Traditional detection methods need to consume a lot of manpower and material resources, causing great difficulties in the research of malicious code. Aiming at the problem of malicious code family classification, this paper proposes an improved convolutional neural network model based on Xception (Simplified-Xception). The model uses the dataset from the 2015 Kaggle Microsoft Malicious Code Classification Contest. Firstly, the malicious code is transformed into the gray image as the model input, the number of modules of the original model is reduced, and a layer of depth-separable convolution with a step size of 2 is added to enhance the generated gray image. The convolutional network model is improved based on the Xception model. The model in this paper is compared with the CNN model, ResNet50, and Inception-related improvement models. The experimental results show that the accuracy of Simplified- Xception is 98%, which is better than that of other inception related models. Compared with the Xception model, the accuracy of the Simplified -Xception model is increased by 1.3%, and the number of parameters is reduced by half.
Random feature maps attempt to approximate the kernel method with low computational complexity, and they are efficient and effective algorithms for dealing with the non-linear structure of data. Nevertheless, the existing random feature-based methods are sensitive to hyper-parameters, and their performances are affected by the randomness in feature maps. To improve the robustness of random feature maps and further speed up kernel-based image segmentation algorithms, a multiple random feature superpixel-based fuzzy clustering is proposed in this paper. First, we generate a group of random features from the original image via a novel multiple random feature scheme. Then, we use the superpixel segmentation to reduce the scale of multiple random features. The final segmentation results are obtained by performing the multi-view fuzzy clustering on the multiple random feature superpixels. Experiments on two shared benchmark data sets verify the superiority of our approach.
The COVID-19 has been posing threats to peoples health around the world due to its simple transmission route, strong mutation ability, and high infection. Currently, the superb vaccine breakthrough ability of mutant strains puts people's lives and health further at risk. Many measures are taken to fight against the epidemic. Keeping a safe social distance to prevent the spread of COVID-19 in public places. In this paper, we propose a method to monitor social distancing from a bird's-eye view. It mainly includes three parts. Firstly, an objector based on YOLOv5m is trained to detect human beings. It shows good performance compared with other models such as Faster R-CNN and YOLOv3. Secondly, we propose to use SORT algorithm to track each human by assigning it an ID. It also ensures the accuracy of multiple-object tracking. More importantly, it makes for tracking those who violate social distance. Thirdly, the Euclidean distance between detected people is used to judge whether human beings keep social distance or not. It is implemented by computing the pairwise distances of the detected bounding box centroids. Experiments conducted on birds view dataset show that our method can monitor social distance well. Hopefully the proposed method and dataset can provide some help in the fighting against COVID-19.
Model predictive control is an effective way to achieve the control of wastewater treatment process (WWTP). However, it is a challenge to control multiple objectives due to the complexity of WWTP. To solve this problem, an eigenvector based multiobjective model predictive control (EMMPC) strategy is developed for WWTP to handle the conflicting objectives. First, a multiobjective control scheme is designed with adaptive fuzzy neural network prediction (AFNNP) model and gradient eigenvector optimization (GEO) algorithm. Then, AFNNP can be used to describe the nonlinear of WWTP to predict the controlled variables. Second, GEO is presented to obtain the control laws of the multiple control objectives. Specifically, GEO can reduce the computational burden by avoiding the determination of the control objective weights. Third, the stability of EMMPC is provided in theory. Finally, EMMPC is implemented on the benchmark simulation platform to demonstrate the effectiveness of the presented multiobjective control method.
