Abstract: To address the challenges of poor dynamic adaptability in feature extraction and the huge computational overhead constraining complex monitoring models during titanium alloy curved surface machining, a cutting-transient modulated s-transform and MobileNetV3-small-GRU-attention network (CTM-MGA) is proposed for ball-end mill wear monitoring. By employing the CTM locally adaptive time-frequency encoding method and introducing the γ-Mapping and K-Boost dual-feature dynamic enhancement strategies, the wear stage classification of ball-end mills is achieved on the constructed efficient MGA architecture. The results demonstrate that this method effectively extracts degradation features under strong alternating loads, achieving a classification accuracy of up to 93.39%, and maintaining stable recognition performance in the cross-condition stress test of concave surface milling. Meanwhile, with an extremely small parameter size of 0.98 M and an inference latency of 12.52 ms, it exhibits excellent real-time inference performance. The study indicates that, under titanium alloy curved surface machining conditions characterized by strong alternating loads, the CTM-MGA model clarifies the performance boundaries for state recognition against strong interference from complex contact boundaries, demonstrating significant potential for online monitoring applications.