Optical Coherence Tomography (OCT) is an emerging imaging tool that is now widely adopted in various medical settings such as cardiology and ophthalmology and is emerging in dentistry. In OCT, light of low coherence is used for image capturing which results in the introduction of speckle noise. Specifically, a degraded signal-to-noise ratio accentuates ambiguity in feature-extraction and contributes to the introduction of artefacts. This ultimately impacts its clinical utility where clear diagnostic detail is essential. In this work a concentrated-unsupervised deep learning denoising framework for OCT images is proposed, incorporating attention gate encoders. Attention gates are utilized to ensure focus on denoising the foreground and to ‘hard-threshold’ the background. Training data was created by processing the image with state-of-the-art denoisers (BM3D and NLM, etc.) to emphasize only essential data removal. The proposed framework was analysed quantitatively and visually, in comparison against state-of-the-art denoising algorithms. The experimental results show that the approaches verifiably remove speckle noise and achieves superior quality to well-known denoisers. The method improved the PSNR by 29.6 dB, CNR by 11.5 dB and, ENL by 1196.6 dB compared to original image and state-of-the-art denoisers.
The use of seismic waves to explore the subsurface underlying the ground is a widely used method in the oil industry, since different kinds of the rocks and mediums have different reflection rate of the seismic waves, so the amplitude of the reflected waves can unraveling the geological structure and lithologic character of a certain area under the ground, but the management and processing of seismic wave data often affects the efficiency of oil exploration and development. Different kinds of the seismic data bulk are always mixed and hard to be classified manually. This paper presents a classification model for four main types of seismic data, and proposed a classification method based on Mel-spectrum. An accuracy of 98.32% was achieved using pre-trained ResNet34 with transfer learning method. The accuracy is further improved compared with the pure fourier transformation method widely used in previous studies. Meanwhile, the transfer learning method and fine-tune strategy to train the neural network by training the first N-1 layers of the network separately and then train the fully connected layers further improves the training efficiency. Our model can also be seen as an efficient data quality control scheme for oil exploration and development. Meanwhile, our method is future-proofed, for further improvement of the seismic data processing quality control system, according to the spectrum characteristics, this model can be further extended into a error data classification model, reduces the workload of the bulk data management.
Autism spectrum disorder (ASD) is an incurable neurodevelopmental disorder with a wide range of clinical symptoms that mainly include social and communication deficits. Unfortunately, there is still no effective method for ASD diagnosis. Recently, researchers have presented a number of machine learning methods for ASD identification based on multi-site data, and these methods have achieved remarkable results. However, multi-site data is directly used, ignoring the heterogeneity between different sites. To address this issue, we propose a low-rank domain adaptive method with inter-class difference constraint (LRDAIC) for multi-site ASD identification based on resting-state functional magnetic resonance imaging (rs-fMRI). Firstly, we treat one site as the target domain and the remaining sites as the source domains. Then, data from these domains is transformed into a common space while considering inter-class difference, and the inter-class difference constraint term is further introduced to maximize the distance between different classes to enhance data discrimination ability. Moreover, each class of data from each source domain is linearly represented by all the data of the corresponding class from the target domain in this space. Finally, we evaluate the performance of our method on the basis of the ABIDE1 dataset, and the results demonstrate that our method is superior to several state-of −the-art low-rank domain adaptation methods.
Constructing visual encoding model is an important way to understand the human visual system. Data-driven encoding models learn features by directly fitting neural data, providing novel insights for visual encoding. Prediction accuracy is an important performance goal of visual encoding. The GaborNet visual encoding (GaborNet-VE) model has both good expressiveness and interpretability, while its prediction performance still needs to be improved. In this work, we proposed an improved GaborNet-VE model by adding data augmentation and nonlinear mapping module. The experimental results show that the improved GaborNet-VE model achieves higher prediction accuracy in V2, V3, V4 and LO. Our research provides a new framework for the data-driven encoding model.
Using communication networks and processing units in physical infrastructures has led to a new generation of systems which is known as Cyber Physical Systems (CPSs). These systems have several applications in different fields such as transportation networks, industrial control processes, water distribution systems, power systems, and etc. As it is shown in the literatures, CPSs are vulnerable to cyber-attack and physical attack in their communication channels and physical infrastructures, respectively. In this study, we will propose an observer which is used to detect false data injection attack in the turbine governor and load frequency controllers systems in the power systems. Then, a bank of attack isolation observers (AIOs) are proposed for identifying the location of the attacks in the generator buses. The numerical simulations on the IEEE 39-bus system show the feasibility and effectiveness of the proposed approach.
