Lying posture recognition is of great significance for pressure ulcer prevention, obstructive sleep apnea syndrome, and sleep quality assessment. Correct lying posture is particularly important for certain special patients as it can shorten the course of the disease and accelerate recovery. This paper discusses the transfer effects of convolutional neural networks based on two different signal processing methods, AlexNet and InceptionTime, for the problem of lying posture recognition using piezoelectric signals collected by the SleepMatrix@ device. Three different transfer strategies are applied to both networks to transfer the knowledge learned from a 5cm mattress thickness to the task of classifying 20cm mattress thickness in a semi-supervised scenario. Ultimately, applying the Maximum Classifier Discrepancy do-main adaptation method on InceptionTime achieved an accuracy of 77.9%, which represents a significant improvement compared to the traditional fine-tuning method.