Human Activity Recognition Algorithm Based On Multi-head-siamese Neural Network

With the rapid development and popularity of wearable devices,collecting human behavior data and realizing Human Activity Recognition(HAR)based on wearable sensors has gradually attracted the attention of many scholars.In recent years,deep learning has been able to automatically learn features at different scales and levels from a large amount of data,to a certain extent avoiding the human consumption of machine learning in feature extraction.However,in order to extract features more fully,deep learning will inevitably adopt a relatively deeper and more complex network architecture,resulting in an increase in the number of network training parameters,which will make the network difficult to train and overfitting.Therefore,based on the twin network theory,this paper designs corresponding multi-head connected neural networks for obstacle behaviors and daily behaviors involved in human behavior recognition in order to achieve better behavior recognition and reduce the number of training parameters accordingly.Aiming at the problem that there are few data sets in the field of impired behavior recognition,this paper simulates five kinds of behaviors by 36 experimenters and collects selfcollected data sets.Secondly,the occurrence of behavior requires the coordination of various body parts.For the impired behavior,the sensor is placed on the corresponding body part,indicating that there is a correlation between the sensors.As the medium of data acquisition,the sensor further explains the relationship between the sensor data.By analyzing the sensor data with normalized mutual information,the similarity between the data is judged.In addition,to solve the problem that there are few relevant data for modeling impired behavior and the recognition accuracy is low,a multi-head connected neural network is proposed based on the twin network theory.The network consists of three sub-networks,each of which has the same structural weight,and the feature extraction logic of different sensor data is consistent,so it is easy to process similar sensor data.The multi-head connected neural network applied in the field of obstacle behavior is built in convolutional mode,and the network can be called MtiSiam CNN model.However,the network has many hyperparameters,which is not convenient for manual parameters,so Bayesian optimization is introduced to set the hyperparameters.In addition,considering the relatively small amount of data in the current self-collected data set,the Adam optimizer is prone to overfitting.Therefore,Adam W optimizer is used to solve the problem of unsatisfactory L2 regularization of the Adam optimizer,so as to solve the overfitting phenomenon.Finally,experimental results show that the behavior recognition accuracy of the model reaches 96.0%.In order to further improve the recognition accuracy,the convolutional architecture of the original model is improved to a deep separable convolutional architecture,and the model accuracy reaches 97.3%.Compared with foreign literature models,the effectiveness of the Mti-Siam CNN model in the field of obstacle behavior is further proved.In the field of daily behavior,the MHEALTH dataset(which was divided into MHEALTHA dataset and MHEALTH-B dataset during the experiment)and PAMPA2 dataset have the characteristics of large data volume and many kinds of behaviors,so they are usually used by many scholars in model algorithm research.So the research on normal behavior recognition is divided into two aspects.Firstly,an algorithm for feature extraction and classification recognition based on Mti-GCATT model is proposed.This model is also composed of three sub-networks,so each sub-network has the same architecture and weight,and the feature extraction logic is same for each sensor data with similar characteristics.On the one hand,it is beneficial to the interpretability of the model,on the other hand,the number of training parameters of the Mti-GCATT model is reduced.Secondly,the Temporal Convolutional Network(TCN)is introduced into the model,and the TCN module is improved in the field of behavior recognition.Given that both the MHEALTH data and the PAMPA2 data set use 256 data volumes in a single training,Weight Normalization is not as easy to use for bulk data processing in favor of the more appropriate Batch Normalization.Secondly,Re LU activation function has the problem that neurons cannot be activated during training,so Leaky Re LU function is used instead.Finally,the improved TCN module is applied to MHEALTH data set and PAMPA2 data set,and it is proved that the improved TCN module has the highest accuracy by comparing multiple data sets respectively.