Human Physical Activity Identification Based On Cascaded Self-Organizing Map Network

A large number of studies have shown that proper physical exercise can help maintain physical health,prevent many chronic diseases,and promote the metabolism and growth and development of the body.However,the high-intensity work and study pressure makes people often have no time to take care of their exercise volume,and the lack of clearly quantified exercise leads to a certain degree of enthusiasm for exercise.Therefore,all-weather real-time recording and recognition of human movements have important research value and significance.The combination of unsupervised learning algorithms that can generate unlabeled data anytime and anywhere and wearable sensor systems that can record human motion data anytime and anywhere has broad application prospects in this field.An unsupervised integrated learning method is proposed for human activity recognition tasks and a pruning strategy is introduced to reduce the cost of model training under the premise of ensuring the performance of the model and achieve effective recognition of daily human motions.The main contents of the study are as follows:1.The human motion signal analysis and preprocessing are studied.Since the self-organizing mapping network needs a lot of iterative training to obtain a stable model,too large feature data will also affect the real-time performance in actual use,so it is necessary to reduce the amount of calculation for subsequent model training and use in the data processing stage to save model training time and improve the real-time performance of the model.In this paper,a total of 123-dimensional features are extracted in the feature extraction stage.In order to reduce the redundancy of feature vectors and improve the speed of model training,this paper uses principal component analysis to reduce the dimension of data,reducing the dimension of feature vectors from 123 to 26 dimensions.The experimental results show that the dimensionality reduction operation will not significantly affect the model recognition accuracy.After feature reduction,the model training time is reduced by about 68%.At the same time,the degree of contribution of the features of different dimensions to the 26 principal components is evaluated according to the coefficients of the principal component conversion matrix,and the extraction of 25 features whose contribution to the principal component is close to 0 is eliminated,reducing the time required for feature extraction by about 16%.2.Introduce the cascade structure into the self-organizing map network,and propose a new clustering integration method.The proposed cascading self-organizing map network uses one-hot-vector coding method to fuse the output of multiple base self-organizing map networks as the input of the next layer of selforganizing map network.It can effectively improve the performance on human activity recognition tasks.The experimental results show that the cascade structure can improve the accuracy of the self-organizing map network in human activity recognition tasks by about 5%.From the perspective of specific actions,the cascade model can effectively reduce the probability of misidentification between confusing actions.In addition,since the cascaded self-organizing map networkis completely composed of self-organizing map networks,it inherits the characteristics of self-organizing map networks that are easy to improve.3.A growth and pruning strategy is proposed to improve the inherent defects of the original selforganizing map network,so that the improved model has a more flexible network structure to adapt to the pattern distribution of the input space.After experimental verification,the introduction of the pruning mechanism into the growing self-organizing map network is proved to be able to effectively control the number of neurons in the model,which reduces the network training time by about 28%,and the storage space required to store the trained model is also reduced by 14.93 %.In addition,the SOM network with the growth and pruning mechanism is as accurate asthe original SOM network in terms of overall activity recognition or specific activity recognition.At the same time,due to its growth characteristics,the improved model can also learn dynamic data sets in incremental learning scenarios.Experiments show that when the model is used for incremental learning,while the new activity categories can be effectively recognized,the recognition accuracy of the original activity categories will not be seriously reduced.