A thorough analysis and cross-comparison of recent relevant works was provided, outlining a closed-loop process for EEG data analysis based on deep learning. EEG data were introduced, and the application of deep learning in three key stages: preprocessing, feature extraction, and model generalization was unfolded. The research ideas and solutions provided by deep learning algorithms in the respective stages were delineated, including the challenges and issues encountered at each stage. The main contributions and limitations of different algorithms were comprehensively summarized. The challenges faced and future directions of deep learning technology in handling EEG data at each stage were discussed.

For the poor real-time detection capability of the current object detection model in the production environment of electronic components, GhostNet was used to replace the backbone network of YOLOv5. And for the existence of small objects and objects with large scale changes on the surface defects of electronic components, a coordinate attention module was added to the YOLOv5 backbone network, which enhanced the sensory field while avoiding the consumption of large computational resources. The coordinate information was embedded into the channel attention to improve the object localization of the model. The feature pyramid networks (FPN) structure in the YOLOv5 feature fusion module was replaced with a weighted bi-directional feature pyramid network structure, to enhance the fusion capability of multi-scale weighted features. Experimental results on the self-made defective electronic component dataset showed that the improved GCB-YOLOv5 model achieved an average accuracy of 93% and an average detection time of 33.2 ms, which improved the average accuracy by 15.0% and the average time by 7 ms compared with the original YOLOv5 model. And the improved model can meet the requirements of both accuracy and speed of electronic component surface defect detection.

The effects of leg structure design, foot-end design and sensor design on touchdown detection were comprehensively discussed by analyzing the existing legged robot touchdown detection methods. The touchdown method for direct detection of external sensors, the touchdown detection method based on kinematics and dynamics, and the touchdown detection method based on learning were summarized. Touchdown detection methods were summarized in three special scenarios: slippery ground, soft ground, and non-foot-end contact. The application scenarios of touchdown detection technology were analyzed, including the three application scenarios of motion control requirements, navigation applications, and terrain and geological sensing. The development trends were pointed out, which related to the four major touchdown detection methods of hardware improvement and integration, multi-mode touchdown detection, multi-sensor fusion touchdown detection, and intelligent touchdown detection. The specific relationships between various touchdown detection algorithms were summarized, which provided guidance for the development of follow-up technology for touchdown detection and specific applications of touchdown detection.

The improved YOLOv5 object detection algorithm was used to detect the facial region of the driver and a multi-feature fusion fatigue state detection method was established aiming at the problem that existing fatigue state detection method cannot be applied to drivers under the epidemic prevention and control. The image tag data including the situation of wearing a mask and the situation without wearing a mask were established according to the characteristics of bus driving. The detection accuracy of eyes, mouth and face regions was improved by increasing the feature sampling times of YOLOv5 model. The BiFPN network structure was used to retain multi-scale feature information, which makes the prediction network more sensitive to targets of different sizes and improves the detection ability of the overall model. A parameter compensation mechanism was proposed combined with face keypoint algorithm in order to improve the accuracy of blink and yawn frame number. A variety of fatigue parameters were fused and normalized to conduct fatigue classification. The results of the public dataset NTHU and the self-made dataset show that the proposed method can recognize the blink and yawn of drivers both with and without masks, and can accurately judge the fatigue state of drivers.

A multi-agent reinforcement learning algorithm based on priority experience replay and decomposed reward function was proposed in multi-agent pursuit and evasion games. Firstly, multi-agent twin delayed deep deterministic policygradient algorithm (MATD3) algorithm based on multi-agent deep deterministic policy gradient algorithm (MADDPG) and twin delayed deep deterministic policy gradient algorithm (TD3) was proposed. Secondly, the priority experience replay was proposed to determine the priority of experience and sample the experience with high reward, aiming at the problem that the reward function is almost sparse in the multi-agent pursuit and evasion problem. In addition, a decomposed reward function was designed to divide multi-agent rewards into individual rewards and joint rewards to maximize the global and local rewards. Finally, a simulation experiment was designed based on DEPER-MATD3. Comparison with other algorithms showed that DEPER-MATD3 algorithm solved the over-estimation problem, and the time consumption was improved compared with MATD3 algorithm. In the decomposed reward function environment, the global mean rewards of the pursuers were improved, and the pursuers had a greater probability of chasing the evader.

