Accurate and non-intrusive patient identification is vital for ensuring safety in clinical environments, particularly in scenarios involving bedridden or unconscious individuals. This paper proposes a novel in-bed identity recognition framework based on multi-channel piezoelectric sensor arrays and deep learning. A comprehensive dataset was constructed using an 8 \(\times \) 8 piezoelectric sensor grid, capturing 5,040 samples across 30 individuals in 168 distinct in-bed action classes covering four standard postures and multiple bed zones. Two data processing strategies were explored—image-based and time-series modeling. Experimental comparisons among AlexNet, ResNet, InceptionTime and XceptionTime demonstrated that temporal models, particularly XceptionTime, offered superior performance, achieving 98.73% average accuracy and a peak of 99.50%. Furthermore, a channel-wise averaging replacement data augmentation strategy was introduced to improve generalization under limited sample conditions, leading to performance gains of up to 9.3%. The study also investigated the influence of time window lengths, confirming that a 10-second window optimally captured temporal identity features. Overall, the proposed framework offers a robust, privacy-preserving, and real-time solution for in-bed individual recognition. Future work will focus on expanding recognition capabilities to include unregistered individuals and integrating the system into broader smart healthcare infrastructures.