Floating raft aquaculture is an important part of mariculture, and the use of single source data is easy to inhibits the extraction of floating raft aquaculture, so the advantages of integrating multisource data are particularly important. However, multisource remote sensing fusion still has the problem of data information imbalance and feature redundancy. In this paper, a channel exchanging bottleneck attention network (CEBANet) is proposed to extract features from multisource remote sensing data, determine whether features are redundant by using the scaling factor of batch normalization (BN) layer, replace the current redundant features with another modal feature. Bottleneck attention module (BAM) and feature calibration module are added to improve the capability of feature extraction and receptive field processing. The CEBANet is optimized by the sparse constraint on channel exchanging condition and cross-entropy loss and consistency loss. Using GF-5 and GF-3 data from Jinzhou District of Dalian City, it is proved that the proposed model can realize the complementary advantages of multisource remote sensing data and improve the accuracy of extracting floating raft aquaculture areas.
In this paper, the 5-instant discretization formula (5IDF), renowned for its high computational accuracy in approximating the first-order derivative, is employed for the discretization of the continuous-time zeroing neural network (CTZNN). The formula is then compared with other common difference formulas of varying precision. Based on this formulation, the 5IDF-type discrete-time zeroing neural network (DTZNN) model is introduced and studied for solving discrete-form time-varying matrix inversion. The feasibility, effectiveness, and superiority of the proposed 5IDF-type DTZNN model for addressing discrete-form time-varying matrix inversion are further substantiated through numerical experiments.
Idiopathic pulmonary fibrosis (IPF) is a chronic lung disease that significantly impacts individuals' health. Early diagnosis of IPF can enhance patient survival rates. Considering the limited research on automatic detection of IPF in academic circles and the scarcity of labeled data, this paper proposes a transfer learning-based algorithm for detecting IPF lesion areas. The proposed method consists of two stages: firstly, construct an IPF-like dataset comprising natural texture images with similar visual features to IPF, and pre-train the U-net network using this dataset; secondly, fine-tune the network using both the pre-trained weights and actual IPF data, followed by utilizing the trained network for lesion area detection. Experimental results demonstrate substantial improvements in terms of accuracy and sensitivity compared to corner distribution-based methods for IPF detection.
The frequent occurrence of emergencies in recent years has seriously affected the stability and development of society and economy. Effectively responding to emergencies has become a crucial research topic. A large number of emergency cases contain potential rules and valuable response experience in the process of occurrence, development and response, and the association rule mining can discover the fine-grained and valuable "if scenario then response" association rules in the cases. The knowledge of emergency response rules obtained from mining can assist the generation of response programs under actual emergency scenarios and provide reference basis for emergency decision-making. The existing methods only consider the support and confidence indicators but ignore the lift and comprehensibility of the rules and the lack of domain knowledge guidance in the mining process. Aiming at the problem, this paper proposes a response rule mining method based on knowledge guidance and improved genetic algorithm, which effectively reduces the number of low-quality rules and attributes of the mined rules and improves the quality and efficiency of association rule mining.
On the basis of the gradient-based differential k-winners-take-all (GD-kWTA) network, this paper focuses on its applications on multi-robot competition and coordination under noise-free condition. The GD-kWTA network offers improved handling of temporal lag in control variables under noise-free condition. At the same time, theorems and relevant validations are presented to ensure the exponential convergence and antinoise elasticity of the proposed GD-kWTA network. Finally, the GD-kWTA network is integrated with a consensus filter as a robust control strategy for the simulation of the coordinated competitive behavior of various robots. Even under noise-free condition and when handling complex target trajectory curves, the GD-kWTA network successfully achieves satisfactory results, which underscores its effectiveness in tackling the issue of temporal lag.
