By combining gradient-based reinforcement learning (RL) with gradient-free evolutionary algorithms (EA), evolutionary reinforcement learning (ERL) algorithms have shown effectiveness in continuous high-dimensional tasks. However, traditional ERL faces challenges such as distributional differences between EA and RL samples affecting critic evaluation, limited exploration due to reward-based metrics, and high interaction costs from low-quality samples. To address these issues, we propose MCE-ERL, an ERL algorithm with multi-distribution critics. We uses a value group and multi-distribution sampling to better cover the sample space, leverages uncertainty for exploration, and applies importance sampling to filter low-quality samples. Experiments on continuous control tasks show that MCE-ERL achieves faster convergence and better performance than baselines, offering insights for reducing sample collection loss in real-world robotic applications.
Pantograph is the crucial component of rail transit vehicles. Detecting and resolving faults swiftly is important in the pantograph. Existing fault diagnosis technologies with artificial intelligence have difficulties in dealing with some rare faults due to the scarcity of training samples. This paper proposes a new algorithm for few-shot fault diagnosis of pantograph. The proposed method constructs a reconstructive adversarial network with hierarchical structure to mine multi-scale feature from few-shot samples. A hierarchical data generation module is designed leveraging adversarial learning to generate scarce category samples. An initial classifier is employed to constrain the generated results, while the classifier is dynamically reconstructed to enhance overall performance. This paper designs detailed experiments to demonstrate that, compared with other existing methods, the proposed algorithm can better solve the problem of pantograph fault diagnosis.
The tamper-proof feature of blockchain makes the correctness of smart contracts crucial before they are run on the chain. Using model checking methods to verify smart contracts helps ensure their correctness. Methods for converting smart contracts into semantically rigorous formal models that can be used in model checking methods still face the challenges of poor usability and difficulty in breaking away from manual modeling. Colored Petri nets, as discrete state-based systems, have similarities to smart contracts and are suitable for analyzing interactions between transactions. The analysis of control flow and data flow interactions, and transactions on the blockchain is missing in the research on converting smart contracts into colored Petri nets. In this paper, we propose an automated modeling method based on colored Petri nets, using template modeling and combination with data flow interactions and transaction modeling. Firstly, the data and control flows are modeled and combined, and secondly, the external user invocation is modeled to be able to correctly portray the on-chain contract invocations. The correctness in the smart contract is then verified based on the colored Petri net. Finally a prototype tool is implemented to demonstrate the usability and correctness of the modeling algorithm.
Chinese Train Control System level 3 (CTCS-3) train control onboard equipment plays a crucial role in ensuring train safety and improving operational efficiency. And the on-board interface equipment enables the interaction between the on-board Automatic Train Protection (ATP), ground equipment, drivers, and trains. While its fault accounts for a relatively high proportion of faults, therefore, this paper proposed a fault diagnosis method for on-board interface equipment based on temporal knowledge graph completion. Firstly, this method introduces temporal series to integrate travel logs and fault statistics, which extract fault phenomena and perform entity alignment, construct the temporal knowledge graph. Secondly, on this basis, we constructed a fault diagnosis network based on knowledge graph completion, which incorporates the Temporal-Translating Embedding (T - TransE) vectorization algorithm, Bidirectional Long Short-Term Memory (Bi-LSTM), and Self-Attention (SA) mechanism for temporal feature extraction. Finally, the T - TransE vectorization model was pretrained using the on-board interface equipment fault data from a railroad station in recent years for selecting the most effective temporal integration method. In order to validate the superiority of the proposed fault diagnosis method and the effectiveness of the data combination method, the diagnostic network without data combination and temporal relationship introduction, and other common fault diagnostic networks, were tested using the on -board fault data. Compared with the following fault diagnosis methods, the model proposed in this paper achieved the highest fault diagnostic accuracy of 96.8%.
