Understanding the evolution law of rail service performance is of great significance for reducing the operation and maintenance costs of heavy-duty railway rails. Due to the complex and variable operating environment of rail tracks, which makes it difficult to construct scientifically effective damage evolution indicators to reflect objective development patterns, a method based on t-SNE(t-distributed stochastic neighbor embedding) was proposed for constructing the evolution law of corrugation damage. Firstly, the time-domain, frequency-domain, statistical, and entropy features were extracted from the original rail corrugation vibration signal. The random forest algorithm was then used to rank the features by importance, and the top-ranked features were selected to construct the feature vector. Dimensionality reduction was performed using t-SNE and other methods, and it is found that t-SNE demonstrates superior performance. The final temporal damage degradation index is obtained through Euclidean distance metric and median filtering for smoothing. The results indicate that this method provides good discrimination, anti-interference capability, and practical applicability for damage stages classification.