Search

Wind turbine (WT) surface damage detection based on deep learning-aided drone inspection is an important emerging technology. Traditional deep learning algorithms have the issues of low global search capability, low damage detection accuracy, and long inference time. This paper proposes a new real-time detection and semantic segmentation-you only look once (RDSS-YOLO) neural network (NN) for both real-time and accurate detection and semantic segmentation of WT surface damage. A damage size quantification method is further created to calculate the real size of WT surface damage using segmentation results. Moreover, a digital twin (DT) of WT surface damage detection based on the drone inspection aided by the proposed deep learning methods is developed, which can ultimately realize real-time surface damage detection for both standstill and rotating WTs. The proposed RDSS-YOLO NN is tested on an augmented drone inspection dataset and obtains mean average percentage, precision, recall, and mean intersection over union of 95.7 %, 93.9 %, 96.8 %, and 81.5 %, respectively, which are superior to those obtained by some state-of-the-art surface damage detection NNs. The proposed damage size quantification method is tested on another dataset generating an average relative error of 7.24 %, 13.06 %, and 9.44 % for WT blade crack, leading edge erosion, and paint peeling, respectively. The developed DT has been successfully applied in three different wind farms and achieved the real-time detection and semantic segmentation of WT surface features. [Display omitted] [ABSTRACT FROM AUTHOR]

The rain-induced fatigue damage in the wind turbine blade coating has attracted increasing attention owing to significant repair and maintenance costs. The present paper develops an improved computational framework for analyzing the wind turbine blade coating fatigue induced by rain erosion. The paper first presents an extended stochastic rain field simulation model that considers different raindrop shapes (spherical, flat, and spindle), raindrop sizes, impact angles, and impact speeds. The influence of these raindrop characteristics on the impact stress of the blade coating is investigated by a smoothed particle hydrodynamics approach. To address the expensive computational time, a stress interpolation method is proposed to calculate the impact stress of all raindrops in a random rain event. Furthermore, coating fatigue analysis is performed by including the fatigue crack initiation in the incubation period and the fatigue crack propagation in the mass-loss-rate increasing period due to raindrop impact. Finally, the proposed computational framework is verified by comparing the estimated fatigue life with those obtained in literature. The results from the study show that by incorporating the statistics of rainfall data, the proposed framework could be used to calculate the expected fatigue life of the blade coating due to rain erosion. • Rain field simulation considers varied raindrop shapes, sizes, and distributions. • Raindrop impact stress is calculated based on smooth particle hydrodynamics. • Rain erosion of blade coating includes fatigue incubation and increasing periods. [ABSTRACT FROM AUTHOR]

Rain erosion is one of the most detrimental factors contributing to wind turbine blade (WTB) coating fatigue damage especially for utility-scale wind turbines (WTs). To prevent rain erosion induced WTB coating fatigue damage, this paper proposes a deep reinforcement learning (DRL)-based optimization method for finding the optimal rotor speed under different rain intensities and wind speeds. First, an efficient physics-based model for predicting WTB coating fatigue damage considering the comprehensive blade coating fatigue mechanism, rain intensity distribution, and wind speed distribution is presented. Then, a WT rotor speed design optimization problem is constructed to search for the optimal rotor speed under different rain intensity and wind speed conditions. To address the challenge of optimizing the efficiency, the original design optimization problem is converted into a DRL-based design optimization model. A hybrid reward is proposed to enhance the DRL agent trained by a deep deterministic policy gradient algorithm. Finally, the proposed DRL-based design optimization method is utilized to guide the optimal rotor speed scheduling of a 5-MW WT under given wind speed and rain intensity conditions. The results show that the proposed method could extend the predicted WTB blade coating fatigue life by 2.55 times with a minor reduction in the energy yield (0.027%) compared to the original rotor speed schedule that only considers maximum power capture. The computational time of the proposed method is reduced significantly compared to that of the traditional gradient and evolutional design optimization methods. • A fast and accurate physics-based surrogate model is created for predicting the wind turbine blade coating fatigue. • A deep reinforcement learning-based optimization method is proposed for designing high-dimensional wind turbine rotor speeds. • A new hybrid reward is proposed to facilitate the agent training by the deep deterministic policy gradient algorithm. [ABSTRACT FROM AUTHOR]

