Search

In order to study the pressure change with time at the blast face of the blast resistant wall when the shock wave acts on the blast resistant wall after the explosion of TNT explosive, the equivalent model of the blast resistant wall is designed based on the similarity theory, and the TNT near ground explosion test is carried out. During the test, the pressure change data of the design measuring point on the blast resistant wall with time are collected by high-speed data acquisition instrument and pressure sensor, The time history curve of shock wave overpressure was obtained by processing the data. Then, LS-DYNA finite element analysis software is used to establish a separate finite element model of the blast wall. At the same time, the relevant parameters of air, explosive, concrete and ground in the explosion field are set. The numerical simulation of the explosion field is carried out, and the overpressure time history curve at the measuring point is extracted by the finite element analysis post-processing software. Compared with the numerical simulation results, it is found that the numerical simulation results are in good agreement with the experimental data, which indicates that the selection and setting of parameters in the simulation calculation are reasonable. The model and algorithm provide an effective means for the numerical simulation of blast wall in the explosion field.

Convolutional neural networks (CNNs) have shown their power on many computer vision tasks. However, there are still some limitations, including their sensitivity to weight initialization and dependency to large scale labeled data. In this paper, we try to address these two problems by proposing a simple yet powerful CNN based denoising auto-encoder network which can be trained end-to-end in an unsupervised manner. The network architecture we use is a fully convolutional auto-encoder with symmetric encoder-decoder connections. The proposed method can not only reconstruct clean images from corrupted ones, but also learn abstract image representation through the reconstruction training. The encoder part of the network can be a good all-convolution network for classification. With the help of unsupervised pre-training, it achieves very competitive results even without extra unlabeled data. Further more, we show experimentally that our network also performs well in semi -supervised learning tasks.

In this paper, we propose a two-stage unsupervised deep hashing method to map the high-dimensional images to compact binary codes for large-scale image retrieval. We employ a neural network to look for the non-linear complex transformations to learn the binary representations in an unsupervised way which is suitable for the unlabeled data. Our method consists of two stages. The first stage trains a stacked convolutional autoencoder with the binary relaxation criterion. Then the second stage adopts the encoder part of the first-stage autoencoder and iteratively updates the parameters with back-propagation algorithm under the following criterions: (1) preserving original information, (2) producing balanced codes, (3) keeping bits independent and (4) the above binary relaxation criterion. For the new-coming image, we can easily obtain its compact binary codes by propagating it through the trained network and quantizing the network’s output. We evaluate our proposed method on two widely-used datasets and achieve very competitive result compared to the state-of-the-arts.

Although deep neural networks have brought impressive advances in a variety of machine learning tasks, it is more difficult to train a top-performing model in the absence of the labeled data. To alleviate this issue, domain adaptation has been extensively researched, which aims to reduce the difference between the distributions of the source and target domain by imposing restrictions on features. Adversarial learning method is the most promising approach to generate data that obeys a complex distribution. However, generator model often sinks into partial or full collapse. In this paper, we transform the complex data into a simple distribution, then calculate KL divergence (KL-MMD). We combine the Matching Gate with Attention Mechanism and put forward Matching Attention to learn feature vectors. Extensive experiments and analysis are conducted on three different digits datasets: MNIST, USPS, SVHN. To our knowledge, our method achieves state-of-the-art digit recognition performance on three unsupervised adaptation results.