• Journal of Power Electronics
• /
• v.18 no.2
• /
• pp.445-455
• /
• 2018

A novel algorithm is proposed to regulate the neutral point potential in neutral point clamped three-level inverters. Oscillations of the neutral point potential and an unbalanced dc-link voltage cause distortions of the output voltage. Large capacitors, which make the application costly and bulky, are needed to eliminate oscillations. Thus, the algorithm proposed in this paper utilizes the finite-control-set model predictive control and the multistage medium vector to solve these issues. The proposed strategy consists of a two-step prediction and a cost function to evaluate the selected multistage medium vector. Unlike the virtual vector method, the multistage medium vector is a mixture of the virtual vector and the original vector. In addition, its amplitude is variable. The neutral point current generated by it can be used to adjust the neutral point potential. When compared with the virtual vector method, the multistage medium vector contributes to decreasing the regulation time when the modulation index is high. The vectors are rearranged to cope with the variable switching frequency of the model predictive control. Simulation and experimental results verify the validity of the proposed strategy.

• Journal of Power Electronics
• /
• v.14 no.6
• /
• pp.1178-1188
• /
• 2014

This study proposes a finite-state model predictive controller to regulate the load current and balance the DC-link capacitor voltages of a four-leg neutral-point-clamped converter. The discrete-time model of the converter, DC-link, inductive filter, and load is used to predict the future behavior of the load currents and the DC-link capacitor voltages for all possible switching states. The switching state that minimizes the cost function is selected and directly applied to the converter. The cost function is defined to minimize the error between the predicted load currents and their references, as well as to balance the DC-link capacitor voltages. Moreover, the current regulation performance is improved by using a two-step prediction horizon. The feasibility of the proposed predictive control scheme for different references and loads is verified through real-time implementation on the basis of dSPACEDS1103.

• Transactions of the Korean Society of Mechanical Engineers A
• /
• v.24 no.11
• /
• pp.2836-2844
• /
• 2000

Thickness control of hot-rolled strips has become an important issue in recent years because of the need for improving the quality of the hot-rolled strip. In this study, a modifying method of rolling force set-up with consideration of spread was developed to improve the thickness uniformity at the finishing rolling units in hot rolling. Through the analysis of real production data it was found that the accuracy of the rolling force determined from the finishing mill set-up (FSU) model dominantly governed the thickness uniformity in rolled plates at the front. Based on this analysis , several examples were selected to calculate the spread of rolled plate using three dimensional rigid thermo-viscoplastic finite element program. FE analysis results were used to train the neural network system that can predict the spread hot-rolled plate and the rolling force was modified based on the predicted value of spread. The modified rolling forces were closer to the measured rolling force so it can be expected that the accuracy of thickness uniformity of hot-rolled plate will be improved.

• Structural Engineering and Mechanics
• /
• v.81 no.6
• /
• pp.665-675
• /
• 2022

The safety problems of giant hydraulic structures such as dams caused by terrorist attacks, earthquakes, and wars often have an important impact on a country's economy and people's livelihood. For the national defense department, timely and effective assessment of damage to or impending damage to dams and other structures is an important issue related to the safety of people's lives and property. In the field of damage assessment and vulnerability analysis, it is usually necessary to give the damage assessment results within a few minutes to determine the physical damage (crack length, crater size, etc.) and functional damage (decreased power generation capacity, dam stability descent, etc.), so that other defense and security departments can take corresponding measures to control potential other hazards. Although traditional numerical calculation methods can accurately calculate the crack length and crater size under certain combat conditions, it usually takes a long time and is not suitable for rapid damage assessment. In order to solve similar problems, this article combines simulation calculation methods with machine learning technology interdisciplinary. First, the common concrete gravity dam shape was selected as the simulation calculation object, and XFEM (Extended Finite Element Method) was used to simulate and calculate 19 cracks with different initial positions. Then, an LSTM (Long-Short Term Memory) machine learning model was established. 15 crack paths were selected as the training set and others were set for test. At last, the LSTM model was trained by the training set, and the prediction results on the crack path were compared with the test set. The results show that this method can be used to predict the crack propagation path rapidly and accurately. In general, this article explores the application of machine learning related technologies in the field of mechanics. It has broad application prospects in the fields of damage assessment and vulnerability analysis.

• The Journal of Engineering Geology
• /
• v.26 no.2
• /
• pp.187-196
• /
• 2016

We conducted three-dimensional finite element analysis to predict the presence of upcoming fault zones during tunneling. The analysis considered longitudinal displacements measured at tunnel face, and used 28 numerical models with various fault attitudes. The x-MR (moving range) control chart was used to analyze quantitatively the effects of faults distributed ahead of the tunnel face, given the occurrence of a longitudinal displacement. The numerical models with fault were classified as fault gouge, fault breccia, and fault damage zones. The width of fault cores was set to 1 m (fault gouge 0.5 m and fault breccia 0.5 m) and the width of fault damage zones was set to 2 m. The results, suggest that fault centers could be predicted at 2~26 m ahead of the tunnel face and that faults could be predicted earliest in the 45° dip model. In addition, faults could be predicted earliest when the angle between the direction of tunnel advance and the strike of the fault was smallest.