Research On Neural Network Method And Its Improvement Oriented To Cement And Cement-based Material Classification

Along with the development of industry, people have higher requirements on the performance and quality of cement and cement-based materials. The compressive strength is an important indicator to evaluate the quality of cement and cement-based materials. As an important cement-based material, the concrete compressive strength is directly affected by the cement hydration process and cement strength. The samples must be maintained 28 days under the standard environment to test the concrete strength in the traditional method, which requires a lot of raw materials and has a high cost and time consuming. With the development of computer hardware and software and the rise of computational materials science, people try to use the computer to quickly solve related issues in the field of materials. Along with the development of digital image processing technology, people try to use the computer to realize the visualization of complex cement hydration process. The image is an important research content of cement hydration process in the computer simulation. In this paper, the main research content is the neural network classification method which is used to solve the classification problems of cement and cement-based materials. This paper makes a research on existing neural network classification methods at first. And this paper is mainly focus on the research and improvement of FCM neural network classifier. Then, some neural network methods and the improved FCM neural network classifier based on DBSCAN are used to finish the concrete and cement hydration image classification tasks. The main research content in this paper is showed in below:(1) Original FCM neural network classifier uses K-Means algorithm in the partition space to get the colored partition space. But K-Means method is sensitive to the outliers which will affect the accuracy of obtained the colored partition space. So, we propose an improved FCM neural network method based on K-medoids algorithm to weaken the effect of outliers in partition space on the classification accuracy. Meanwhile, we also choose several commonly used datasets to test the performance of this improved method. The experimental results show that the improved FCM method based on K-medoids can obtain a better classification performance.(2) Considering that the shape of partitions can make an influence on the performance of FCM classifier, we use DBSCAN algorithm to improve the original FCM method. We use DBSCAN algorithm to divide color points in the partition space into arbitrary shape partitions. In this improved method, there doesn’t exist the concept of centroid, we define a new optimization objective function and then invoke the PSO algorithm to optimize the parameters of neural network to get an optimal classification model. In addition, the DBSCAN algorithm based on density is employed instead of K- Means in the process of cluster formation, so we respectively record the running time of original FCM method and this improved FCM method based on DBSCAN in the training process. The experimental results show that the improved FCM method based on DBSCAN has a shorter running time, which further verifies the feasibility of this improved FCM neural network classifier from the view of time.(3) We use several current neural network methods and the improved FCM method based on DBSCAN to classify concrete strength and cement hydration images. We also record the generalization accuracy and the average F-measure in the 10-fold cross-validation process. The experimental results show that using neural network classifier to solve related issues in the field of Portland cement and cement-based materials is feasible. Meanwhile, it also further verifies that the improved FCM method based on DBSCAN we proposed in this paper has better classification performance on some related experimental data.