Generalized Voronoi Diagrams (GVD) are extensively utilized in path planning, particularly in algorithms based on RRT*, where they guide the sampling process and thereby enhance the search efficiency of RRT*. Traditional methods for constructing GVD, however, often demand substantial computational resources, especially in complex and large-scale environments. In this article, we propose an efficient GVD generation algorithm leveraging Convolutional Neural Networks (CNN). By harnessing the learning capabilities of CNN, the model can swiftly predict the corresponding GVD probability map from an input image containing obstacle information. The training dataset is derived from traditional GVD computation methods, with GVD pruning applied to complex environmental scenarios. The training outcomes indicate that for simple environments, the CNN-based GVD generation aligns closely with the results obtained via traditional methods. In complex environments, the model is also capable of accurately predicting smooth and efficient GVD, consistent with manually pruned GVD results. Simulation results demonstrate that the model's predictions can be seamlessly integrated into the RRT* algorithm, providing effective sampling heuristics. This integration leads to a significant improvement in GVD computation speed compared to traditional methods and exhibits robust generalization capabilities in complex environments.
To enhance diagnostic efficiency and precision, this paper proposes a novel fault diagnosis method that combines multi-domain feature extraction with a Back Propagation Neural Network (BPNN). First, the characteristic curves of photovoltaic (PV) array current-voltage (I-V) and power-voltage (P-V) are analyzed to identify key feature parameters such as voltage, current, and power. Traditional fault detection methods often rely on unimodal information from either the time domain $(D_{t})$ or frequency domain $(D_{f})$ , resulting in incomplete characterizations. To address this limitation, this study utilizes multi-domain feature extraction from the time domain, frequency domain, and time-frequency domain $(D_{tf})$ to construct a multi-domain feature set. Variational Mode Decomposition (VMD) is used to prevent over-decomposition during feature extraction in the $D_{tf}$ . Second, a fault diagnosis method is developed by optimizing the neural network weights and thresholds using the Circle Adaptive-Grey Wolf Optimizer (CA-GWO) algorithm. Experimental re-sults demonstrate that the proposed multi-domain BPNN-based approach effectively and accurately detects faults such as short-circuit $(F_{sc})$ , open-circuit $(F_{oc})$ , abnormal aging $(F_{aa})$ , and local shading $(F_{ls})$ in PV arrays. Furthermore, extensive testing shows that the proposed method significantly improves the accuracy and speed of fault diagnosis compared to classical BPNN methods.
Today, numerous web services with similar functionalities are available on the Internet. Users often evaluate the Quality of Service (QoS) to choose the best option among them. Predicting the QoS values of these web services is a significant challenge in the field of web services. A Canonical Polyadic (CP)-based tensor network model has proven to be efficient for predicting dynamic QoS data. However, current CP-based tensor network models do not consider the correlation of users and services in the low-dimensional latent feature space, thereby limiting model's prediction capability. To tackle this issue, this paper proposes an Extended Canonical polyadic-based Tensor Network (ECTN) model. It models the correlation of users and services via building a relation dimension between user feature and service feature in low-dimensional space, and then designs an extended CP decomposition structure to improve prediction accuracy. Experiments are conducted on two public dynamic QoS data, and the results show that compared with state-of-the-art QoS prediction models, the ECTN obtains higher prediction accuracy.
This work studies the distributed permutation flow-shop scheduling problems (DPFSP) for minimizing the maximum completion time (makespan). First, a mathematical model is established for the DPFSP. Second, four meta-heuristics, named genetic algorithm (GA), particle swarm optimization (PSO), artificial bee colony (ABC), and differential evolution (DE), are used and improved to solve the problems. Third, the Nawaz-Enscore-Ham (NEH) heuristic is employed to initialize the factory allocation scheme. Then, six local search operators are designed based on problem-specific knowledge. Two Sarsa-based strategies are developed to enhance the exploitation of meta-heuristics by optimizing factory allocation and choosing appropriate local search operators, respectively. Finally, to validate the effectiveness of the proposed algorithms, we solve 72 benchmark instances. The performance of the two Sarsa-based strategies is evaluated. Comparisons and discussions show that the improved ABC with the Sarsa strategy for local search selection is the best one among the 16 algorithms.
