TSPulse is a compact time series pre-training model with only 1 million parameters. It is specialized for powerful performance on classification, anomaly detection, missing data compensation, and retrieval tasks. It learns in both time and frequency domains through dual space mask reconstruction to capture complementary signals. It also uses dual embedding separation to generate detailed embeddings for detailed analysis and high-dimensional semantic embeddings for broad task understanding. At the task level, it integrates TSLens, a fine-tuning component that enables task-specific feature attention, and multi-head triangulation techniques that improve anomaly detection by correlating the biases of multiple prediction heads and fusion of complementary model outputs. It also proposes hybrid mask pre-training to improve zero-shot missing data compensation by reducing pre-training bias. With these architectural and task innovations, we achieve performance gains of 5–16 % on the UEA classification benchmark, +20% on the TSB-AD anomaly detection leaderboard, +50% on zero-shot missing data compensation, and +25% on time series retrieval. We achieve these results with just 1 million parameters, which is 10-100 times smaller than existing state-of-the-art models, enabling inference without GPUs.