• Select an optimal TBM performance prediction model with Bayesian optimization. • Optimization of an ANN model using neural architecture search. • The automated machine learning method for TBM performance prediction is proposed. • The state-of-the-art prediction results of TBM performance are obtained. Accurately predicting the performance of a tunnel boring machine (TBM) is important to safe and efficient tunneling. The application of machine learning algorithms to TBM performance prediction creates several challenges. Such prediction is a nontrivial task involving procedures such as data preprocessing, selection of a machine learning algorithm and optimization of the related hyperparameters. The demand for expert knowledge has restricted the application of machine learning methods to TBM performance prediction, and it is meaningful to study predicting TBM performance automatically. In this paper, we explore three approaches to TBM performance prediction using Bayesian optimization and automated machine learning (AutoML). In the first study, Bayesian optimization is used to determine the optimal hyperparameters of various machine learning algorithms, including support vector regression (SVR), decision tree, bagging tree, random forest and AdaBoost. We attain the minimum mean squared error (MSE) values of 3.135 × 10 - 2 and 3.177 × 10 - 2 for a decision tree and SVR, respectively. In the second approach called the neural architecture search (NAS), the optimal combination of architecture, hyperparameters and the training procedure of an artificial neural network is found in a single operation. We obtain the optimal results of 3.514 × 10 - 2 and 3.237 × 10 - 2 if complete and simplified NAS are used, respectively. In the third method, the best combination of a data preprocessing method, a machine learning model and the related hyperparameters is found, and an optimal MSE value of 3.148 × 10 - 2 is obtained using AutoML. In all three studies, we obtain state-of-the-art prediction results that are superior to a previous best prediction result of 3.500 × 10 - 2 . The prediction results prove that Bayesian optimization and AutoML are powerful tools that can not only effectively predict TBM performance but also reduce the demand for expert knowledge of machine learning. [ABSTRACT FROM AUTHOR]

In this article, an efficient collaborative robust topology optimization (CRTO) method is proposed for functionally graded materials (FGMs) considering hybrid bounded uncertainties (HBU). Firstly, to realize the collaborative optimization of the structural topology and volume fraction of reinforcement in FGMs, two sets of design variables are defined following the SIMP framework. Secondly, with the uncertain material properties and external loads mathematically modeled as random and interval uncertain parameters respectively, the objective performance of FGMs are described as the implicit function of two sets of design variables and uncertain parameters. In contrast to the classical method of weighting the mean and standard deviation of objective performance, the robust objective function is constructed based on a novel distance to ideal solution to avoid the manually setting of weight parameters. The sensitivity analysis of the objective and constraint function with regard to the two sets of design variables is derived explicitly and a realization vector set is defined for bounded probabilistic uncertainties to parallelize the sensitivity analysis. In addition, an adaptive density shift algorithm is proposed to produce a clear topological profile and accelerate the convergence of the optimization. The effectiveness of the proposed method is demonstrated by both numerical and engineering examples. [ABSTRACT FROM AUTHOR]

Due to the incapability of one-dimensional (1D) and two-dimensional (2D) models in simulating the frontal polymerization (FP) process in laminated composites with multiple fiber angles (e.g., cross-ply, angle-ply), modeling a three-dimensional (3D) domain, which is more representative of practical applications, provides critical guidance in the control and optimization of the FP process. In this paper, subroutines are developed to achieve the 3D modeling of FP in unidirectional and cross-ply carbon fiber laminates with finite element analysis, which are validated against the experimental data. The 3D model is employed to study the effect of triggering direction in relevance to the fiber direction on the FP process, which cannot be studied using traditional 1D/2D models. Our findings suggest that triggering in the fiber direction leads to a higher front velocity, in comparison to cases where front was triggered in the direction perpendicular to the fiber. Moreover, the average front velocity in cross-ply laminates is on average 20–25% lower than that in unidirectional laminates. When triggered using two opposite fronts in the in-plane direction, the maximum temperature of the thermal spike in the cross-ply laminate, when two fronts merge, is about 100 °C lower than that in the unidirectional laminate. In cross-ply laminates, a sloped pattern forms across the thickness direction as the front propagates in the in-plane direction, as opposed to the traditionally observed uniform propagation pattern in unidirectional cases. Furthermore, the effect of thermal conductivity is studied using two additional composite laminates with glass (1.14 W/m·K) and Kevlar fibers (0.04 W/m·K). It is shown that the frontal velocity, degree of cure, and the thermal spike temperature decrease as the thermal conductivity reduces. [ABSTRACT FROM AUTHOR]

