Abstract: To address the issues of high cost and strong subjectivity associated with manual observation in dairy cow estrus detection, as well as the high computational power consumption caused by high-dimensional model features, this study proposes a dairy cow estrus detection method based on unsupervised clustering and interpretable feature selection. The method utilized the K-means algorithm to cluster raw activity time series data collected by neck-mounted accelerometer on dairy cows to automatically generate estrus behavior labels. The study compared three models: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Random Forest (RF). These models were evaluated under time windows of 10 s, 30 s, and 50 s. Finally, the optimal model was determined by the Friedman test and the Nemenyi post-hoc test. And the Shapley Additive Explanations (SHAP) method was introduced to quantify feature contributions and screen for the key feature. The results indicated that the RF model based on a 10 s time window is the best. Through SHAP selection, the feature was reduced from 25 to 3. The performance of the dimension-reduced model showed no significant deviation, with accuracy, precision, recall, and F1 score reaching 0.9911, 0.9370, 0.9955, and 0.9654 respectively. In independent trials based on three-axis acceleration data from dairy cows, the RF model achieved an estrus detection accuracy of 0.9090 using a 10-second window following dimensionality reduction. The proposed method enables both estrus label generation and feature dimensionality reduction, whilst maintaining model interpretability. This provides technical support for developing low-power, long-endurance estrus monitoring devices for dairy cows.