Cracks are one of the forms of damage to concrete structures that debase the strength and durability of the building material and may pose a danger to the living being associated with it. Proper and regular diagnosis ...
详细信息
Cracks are one of the forms of damage to concrete structures that debase the strength and durability of the building material and may pose a danger to the living being associated with it. Proper and regular diagnosis of concrete cracks is therefore necessary. Nowadays, for the more accurate identification and classification of cracks, various automated crack detection techniques are employed over a manual human inspection. convolutionneuralnetwork (CNN) has shown excellent performance in image processing. Thus, it is becoming the mainstream choice to replace the manual crack classification techniques, but this technique requires huge labeled data for training. Transfer learning is a strategy that tackles this issue by using pre-trained models. This work first time strives to classify concrete surface cracks by re-training of six pre-trained deep CNN models such as VGG-16, DenseNet-121, Inception-v3, ResNet-50, Xception, and InceptionResNet-v2 using transfer learning and comparing them with different metrics, such as Accuracy, Precision, Recall, F1-Score, Cohen Kappa, ROC AUC, and Error Rate in order to find the model with the best suitability. A dataset from two separate sources is considered for the re-training of pre-trained models, for the classification of cracks on concrete surfaces. Initially, the selective crack and non-crack images of the Mendeley dataset are considered, and later, a new dataset is used. As a result, the re-trained classifier of CNN models provides a consistent performance with an accuracy range of 0.95 to 0.99 on the first dataset and 0.85 to 0.98 on the new dataset. The results show that these CNN variants can produce the best outcome when finding cracks in the real situation and have strong generalization capabilities.
Recently, deep learning (DL) based spectrum sensing (SS) has drawn much attention due to its better capacity of feature extraction and superb performance. However, the model robustness of the DL based scheme is limite...
详细信息
Recently, deep learning (DL) based spectrum sensing (SS) has drawn much attention due to its better capacity of feature extraction and superb performance. However, the model robustness of the DL based scheme is limited by reason of the dynamic radio environment, leading to the floating of sensing performance. Motivated by this, adversarial transfer learning is applied to SS here, where the model is pre-trained at the central node firstly and fine-tuned at the local nodes. More specifically, a 2D dataset of the observed signal is constructed under various signal-to-noise-ratio (SNRs) and a convolutionneuralnetwork (CNN) model is designed. Then a part of samples with various SNRs in the constructed dataset are employed to pre-train the proposed CNN model. After that, the pre-trained CNN model is distributed to local nodes with different SNRs and the pre-trained CNN model is fine-tuned. The proposed CNN model is pre-trained based on the samples under various SNRs, resulting in its stronger adaptability at the local node. The simulation experiments validate the effectiveness of the proposed scheme.
Intelligent operation and green distribution have become a mainstream task to enable fast development of urban distribution applications. However, how to improve the distribution efficiency with low operating costs, a...
详细信息
Intelligent operation and green distribution have become a mainstream task to enable fast development of urban distribution applications. However, how to improve the distribution efficiency with low operating costs, and mitigate environmental pollution with high service quality is still a significant challenge in the practical industry applications. To address the above challenge, in this paper, we take into account both the economic cost and environmental cost, and propose a joint distribution path planning model based on neural architecture search (NAS) for electric vehicles with double-decked drones. More specifically, in our design, the factors such as energy consumption and carbon emissions of vehicles and drones during different distribution stages are considered. Then, a mixed integer linear programming model is established under the constraints of customer time window, vehicle capacity and vehicle battery capacity. Based on this model, a hybrid genetic algorithm is proposed to solve the optimization problem, where the carbon emission cost is estimated by the convolution neural network model, which is optimized by the neural architecture search technique. We conduct extensive experiments to validate the effectiveness of the proposed method. The experimental results show that, compared with CPLEX, the proposed method excelled in both solution quality and speed, which verify the effectives of our hybrid algorithm in dealing with the 2E-VRPD problem, where delivery routes can be well optimized, the efficiency of vehicle-machine collaborative delivery can be improved, and delivery costs are reduced.
Monitoring fouling behavior for better understanding and control has recently gained increasing attention. However, there is no practical method for observing membrane fouling in real time, especially in the forward o...
详细信息
Monitoring fouling behavior for better understanding and control has recently gained increasing attention. However, there is no practical method for observing membrane fouling in real time, especially in the forward osmosis (FO) process. In this article, we used the optical coherence tomography (OCT) technique to conduct real-time monitoring of the membrane fouling layer in the FO process. Fouling tendency of the FO membrane was observed at four distinguished stages for 21 days using a regular membrane cleaning method. In this method, chemical cleaning, which extracts two to three times as much organic matter (OM) as physical cleaning, was used as an effective method. Real-time OCT image observations indicated that a thin, dense, and flat fouling layer was formed (initial stage). On the other hand, a fouling layer with a thick and rough surface was formed later (final stage). A deep learning convolutional neuralnetworkmodel was developed to predict membrane fouling characteristics based on a dataset of real-time fouling images. The model results show a very high correlation between the predicted data and the actual data. R-2 equals 0.90, 0.86, 0.92, and 0.90 for the thickness, porosity, roughness, and density of the fouling layer, respectively. As a promising approach, real-time monitoring of fouling layers on the surface of FO membranes and the prediction of fouling layer characteristics using deep learning models can characterize and control membrane fouling in FO and other membrane processes. (C) 2021 Elsevier Ltd. All rights reserved.
暂无评论