The research on time derivativer (TD, a.k.a., tracking differentiator) is a hot topic in the field of industrial automation control. In this paper, a new kind of time derivativer (called integral-aided denoising Zhang time derivativer, IDZTD) is proposed with the aid of Zhang neural dynamics (ZND). The comparative study is conducted among Han tracking differentiator (HTD, also called classic tracking differentiator), standard Zhang time derivativer (ZTD), and IDZTD. The research indicates that the IDZTD has better performance and noise immunity in the numerical experiments with different initial activation functions and design parameters. In addition, to some extent, the larger the design parameters are, the better the performance of IDZTD is. Besides, the noise immunity of IDZTD with nonlinear activation is better than that with linear activation.
The rapid spread of information through online platforms and social media networks has led to an increase in the propagation of rumors, which can have detrimental effects. Researchers have proposed various deep learning models for multimodal rumor detection. However, these models often handle each modality individually, limiting the abilities of information complementation and modal enhancement. To address this challenge, we propose a Cross-modal Information-enhanced Fusion Network (CIFN) for rumor detection on social media platforms. CIFN enhances the representation of different modalities in a unified framework before effectively combining textual and visual information to accomplish the rumor detection task. Specifically, CIFN introduces the Feature Information Enhancement (FIE) module, which enhances different modal information by selectively focusing on relevant features and capturing interdependencies between modalities. Additionally, CIFN introduces a Review-based Fusion Mechanism (RFM) to integrate textual and visual features, considering the weight allocation of different modalities at the feature level. Extensive experiments conducted on two public datasets show that the proposed CIFN outperforms existing methods in rumor detection.
This article presents a cascading neural network model to implement one-to-many associative memories, where the output of the preceding network serves as the input of the next network, and the model is driven by external input patterns. The design procedure is provided by considering the existence and global asymptotical stability of equilibrium point. Compared to traditional multidirectional associative memory models, the model proposed in this article requires fewer neurons to achieve the same functionality and avoids the issue of pseudo-equilibrium points. The illustrative example is given to support the effectiveness of the theoretical results.
Crash data is the foundation of traffic safety analysis. In China, the accident reporting form (ARF) only allows reporting one cause for each crash based on the prespecified crash cause code, which may lead to inaccuracy in recording crash data, especially for state-related crashes. This study investigated the directly contributory factors behind the state-related crashes through the development of natural language processing (NLP) and deep learning models based on 1,625 state-related crash narratives. Based on the directly causal factors described in the crash narratives, the state-related crashes were labeled by speed-related, turning-related and other causes. Then the crash narratives were vectorized for model training and frequent analysis. The text-CNN, LSTM, and GRU, and SVM models were applied to reclass the vectorized crash. The results showed the text-CNN model demonstrated the best performance in text classification, with an AUC value of 0.90 for micro-average curves. The results from this study can engage the usage of crash narratives and help identify the actual causative reason behind some inaccurate crash value designation.
The 5G network provides low latency, high speed, and large bandwidth characteristics, shorting the transmission time and providing efficient services to users. In a typical 5G network architecture, multiple resources, i.e., communication and computation resources, are provided to meet various requirements of the services. Therefore, proper resource allocation is important to improve the quality of services (QoS) of users while network slicing can properly allocate the resources according to their service requirements. In this paper, we propose a novel approach, deep reinforcement learning (DRL)-based network slicing with maximum QoS satisfaction (DRL-MQS), for multi-resource allocation. The concept of this approach is to first calculate the delay distribution of each service to obtain the overall QoS satisfaction ratio, and then use the DRL approach to find the resource allocation with an optimal QoS satisfaction ratio. The evaluation results show that the DRL-MQS can improve the QoS satisfaction ratio by 10.54% compared with the previous approach which tries to maximize resource utilization. Moreover, no matter what packet arrival rate, DRL-MQS always has the best performance.
The improvement of deep learning algorithms for small object detection in low-resolution images remains a significant challenge. Time-of-Flight (TOF) sensors can replace cameras for indoor use, offering privacy protection but facing limitations in accurately detecting small objects. This paper proposes an algorithm that combines Adaptive Histogram Equalization (AHE) and Contrast Limited Adaptive Histogram Equalizer (CLAHE) for image enhancement, further optimizing the Infrared (IR) images and depth maps collected by TOF sensors. Simultaneously, using the enhanced IR images and fused depth maps data based on the improved lightweight YOLOv5n algorithm, the performance of the proposed algorithm is validated. Experimental results demonstrate that the novel algorithm outperforms existing methods, with average precision scores of 98.4% and 72.1%, respectively.
