Existing methods for audio-visual cross-modal association learning often adopt a dual-stream network structure, but they still face challenges in reducing computational complexity, model light weighting, and efficient feature fusion. To improve model performance and enhance the efficiency of cross-modal learning, a single-stream network-based approach for audio-visual cross-modal learning was proposed. Firstly, preprocessed data from both modalities were fed into a single-stream feature extraction network, where a class-information-based loss function was employed to learn and extract feature vectors from both modalities. Subsequently, attention-based feature fusion was performed on the extracted feature vectors from both modalities. Finally, a combination of cosine similarity algorithm and cross-entropy loss was used to learn the association between the two modalities, thus completing the cross-modal association learning task. Experimental results demonstrate that the proposed method achieves promising performance in audio-visual cross-modal verification, matching, and retrieval tasks, ensuring excellent performance while considering the lightness and flexibility of the network structure.

An open world object detection method based on shape perception and class balance optimization was proposed to address the issue of poor prediction performance of unknown class objects in open world object detection. Unknown classes referred to classes that were not labeled during the training phase. Due to the lack of guidance from labels, detecting unknown class objects was a challenging task. An unknown class enhanced detector has been constructed as an unknown class detection branch. During training, this detector was supervised using only known class labels, allowing it to learn the similarities in features of known class objects and generalize to unknown class objects. To improve the detector's sensitivity to unknown classes, the region proposal network (RPN) module's ability to distinguish between foreground and background was utilized. A specific filtering method was employed to select results with “unknown class potential” from the RPN output, which were then used as pseudo labels in the training process. Due to the absence of confidence scores, traditional non-maximum suppression (NMS) methods were difficult to apply for post-processing unknown objects. Therefore, a redundant unknown object suppression mechanism was designed, consisting of a center point-based grouping strategy and a redundancy score matrix based on shape perception. The center point-based grouping strategy included three methods based on the unknown class center points to determine the suppression range. Subsequently, a redundancy score matrix was constructed based on the redundancy scores of each prediction box within the group to suppress highly redundant predictions. Experimental results on open world object detection datasets demonstrated that the open world object detection based on shape perception and class balance optimization maintained high recall rates for unknown classes while achieving high prediction accuracy. This method effectively addressed the challenges of open world scenarios and avoided generating a large number of useless predictions.

Addressing the limitation of existing sequence labeling approaches in effectively recognizing nested entities within Chinese electronic health records (EHRs), a novel named entity recognition model that integrates MacBERT and a global pointer network was proposed. Initially, the MacBERT-large pre-trained model transformed the text into context-sensitive dynamic vectors. Subsequently, the fast gradient method (FGM) was employed to generate adversarial samples, which were incorporated into the original vectors and fed into a BiLSTM (bi-directional long short-term memory) network to capture contextual features. To enhance the capture of long-distance semantic features, an attention mechanism was introduced. Finally, a global pointer network model was leveraged to decode simultaneously considering both head and tail feature information, thereby achieving superior prediction performance for medical nested entities. Experimental results demonstrate that compared to the state-of-the-art global pointer model, the proposed model achieves an improvement of 1.8%, 1.37%, and 1.72% in F₁-score on the CCKS2019 dataset and two versions of the CMeEE Chinese EHR dataset, respectively, validating the effectiveness of the proposed approach.

Defect detection is regarded as an indispensable step in the industrial production process. At present, manual detection is faced with the problems of low efficiency and high cost. A ceramic small target defect detection algorithm based on deep learning was proposed. For small target defects, a slice pre-training layer was first added to reduce the loss of graphics memory resources by large-size images. Secondly, a small target detection layer was added for the detection of small target defects, and a large target detection layer was removed to reduce the number of parameters. In addition, a feature selection fusion module based on MLCA (mixed local channel attention) was proposed to improve the perception of small target defects. Finally, a detection head with shared parameters was designed to further reduce the number of learnable parameters of the algorithm. By comparing with the baseline model, taking the ceramic cup as an example, the detection accuracy of this algorithm has been improved by 20.9%. Combined with the developed detection software and experimental platform, the detection efficiency of the ceramic cup has been enhanced by about 46.9%.

In order to address the issues of incomplete collection, vulnerability to attacks, and limitations in specific recognition scenarios in single modal biometric information, a multi-level fusion recognition model for faces and iris was proposed, a multi-modal biometric recognition system was designed and implemented to integrate the proposed model in a modular manner. The lightweight convolutional neural networks was used as feature extractors, intra class correlations between different modal features was utilized on the feature level, normalizing and concatenating the features of different modalities. The minimum strategy was used to fuse left and right iris scores on the score layer, the average strategy was used to fuse iris scores and face scores. Homologous multi-modal datasets was extracted from the CASIA-IrisV4-Distance dataset for experiment verification, feature layer fusion algorithm and score layer fusion algorithm both achieves an accuracy of 99.8%. It is observed in the experiment that this system has robustness and generalization.

