Research And Implementation Of Behavior Detection Based On Fuzzy Sets

With the maturation of sensing technology and mobile computing,sensor-based behavior detection becomes feasible and has become the focus of research in activity recognition and behavior detection.However,under realistic conditions,there are still many problems that need to be overcome in sensor-based behavior detection.Such as selecting the attributes to be measured,constructing a portable,inconspicuous and inexpensive data acquisition system,collecting data under realistic conditions,and supporting new users' flexibility without retrain the system.This thesis studies the behavior detection algorithm to solve these problems,and proposes a Neural Turing Machine model based on Capsule.By constructing capsule,neurons in NTM are transformed into capsule.Using capsule to represent some of the properties of the entity,through the transformation matrix linear transformation to predict the next layer of the capsule.This makes the NTM's long timing processing ability stronger and accelerates the training process of the algorithm.At the same time,in the process of using Capsule theory,we found that the coupling coefficient evaluation algorithm is not accurate enough.This thesis proposes a Bayesian-based dynamic routing algorithm to improve the calculation of the coupling coefficient and make the weights between the capsules more accurate.Finally,this thesis improves the memory matrix module in the NTM.The matrix structure is improved from two-dimensional to threedimensional.With the same amount of data,each address can store more data and the total address becomes less,making the read/write head faster to access and read the memory matrix,reducing the interaction frequency of the controller and the memory matrix.Meanwhile,training time is reduced.In order to solve the problem of missing tag data,and make behavior detection to adapt to various sensor environments,this thesis proposes a data transfer algorithm based on fuzzy sets.It combines radial basis functions and fuzzy sets to map data samples from high-dimensional to low-dimensional space.It not only utilizes the characteristics of the original spatial features,but also utilizes the information of high-dimensional features to correct the propagation of tags.This algorithm can obtain high-quality tagged data from a large number of unlabeled sensor data,providing more samples for NTM training.Finally,combining the NTM based on Capsule and the transfer learning based on fuzzy sets,this thesis constructs a dynamic adaptive sensor behavior detection system for practical application.Through the continuous development of the model,it can adapt to more different environments.