Efficient human-robot collaboration necessitates bidirectional perception between humans and robots. For robots, understanding the operator's behavior is crucial for enhancing safety and work efficiency in human-robot collaboration. Human limb movements encompass diverse behavioral information, and accurate human motion prediction has attracted significant attention. Existing human motion prediction methods can extract some behavioral features, but they fall short of fully capturing the dynamic and complex interactions across different time points and body joints in human motion sequences. In this study, a Spatio-Temporal Transformer Network model (STTFN) is proposed to automatically learn the spatio-temporal dependency relationships in human motion sequence data for prediction. The spatio-temporal knowledge embedding block employs an attention mechanism and a graph attention network to extract spatio-temporal behavioral features from raw data. An encoder-decoder network based on Transformer and Long Short-Term Memory (LSTM) is constructed for further analysis to acquire motion prediction data. This study conducts an experiment utilizing a human-robot collaborative assembly dataset to validate the effectiveness of the proposed model. The results demonstrate that our model outperforms classical models, thereby advancing the field of human motion prediction.
Water quality prediction is an important tool for environmental protection and public health. It can prevent water pollution, maintain ecological balance, and provide crucial scientific evidence for water resource management. To address the low prediction accuracy of traditional methods, this study uses data from the water quality monitoring station on the Huangpu River in the Taihu Basin of Shanghai to predict dissolved oxygen levels. We propose a water quality prediction method based on feature engineering and the BiGRU-Attention-TCN model. The proposed method employs feature engineering to extract monitoring station status information, improving model robustness and stability. It uses Bidirectional Gated Recurrent Unit (BiGRU) to process time series data, capturing bi-directional dependencies, and applies the Attention mechanism to enhance the model's focus on key time steps and features. Additionally, Temporal Convolutional Network (TCN) is introduced to further extract long-term dependencies and global features. Our method was compared with several commonly used baseline methods, and experimental results show that the proposed method outperforms other advanced baseline methods.
Recent deep learning-based video object tracking approaches have shown promising performance on standard tracking benchmarks. However, the scarcity of low-light scenes in common tracking datasets poses a significant challenge for developing robust low-light deep trackers, especially for the Unmanned Aerial Vehicle (UAV) tracking scenario with more difficulty for data collection, thus severely degrading the performance of existing deep trackers in low-light UAV conditions. In this paper, we propose a simple yet effective approach to enhance UAV tracking in low-light conditions. Different from previous works that need annotated pairwise images for lightness enhancement, our proposed approach is annotation free, which can be generally used to improve existing trackers in low-light scenario. Specifically, we first simulate low-light conditions by performing simple data augmentation on normal tracking videos. We then conduct the feature alignment between the normal and low-light video frames, in order to facilitate the trackers to adapt well in low-light conditions. Surprisingly, we find that our simple approach can effectively improve existing deep trackers, including both typical CNN-based SiamFC and transformer-based OSTrack. Experiments on UAVDark135 show that our proposed trackers can achieve favorable performance with simple design, which has the potential to be served as a simple yet effective baseline in low-light UAV tracking.
With the continuous advancements in science, technology, and production methods, the digitalization process of the chemical industry is rapidly progressing. This paper addresses the limitations of existing hierarchical system structures in chemical process virtual factory construction, which often result in lengthy development cycles and limited scalability, failing to accommodate real-time user needs. Additionally, conventional virtual factories primarily presented through computer screens lack immersion and interactivity. To overcome these challenges, this study proposes a de-layered and de-centralized virtual factory construction scheme. It integrates projected virtual reality technology to enhance personnel training and work experience. This framework empowers users to develop customizable functional modules, thereby catering to personalized training requirements.
Dynamic Visual Sensors (DVS) output pixel information asynchronously as an address event, only transmitting information for pixels where the intensity change exceeds a threshold, thereby reducing redundant data generation at the source. They possess advantages such as low-latency response and high efficiency with low power consumption. However, factors such as thermal noise and leakage current can lead to noise events even when the light intensity remains constant, which affects image quality and subsequent processing. In this paper, we propose a new method for calculating event density called Local Density Segmentation (LSD), which divides events and their surrounding neighborhoods into different regions for independent event density calculation. Based on LSD-calculated event density, we further propose a denoising method named Local Density Segmentation K-means (LSD-K), which utilizes event density as a feature to select and remove noise events. Finally, through experiments, we validate the performance of our algorithm.
