Search

Recently, records on stereo matching benchmarks are constantly broken by end-to-end disparity networks. However, the domain adaptation ability of these deep models is quite poor. Addressing such problem, we present a novel domain-adaptive pipeline called AdaStereo that aims to align multi-level representations for deep stereo matching networks. Compared to previous methods for adaptive stereo matching, our AdaStereo realizes a more standard, complete and effective domain adaptation pipeline. Firstly, we propose a non-adversarial progressive color transfer algorithm for input image-level alignment. Secondly, we design an efficient parameter-free cost normalization layer for internal feature-level alignment. Lastly, a highly related auxiliary task, self-supervised occlusion-aware reconstruction is presented to narrow down the gaps in output space. Our AdaStereo models achieve state-of-the-art cross-domain performance on multiple stereo benchmarks, including KITTI, Middlebury, ETH3D, and DrivingStereo, even outperforming disparity networks finetuned with target-domain ground-truths., Comment: Accepted by CVPR 2021

A three-phase three-level phase-shifted pulsewidth modulation (PSPWM) dc-dc converter is proposed for electric ship medium-voltage direct current application. Output voltage is controlled by incorporating PSPWM. Gate signals of each phase are phase shifted by 2π/3 from each other. This results in a decrease in output voltage ripple. Major features of the converter include: 1) good dynamic performance under low switching frequency; 2) fault tolerance capability; and 3) plug and play capability. In order to obtain behavioral and performance characteristics of the proposed converter, detailed analytical and hardware in loop simulation studies are carried out. Finally, the viability of the scheme is confirmed through detailed experimental studies on a MW level, 1-kHz prototype developed for the purpose.

A three-level boost-buck converter is proposed for ultracapacitor charging in vessel medium voltage direct current (MVDC) integrated power system (IPS). The topology, filters design, and control aspects are analyzed and discussed. A constant current control combining constant power control scheme is proposed to charge the ultracapacitor. The effectiveness of the proposed control scheme is verified by the simulation and hardware in the loop (HIL) experimental results.

Great progress has been made on estimating disparity maps from stereo images. However, with the limited stereo data available in the existing datasets and unstable ranging precision of current stereo methods, industry-level stereo matching in autonomous driving remains challenging. In this paper, we construct a novel large-scale stereo dataset named DrivingStereo. It contains over 180k images covering a diverse set of driving scenarios, which is hundreds of times larger than the KITTI Stereo dataset. High-quality labels of disparity are produced by a model-guided filtering strategy from multi-frame LiDAR points. For better evaluations, we present two new metrics for stereo matching in the driving scenes, i.e. a distance-aware metric and a semantic-aware metric. Extensive experiments show that compared with the models trained on FlyingThings3D or Cityscapes, the models trained on our DrivingStereo achieve higher generalization accuracy in real-world driving scenes, while the proposed metrics better evaluate the stereo methods on all-range distances and across different classes. Our dataset and code are available at https://drivingstereo-dataset.github.io.

Lithium-ion batteries (LiBs) have been widely applied in energy storage systems, and accurate State of Charge (SOC) estimation of LiBs has become a key issue. However, SOC is an important LiBs state that cannot be directly observed. It needs to be indirectly estimated by the observable variables, and SOC estimation is a challenging task due to strong nonlinear and complex electrochemical reactions change with temperature in the battery. In this paper, a Recurrent Neural Network with Long Short-Term Memory (LSTM-RNN) for SOC estimation of LiBs is introduced. By training on the observable variables of LiBs, the voltage, current and temperature can be mapped directly to SOC. Based on the dataset of LiB measured under drive cycles of electric vehicles, the effectiveness of the LSTM-RNN for SOC estimation of LiBs proposed in this paper is validated.

Stereo matching task has been greatly improved by convolutional neural networks, especially the fully-convolutional network. However, existing deep learning methods always overfit to specific domains. In this paper, focus on domain adaptation problem of disparity estimation, we present a novel training strategy to conduct synthetic-realistic collaborative learning. At first, we design a compact model that consists of shallow feature extractor, correlation feature aggregator and disparity encoder-decoder. Our model enables end-to-end disparity regression with fast speed and high accuracy. To perform collaborative learning, we then propose two distinct training schemes, including guided label distillation and semi-supervised regularization, both of which are used to alleviate the lack of disparity labels in realistic datasets. Finally, we evaluate the trained models on different datasets that belong to various domains. Comparative results demonstrate the capability of our designed model and the effectiveness of collaborative training strategy.

Accurate and robust state of charge estimation of lithium-ion battery is a challenging task in battery management system. In this paper, a novel data-driven SOC estimation approach for Lithium-ion (Li-ion) batteries is proposed based on the Gaussian process regression framework. Kernel function selection and hyperparameters optimization are critical for Gaussian process regression due to the reason that kernel function could capture rich structure of data. By integrating the structural properties of deep learning with the flexibility of kernel methods, a new deep learning technology called deep recurrent kernel that fully encapsulates GRU structure is introduced to capture ordering matters and recurrent structures in sequential data. The proposed method could not only learn the mapping relationship from one sequence of measured quantities such as voltage, current, temperature to SOC, but also quantify estimation uncertainty which is essential for making informed decisions for battery management system. The performance of proposed methods is evaluated by two experimental datasets, one under a series of electric vehicle drive cycles and another under high rate pulse discharge test. We demonstrate the proposed method achieves satisfactory performance, as well as performs strong robustness against unknown initial SOC and outliers occurred in voltage, current and temperature.

