Sleeping posture significantly impacts human health. To address the challenges of high cost and high complexity in existing sleeping posture recognition techniques, we propose a novel method leveraging attention mechanisms and spatio-temporal extraction of body pressure features from an air cushion. The method utilizes a partitioned structured air cushion to unobtrusively collect non-image pressure data from various body parts. Spatial structural features are extracted by a one-dimensional convolutional neural network (1DCNN), and the key time steps are dynamically weighted using the attention mechanism, and the effective fusion of spatio-temporal features is achieved by combining with long short-term memory (LSTM). The recognition results are output through softmax classifier, while the air cushion softness is adjusted to fit the spine curve. Cross-validation results indicate that the proposed method integrates shoulder-to-hip body pressure ratios and spatio-temporal features, achieving 93% accuracy on the self-acquired dataset. The small number of parameters and easy portability provide an innovative and scalable solution for health monitoring and sleep intervention.