A new multimodal sentiment analysis model (MTSA) was proposed on the basis of cross-modal Transformer, aiming at the difficult retention of the modal feature heterogeneity for single-modal feature extraction and feature redundancy for cross-modal feature fusion. Long short-term memory (LSTM) and multi-task learning framework were used to extract single-modal contextual semantic information, the noise was removed and the modal feature heterogeneity was preserved by adding up auxiliary modal task losses. Multi-tasking gating mechanism was used to adjust cross-modal feature fusion. Text, audio and visual modal features were fused in a stacked cross-modal Transformer structure to improve fusion depth and avoid feature redundancy. MTSA was evaluated in the MOSEI and SIMS data sets, results show that compared with other advanced models, MTSA has better overall performance, the accuracy of binary classification reached 83.51% and 84.18% respectively.

There has been a trend of intelligent development using artificial intelligence technology in order to improve the efficiency of software development. It is important to understand program semantics to support intelligent development. A series of research work on program representation learning has emerged to solve the problem. Program representation learning can automatically learn useful features from programs and represent the features as low-dimensional dense vectors in order to efficiently extract program semantic and apply it to corresponding downstream tasks. A comprehensive review to categorize and analyze existing research work of program representation learning was provided. The mainstream models for program representation learning were introduced, including the frameworks based on graph structure and token sequence. Then the applications of program representation learning technology in defect detection, defect localization, code completion and other tasks were described. The common toolsets and benchmarks for program representation learning were summarized. The challenges for program representation learning in the future were analyzed.

Most of the compound fault diagnosis methods regard the compound fault as a new single fault type, ignoring the interaction of internal single faults, and the fault analysis is vague in granularity and poor in interpretation. An improved Transformer-based compound fault decoupling diagnosis method was proposed for industrial environments with very little compound fault data. The diagnosis process included pre-processing, feature extraction and fault decoupling. With introducing the decoder of the Transformer, the cross-attention mechanism enables each single fault label to adaptively in the extracted feature layer focus on the discriminative feature region corresponding to the fault feature and predicts the output probability to achieve compound fault decoupling. Compound fault tests were designed to verify the effectiveness of the method compared with the advanced algorithms. The results showed that the proposed method had high diagnostic accuracy with a small number of single fault training samples and a very small number of compound fault training samples. The compound fault diagnosis accuracy reached 88.29% when the training set contained only 5 compound fault samples. Thus the new method has a significant advantage over other methods.

A multi-behavior aware service recommendation method based on hypergraph graph convolutional neural network (MBSRHGNN) was proposed to resolve the problem of insufficient high-order service feature extraction in existing service recommendation methods. A multi-hypergraph was constructed according to user-service interaction types and service mashups. A dual-channel hypergraph convolutional network was designed based on the spectral decomposition theory with functional and structural properties of multi-hypergraph. Chebyshev polynomial was used to approximate hypergraph convolution kernel to reduce computational complexity. Self-attention mechanism and multi-behavior recommendation methods were combined to measure the importance difference between multi-behavior interactions during the hypergraph convolution process. A hypergraph pooling method named HG-DiffPool was proposed to reduce the feature dimensionality. The probability distribution for recommending different services was learned by integrating service embedding vector and hypergraph signals. Real service data was obtained by the crawler and used to construct datasets with different sparsity for experiments. Experimental results showed that the MBSRHGNN method could adapt to recommendation scenario with highly sparse data, and was superior to the existing baseline methods in accuracy and relevance.

Physical information was divided into rule information and numerical information, in order to explore the role of physical information in training neural network when solving differential equations with physics-informed neural network (PINN). The logic of PINN for solving differential equations was explained, as well as the data-driven approach of physical information and neural network interpretability. Synthetic loss function of neural network was designed based on the two types of information, and the training balance degree was established from the aspects of training sampling and training intensity. The experiment of solving the Burgers-Fisher equation by PINN showed that PINN can obtain good solution accuracy and stability. In the training of neural networks for solving the equation, numerical information of the Burgers-Fisher equation can better promote neural network to approximate the equation solution than rule information. The training effect of neural network was improved with the increase of training sampling, training epoch, and the balance between the two types of information. In addition, the solving accuracy of the equation was improved with the increasing of the scale of neural network, but the training time of each epoch was also increased. In a fixed training time, it is not true that the larger scale of the neural network, the better the effect.

