Wheel-rail force is a core indicator for measuring railway safety. The most direct and effective way to obtain wheel-rail force data is to use the force-measuring wheel pairs on inspection vehicles. However, the data collected from the force-measuring wheel pairs often contains multiple interference factors, which makes the accurate assessment of railway conditions complex. Moreover, existing algorithms are difficult to separate the required wheel-rail force signals from the complex interferences in reality. To this end, this paper proposes a robust wheel-rail force detection method that integrates an improved complexity tracking algorithm with signal feature extraction. Firstly, a new small-batch iterative strategy is adopted to extract subsets from the total wheel-rail force data set as small-batch samples, which enhances the global optimization ability of the algorithm and avoids getting trapped in local extremum. Secondly, the gradient descent algorithm based on the adaptive learning rate scheduler is used for complexity tracking, which effectively optimizes the convergence speed and overall performance of the model, making it more suitable for practical engineering. Then, the Hilbert-Huang transform method is utilized to extract the characteristic parameters of the separated wheel-rail force source signals. Finally, through the experimental verification of actual wheel-rail force data, the results show that this detection method can effectively separate the wheel-rail force signals from the mixed signals and accurately extract the characteristic parameters, providing strong data support for the monitoring of railway safety conditions.