A fiber-reinforced composite wind turbine blade (WTB) is exposed to numerous impact threats during its service life causing damages that can be detrimental to its structural integrity. Currently, impact loads are not considered during blade design, so high safety factors are introduced, which result in a conservative design. However, as wind turbine blades become stiffer and lighter and health monitoring systems become more sophisticated, the design process is shifting toward damage-tolerant approaches. The design philosophy accepts damages to the structure, but it also requires that the damaged blade still meet structural and functional requirements. This design procedure requires a comprehensive understanding of different impact threats and their characteristics, which is currently unavailable in the public domain. This paper is a first attempt to review the impact loads on composite wind turbine blades. The aim of the current paper is to (a) identify different sources of impact threats on wind turbine blades during different stages of their service life, (b) describe their qualitative (causes and vulnerable regions) as well as quantitative characteristics (size, mass, and velocity of impactor), and to (c) provide modeling guidelines by comparing these impact threats using five different criteria - (i) relative deformability of projectile and wind turbine blade, (ii) impact velocity, (iii) kinetic energy of impact, (iv) repeatability of impacts and (v) nature of the impact. The review paper will be of special interest to researchers working on wind turbine blades and will serve as a baseline report for designing damage-tolerant blades. Recommendations are also provided for future research. • Different sources of impact threats are identified on wind turbine blades during different stages of their service life. • Qualitative and quantitative characteristics of impact Loads are comprehensively described. • Modeling guidelines are provided by comparing the impact threats using five different criteria. [ABSTRACT FROM AUTHOR]

Artificial neural networks (ANNs) have been widely adopted as general computational tools both in computer science as well as many other engineering fields. Stochastic gradient descent (SGD) and adaptive methods such as Adam are popular as robust optimization algorithms used to train the ANNs. However, the effectiveness of these algorithms is limited because they calculate a search direction based on a first-order gradient. Although higher-order gradient methods such as Newton's method have been proposed, they require the Hessian matrix to be semi-definite, and its inversion incurs a high computational cost. Therefore, in this paper, we propose a variable three-term conjugate gradient (VTTCG) method that approximates the Hessian matrix to enhance search direction and uses a variable step size to achieve improved convergence stability. To evaluate the performance of the VTTCG method, we train different ANNs on benchmark image classification and generation datasets. We also conduct a similar experiment in which a grasp generation and selection convolutional neural network (GGS-CNN) is trained to perform intelligent robotic grasping. After considering a simulated environment, we also test the GGS-CNN with a physical grasping robot. The experimental results show that the performance of the VTTCG method is superior to that of four conventional methods, including SGD, Adam, AMSGrad, and AdaBelief. [ABSTRACT FROM AUTHOR]

A critical problem in engineering reliability analysis is obtaining an accurate failure probability with a high computational efficiency. This study aims to present failure probability estimation under the conditions of uncertain input variables and their uncertain distribution parameters. An adaptive Kriging model of failure probability with respect to distribution parameters (FP-DP model) is developed, which avoids coupling modeling among the distribution parameters, input variables, and failure probability. An improved U -learning function that simultaneously considers the statistical information of uncertain distribution parameters and failure probability is proposed to select new sampling points for the FP-DP model. The stopping criteria based on sample distances and relative errors of the predicted failure probability are constructed to improve the convergence performance around the limit state function. Three numerical and four engineering examples with different complexities are considered to verify the effectiveness of the proposed adaptive FP-DP Kriging metamodel. The results show that the proposed method can obtain an accurate failure probability with fewer sampling points of uncertain distribution parameters than some existing methods, indicating that the proposed method can be efficiently integrated into reliability-based design optimization problems considering both the uncertainties of input variables and their distribution parameters. [ABSTRACT FROM AUTHOR]