Traditional and new geotechnical engineering problems such as compressed air energy storage, intercepting water with compressed air, carbon dioxide sequestration and oil and gas underground reserve project are all involving air-water two-phase flow and stress coupling problems. For this engineering reality, based on the weak coupling theory of gas-water two-phase seepage and stress in unsaturated soil, a air-water two-phase percolation-stress coupling calculation program based on coupled TOUGH2 and FLAC3D was developed. The calculation program can simulate real air-water two phase flow, and can investigate the gas-water interaction of seepage process. The calculation program considers the direct interaction between gas-water two-phase seepage and soil skeleton deformation, reflects the process of porosity, permeability, capillary pressure and the change of soil physical and mechanical parameters, and achieve a more perfect gas-water two-phase seepage-stress coupling analysis. Furthermore, by comparing with classical drainage test and model test, it is verified that the program can accurately simulate the gas-water two-phase flow-stress interaction.

Few existing studies cover the state-of-the-art multi-objective particle swarm optimization (MOPSO) algorithms. To fill the gap in this area, the research background of multi-objective optimization problems (MOPs) was introduced, and the fundamental theories of MOPSO were described. The MOPSO algorithms were divided into three categories according to their features: Pareto-dominated-based MOPSO, decomposition-based MOPSO, and indicator-based MOPSO, and a detailed description of their existing classical algorithms was also developed. Next, relevant evaluation indicators were described, and seven representative algorithms were selected for performance analysis. The experimental results demonstrated the strengths and weaknesses of each of the traditional MOPSO and three categories of improved MOPSO algorithms. Among them, the indicator-based MOPSO performed better in terms of convergence and diversity. Then, the applications of MOPSO algorithms in production scheduling, image processing, and power systems were briefly introduced. Finally, the limitations and future research directions of the MOPSO algorithm for solving complex optimization problems were discussed.

The problem of structured query language (SQL) generation from natural language queries (Text-to-SQL) in financial domain was investigated. First, SOFT, a Text-to-SQL dataset in the financial domain was constructed. The dataset covered common queries in the financial domain with distinctive features and presented challenges to Text-to-SQL research. Then, FinSQL, a Text-to-SQL model, which optimized the support for complex queries in the financial domain, was proposed. In particular, by analyzing the characteristics of row calculation queries, a class of queries with complex numerical calculations, a divide-and-conquer based method was proposed. A row calculation query was divided into several subqueries, the SQL statement for each subquery was generated, and the SQL statements were finally combined into together to get the SQL statement for the original query. Experimental results on SOFT dataset show that the proposed FinSQL model outperforms existing methods for the hard queries, and performs well for row calculation queries.

An improved method for the blockchain Kademlia network based on small world theory was proposed aiming at the issue of sacrificing security to improve scalability in the current research of the blockchain Kademlia network. The idea of the small world theory was followed, and a probability formula for replacing expansion nodes was proposed. The probability was inversely proportional to the distance between nodes. The number of node replacements and additional nodes could be flexibly adjusted according to actual conditions. The theoretical analysis and experimental verification demonstrate that the network transformed by this method can reach a stable state. The experimental results showed that the transmission hierarchy required for broadcasting transaction messages throughout the network was reduced by 15.0% to 30.8% and the rate of locating nodes was increased. The level of network structure was reduced and network security was enhanced compared to other optimization algorithms that modify the network structure.

Ensuring that the flue gas in the furnace stays within the temperature range of no less than 850 ℃ for at least 2 s contributes to the steady municipal solid waste (MSW) incineration, and the reduction of secondary pollution. However, at present, it is difficult to quantitatively calculate and predict the residence time of flue gas in the high temperature area by only using the thermocouple for qualitative evaluation. Based on the thermodynamic calculation, correlation analysis of practical operation parameters, and a variety of machine learning algorithms (backpropagation neural network, recurrent neural network, and random forest regression), the residence time of flue gas in high-temperature areas (>850 ℃) was calculated, correlation analysis of key operation parameters was conducted, and the prediction model of residence time was constructed, aiming at a typical MSW circulating fluidized bed boiler in China. Results revealed that 10 key operating parameters, e.g. section temperature of the furnace, temperature and pressure of primary air and secondary air, etc., had a strong correlation and predictability with the high-temperature flue gas residence time. Moreover, the model of the recurrent neural network was relatively optimal, with a higher fitting degree and accuracy. Specifically, the mean square error (MSE) was 0.11626, and the average absolute error between the predicted value and real value was 1.174%. Research enabled the prediction of flue gas temperature variation in high-temperature areas, helped optimize the MSW incineration, and contributed to the advanced control of pollutant emission reduction.