Huge industries and companies need to save energy while delivering various products to customers. This paper studied the Vehicle Routing Problem with Time Window (VRPTW). Using VRPTW optimization methods, product distribution companies would be able to choose almost the best routes for a specific number of vehicles to minimize the cost of fuel, driver's wage, driving range, and the time needed to provide the products to the customers. This study has focused on suggesting a programing method using the Simulated Annealing (SA) algorithm to minimize the total distance traveled by vehicles in a distribution chain and set the desired timetable in which vehicles are supposed to serve specific customers in the network. The result of this study is that problems as big as 350 nodes are calculated with zero-time window violation and zero capacity violation in less than 10 minutes, which is much less than many similar approaches.
Command information system can realize flexible matching between platform resources and tasks on the basis of platform combat capability as a service. In order to satisfy the intelligent decision-making need of command information system, the platform resource scheduling problem under the condition of capability as a service is described mathematically. Then, a scheduling model aiming to maximize the task processing quality is constructed and a model solving algorithm with heuristic strategy is designed. Finally, the method is simulated and verified with detailed examples. The experiment results show that the method proposed is effective and efficient, which can support intelligent decision-making of command information system.
The double inverted pendulum is a nonlinear system with unstable balance and rapid response. When two pendulums are allocated vertically, and there is no movement, the system is stable. In this article, we propose a control mechanism based on reinforcement learning. The system can improve its performance and balance by implementing the Q-Learning techniques besides the PID control scheme. The Q-Learning algorithm balances the pendulum to its inverted vertical position in the first step. In the second step, the Q-Learning algorithms are combined with classical PID to balance the double inverted pendulum and perform swing-up control. Eventually, simulation findings showed the efficiency of the proposed controller.
Quadruped robots are generally more stable and agile than biped robots. In this paper, a model of a quadruped robot similar to Mini-Cheetah has been presented and investigated in terms of dynamic gait planning. First, the model has graphically been designed by SolidWorks which is composed of four legs with twelve rotational joints (three joints for each leg). Then, the robot has been dynamically modeled in ADAMS, by introducing appropriate constraints and contacts. An intuitive graphical kinematic analysis has been conducted and a proper dynamically stable gait for the robot locomotion, i.e. trotting gait, has been planned. To extend and parameterize the gait study, the robot has further been simulated by linking ADAMS and MATLAB. The results showed the feasibility and effectiveness of the approach for quadruped gait planning and simulation.
The purpose of this paper is to linearize given non-linear differential equations and design a complete automatic control system for the vertical and horizontal motions of an underwater vehicle like submarine using a steady state decoupling scheme. The method of solution is linearizing the DSL (Dynamic Language Simulation) non-linear equations of motion by assumption of constant axial velocity and...
Aiming at the problem that truck brand type recognition is easily affected by complex scenes and illumination changes, this paper studies the truck brand type recognition method based on machine learning. The feature extraction method-Scale-Invariant Feature Transform (SIFT) and Histogram of Oriented Gradient (HOG) are compared and analyzed. Then different kernel functions of support vector machine classifier such as Linear Kernel, Gaussian Kernel, Sigmoid Kernel, Polynomial Kernel and Laplace Kernel are proposed and analyzed. Based on the truck face image set of Southeast University, a comparative experimental study is carried out. The theoretical analysis and experimental results show that: Based on HOG feature and Linear Kernel function, the performance of truck brand type recognition method based on support vector machine is better than the other four methods, and its recognition accuracy reaches 89.63%.
Aiming at the complexity and instability of blood glucose data of diabetic patients, this paper introduces an extreme learning machine algorithm (ELM) based on improved particle swarm optimization (IPSO) into the prediction of blood glucose concentration in patients with type I diabetes. First, the blood glucose concentration time series of diabetes patients collected by dynamic blood glucose monitoring is smoothed and normalized to improve the smoothness of the blood glucose data sequence and weaken the randomness of the original blood glucose data sequence. Then, the extreme learning machine is optimized, the improved particle swarm optimization algorithm is introduced to select the appropriate parameters required by the extreme learning machine, and the optimized algorithm is applied to the prediction of blood glucose concentration of diabetic patients. The experimental results show that the extreme learning machine algorithm based on improved particle swarm optimization has higher accuracy for short-term blood glucose concentration prediction of patients.
In this paper, we focus on image classification based on dictionary learning. The structured analysis dictionary learning (SADL) algorithm introduces a mapping matrix to the representation coefficient matrix and imposes a structural constraint to enhance the capability of discrimination, and it achieves promising classification results. However, in the formulation of SADL, the discrimination of the model is mainly based on the transformed coefficient matrix rather than the original coefficients, and thus the sparsity of the original coefficients has little impact on classification. In addition, classification is based on a linear classifier learned simultaneously with the dictionary, which may also restrict the performance of the algorithm. To address these issues, we propose a simplified SADL (SSADL) algorithm by simplifying the original formulation of SADL and introducing a new classification approach based on support vector machine (SVM). Simulation results on widely used databases demonstrate that the proposed SSADL algorithm achieves better performance than several state-of-the-art classification algorithms based on dictionary learning.
