Traffic flow prediction is an important problem in the field of intelligent transportation systems. Most existing traffic-flow prediction methods have made good progress, which however still face the following two key challenges. (1) The underlying pattern of traffic flow depends on not only the historical information along the timeline, but also the information of spatially adjacent areas, making it a challenging problem on how to balance the two temporal-spatial patterns; (2) Due to fact that time information has the characteristic of multiple granularity, such as hour, day and week, how to capture the multi-grained temporal patterns is another challenge problem. In this paper, we design a multi-grained spatio-temporal deep regression model (MGSTDR) to address the above challenges. By extending the typical autoregressive integrated moving average model (ARIMA) on the basis of multi-grained spatio-temporal traffic flow information, the proposed model can effectively use historical information along the timeline as well as the information of adjacent regions, such that the prediction of multi-grained spatio-temporal traffic flow can be performed. Experimental results on several datasets demonstrate that the proposed model outperforms existing benchmark methods on the task of multi-granularity, and particularly obtains an approximately 5.66% improvement in the hourly traffic flow prediction.

As a pivotal part of the task-oriented dialogue system, dialogue policy can be trained by using the discriminative deep Dyna-Q framework. However, the framework uses vanilla deep Q-network method in the direct reinforcement learning phase and adopts MLPs as the basic network of world model, which limits the efficiency and stability of the dialogue policy learning. In this paper, we purpose an improved discriminative deep Dyna-Q method for task-oriented dialogue policy learning. In the improved direct RL phase, we first employ a NoisyNet to improve the exploration method, and then combine the dual-stream architecture of Dueling Network, Double-Q Network and n-step bootstrapping to optimize the calculation of the Q values. Moreover, we design a soft-attention-based model to replace the MLPs in the world model. The experimental results show that our proposed method achieves better results than other baseline models in terms of task success rate, average dialog turns and average reward. We further validate the effectiveness of proposed method by conducting both ablation and robustness analysis.

In this paper, we propose a channel-split-based dual-branch block (CDBlock) based on the PointMLP method, which splits input features into two groups along the channel dimension and feeds them into different branches of the dual-branch network module. Specifically, our proposed CDBlock consists of a lightweight branch and a heavyweight branch. The lightweight branch, which uses the structure of the residual MLP, is designed to extract coarse semantic features. The heavyweight branch, which uses the bottleneck network as backbone, is responsible for extracting more deeper and distinguishable information. By doing this, our proposed CDBlock effectively improves the network's feature extraction and learning capabilities. Experimental results show that our proposed method outperforms the existing PointMLP by achieving an overall accuracy of 86.2% and a class average accuracy of 84.97% on ScanObjectNN dataset. In particular, our approach achieves better results with using fewer parameters and computational cost than the PointMLP. Additionally, our approach also achieves encouraging performance on the ShapeNetPart dataset.

Due to the fact that features are not category-deterministic and the equipment resources are usually limit to support the category structure learning of samples, existing knowledge distillation methods possibly ignore the category knowledge distillation of samples. Therefore, this paper proposes a distillation method based on incremental class activation knowledge (ICAKD). First, this paper uses the class activation gradient map to extract class-discriminative sample features and proposes a class-activation constraint loss. Then, an incremental memory bank is built to store class-deterministic features, and multiple training batch samples are saved and updated iteratively. Finally, our proposed method calculates the quasi-quality center of the samples in the memory bank to construct the category structure relationship, and further performs the category knowledge distillation according to the class-activation constraint and the category structure relationship. Experimental results on the Cifar10, Cifar100, Tiny-ImageNet, and ImageNet datasets show that the proposed method achieves a 0.4%～1.21% improvement in term of accuracy when compared with the Category Structure Knowledge Distillation(CSKD) methods, demonstrating the promising effectiveness of the characteristics and increment of category judgment for category knowledge distillation.