Glaucoma is one of the leading causes of blindness worldwide. Patients with open-angle glaucoma have difficulty realizing the graduate vision loss. However, the optic nerve head's morphological change can manifest in the retinal image. Therefore, many studies offer computer-aided glaucoma diagnosis based on retinal image analysis. With the trend of deep learning, there have been several attempts to utilise deep-learning-based glaucoma classification in practice. However, the limited size of the dataset is a problem when adapting deep-learning models to glaucoma classification. Therefore, we test three vision transformer models to handle this problem and demonstrate their performance in glaucoma classification. Also, based on their performance, we give suggestions for their usage in clinical practice. In this paper, we fine-tuned three transformer models for glaucoma classification using two datasets, RIM-ONE_DL and Drishti-GS. Based on our experiments, the pyramid vision transformer (PVT) model achieves the best performance, with good scores in terms of accuracy and sensitivity. This suggests that the PVT can catch the missing detail and can be used to support the glaucoma diagnosis.
To improve the energy efficiency of the dryer section in the papermaking process, it is essential and significant to accurately predict the key parameters of the dryer section. Compared with conventional mechanism-based prediction models, data-driven methods effectively establish the parameter prediction models. This work selects the feature variables from the production data, and then adopts four kinds of data-driven prediction models for comparative analysis. All the prediction models are performed on two testing cases with high and low correlation degrees respectively. The testing results provide useful experience and insights for the operators and managers in the pulp and paper industries.
Although moving block mode “hard-wall” safety protection principle adopted by the existing urban rail train control system has shortened the following interval of trains to a certain extent, it fails to meet the operational demand during peak hours. In this paper, a train following safety protection method based on stochastic reachable set is proposed, which upgrades the safety protection criterion from position non-overlapping in spatial dimension to trajectory non-overlapping in space-time dimension. Firstly, the whole process predicted trajectory of the preceding train and the safety braking trajectory of the following train are represented. Secondly, the efficient calculation model of collision probability is established by combining the probability distribution functions of the preceding train and the following train. Then, the corresponding safety protection curve is generated according to the safety verification results of the urban rail train. Finally, the method is verified by using the data of Beijing Metro Line 11. The experimental results show that, compared with the short-time protection, the method proposed in this paper can extend the reverse calculation starting point of the safety protection curve for following train and shorten the following interval on the basis of collision avoidance, which proves the necessity and applicability of this method.
Deep neural networks typically impose significant computational loads and memory consumption. Moreover, the large parameters pose constraints on deploying the model on edge devices such as embedded systems. Tensor decomposition offers a clear advantage in compressing large-scale weight tensors. Nevertheless, direct utilization of low-rank decomposition typically leads to significant accuracy loss. This paper proposes a model compression method that integrates Variational Bayesian Matrix Factorization (VBMF) with orthogonal regularization. Initially, the model undergoes over-parameterization and training, with orthogonal regularization applied to enhance its likelihood of achieving the accuracy of the original model. Secondly, VBMF is employed to estimate the rank of the weight tensor at each layer. Our framework is sufficiently general to apply to other convolutional neural networks and easily adaptable to incorporate other tensor decomposition methods. Experimental results show that for both high and low compression ratios, our compression model exhibits advanced performance.
Neuromorphic electronic devices have already exhibited tremendous potential in mimicking the basic sensory and neuro-memory functions of living beings. However, while olfactory sensing electronics with ideal sensitivity and high selectivity achieved significant breakthroughs, synaptic devices that mimic the olfactory system have rarely been reported. Herein, a stretchable resistive olfactory neuromorphic device is proposed, which combines the common ductility of stretchable gold nanowires and porous polymer films to enable the olfactory synaptic function under different stretching states. The device achieves NO2-mediated synaptic plasticity and realize typical synaptic behaviors, including functions such as paired-pulse facilitation (PPF), short-term potential (STP) and long-term depression (LTD), and is able to mimic the cumulative damage that can be caused by exposure to NO2 even in the tensile strains up to 50%. The proposed stretchable electrons pave a new way for next-generation olfactory biomimetic systems.
Petri-net-based information flow analysis is an effective approach to check non-interference and analyze information security. However, Petri nets generally utilize the interleaving semantics, and easily suffer from the state space explosion problem. In this paper, we utilize unfolding techniques in checking non-interference. Specifically, we introduce some theorems to depict the structural behaviors of non-interference in the unfolding, and propose the finite interference-complete prefix (FICP) for checking non-interference. The theorems and FICP model benefit the checking of non-interference, which can alleviate the state space explosion problem.