Most of existing structural image super-resolution reconstruction algorithms can only solve a specific single type of structural image super-resolution problem. A structural image super-resolution network based on improved Transformer (TransSRNet) was proposed. The network used the self-attention mechanism of Transformer mine a wide range of global information in spatial sequences. A spatial attention unit was built by using the hourglass block structure. The mapping relationship between the low-resolution space and the high-resolution space in the local area was concerned. The structured information in the image mapping process was extracted. The channel attention module was used to fuse the features of the self-attention module and the spatial attention module. The TransSRNet was evaluated on highly-structured CelebA, Helen, TCGA-ESCA and TCGA-COAD datasets. Results of evaluation showed that the TransSRNet model had a better overall performance compared with the super-resolution algorithms. With a upscale factor of 8, the PSNR of the face dataset and the medical image dataset could reach 28.726 and 26.392 dB respectively, and the SSIM could reach 0.844 and 0.881 respectively.

A news recommendation method based on Transformer and knowledge graph was proposed to increase the auxiliary information and improve the prediction accuracy. The self-attention mechanism was used to obtain the connection between news words and news entities in order to combine news semantic information and entity information. The additive attention mechanism was employed to capture the influence of words and entities on news representation. Transformer was introduced to pick up the correlation information between clicked news of user and capture the change of user interest over time by considering the time-series characteristics of user preference for news. High-order structural information in knowledge graphs was used to fuse adjacent entities of the candidate news and enhance the integrity of the information contained in the candidate news embedding vector. The comparison experiments with five typical recommendation methods on two versions of the MIND news dataset show that the introduction of attention mechanism, Transformer and knowledge graph can improve the performance of the algorithm on news recommendation.

In order to portray the public emotion orientation caused by the public opinion on traffic incidents disseminated in short videos, a physiological feature graph was constructed by the text sentiment analysis and the multimodal physiological signal feature extraction. This work collected 136 highly-liked videos with 38 805 comments on TikTok. Considering all videos as a document set, with each video treated as a document and comments as words, the latent Dirichlet allocation topic model was adopted to obtain the distribution of comments under different topics and the distribution of topics under different videos. Naive Bayes-based SnowNLP was utilized to calculate the sentiment scores of comments and analyze the sentiment tendencies expressed by different opinion topics. Neuroscience experiments were carried out to collect multimodal physiological signals such as EEG, eye movement, ECG, and respiration as well as emotion ratings. Statistical test results show that videos with different sentiment tendencies induce different emotions, and the multimodal physiological features such as the relative spectral power of EEG, blinking frequency, respiration standard deviation, and the very low-frequency power of ECG are specific under different emotions. The emotional semantics embedded in the comments influence public emotion in various ways beyond that evoked by videos.

A spatial-temporal multi-graph convolution traffic flow prediction model by integrating static and dynamic knowledge graphs was proposed, as current traffic flow prediction methods focus on the spatial-temporal correlation of traffic information and fail to fully take into account the influence of external factors on traffic. An urban traffic knowledge graph and four road network topological graphs with distinct semantics were systematically constructed, drawing upon the road traffic information and the external factors. The urban traffic knowledge graph was inputted into the relational evolution graph convolutional neural network to realize the knowledge embedding. The traffic flow matrix and the knowledge embedding were integrated using the knowledge fusion module. The four road network topology graphs and the traffic flow matrix with fused knowledge were fed into the spatial-temporal multi-graph convolution module to extract spatiotemporal features, and the traffic flow prediction value was outputted through the fully connected layer. The model performance was evaluated on a Hangzhou traffic data set. Compared with the advanced baseline, the performance of the proposed model improved by 5.76%-10.71%. Robustness experiment results show that the proposed model has a strong ability to resist interference.

A method for creating binocular dataset and a 6D pose estimation network called Binocular-RNN were proposed, in response to the problem of low accuracy in the current task of 6D pose estimation for objects. The existing images in the YCB-Video Dataset were used as the content captured by the left camera of the binocular system. The corresponding 3D object models in the YCB-Video Dataset were imported using Open GL, and the parameters related to each object were input to generate synthetic images captured by the virtual right camera of the binocular system. A monocular prediction network was utilized in the Binocular-RNN to extract geometric features from the left and right images in the binocular dataset, and recurrent neural network was used to fuse these geometric features and predict the 6D pose of the objects. The evaluation of Binocular-RNN and other pose estimation methods was based on the average distance of model points (ADD), average nearest point distance (ADDS), translation error and angle error. The results show that when the network was trained on a single object, the ADD or ADDS score of Binocular-RNN was 2.66 times that of PoseCNN and 1.15 times that of GDR-Net. Furthermore, the Binocular-RNN trained by the physics-based real-time rendering (Real+PBR) outperformed the DeepIM method based on deep neural network iterative 6D pose matching.