Machine learning tasks are usually single-task learning. However, in practical applications, the learning tasks are often related. As a result, the correlation between tasks is often ignored, and the complexity of samples is sometimes not considered in single-task learning. To address this, this paper proposes a new inductive transfer learning model based on the self-paced learning strategyby constructing a prediction model to learn the shared parameters of multiple related source tasks. First, the proposed model uses the strategy of self-paced learning to jointly learns the multiple related source tasks, and weights the learning samples according to the loss and difficulty degree of the samples in the source tasks. The parameters of the self-paced learning are iteratively updated to select the optimal samples with the less loss. Then, the knowledge learned from multiple related source tasks guide the learning of target tasks to construct multiple models for related transfer learning target tasks, and transfer these models to related target tasks to improve their generalization ability. Finally, we optimize the target model by using the Lagrange function to improve the performance of the classifiers. Experimental results show that the proposed model is superior to the existing transfer learning model under the same experimental conditions.

Users' long-term preferences and the relationship between historical information and current situation are usually ignored in existing user preference modeling and user record unified modeling. To address this, in this paper, we propose a music recommendation model based on users' long-term and short-term preferences and music emotional attention. Firstly, we divide the user's listening record into multiple historical and current sequences, and learn the features by multiple long and short-term memory networks, respectively, to obtain user's long and short-term preferences. For the historical music sequences, we propose the concept of sequence period is proposed to obtain the long-term preferences by calculating the weights of the sequence period. For the current sequence, we use the average pooling to extract the music features of the current scene to obtain the short-term preference. Secondly, we learn the emotional characteristics of music from the acoustic signals, and use the attention mechanism to calculate the emotional factors of music. Finally, we integrate the music emotion factor into the user's long-term and short-term preference to generate a music recommendation list. The experimental results on the last.fm real data set show that the proposed model achieves 0.5435 in term of the NDCG@10, which is better than the existing methods. The ablation experiment and characteristic contribution analysis further demonstrate the effectiveness of the model.

Graph convolutional networks (GCN), with natural advantages for skeleton-based action recognition, has attracted more and more attention. The key lies in how to obtain richer feature information and the design of the skeleton topology. In this research, the feature fusion method of joint and semantics (joint type and frame index) is improved, and integrated into a Semantics Coding Module (SCM), which is more applicable for complex multi-layer networks. Guided by the SCM, the network can obtain more feature information of skeleton. Secondly, a skeleton Topology Inference Network (TIN) is proposed, which adaptively learns different adjacency matrices according to the context information of different samples with the efficient feature learning ability of CNN, so that the network can get rid of the limitation of fixed topology. By applying the SCM and TIN to 2s-AGCN, we propose a semantics-guided multi-stream adaptive topology inference graph convolutional network for skeleton-based action recognition. Extensive experiments on datasets, NTU RGB+D and NTU RGB+D 120, demonstrate that our methods obviously improve the accuracy of network and our model has achieved the state-of-the-art performance.

Falling behavior will bring serious injury to the elderly, especially to the elderly living alone. How to correctly identify the falling behavior and issue a warning is an important factor for reducing the injury. A fall detection algorithm is proposed, which is based on TSSI (tree structure skeleton image) and learnable STB-CNN (spatio-temporal block convolutional neural network). Firstly, human joint point is extracted by the three-dimensional pose estimation algorithm, and the corresponding skeleton sequence can be obtained. Secondly, the skeleton sequence is encoded into TSSI by the algorithm based DFS (depth first search) method. Finally, a learnable STB-CNN is proposed to classify TSSI and detect the fall behavior, which consists of spatio-temporal difference module, learnable spatio-temporal framework and spatio-temporal multi-branch convolution module,. Experiments are carried out on the public datasets UR FALL Detection Datasets and the simulation datasets. Experimental results are shown that our fall detection algorithm is more accurate than other related algorithms, especially the accuracy to 98.6% and 98.3% respectively.