Fuzzy transfer learning has the potential to enhance the performance of a task in the target scene by leveraging knowledge from the source scene. However, the existence of scene discrepancies between the two scenes poses two major problems for fuzzy transfer learning: 1) the mismatch between the distribution of source knowledge and the target task, and 2) the mismatch between the learning ability of the target task and the source knowledge. To address these challenges, we propose a fuzzy transfer learning algorithm called Knowledge-Task Matching Fuzzy Transfer Learning (KTM-FTL), which aims to adaptively match the source knowledge and the target task. Firstly, we develop a knowledge matching mechanism that balances matching accuracy and diversity to reconstruct the knowledge from the source scene. This process enables the source knowledge to align more accurately with the requirements of the target scene's task. Secondly, we design a task matching mechanism that allows for learning the source knowledge at different levels of granularity. This approach enables the learning ability of the target Feedforward Neural Network (FNN) to match the complexity of the source knowledge, thereby improving its generalization performance. Thirdly, we conduct theoretical analysis of the proposed KTM-FTL, including an examination of its computational complexity and error bound. These analyses provide valuable insights for the successful application of KTM-FTL. Finally, we evaluate the performance of the proposed KTM-FTL on several benchmark problems as well as real-world problems. The experimental results demonstrate the significant improvements achieved by KTM-FTL when compared to state-of-the-art algorithms.
The current research on biped robots is mostly concentrated on ideal conditions, while in reality, the robot system is always subject to many external disturbances and internal parameter perturbations, which will lead to such defects as weak anti-interference ability, low real-time performance, low accuracy of controller trajectory tracking, weak balance control ability, and solving its standing posture balance problem is the premise to solve the robot walking stability. Combining the advanced depth deterministic strategy gradient algorithm (DDPG) and extended state observer (ESO), a control strategy for biped robot upright disturbance rejection balance is designed, which can quickly and timely adjust the pitch angle of the robot hip joint to restore the balance state. The DDPG controller adjusts the joint angle in real time through the difference between the actual zero torque point (ZMP) and the expected value to suppress the disturbance; ESO estimates and compensates the disturbance of unmodeled dynamics and internal parameter perturbation, and outputs joint angle compensation to improve control accuracy and rapidity. In order to verify the effectiveness of ESO-DDPG controller, it is applied to the interference suppression experiment of a stationary NAO robot. When the biped robot is disturbed by external disturbances and internal parameter perturbations, it can quickly restore its standing posture balance in real time.
In the realm of linear regression, the concept of utilizing relaxed regression targets for classification has shown considerable success. However, it suffers from a lack of strong discriminative ability. In this study, we build upon the theory of relaxed regression targets and propose a more concise and discriminative model for multi-class classification. Our approach introduces the £ 2,1 -norm regularized term to enhance the efficacy and compactness of the learned projection matrix. This regularization term not only generates a sparse row structure but also facilitates feature selection during the training phase. Consequently, both the accuracy of classification and the convergence speed of the model are improved. Furthermore, when compared to various variants of the linear regression model, our designed method exhibits superior performance in image classification. Through extensive experiments conducted on image databases, we provide evidence that our proposed approach excels in recognition capability, surpassing the performance of state-of-the-art methods in terms of classification accuracy.
In chemical processes, reliable soft sensors are generally established by enough labeled data. However, in most multimode processes, the collection of sufficient labeled data is difficult due to the high cost and complexity. In this work, transductive transfer broad learning (TTBL) is proposed for multimode quality prediction. By transferring the useful information from the related domain, unlabeled data in the prediction domain is utilized for modeling. First, the data feature is extracted by the feature and enhancement nodes. The similarity information of current and related domain data is captured by the $k$ nearest-neighbor graph. Then, label information in the related domain can be transferred and similar information in the same domain can be retained by the manifold regularization framework. Finally, the output weight can be effectively calculated by the ridge regression algorithm. Experimental results on continuous stirred tank reactor datasets show the superiority of TTBL, compared with several common methods.
In recent years, spatiotemporal sequence prediction has received increasing attention from researchers and has a wide range of promising applications in the fields of meteorology, traffic flow prediction, and autonomous driving. However, existing spatiotemporal sequence prediction models have some problems, such as slow convergence, training difficulties, and loss of image structural and detail information. We propose a novel end-to-end two-branch spatiotemporal sequence prediction model, which has been improved on these issues. We have compared our model with current advanced models using two datasets and found that our model reached or exceeded the level of the other advanced models in several metrics.
