In the fault detection of railway vehicle gearboxes, quantitative time-domain characteristic parameters and traditional dimensionless time-domain parameters are greatly influenced by working conditions, while the credibility of single-threshold warning is relatively low. To address these issues, a gearbox fault detection method based on k-nearest neighbor algorithm has been proposed. This method selects new dimensionless parameters for gearbox safety monitoring, establishes vehicle dynamics model, builds vehicle detection system, and uses wavelet decomposition and reconstruction technology to calculate and compare the sensitivity of various parameters to vibration signals in different working conditions of gearbox, and obtains the best insensitivity characterization of new dimensionless parameters. Subsequently, the k-nearest neighbor-based anomaly detection technique is applied to the safety monitoring of gearboxes. Upon training with historical data that encompasses varying fault severities, graded warning thresholds are attained and continually refined in real-world application. This ultimately achieves staged warnings during the development of gearbox faults, effectively addressing the drawback of low credibility associated with traditional single-threshold warning methods.