In this paper, we introduce an extrapolation method for solving two-dimensional linear and nonlinear fuzzy Volterra integral equations. The iterative algorithm based on a two-dimensional fuzzy trapezoid quadrature formula is constructed and obtains the error asymptotic expansion of the numerical method. Then, an extrapolation algorithm based on error asymp-totic expansion is built. By the first extrapolation, we can raise the order of convergence from O(h 2 0 ) to O(h 3 0 ), and by the second extrapolation, the order of convergence can reach O(h 4 0 ) . Finally, two numerical experiments demonstrate the accuracy and efficiency of the method.
Most of the causes of death are related to cardiovascular disease. Heart sound classification plays a key role in the early detection of cardiovascular disease. The distinction between normal and abnormal heart sounds is not obvious, making heart sound classification a challenging task. This paper proposes a method based on statistical analysis and feature engineering to analyze the features extracted from heart sound signals. The discriminative features are selected and input into the classifier to classify the heart sound. While ensuring accuracy, the method reduces the scale of the input data and computational time for classification. The selected distinguishable features achieved an accuracy of 96.5 % for the classification of abnormal and normal heart sounds. The experimental results show that the selected features can achieve a high accuracy rate in the classification of heart sound, which is suitable for real-time applications.
Network attacks are increasingly being paid attention to. People are constantly studying effective attack recognition and prevention methods. This paper presents a method of network attack traffic recognition based on quantum neural network for the first time. The problem of gradient explosion is avoided by making full use of the bounded gradient function of quantum neural network. A method of transforming the classical traffic characteristic value data into quantum state representation is proposed. And the preliminary experimental verification is carried out on Google Tensorflow Quantum platform. The feasibility of the method is verified.
Multi-hop reasoning is an effective and explainable approach to predicting missing facts in Knowledge Graphs (KGs). It usually adopts the Reinforcement Learning (RL) framework and searches over the KG to find an evidential path. However, there are few RL based reasoning methods in Knowledge Forests (KFs), the existing approach is usually to transfer multi-hop reasoning methods in the KG to the KF. However, this way often leads to a long reasoning path with a lot of redundant information. And the existing reasoning methods tend to capture vector representations from local structures, which will lead to get meaningless paths. In order to solve these two problems, we present, in this paper, a deep reinforcement learning based model named by MultiKR for KFs and KGs, which reduces redundant information in the path by learning global vector representations. In addition, we construct a novel reward function that guides the agent's choice of actions by considering accuracy, diversity and rationality to get valid and short path. We test our model on two KF datasets, that is, COMPUTER SCIENCE (CS) and MEDICINE (MED), and two KG datasets FB15K-237 and NELL-995. Extensive experiments show that our model is effective and competitive with many current state-of-the-art methods, and also performs well in practice.
t-distributed stochastic neighbor embedding (t-SNE), is a famous supervised, nonlinear dimensionality reduction and data visualization method in manifold learning, its core idea is to pursue the probability isomorphism of data points from high-dimensional space to low-dimensional space, that is, it requires the points in high-dimensional space to meet a certain probability distribution, and still meet a similar probability distribution after projecting them into low-dimensional space. Generally, t-SNE algorithm will be eventually transformed into a problem of solving Kullback-Leibler (KL) divergence by gradient descent (GD) method or stochastic gradient descent (SGD) method. However, gradient dependent methods are easy to fall into the trap of local optimum, and the closer they are to the optimal value, the more oscillatory sawtooth effect will occur. Therefore, in order to overcome the shortcomings, in this paper, a novel Sobol sequence initialized archerfish hunting optimizer (SSAHO) has been proposed, which can increase the randomness and robustness of the algorithm, and via numerical experiment, its performance in improving the optimization ability has been verified.
The daily average price of natural gas has an important impact on the business decisions of natural gas companies, so the accurate prediction of future natural gas prices has become a hot topic. The current single forecasting model has low accuracy for complex time series forecasting and cannot predict the future natural gas prices well. Based on the in-depth analysis of the properties of Prophet additive model and GRU neural network model, a Prophet-GRU nonlinear combined forecasting model based on improved BP neural network is designed based on the trend of natural gas price change from 1997 to 2020, compared with the pre-combined GRU, Prophet single model and the current more popular Long short-term memory LSTM model, the improved BP neural network-based Prophet-GRU nonlinear combined forecasting model has higher accuracy and is more suitable for complex time series forecasting, which provides a powerful help for natural gas enterprises' business decisions.
Disconnectors are extremely important equipments in a substation. Long-term use may lead to abnormal structure, resulting in inadequate closing, which means a serious accident. Therefore, the judgment of disconnector closing in place has become a very important problem. Judging whether the closing is in place requires very accurate state identification, which is difficult to judge by the current method. Firstly, this paper used mask R-CNN network for target recognition, and obtained a model with high recognition accuracy. Then, this paper used HED method to extract the edge, and proposed an algorithm to detect the closing state. Experiments show that this method can effectively identify the closing state of a horizontal rotary disconnector.