The identification of differences between 3D CAD models is of great significance for engineering change management. However, existing methods are unable to accurately identify the differences of CAD models under various poses. To overcome the limitation, a novel difference identification method for 3D CAD models with arbitrary locations and arbitrary orientations is proposed in this paper. Firstly, two CAD models to be compared are represented by their NURBS-based key-points, which are then matched by determining an optimal rigid transformation between the models. The parameter calculating process of the optimal rigid transformation is accelerated by a graph-based vertex pre-matching method while the calculating accuracy is guaranteed by the iterative mismatching elimination algorithm (IMEA). Finally, the differences between CAD models are identified hierarchically based on the matching sequences of three kinds of key-points. Experimental results show that the proposed method can accurately identify the geometric differences of CAD models located at various positions and orientations. [ABSTRACT FROM AUTHOR]

The uncertainty quantification of the input statistical model in reliability-based design optimization (RBDO) has been widely investigated for accurate reliability analysis, and it could be estimated through its characteristics, cumulative experiences, and available data. However, uncertainty quantification of random variables in existing RBDO studies has exploited parametric distributions quantifying the uncertainty through the Bayes' theorem. In addition, a correlation between random variables is often underestimated due to a lack of knowledge and difficulty to describe the high-dimensional correlation. Hence, it has been a challenge to properly quantify input statistical model and its uncertainty. Therefore, a multivariate kernel density estimation (KDE) is employed to perform data-driven confidence-based design optimization (CBDO) for effective quantification of input model uncertainty. Any assumption on input distribution is not necessary since it is established only with the given input data. Moreover, the input model uncertainty due to insufficient data is quantified using bootstrapping and optimal adaptive bandwidth matrices through the Bayes' theorem using cross-validation error. Consequently, the proposed CBDO with given input data is capable of finding a conservative optimum of RBDO accounting for both aleatory uncertainty of random variables and epistemic uncertainty induced by a limited number of input data through the multivariate KDE. [ABSTRACT FROM AUTHOR]

The use of marine current energy to provide a stable supply of energy and fresh water to islands and remote areas is a promising prospect. In this study, an integrated marine current-powered seawater desalination and hydrogen production system was proposed. For higher integration and efficiency, the generator and seawater pump were integrated into the nacelle and mechanically connected to the main shaft. Four operating modes and the corresponding start-stop and maximum power point tracking (MPPT) control strategy were designed to achieve energy distribution between the two loads. Considering the frequent start-stop caused by the sinusoidal characteristic of marine current energy, a soft start-up process was designed for the RO elements, including dual hydraulic accumulators and a smooth transition process to avoid abrupt changes in feed pressure and feed flow rate. Finally, the performance was obtained by simulation. During the 6 h of the ebb tide, the desalination system produced a total of 841.77 L of fresh water, while the electrolyzer produced 1038.5 SL of hydrogen and consumed 0.95 L of water. The average power coefficient of the turbine was 0.434, and the specific energy consumption of the desalination system was 2.03 kWh/m3, indicating good MPPT performance and high efficiency. [ABSTRACT FROM AUTHOR]

The ocean contains a large amount of marine current energy which is considered to be one of the most promising renewable energy sources for commercialization. In contrast to the high-velocity currents in straits close to the mainland or islands, the velocities of the currents in most of the sea are low. The exploitation of low-velocity currents is of great significance to the regions with such energy sources. Although the technology for harvesting marine current energy is developing rapidly, the exploitation of low-velocity current energy is often neglected due to its low power density. When the existing marine current turbine (MCT) technology developed for high-velocity currents is used for low-velocity currents, it will result in large rotor diameters and low performance of starting current velocities, which is subject to the water depth and the levelized cost of energy. To solve those problems, a direct-drive parallel-axis twin-rotor marine current turbine (TRMCT) was developed in this research, with a focus on the design of the transmission and generator of the MCT, as well as the improvement of the control strategy. Firstly, the structure of the TRMCT and its characteristics were given. Secondly, the improved optimal torque control strategies for the TRMCT were designed. Finally, to verify the feasibility of the TRMCT and its control strategy, the mathematical model was built and simulated, and the starting performance and operating performance were analyzed. The results show that the proposed TRMCT can start to work at about 0.2 m/s and the conversion efficiency of the whole system can reach about 25%. Compared to the scheme using two identical single-rotor MCTs, the TRMCT's output power can increase by up to 26%. Besides, the torque imbalance caused by the difference in the current velocities on the two rotors can be well coped with. [ABSTRACT FROM AUTHOR]

