Research And Application Of Closed Surface Nearest Neighbor Partitioning Neural Network Classifier

At present,neural network has proven to be an important classification technique in data mining.However,the new space obtained after neural network mapping affects the distribution of points and further affects the classification performance.In the previous study,the space formed was called "partition space",which may be an irregular area or an inner area of a hypersphere.Therefore,the quality of the partition space has also become part of the neural network classifier evaluation.As a neural network classifier that can subdivide partition space,the nearest neighbor partitioning method improves the neural network classifier by generating boundaries of arbitrary shapes in the partition space.However,there are some problems with the nearest neighbor partitioning.On the one hand,using a surface with boundary in a partition space leads to the boundary problem and limits the scope of applicable models.On the other hand,the feedforward neural network with lower expressive capabilities cannot take full advantage of the flexibility of mapping.Therefore,this study describes a novel closed surface nearest neighbor partitioning neural network classifier to improve the classification performance.However,the application of the neural network classifier on large-scale data sets reduces the efficiency of model training.In order to solve this problem,we accelerate the model based on the GPU to improve efficiency and reduce training time.In addition,due to the complexity of the cement hydration process,a large amount of data is needed during the training process to increase the robustness of the model.Therefore,the closed surface nearest neighbor partitioning is used to evaluate the cement microstructure performance and parallel methods are used to accelerate the training processIn this paper,we study the closed surface nearest neighbor partitioning neural network classifier from the following aspects:(1)Using the closed surface nearest neighbor partitioning method to improve the neural network classifierIn this study,a closed surface nearest neighbor partitioning neural network classifier is proposed.The closed surface is adopted to accommodate mapped points in a partition space,optimizing their distribution without boundary constraint along the entire closed surface.Furthermore,the recurrent neural network is also introduced to exploit closed surface partition space by increasing the flexibility and complexity of the mapping relationship.On the one hand,the distribution of sample points is optimized without boundary on the closed surface,which makes the same classes close to each other,and the different classes are far apart from each other.On the other hand,the recurrent neural network is a fully connected network with strong expressive power because of the existence of recurrent links,thereby increasing the chance of finding the optimal neural network.The experimental results show that the method achieves better performance in terms of F-measure and accuracy.(2)Accelerate closed surface nearest neighbor partitioning neural network classifierIn this study,a parallel closed surface nearest neighbor partitioning(PCSNNP)method based on computational unified device architecture(CUDA)is proposed.In this approach,blocks and threads are used to evaluate potential neural network and perform parallel sub-tasks,respectively.The results show that the PCSNNP improves the training efficiency while ensuring performance,and has superior performance compared to other parallel methods in accuracy.(3)Evaluation of cement microstructure performance by closed surface nearest neighbor partitioning methodThe cement image is first acquired by Micro-CT,and the features are extracted from the image through the gray histogram and the gray level co-occurrence matrix,which used as the input of the neural network classifier.The research mainly uses the closed-surface nearest neighbor partition method to evaluate the performance of cement microstructure,and the parallel process is used for training.Compared with other classification methods,the closed surface nearest neighbor partitioning is superior to other classification methods in evaluating cement performance.