In order to avoid the problems of life shortening or safety accidents caused by the surface defects of lithium batteries, it is necessary to study an efficient and accurate methods for lithium battery electrode plate defect detection. In this research, the simple Unet semantic segmentation network is used to detect defects of lithium battery. In order to improve the segmentation accuracy, first the coding structure in the original network is replaced with VGG16, which is similar to the Unet coding structure, to obtain the pre-training weights had been trained. Then, feature fusion module of the simply fusion pyramid network (SFPN) is added to the skip connection of the Unet network to avoid large information differences between feature maps. Finally, label smoothing is applied to optimize the loss function to prevent the network from overfitting. Through experimental verification, the accuracy of the semantic segmentation network optimized by the proposed method is improved to 93.70%, and the probability of false segmentation and segmentation discontinuity is significantly reduced. This optimization process has certain practical value.

To address the problems of missing detection and low detection accuracy of the existing helmet wearing detection algorithms for small and crowded targets detection, this paper proposes a helmet wearing detection method based on improved YOLOv5 and transfer learning. First, different from the default priori frame that is not suitable for the task, we use the K-means algorithm to cluster the suitable priori frame size for the detection task. Then, in the back of the feature extraction network, we introduce a spatial channel mixed attention module to strengthen the learning of relevant weights and suppress the weights of irrelevant backgrounds, respectively. Further, we improve the judgment metric of the non-maximum-suppression (NMS) algorithm in the post-processing stage of YOLOv5 to reduce the phenomenon of false deletion and missing of prediction boxes. After that, the proposed network is trained based on the strategy of transfer learning, which can overcome the scarcity of limited existing data sets and improve the generalization ability of the model. Finally, we build a cascade judgment framework for helmet wearing deployed in visual sensor networks. The experimental results show that our proposed method improves the average accuracy (IOU=0.5) to 93.6%, which is 5% higher than the original model in the helmet wearing data set. The proposed model also outperforms other state-of-the-art algorithms by obviously improving the accuracy of helmet wearing detection in the construction scenarios.

With the advantages of high flexibility and lightweight, unmanned aerial vehicles (UAVs) have been widely used in data collection of wireless sensor networks. For a multi-UAV-assisted wireless sensor network with randomly distributed and moved users, how to plan the flight paths of the UAVs to effectively collect data from the users remains a challenging problem. This paper aims to maximize the average throughput of data collection in a dynamic environment where the user's location cannot be predicted by optimizing the flight path of the UAVs, which is subject to the shortest flight time and range constraints of UAVs, the constraints of UAV start and end points, the communication distance constraints, the user communication constraints, and the UAV collision avoidance constraints. The resultant problem can be solved by using existing optimization methods with high complexity, which however is difficult to obtain the globally optimal solution. To address this problem efficiently, this paper proposes a deep reinforcement learning algorithm based on Dueling Double DQN (Dueling-DDQN). The proposed algorithm adopts the Dueling network architecture, which enhances the learning ability of the algorithm and improves the robustness and convergence speed of tracked in suboptimal solutions due to the over-estimation on the $ Q $ value. Simulation results show that the proposed algorithm can efficiently obtain the flight paths of multiple UAVs under all constraints. In particular, our proposed algorithm has encouraging convergence and stability performance in comparison with the existing benchmark algorithms.

With the rapid development of mobile communication technology, the contradiction between the limited spectrum utilization resources and the demand of a lot of spectrum communication is increasingly aggravated. New intelligent methods are needed to improve the utilization rate of spectrum. A multi-user dynamic spectrum access method based on distributed priority experience pool and double deep Q network is proposed. This method can help the secondary users to continuously learn according to their perceived environment information in the dynamic environment, and choose the idle channel to complete the spectrum access task for improving the spectrum utilization rate. In this method, a distributed reinforcement learning framework is adopted, and each secondary user is regarded as an agent. Each agent learns by using standard single-agent reinforcement learning method to reduce the underlying computing overhead. In addition, the method adds priority sampling on the basis of neural network training, and then optimizes the training efficiency of neural network to help sub-users choose the optimal strategy. The simulation results show that this method can improve the success rate, reduce the collision rate and improve the communication rate.

