Despite the increasing prevalence of wearable flexible sensors for human joint biomechanics, challenges, such as limited intelligence capacity, insufficient sensitivity, and power constraints, limit their practical application. To address these issues, this work presents a low-power and low-latency edge computing algorithm that incorporates interpretable machine learning to guide dimensionality reduction for direct on-sensor signal processing. Compared to traditional wireless transmission methods, the deployed tiny machine learning (tinyML) model achieves a prediction latency of only 9 ms and reduces power consumption by 75.6%. Furthermore, utilizing a triboelectric sensor based on MXene and featuring a micro-conical structure demonstrates excellent self-powered sensing capability, with output voltage and charge increased by 27.4% and 52.9%, respectively, and a high-sensitivity monitoring performance of 16 mV/Pa. The synergy between the efficient algorithm and the high-performance sensor is validated in knee joint biomechanics scenarios, showing advantages over conventional approaches in power consumption, cost, response speed, size, and accuracy. These combined strengths indicate broad application prospects in portable intelligent healthcare.