Search

Recently, DeepNorm scales Transformers into extremely deep (i.e., 1000 layers) and reveals the promising potential of deep scaling. To stabilize the training of deep models, DeepNorm (Wang et al., 2022) attempts to constrain the model update to a constant value. Although applying such a constraint can benefit the early stage of model training, it may lead to undertrained models during the whole training procedure. In this paper, we propose BranchNorm, which dynamically rescales the non-residual branch of Transformer in accordance with the training period. BranchNorm not only theoretically stabilizes the training with smooth gradient norms at the early stage, but also encourages better convergence in the subsequent training stage. Experiment results on multiple translation tasks demonstrate that BranchNorm achieves a better trade-off between training stability and converge performance., Comment: Long paper, 9 pages

We participate in the WMT 2020 shared news translation task on Chinese to English. Our system is based on the Transformer (Vaswani et al., 2017a) with effective variants and the DTMT (Meng and Zhang, 2019) architecture. In our experiments, we employ data selection, several synthetic data generation approaches (i.e., back-translation, knowledge distillation, and iterative in-domain knowledge transfer), advanced finetuning approaches and self-bleu based model ensemble. Our constrained Chinese to English system achieves 36.9 case-sensitive BLEU score, which is the highest among all submissions., Comment: Accepted at WMT 2020. Our Chinese to English system achieved the highest case-sensitive BLEU score among all submissions

This paper presents a speed tracking control method based on the model predictive control (MPC) algorithm for the four-wheel independently actuated (FWIA) electric vehicle. In the upper level controller, a switching control law is designed to reduce the impact when the car from acceleration to deceleration or from deceleration to acceleration. To avoid high frequency oscillations in the speed control, acceleration increment input law is designed in the lower level controller. The lower level controller calculates the desired torque using Newton's second law. The validity of the proposed control algorithm is verified using the CarSim/Simulink simulation, and the simulation results of the proposed method under extreme conditions are compared with PI controller, which shows the superiority of the proposed method. Finally, by the vehicle test, the experimental results verify the effectiveness and robustness of the proposed control approach.

Performing line commutation at an appropriate distribution switch can better balance the inductive reactance between three-phase lines and reduce transmission line losses. However, due to the large number of distribution lines and the complex operating environment, the management of distribution line is very difficult, and after the commutation some distribution lines may be in negative sequence operation because of failing to plan reasonably and no checking phase sequence in time, that affect the normal operation of the feeder terminal unit(FTU). This paper first analyzes the FTU's working principle of the dual power supply, then studies the FTU's protection function configuration, and then obtains the influence of negative sequence operation of the line on the power supply and protection of the FTU. And then, according to the negative sequence operation characteristics of the line, this paper puts forward some suggestions, such as adding phase sequence alarm function, phase sequence adjusted by software, etc. Finally, combined with a negative sequence operation case in Shandong Province, it is verified that the negative sequence operation of the line can affect the operation of the FTU and the proposed method can effectively improve the distribution line management efficiency and the reliability of the FTU.

Because it has a large, growing population but only a small share of land that can be cultivated, it is important for China to enhance its agricultural productivity through technological progress. Using data envelopment analysis, we decompose productivity into pure technical efficiency change, scale efficiency change, and technological progress. We thereby find that annual growth of agricultural productivity in China is about 2.2 percent. Technological progress improved agricultural productivity at a rate of 4.2 percent annually from 1980 to 2005, but the technology efficiency dampened it by an average of 1.9 percent per year. TFP growth and technological progress are faster in eastern provinces than for those in central and western regions. Relative technology efficiency was stable in eastern provinces but declined in the central and western provinces during the study period. Thus, it was technological progress that boosted the TFP growth in china's agriculture. Tests also reveal that sigma convergence existed in Chinese agricultural productivity.

With the medium and low speed maglev train plays the role of commercial operation gradually, people put forward higher requirements for train safety reliability, which makes train security assessment even more prominent. Aiming at the characteristics of the maglev train equipments as well as the limitations of traditional security assessment, the establishment of a multi-state security assessment based on Bayesian network model has better diagnostic reasoning and causal reasoning ability. Finally, using the model to analysis the train traction system quantitatively, finding the weaknesses of the system and the relationship between the equipments to make rational maintenance decision. This will provide a basis to improve the reliability of train equipment and repair and maintenance work.