• A reliability-based design optimization under multiple simulation models is proposed. • An active learning function is proposed by estimating how updates on simulation model affect reliability. • A concept of activity function is proposed, which estimates how active constraint functions are. • The proposed active learning function is modified to apply to problems where the cost of each simulation model is different. • Two numerical and one engineering examples are investigated to validate the proposed method. This paper proposes an effective active learning strategy for reliability-based design optimization (RBDO) problems in which the constraint functions are acquired from multiple simulation models. To achieve this goal, a new active learning function (ALF) is derived by estimating the increased reliability of active constraint functions after adding one point to the train points of constraint functions in each simulation model. The proposed ALF distinguishes possibly active constraint functions that seem active near the current optimum and considers how the constraint functions are active. In the proposed RBDO method, a Kriging model is iteratively updated by adding the best point to the train points of constraint functions included in the crucial simulation model until the optimum converges and the Kriging model is sufficiently accurate. The best point and the crucial simulation model are obtained by comparing the proposed ALF. The ALF is further modified to apply to problems where the cost of each simulation model is different. To verify the effectiveness of the proposed method, two numerical and one engineering examples are analyzed. The results show that the proposed method efficiently and accurately obtains the RBDO optimum involving multiple simulation models, regardless of simulation cost. [ABSTRACT FROM AUTHOR]

• Design and implementation of the marine current powered hydrogen production system are presented. • A novel dynamic model of the PEM electrolyzer is presented and validated by experiments. • The feasibility of the system was demonstrated by simulation and sea trials. • The performance degradation of the electrolyzer is discussed. In this work, an off-grid hydrogen production system powered by marine current energy was studied, which employed a horizontal axis marine current turbine (HAMCT) and a polymer electrolyte membrane (PEM) electrolyzer fed with ultrapure water. The fluid kinetic energy of the marine current will be captured by the turbine and finally stored as hydrogen energy through the electrolysis reaction in the electrolyzer. It is important to fully understand the characteristics of the electrolyzer for the stable and efficient operation of the system. Here, a dynamic model of PEM electrolyzers was developed, which is based on the Hammerstein structure. The particle swarm optimization (PSO) method and the least squares method were used to fit the static part and the dynamic part of the model, respectively. The experimental validation shows enough precision for engineering applications and the ability to characterize the transient behavior of the electrolyzer. Faraday's efficiency of the stack was modeled using an empirical formula. The simulation of the proposed system was then carried out using the measured current velocity data as input. The results demonstrate that the system achieved the designed operating performance with the power coefficient of 0.42 and the estimated average energy conversion efficiency from marine current-to-hydrogen of 16.4%. Then, the sea trial was conducted in Zhoushan Archipelago. The power coefficient and the average energy conversion efficiency were found to be 0.35 and 11.9% respectively, with a decrease compared to the simulated results, which was attributed to the idealization of the simulation model and the degradation of the PEM electrolyzer. The performance degradation of the PEM electrolyzer throughout experiments and its effects were discussed. The principle and feasibility of the marine current-hydrogen system were successfully demonstrated. [ABSTRACT FROM AUTHOR]

• An integrated optimization methodology is proposed for design of mooring systems. • The methodology integrates design of experiments, screening analysis, dynamic analysis and a metamodel-based optimization. • Mooring system of a vessel-shaped offshore fish farm is designed using a fully-coupled numerical model. • Validations are performed on the metamodels and the global optimal solutions. Design optimization of mooring systems of offshore floating structures is a challenging task, partly because of the large number of design variables, complicated design constraints, nonlinear system behavior, and time-consuming numerical simulations. For engineering designs, efficient yet accurate approaches are needed. This paper proposes an integrated optimization methodology for design of mooring systems. The methodology integrates the design of experiments, screening analysis, time-domain simulations, and a metamodel-based optimization procedure. To demonstrate the methodology, the mooring system of a vessel-shaped offshore fish farm was designed considering the ultimate limit state. The fully-coupled numerical model includes a floater, flexible fish cages and a single-point mooring system. The Kriging metamodels were applied as surrogates for the responses of time-domain simulations. The optimal solutions were found by exploring the design space using a gradient-based search algorithm. Validations were performed on the metamodels and the global optimal solutions. The proposed methodology is also applicable to design optimization of other marine structures. [ABSTRACT FROM AUTHOR]