Search

To strengthen networks’ security performance, here pose suggests a fused full convolutional approach to monitoring computer networks. This paper first analyzes the various performances of the full convolutional model for error problems rate, error reporting, and monitoring effects on various attack categories then proposes a network monitoring scheme for the full convolutional model and introduces the workflow of the full convolutional model in computer network security protection. In terms of accuracy, the full convolutional error rate of the blueprint meets the requirements rate of 96.8%, which is better than the classical network models of Lenet-5 and AlexNet, with 86.2% and 91.6%. The false alarm rate is only 2.37%, which is lower than the 5.74% MLP algorithm and 4.23% SVM algorithm. By comparison, the full convolutional calculation method is more efficient than other calculation methods in the detection rate of attack types such as Dos, Probe, U2R, and R2L. Therefore, the calculation method here is well adapted to computer network security protection requirements.

As an efficient geophysical exploration tool, the airborne electromagnetic (AEM) method has been widely used in mineral exploration, geological mapping, environmental and engineering investigation, etc. Currently, the imaging and 1D inversions are the mainstream means for AEM interpretation as the amount of AEM data is huge and 2D and 3D inversions are not efficient. In this paper, we propose a 2D fast imaging method for frequency-domain AEM data based on U-net network. The U-net is a symmetric full-convolution neural network, in which the partial pooling operation between the convolution layers is replaced by the up-sampling operation, while the target location is achieved by skipping connection. This method does not need to consider the complex coupling between the EM responses and underground structures, but instead it establishes a mapping relationship between EM responses and the resistivity model and can quickly achieve accurate imaging of AEM data. We use this network to image both synthetic and field survey data and compare the results with the traditional inversion algorithms. The results show that the U-net imaging have high resolution at high speed that provides a new way for interpreting large amounts of AEM data.

Due to the influence of air particles such as fog, haze, dust and so on, the image of outdoor photography was gray and white. However, the existing image de fogging algorithm had the problems of over dependence on prior information and inaccurate transmission calculation. In order to solve the above problems, an adaptive image defogging algorithm based on the depth convolution neural network was proposed in this paper. The algorithm realized the defogging of the foggy image based on the atmospheric scattering model. Three full convolution networks including shallow extraction, parallel extraction and deep fusion, were designed to realize the fusion of the shallow and deep features of the image, which greatly improved the accuracy of the transmittance image. The experimental results showed that the algorithm proposed in this paper had a good defogging effect on outdoor fog map, and the effect of defogging details was better. [ABSTRACT FROM AUTHOR]

• A transfer learning fault diagnosis model for bearing under different working conditions is proposed. • The proposed model is a deep Fully Convolutional Conditional Wasserstein Adversarial Network (FCWAN). • A difference classifier improves the diagnosis accuracy of model. The diagnostic accuracy of existing transfer learning-based bearing fault diagnosis methods is high in the source condition, but accuracy in the target condition is not guaranteed. These methods mainly focus on the whole distribution of bearing source domain data and target condition data, ignoring the transfer learning of each kind of bearing fault data, which may lead to lower diagnostic accuracy. To overcome these limitations, we propose a transfer learning fault diagnosis model based on a deep Fully Convolutional Conditional Wasserstein Adversarial Network (FCWAN). The proposed model addresses the described problems separately: (1) A random-sampling map classification and difference classifier are used to handle the first limitation. (2) A label is introduced into the domain of adversarial learning to strengthen the supervision of the learning process and the effect of category field alignment, thus overcoming the second limitation. Experimental results demonstrate the superiority of this method over existing methods. [ABSTRACT FROM AUTHOR]