Accurate joint angle estimation based on physiological signal-driven wearable technologies remains challenging due to their limitations, such as sensitivity to motion artifacts and signal drift, as observed in electromyography (EMG) and inertial measurement units (IMU). A-mode ultrasound signals provide real-time, non-invasive, and muscle-specific dynamic information, making them a promising alternative for kinematic parameter estimation. However, relevant research remains limited. This study focuses on extracting kinematic parameters by addressing the high-dimensional and time-dependent nature of A-mode ultrasound signals, aiming to enhance their accuracy in predicting lower limb joint angles. We propose an algorithm combining Long Short-Term Memory (LSTM) networks with a multi-path decoupled feature mapping module and a Dilated Convolutional Block Attention Module (DCBAM). The DCBAM-LSTM network captures temporal features of ultrasound signals, while the mapping module translates high-dimensional features into specific joint angles with reduced complexity. Experimental results show high prediction accuracy for hip, knee, and ankle joints, with robust performance across different prediction horizons. Prediction error increases nonlinearly with longer lead times, primarily due to posture adjustments and center-of-mass shifts. The proposed algorithm demonstrates strong accuracy, real-time capability, and generalizability, offering reliable support for motion intent recognition and human-machine interaction development based on A-mode ultrasound signals.