In this paper, we propose a multi-loss network (PFME) based on progressive fusion and mixtures of experts for multimodal sentiment analysis. PFME comprises a progressive attention fusion (PAF) module and a module based on mixtures of attention experts (MAE). The PAF module leverages a learnable shared query to extract modal-shared representations through cyclic iterations. In each iteration, the query continuously reinforces the sentiment dynamics of multilevel features under double cross-attention. The MAE module employs multiple attention experts to complement various aspects of intra-modal information. A router assigns multi-level semantic features to one of the experts, resulting in an exclusive routing line beneath several stacks that generates a modality-specific representation. In particular, we develop three loss functions to improve the performance of these two modules. At first, we exploit contrastive loss on high-level semantic features between modalities to deepen inter-modal associations; next, we utilize orthogonal loss in the PAF module to preserve the shared query paradigm invariant; and finally, we deploy balance loss on the MAE module to equalize assignment probability across experts. Extensive experiments on the CMU-MOSI and CMU-MOSEI datasets show that PFME achieves state-of-the-art performance.
The scientific research on issues related to numerical linear algebra constitutes an interesting topic for study in the last decades. In this manuscript, a general and fast recurrent neural network (RNN) model based on an extended activation function (AF) will be designed and applied for the solution of different numerical linear algebra problems. Theoretical convergence analysis and simulation examples, in simulink will show the efficiency and the accelerated convergence ability of the new dynamical system.
In this paper, an adaptive neurodynamic algorithm based on multi-agent system is proposed to solve the multi-constraint matrix-valued optimization problem. The optimization problem with linear matrix constraints AXB = H and specific inequality constraints can be equivalently reformulated as a distributed optimization problem using the penalty method and matrix decomposition. Furthermore, the adaptive penalty technique is introduced to enhance robustness and reduce computational cost. The problem is then solved using the improved Lagrangian function. Through theoretical analysis, we have proven that the equilibrium point of the proposed neural dynamics algorithm is the optimal solution to the matrix-valued optimization problem. Moreover, the equilibrium point of the neural dynamics algorithm is bounded and Lyapunov stable, with the algorithm state solution converging to the optimal solution of the matrix-valued optimization problem. Finaly, we provide a numerical example based on switching topologies in multi-agent systems. Numerical simulations validate the correctness of the theoretical results and demonstrate the capability to effectively solve the matrix-valued optimization problem under switching topology conditions.
The ability to predict the future states of nearby traffic agents is critical for autonomous vehicles. Recently, it has become a new paradigm to sense and predict the future occupancy of the surrounding targets from the Bird’s Eye View (BEV) perspective, utilizing information captured by multiple cameras mounted on the vehicle. However, modeling the underlying spatiotemporal interactions between traffic agents is a challenging part. This paper proposes a novel spatiotemporal BEV pyramid network which employs Swin Transformer to extract BEV features transformed by images and predict across multiple scales. This network is designed to preserve spatial features at low resolution and capture semantic features embedded at high resolution. In addition, a Feature Alignment Module (FAM) is introduced to aggregate information at multiple scales and reduce mispredictions caused by feature misalignment. Through validation on the nuScenes dataset, the proposed method improves on the compared previous approaches in accuracy and demonstrates an enhancement in predicting the occupancy of various targets in the BEV.
In this paper, we propose a selective ensemble-based EEMD-ROCKET interval forecasting model and apply it to the task of power load interval prediction. Initially, we decompose the power load sequence using modified ensemble empirical mode decomposition (EEMD), obtaining a set number of Intrinsic Mode Function (IMF) components and a trend term. Subsequently, we employ a bootstrap-based approach and selective ensemble technique to construct an interval forecasting model based on ROCKET for predicting each IMF. Additionally, a point forecasting ROCKET model is utilized to predict the residual trend. The final prediction interval results are obtained by directly summing the predicted results of each IMF component and the trend term. We conduct experiments on a publicly available electricity load dataset, validating the performance of the proposed model.
