| Sleep time can occupy one third of a person’s life.At present,over thirty percent of adults are suffering from varying degrees of sleep problems.Conventional polysomnography monitoring is the golden standard in diagnosis and treatment of sleep disorders,but it has been restricted in use severely due to its complicated checking procedures and poor comfort.Ballistocardiogram could realize sleep monitoring by a non-intrusive approach based on non-contact sensors,which contains large amounts of physiological information related to sleep.Moreover,mining sleep postures is significant for monitoring positional sleep disorders.This paper developed the sleep BCG measuring system,studied the noise filtering and motion artifacts recognition algorithms for sleep BCG signals,and proposed two machine learning methods based on BCG signals to recognize sleep postures.The main research contents and innovative points were summarized as follows:1)The system for measuring BCG signals with high precision and low power consumption was developed,which was used for sleep posture recognition later.High-precision signal measuring and analog-digital converter circuits were designed based on a flexible polyvinylidene fluoride film sensor,and embedded software of the signal acquisition system was developed based on a low-power microcontroller.The whole system including hardware part and software part has been successfully aligned,and its reliability and security has been confirmed by the testing data.64 volunteers of different ages and genders participated in the project,and 80 cases of sleep data were collected.Experimental data sets were built by rigorous category annotation,which were used to verify effectiveness and accuracy of the proposed algorithms.2)A novel algorithm for automatic recognition of BCG motion artifacts based on statistical signal processing was proposed.By the analysis and comparison of target features,windowed range of the BCG difference sequence was selected as the statistical variable,and strong artifacts with obvious features were recognized preferentially based on Neyman-Pearson principle.On this basis,a comparison algorithm based on the combination of kurtosis coefficient and hard threshold was proposed for fine-grained detection of weak artifacts.Verified by the experimental data sets,the algorithm obtained great performance:sensitivity 99.84%,specificity 97.83%.3)Two machine learning methods for recognizing sleep postures were proposed,which foeused on five morphologieal features of BCG signals.One method was non-linear support vector machine base on adaptive acceleration extended particle swarm optimization(AAE-PSO-SVM),which improved the searching precision and efficiency of the model parameters.The other one was semi-supervised extended fuzzy k-means clustering(SSE-FKM),which set the calibration procedure to optimize initial clustering,weighted the Euclidean distance based on Shannon entropy,improved the constraint of membership function and added interval information.The results from experiments indicate that the whole precision for these two methods is 96.28%and 98.03%,respectively.Compared with other current methods for recognizing sleep postures,the methods proposed in this paper can not only achieve high accuracy,but also realize a real sense of non-intrusive sleep monitoring.In summary,this paper developed the system for measuring BCG signals with high precision and low power consumption,solved the motion artifact disturbance problem,and proposed two machine learning methods based on BCG signal to recognize sleep postures.This study could offer a new way of thinking for portable sleep monitoring,and has significant meaning for promoting screening,diagnosis and treatment of sleep disorders. |
PDF Full Text Request