Research And Application Of The Combination Of Deep Learning And Principal Component Analysis

At Present,The data volume, data model and the relevance between data become more and more complex, classification and prediction tasks are becoming more and more difficult, resulting in some areas of traditional shallow machine learning algorithms are difficult to show good performance. Deep learning is a kind of network structure which is more and more close to the human model, which is developed by adding the hidden layers in the traditional shallow structure. In 2006, Hinton proposed the establishment of a network of confident degree through the greedy cascade method, and its training test results achieved encouraging results, which attracted wide attention from people of all walks of life.Although deep learning is widely used in many fields, but its final result isn't the most optimal. Fusion of traditional method is one direction of deep learning to increase the accuracy and velocity and to reduce the complexity of the model. Principal component analysis can retain the information features of the original data, it is a commonly used method of dimension reduction of the original data, converting the data to relatively low dimension space, so as to reduce the dimension. Considering the deficiency of deep learning and the advantages of principal component analysis, this paper combines deep learning and principal component analysis algorithm to enhance the performance of deep learning. The main contents of this paper are as follows: first, the analysis and testing of the fusion algorithm of deep learning and principal component analysis, second, using the deep learning in the geochemical analysis. Before input layer of the network and reducing the input dimension of input layers by principal component analysis; deep belief network output advanced and abstract features, using principal component analysis method to reduce the high abstract features dimension which is invalid provide lower dimensional for he supervised learning method, reduce the time complexity of the algorithm supervised exploration; geochemical is an important means for exploration of mineral resources, chaos, fractal and neural networks are widely used, however deep learning in the use of the subject is very few, so the application of deep learning recognition the mineralization information related to mineralization in geochemistry and can construct a reasonable block model. Firstly according to the deep learning and principal component analysis method, through the organic combination and the relative literature theory, set the network initialization parameters, and through repeated experiments and the average, ultimately determine the various network parameters, and using a publicly available database of MNIST handwritten data set to test the fusion algorithm. Finally, this paper will use this method in geochemistry by all above testing steps and get the final results.Through this analysis, this paper organically combined with the advantages of principal component analysis and deep learning and use this method in the MNIST handwritten database testing. This paper creates 3-layers networks, learning rate is 0.5, momentum is 0.1 and get the error rate was 1.1% and training time was 4.01 s. This results compared to traditional networks under the same parameters get the error rate was 1.2%, the training time was 4.3s. the results are faster and more accurate results than traditional deep belief networks. Finally, Using the fusion algorithm in a mining area in the Panzhihua soil geochemical analysis data. This paper uses the application which is developed by ourselves to show the local distribution of ore deposits, illustrate the result of geochemical data and construct the three-dimensional model. This paper shows that the improved method can be used in geochemical analysis.