarity of fault impulses to enhance the harmonic prominence index, enabling precise identification of the fault resonance band while effectively suppressing noise and random disturbances. Additionally, to overcome the limitations of fixed iteration step sizes, a variable-step iteration adjustment mechanism is introduced. By integrating energy spectrum trend analysis, the approach facilitates rapid localization and dynamic adjustment of the iteration step size, improving fault resonance band identification accuracy while reducing computation time and enhancing efficiency. This fault-driven adaptive frequency band division method addresses the shortcomings of traditional data-driven techniques, proving to be effective and superior in dealing with random impacts and strong noise interference. Simulation and experimental analyses show that the proposed method can quickly and accurately identify the fault resonance band under complex working conditions. Compared to fixed band division methods, improved power spectral density methods, and fixed-step adaptive division techniques, the proposed method offers significant advantages in signal-to-noise ratio enhancement, fault feature extraction accuracy, and computational efficiency.