The state-of-health (SOH) estimation is a challenging task for lithium-ion battery, which contribute significantly to maximize the performance of battery-powered systems and guide the battery replacement. The complexity of degeneration mechanism enables data-driven methods to replace mechanism modeling methods to estimate SOH. The insight that motivates this study is that the charging curve of constant current-constant voltage charging mode could reflect the magnitude of SOH from the perspective of capacity. The proposed approach is based on a hybrid neural network called gate recurrent unit-convolutional neural network (GRU-CNN), which can learn the shared information and time dependencies of the charging curve with deep learning technology. Then the SOH could be estimated with the new observed charging curves such as voltage, current and temperature. The approach is demonstrated on the public NASA Randomized Battery Usage dataset and Oxford Battery Degradation dataset, and the maximum estimation error is limited to within 4.3%, thus proving its effectiveness.

A three-level (TL) phase-shift pulse width modulation (PSPWM) dc-dc converter is proposed for electric ship medium voltage direct current (MVDC) application. A dc-ac-dc topology is adopted for the converter considering galvanic isolation between input MVDC and output low voltage direct current (LVDC). Output voltage regulation of the converter is obtained by the use of PSPWM. In order to achieve fixed frequency and zero steady-state error during steady-state conditions and fast response during transient conditions for the converter, a linear and nonlinear composite control strategy is presented in this paper. The proposed composite control strategy is based on the small-signal model of the converter. To verify the feasibility of the control strategy, detailed simulations have been made. The control strategy has also been validated through the hardware-in-loop RT-LAB system. The experimental and simulation results prove the proposed composite control strategy to be superior in dynamic response over that of a conventional linear control strategy.

The Grounding fault location and protection is a key technology for the DC power system and is important to the normal operation of the power system and the safety of operators. The grounding fault location protection method based on additional grounding resistance is introduced in this paper, and its controller is made. It takes high risk and cost to directly test the controller directly in the DC power system. Therefore, a hardware in loop (HIL) real-time simulation system is established on the real-time simulation platform RT-LAB to verify the function of the controller. The real-time simulation results are compared with those obtained from MATLAB/ simulink off-line simulation, which shows that the grounding fault location and protection method is feasible and effective, and that the controller has successfully implemented its function.

In the field of ADAS and self-driving car, lane and drivable road detection play an essential role in reliably accomplishing other tasks, such as objects detection. For monocular vision based semantic segmentation of lane and road, we propose a dilated feature pyramid network (FPN) with feature aggregation, called DFFA, where feature aggregation is employed to combine multi-level features enhanced with dilated convolution operations and FPN under the framework of ResNet. Experimental results validate effectiveness and efficiency of the proposed deep learning model for semantic segmentation of lane and drivable road. Our DFFA achieves the best performance both on Lane Estimation Evaluation and Behavior Evaluation tasks in KITTI-ROAD and take the second place on UU ROAD task.

Disparity estimation is a challenging task in the field of computer stereo vision. In this paper, we propose a multi-granularity fully convolutional network architecture for end-to-end dense disparity estimation. First, we use single well-pretrained residual network for extraction of multi-granularity and multi-layer features. Second, correlation layers at three different granularities are used to gain hierarchical matching cues between left and right feature maps. Third, we conduct concatenation-deconvolution operations to output disparity maps. Finally, the experimental results show that our method achieves state of the art results, taking the second place on the KITTI Stereo 2012 task.

Integrated power system (IPS) is the future of power system in shipboard. A carrier phase-shift three-level full-bridge converter based on DSP and FPGA is proposed in this paper, which is the core device of the DC zonal electrical distribution module (DC-ZEDM) in the new-generation medium-voltage DC shipboard IPS. DSP and FPGA are main chips in the controller of this converter. The basic principle, system configuration and implementation method of the controller are detailed introduced. The core idea is that FPGA receives the carrier shifting angle from an extern DSP through the data and address bus, and then generates phase adjustable PWM waveform, which can meet the real-time control need of the proposed converter in all conditions. Resources of DSP and FPGA are properly used, so the controller has advantages of fast speed, high control precision, strong expansibility and high reliability. A 750kW protype is implemented based on DSP and FPGA, experimental results match theoretical findings well, which verifies mentioned advantages of the controller.

A novel control scheme called symmetrical chop plus phase-shift (SCPS) control, allowing six switches of the three-level full-bridge converter to achieve zero voltage switching (ZVS) and reducing the total harmonic distortion (THD) of the transformer primary side voltage is proposed in this paper. The traditional control schemes for three-level full-bridge converter to achieve ZVS ignored the impact of the high THD of transformer primary side voltage. The high THD will cause large power loss in transformer and is not conducive to delivery energy for transformer. In the SCPS control, the output voltage of full bridge is symmetrical in half of switching period and the THD is reduced. At the same time, two switches have not achieved zero voltage on. This paper presents switching process and conditions of this control scheme. THD of the output voltage is analyzed and compared with CPS control. The performance of the proposed control scheme was verified on a prototype.

In this paper, an isolated three-level H-bridge DC/DC converter has been designed, which is the core device of the DC zonal electrical distribution module (DC-ZEDM) in the new-generation medium-voltage DC vessel integrated power system (IPS). It supplies isolated low-voltage DC to the second-level converter of DC-ZEDM. Considering the problem of relatively large size and weight of a transformer working at a frequency of 50Hz, the phase-shift PWM was used for controlling the converter to increase its frequency and decrease the size and weight of the transformer. And this paper has analyzed the harmonic waves of the H-bridge's output voltage and presented the static operating point at which all the operating modes in one switching period are under CCM. For the loss of duty ratio due to leakage induction, an appropriate compensation solution has been proposed. Finally, a 20kW, 1kHz prototype was applied to the verification of the operating principle of the converter.