The tension at the suspension point of transmission towers is an important parameter for line renovation and collapse accident warning. Although airborne LiDAR can provide high-precision spatial information for transmission line scenes-LiDAR point cloud data, lacks timely feedback on the operational status information of elements (such as tension of suspension points). A tension extraction method for tower suspension points based on LiDAR point clouds was proposed, and a new approach for studying micro physical parameters of transmission line spatial information was designed. Firstly, a three-dimensional spatial model of the transmission line was reconstructed from point cloud. Then, a mapping model between the spatial curve model and the horizontal stress of the transmission line was constructed. Finally, based on the tension vector relationship of adjacent transmission lines, the tension at the suspension point of the tower was extracted. This method reconstruct the three-dimensional model of transmission lines and extract the tension of tower suspension points in real-time operation based on LiDAR point cloud. Compared with finite element analysis method, the maximum relative error is 3.26%. The research can be extended and applied to the safety inspection and state parameter detection of large-scale transmission lines.

Seasonal segmentation of building electricity consumption time series (BECTS) is of great significance for accurate load forecasting and pattern mining. Aiming at the problem that accurate BECTS seasonal segmentation is difficult to be realized by traditional timing segmentation, fixed-temperature segmentation and adaptive five-days temperature segmentation methods, a new adaptive seasonal segmentation method for BECTS based on Toeplitz inversed covariance-based clustering (TICC) was proposed. The method was based on the binary time series of building hourly electricity load and outdoor dry bulb temperature, and the TICC algorithm was used for real-time segmentation and clustering. A large public building electricity load case in a hot summer and warm winter area was analyzed, and the result showed that the similarity between samples of the same type and the difference between samples of different types were enhanced by the method. Compared with the timing segmentation, fixed-temperature segmentation and adaptive five-days temperature segmentation methods, the average dynamic time warping (DTW) distance of each category after TICC segmentation was improved respectively by 46.54%, 35.73% and 7.59%. This method can be used as data preprocessing to provide data support for single building data mining analysis, such as building electricity consumption pattern mining and load forecasting.

To investigate the flexural bearing capacity of timber-concrete composite beams, Yunnan pine was selected as the base material, and high-strength self-tapping screws were used as shear connectors. Self-tapping screws were drilled into the timber to connect with cast-in-place concrete slabs, with partial slotting of the connection surface on the timber beam as a variable, to study the flexural performance of timber-concrete composite beams. Two groups of four specimens were designed in total, and a stepwise loading method was adopted to conduct four-point bending tests to analyze the mechanical properties of the composite beams. The results indicated that the overall performance of the composite beams was good, with the partially slotted design exhibiting better flexural stiffness and overall performance compared to the unslotted beams. Under the same load conditions, the deflection of the partially slotted beams was reduced by 57% compared to the unslotted beams, and the interface slip was reduced by 50%. Theoretical analysis results were in good agreement with the experimental findings, showing that the effective flexural stiffness and composite effect coefficient of the partially slotted beams were higher than those of the unslotted beams. The partially slotted design of the composite beams demonstrated superior overall performance, as well as improved stiffness and strength.

The stress-strain characteristics of gas hydrate-bearing sediments (GHBS) are essential for achieving safe and secure extraction of marine natural gas hydrate. To quantitatively investigate the influence of hydrate formation on the mechanical properties of GHBS, laboratory tests were conducted under different confining pressures and various hydrate saturations, and it was found that hydrate primarily enhanced the cohesive strength of GHBS, while the internal friction angle remained relatively unchanged. To comprehensively describe the intricate mechanical properties of GHBS, the previously established elastic-plastic constitutive model was employed for predictive purposes. A comparative analysis revealed that the constitutive model exhibited a certain level of applicability. Subsequently, a computer program was developed for the constitutive model by leveraging the UMAT user material subroutine interface provided by ABAQUS. The program was written in FORTRAN language and integrated into ABAQUS, facilitating the implementation of the constitutive model. To validate the effectiveness and stability of the developed model, finite element simulations were conducted on single elements as well as triaxial specimens of GHBS. The results of these simulations confirm the efficacy and reliability of the developed model.

Because Hainan Province is located in the tropics, it is often hit by extreme weather such as typhoons and rains, so it is very prone to geological disasters such as slope collapse and landslides, which eventually cause irreparable losses. In order to improve the stability of tropical soil slope in a green and environmentally friendly way, the slope was strengthened by microbial induced calcium carbonate precipitation (MICP) technology and carpet grass root slope consolidation. Suitable microbial strains were first screened out, and the preparation process of related microbial agents was optimized. Subsequently, carpet grass root system was implanted in the slope soil indoors, and MICP treatment was carried out after the formation of the root-soil complex. Subsequently, a series of laboratory tests and numerical simulation analyses were carried out to evaluate the reinforcement effect of this technology. The results showed that MICP technology and plant root treatment complemented each other in terms of mechanical brittleness and integrity of slope, and significantly improved the unconfined compressive strength and shear strength performance of soil, effectively enhanced the stability of slope, and reduced the risk of slope erosion. Finally, the numerical simulation verification was carried out by using Abaqus finite element software, which enhanced the reliability of the research results. It can be seen that the MICP combined with carpet grass root reinforcement method provides an effective reinforcement method for tropical soil slopes, which not only improves the mechanical properties of slopes, but also promotes ecological restoration and environmental sustainability. This result provides a new technical approach for the ecological reinforcement of slopes in tropical areas, and has certain theoretical and practical significance for protecting the ecological environment and reducing the risk of geological disasters.