For a class of complex dynamic networks composed of nonidentical nodes, a decentralized control strategy is proposed to achieve the finite-time partial state components synchronization. The finite-time partial state components synchronization means that only some, not all, state components of each node in the network can achieve synchronization in finite time. Firstly, for the convenience of the theoretical analyses and derivation, a special diagonal matrix is introduced, which can formulate the desired state components of each node. Secondly, based on this special diagonal matrix, the finite-time partial state components synchronization is defined. Compared to the finite-time synchronization, the finite-time partial state components synchronization is more popular. Thirdly, according to both the control theory and the finite-time stability theorem, a decentralized control strategy is proposed, so that our networks can achieve the finite-time partial state components synchronization. Finally, a simulation example is shown to verify the effectiveness and correctness of the proposed control strategy in this paper.

Aiming at the problem that the existing fault diagnosis method has too high requirements for systems, the weak diagnosability of fuzzy discrete event system (FDES) is studied, and a weak fuzzy diagnosability method is proposed, which extends the weak diagnosability method of classical discrete event system (DES) to the fuzzy system. Firstly, the notion of weak fuzzy diagnosability of FDES is formalized; In order to verify the weak fuzzy diagnosability of fuzzy systems, a verifier automaton is constructed, and the necessary and sufficient conditions for the weak fuzzy diagnosability of FDES are obtained, in which the weak fuzzy fault diagnosis of fuzzy systems is realized. This method is suitable for weak fault diagnosis of both FDES and classical DES.

This paper proposes a new distance metric (NewDM) of categorical data based on the graph structure of ordinal and nominal attributes to address the poor effect of most existing measurement methods for categorical data. In NewDM, it first summarizes the basic framework formula of distance definition of categorical data and analyzes the challenges of measuring categorical data. Then, the graph structures of ordinal attributes and nominal ones are utilized to define the distance between two probability distribution columns. Finally, a new distance metric of categorical data is obtained through simultaneous weight. Experimental results on six public datasets show that the proposed NewDM is superior to the state-of-the-art approaches.

In order to realize the wide application and improve the performance of third-generation semiconductor devices, the development of new and high-performance packaging interconnect materials has become a key initiation. Among them, micro/nano copper material can be sintered into block structures with high electrical conductivity, high thermal conductivity, high stability and electromigration resistance under low temperature conditions due to its surface effect and small size effect, which has become a research hotspot for the development of new packaging interconnect materials. However, the problems of easy oxidation, agglomeration and low yield of micro/nano copper material limit its application in third-generation semiconductor devices, and the oxidation resistance improvement of micro/nano copper material has become a key problem to solve its application. In this research, benzimidazole was used to treat micro/nano copper material, and the coating of micro/nano copper particles by benzimidazole was confirmed by Scanning Electron Microscope (SEM), infrared spectroscopy and other characterization methods. The copper paste was subjected to X-ray diffraction (XRD) and X-ray photoelectron spectroscopy (XPS) to prove that the paste could be left in air for 120 days without oxidation. The strength of the interconnect joint prepared at 300 °C reached 62.3 MPa, and the resistivity of the sintered layer was as low as 6.18×10⁻⁸ Ω·m. The results show that the method of treating micro/nano copper material by benzimidazole can help it to achieve super oxidation resistance and good interconnection performance, which is of profound significance for the research and development of third-generation semiconductor packaging interconnect materials.

In order to ensure the safe production of glyphosate production enterprises and reduce or avoid the occurrence of fire and explosion accidents in the production process, the improved Dow Chemical Fire and explosion risk index evaluation method (Dow method) is proposed to carry out safety evaluation. In view of the inaccurate calculation of safety measure coefficient in the current general Dow method, the personnel safety compensation coefficient is added, and the safety compensation is divided into event prevention type and consequence mitigation type, and then the hazardous process units are quantitatively analyzed and calculated, so as to improve the accuracy of the safety evaluation results of Dow method. The improved Dow method is applied to the safety evaluation of a glyphosate production enterprise, and it is found that compared with the traditional Dow method, the improved fire and explosion risk index (F&EI) and exposure area are more sensitive and can more accurately reflect the fire and explosion safety evaluation results of glyphosate enterprises.