In this paper, a stochastic configured fuzzy system with semantic interpretability is proposed and used for classification. Firstly, transfer entropy is adopted to conduct causal analysis on modeling variables. To reduce computational complexity and achieve effective screening of data variables, the information granule is designed to preprocess the data. Secondly, a research is carried out on the fuzzy modeling method based on the idea of stochastic configuration. Besides, an improved fuzzy system evaluation index for model parameter learning and system structure optimization is proposed. The improved index introduces the axiomatic fuzzy set theory to evaluate the information metric of modeling results, which increases the semantic interpretability of the fuzzy system. Finally, the proposed method is applied in some medical data sets. The experimental results show that it offers better performance in solving the classification problem.
In this paper, we present an adaptive control method based on immersion and invariance (I&I) theory to improve the position tracking control performance of micro-mechanical (MEMS) gyroscopes under external disturbance. Firstly, a dynamic observer of lumped disturbance based on I&I theory helps reduce the influence of external disturbance and mechanical noise on the system and improves the tracking accuracy of the system. Secondly, the position controller is designed based on I&I theory to accurately track the given values of the system and improve stability. Finally, based on Matlab&Simulink, the performance of the proposed adaptive control strategy based on I&I theory under different disturbances was studied and tested. The results indicate that the proposed control strategy based on I&I theory has good stability, can quickly and accurately track nonlinear disturbances at different frequencies, and can quickly and accurately track the position of the gyroscope, verifying the effectiveness and feasibility of the proposed method.
Graphs have become a widely-used tool to model data with relationships in real life for a long time. To discover the important contents in the graph, many graph neural networks (GNNs) have been come up with. Nevertheless, these models tend to adopt ReLU as their activation function for its nonlinearity, effectiveness, and efficiency. Owing to its own deficiency, it would cause many generated feature elements to be zero, which would miss part significant features. To overcome this problem, an adaptive weight vector to tune the features was provided. By limiting the elements of the weight vector, it can be a better substitute for ReLU. Besides, such a weight vector adaptively measures the importance of each feature element to work as a feature selection operator. To show the function of the weight vector, we examine a GCN with the weight vecotor in node classification, and it exhibits overall improvements over three well-known citation networks.
The poor position control precision of a certain part of automatic loading equipment will lead to the failure of the whole loading process. To solve this problem, this paper proposes a method of using deep reinforcement learning to improve the positioning accuracy of the hydraulic forklift, to meet the control requirements of related equipment. Since the motion interval of the forklift is continuous, the normalized dominant function (NAF) model-free algorithm is adopted. to evaluate the scheme, the movement model of the forklift is built in the simulation software, and the corresponding deep reinforcement learning framework is built in the external tool. Finally, simulation is carried out in a virtual environment to prove the effectiveness of the proposed method.
The mobility of nodes makes the topology of wireless ad hoc networks change frequently. The traditional routing protocols are usually based on fixed routing policies, which are difficult to adapt to the dynamic changes of the network. This paper investigates the packet routing optimization problem in wireless ad hoc networks by using deep reinforcement learning (DRL) technique. To seek the optimal transmission path between the source node and the destination node for a packet, average delivery time is applied to assess the transmission path while the routing Markov decision process (RMAP) model is utilized to formulate the routing optimization problem. Furthermore, a DQN-routing algorithm is developed to deal with such an optimization problem. The distinguished features of the proposed algorithm lie in the facts that it can reduce the average delivery time and the network congestion, is also suitable for dynamic topology network environment. Finally, a simulation example is given to verify the validity of the theoretical results.
The bird-oid object (Boids) model proposes a control algorithm to make the positions between agents achieve cooperative stability. By changing the parameters of cohesion and repulsion in the algorithm, the agents in the swarm can be made to converge to different positions, causing expansion and contraction of the formation. But it is often more difficult to select the appropriate parameters to form the ideal formation. Therefore, this paper proposes a method to improve the cohesive and repulsive parameters in the Boids model based on Q-learning network to achieve a simulation scenario with continuous obstacle avoidance and maximum coverage of space.