The high-accuracy geological model of coal seam is an effective guarantee to realize unmanned mining. To reduce errors in the geological model between static and actual, this paper proposes a temporal-spatial attention-based dynamic correction method for the geological model of the coal seam by combining the cutting trajectory of the shearer with the original geological model of the coal seam. The cutting trajectory provides a reasonable extrapolation basis in the temporal aspect, while the original geological model provides a reference for adjusting the target coal seam in the spatial aspect. The method utilizes a convolutional layer to explore the relationship between the targeted coal seam and adjacent geological models; it also uses a recurrent layer to investigate the direct relationship between the targeted coal seam and the known coal seams. Finally, the attention mechanism is combined to determine the temporal and spatial dependencies of the targeted coal seam automatically. The proposed method was validated using the accurate geological model of coal seams and historical data of the cutting trajectory. The results demonstrate that the proposed method effectively improves the local accuracy of the geological model and enhances its practical value and applicability.
Motion sickness is a common affliction that affects nearly half of the global population and poses challenges to comfortable travel experiences, necessitating diverse intervention strategies. Pharmacological interventions have shown promise but present limitations due to side effects. To date, there is no non- pharmacological alleviation method for motion sickness that has been proven effective in a large population. In this study, we developed a novel mindfulness brain-computer interface (mindfulness BCI) for alleviation of motion sickness. Thirty-one subjects who were susceptible to motion sickness participated in our experiments in a real-world car motion environment. The experimental results demonstrated that the proposed mindfulness BCI can effectively alleviate motion sickness symptoms, where more than 90% of subjects reported the effectiveness of the mindfulness BCI system and a significant reduction of Motion Sickness Susceptibility Questionnaire (MISC) scores was obtained when comparing the BCI-based meditation state with the resting state for the subjects (t 30 =4.968, P<0.001). To the best of our knowledge, our mindfulness BCI is the first effective non- pharmacological alleviation method for motion sickness as demonstrated by a strict test.
Wind tunnels are large-scale ground test facilities to conduct aerodynamic characteristic tests on various types of aerospace vehicles, advanced rail transport, and large buildings, are of great significance for the development of aeronautic and astronautic. To address the issues of complex design and analysis processes, as well as the over-design in traditional wind tunnel structure design, an optimized wind tunnel structure design method is developed that embeds the multi-fidelity finite element analysis model into the surrogate model. A primary Kriging model is trained with low-precision finite element samples. Next, a hierarchical Kriging model is constructed based on bridge function and high-precision finite element analysis samples supplemented. Then, the surrogate model could be updated automatically using expected improvement (EI) method. Experiments demonstrates that the new optimized structure for the acceleration section of wind tunnel achieving a 40% exceeded structural mass reduction with the constraints structural stress criterion, first-order natural frequency and deformation limitation.
Hierarchical models have superiority in solving the data dimension explosion problem. An appropriate hierarchical structure is an important factor affecting the robustness of model. A hierarchical structure is constructed, which combines the parallel learning and serial connection manner. Meanwhile, a hierarchical neural network based on parallel and serial structure (PSHNN) is proposed. Firstly, the hierarchical clustering method is used to determine feasible divisions of the data. In this way, the data can be represented as a set of several subsets. These subsets are input into the serial hierarchical structure with layer by layer manner. Each serial hierarchical structure is constructed by several sublayers. These sublayers are connected via serial manner. Secondly, several parallel layers are established to perform the parallel learning, and each parallel layer contains a serial hierarchical structure. Each parallel layer can be considered as a parallel learning channel. Finally, the Bayesian information criterion (BIC) is used to integrate the learning results of each parallel channel, and determine final output of model. Performance of PSHNN was verified on several benchmark datasets, and an actual collected fatty liver medical dataset. Simulation results prove the superiority of the designed PSHNN.