In this paper, fuzzy transformation is considered to modify stochastic configuration network for data regression. Firstly, the mathematical representation for stochastic configuration network based on fuzzy transformation, named FTSCN, is deduced. Then, the parameter learning algorithm for FTSCN is obtained. Further, the ensemble scheme is applied to improve the performance of FTSCN. Some simulation experiments in the benchmark datasets show the effectiveness of the designed model.
Visual Question Answering (VQA) is a multimodal task that answers questions by exploring the content of the images. A comprehensive understanding of both the images and questions is critical for VQA. Although great success has been achieved in VQA recently, the problem of language bias hinders the application of VQA models in real-life situations. Language bias leads to the inability to generalize the VQA model to test sets with different distributions. Therefore, we propose a distance metric learning optimized attention mechanism (DMA), which selects and directly adopts the image information that is related to the question, performing feature selection and compression via a novel unsupervised metric learning scheme. Specifically, we only keep the image features with higher similarities with the question to answer the question. We adopt unsupervised metric learning to separate image features with high similarity to the question from those with low similarity to the question in the feature representation space. By doing this, the difficulty of VQA models understanding of images is reduced and language bias is mitigated. Extensive experiments have proved the effectiveness of our method for both Vanilla Transformer and Cross Transformer, with improved results on VQA v2 and VQA-CP datasets. Particularly, by building on top of model LMH+CSS, we achieve State-of-the-art performance on VQA-CP datasets.
Achieving fully distributed consensus in multi-agent networks (MANs) based on relative output information is of significant importance in the field of distributed cooperative control. Despite the recent development of consensus theory, there is still lack of an efficient approach to constructing fully distributed consensus protocols for MANs using only relative output measurements, especially for the case with inherent nonlinear agent dynamics under directed communication graphs. This paper studies fully distributed minimal-order protocol design for consensus of MANs with Lipschitz-type nonlinear agent dynamics under general directed communication graphs. It is revealed that the minimal order of the observer estimating consensus error equals the difference between the dimension of a single agent's state vector and the rank of the output matrix of the agent. Based on the designed distributed minimal-order observer, fully distributed consensus protocols are developed. The proposed protocols have the abilities of reducing communication burden and also can avoid network attacks on the internal states. Numerical simulations are presented to illustrate the analytical results.
Optimal control is commonly utilized in the wastewater treatment process (WWTP). In WWTP the effluent quality and operational cost should be optimized by solving the optimal set points of dissolved oxygen and nitrate nitrogen. However, due to the unknown disturbances of WWTP, the optimal operation objectives have strong uncertainties. Therefore, it is crucial to solve the optimization problem of the operation process, which has become the major obstacle to the optimal control of WWTP. For overcoming this obstacle, a double closed-loop robust optimal control (DCL-ROC) method is designed in this paper. First, a double-closed-loop scheme, contains a close optimization loop and a close control loop, is formulated to realize the optimal control of WWTP. Second, a closed-loop robust optimization (CL-RO) algorithm is designed to solve the optimal set points with enhanced robustness. Third, an adaptive controller based a fuzzy neural network is proposed to control the process variables with superior accuracy and robustness. The experimental results by using benchmark simulation model No.1 (BSM1) demonstrate the effectiveness of DCL-ROC.
Cement, as an important material, is widely applied in the world. The macroscopic properties of cement are strongly related to its microstructure. To acquire cement microstructure images, scanning electron microscopy (SEM) is a popular scanning technique for obtaining high-quality backscattered electrons (BSE) images of cement microstructure. Nevertheless, this technique generally cannot acquire large-scale microstructural images due to the limited scanning area. Therefore, this paper proposes a BSE texture synthesis framework based on convolutional neural network (CNN) technology. In this framework, the generator can generate BSE images. The framework also optimizes the generator by perceptual loss. Ultimately, experiments demonstrate that our method can generate arbitrary size BSE textures similar to a given BSE texture.
