This groundbreaking research introduces a novel model for classifying heart sound signals, which utilizes Empirical Mode Decomposition (EMD) for detailed signal analysis. The model decomposes these signals into Intrinsic Mode Functions (IMFs) and further categorizes them into four fundamental stages of the cardiac cycle. By employing Hilbert-Huang Transform (HHT) to extract crucial time-frequency features, accurate classification is achieved. Despite the presence of internal and external noise interference, the model maintains significant accuracy, owing to the inherent denoising capability of EMD, thereby circumventing complex preprocessing steps. Through extensive experimental validation, employing a hybrid Convolutional Neural Network-Long Short Term Memory (CNN-LSTM) framework, the model demonstrates outstanding precision, approaching a threshold of approximately 95% in initial stages. Serving as a cutting-edge model for deepening understanding of cardiac activity and early diagnosis of cardiac ailments, it adeptly addresses common noise issues encountered in traditional systems. With wide-ranging applications in clinical and home health monitoring, this pioneering model lays the groundwork for future advancements in cardiac healthcare.

Identifying Self-Admitted Technical Debt (SATD) plays an important role in maintaining software stability and improving software quality; SATDs need to be repaid in time, otherwise they may cause serious vulnerabilities or even crash the software. Identifying SATDs from large project code is a costly task, and although existing methods can detect SATDs, and researchers have identified design debt and requirements debt, there is still a lack of methods to achieve multi-label classification of SATDs. In this paper, we propose a CoTCapNet model based on a deep generative model and capsule network, for both recurrent neural networks and convolutional neural networks have the problems of insufficient textual feature extraction and easy to cause the loss of important feature information, first, we use a text generation model based on CoT co-training to generate new samples by learning the original SATD data, which can increase the number of small and medium SATD samples and reduce the data imbalance, then use graph convolutional neural network to encode syntactic dependency trees, construct multi-head attention to encode dependencies in text sequences, and finally merge with semantic information through capsule network. Experiments on crossitem recognition of 10 items show that our approach is more effective than existing methods such as CNN and text mining. The proposed CoTCapNet method has strong advantages, especially in the case of highly unbalanced data.

The real-time scheduling of automatic guided vehicles (AGV) in flexible manufacturing system (FMS) is observed to be highly critical and complex due to the dynamic variations of production requirements such as an imbalance of AGV loading, the high travel time of AGVs, variation in jobs, and AGV routes to name a few. The output from FMS considerably depends on the efficient scheduling of AGV in the FMS. This paper mainly studied intelligent logistics scheduling of automated guided vehicle(AGV) in job shop. AGV logistics scheduling optimization model was established to minimize the travel time of AGV and to reduce energy consumption of AGV. The multi-objective scheduling is carried out by the application of improved grey wolf optimizer (MPGGWO) with task sequencing as the constraint condition. Finally, the actual logistics scheduling of workshop was taken as example to verify the method proposed in this paper. The calculation results show that the AGV logistics scheduling model proposed can well simulate the AGV scheduling time and energy consumption, and the improved grey wolf optimizer (MPGGWO) presents a faster convergence speed and a better optimization ability.

The problem of large cargo volume and low I/O scheduling efficiency is common in Stereoscopic warehouses. In order to improve the efficiency of warehouse access, on the basis of adopting the two-end single-aisle double stacker model, and taking the running time of the stacker as the evaluation criterion, we establish a double stacker model in which the single-command task and the double-command task coexist and reasonably merge the inbound task and the outbound task into a single process, so as to reduce the number of times that the stacker returns to the inbound platform and the outbound platform. The number of times the stacker returns to the inbound platform and the outbound platform is reduced. In order to solve the problem of scheduling the execution order of tasks, the ocean predator algorithm is used to optimize the solution. In order to evaluate the performance of the model and algorithm, simulation experiments are conducted based on a batch of inbound and outbound tasks, and the results show that this research can reduce the idle running time of the stacker cranes and improve the efficiency of the inbound and outbound scheduling of the warehouse.

With the rapid development of deep learning technology, deep neural networks have achieved remarkable achievements in fields such as computer vision, natural language processing, and speech recognition. The core lies in constructing multi-layer neural networks to learn and represent data, thereby accomplishing complex tasks such as image recognition, speech recognition, and natural language processing.This paper designs a highly parallel systolic array on FPGA to accelerate Vision Transformer networks. For the multi-head attention mechanism, a scalable and variable systolic array architecture is designed. This architecture significantly reduces data transmission latency by directly transferring data between processing units and achieves highly parallel computation. Besides，Designed a simple and efficient data flow for matrix multiplication operations.

Since its birth, convolutional neural network (CNN) has been widely used in many fields. FPGA implementations of CNNs have attracted widespread attention due to their high performance and energy efficiency. However, some current computing architectures do not fully utilize the computing power of FPGAs and can only accelerate a single network. In addition, the traditional method of developing FPGA accelerators using Verilog cannot meet the diverse needs and flexibility of accelerators. Therefore, this paper proposes a parameterized configuration of a general convolutional neural network accelerator. In order to improve the computing throughput and frequency, we adopt a systolic array architecture to implement the computing unit of this accelerator. Furthermore, in order to effectively meet the diverse needs of industry and academia, we adopted an agile development approach using Spi-nalHDL. The accelerator was ultimately deployed on various boards such as VU9P and tested using representative algorithms in convolutional neural networks (YOLOv4-Tiny). Experimental results show that when the accelerator runs at a frequency of 200 MHz and accelerates the YOLOv4-Tiny algorithm, the FPS is 85.09. Has excellent acceleration effect.

In recent years, the detection of targets by unmanned aerial vehicles (UAVs) has gained significant attention in research. However, aerial photography with drones often presents challenges such as object occlusion, large-scale transformations, and the detection of small objects. These difficulties pose significant obstacles for existing deep learning-based target detection algorithms. To address these issues, we propose a novel object detection algorithm called YOLOX-SNH. This algorithm builds upon the transformer structure of the core component, which effectively mitigates object occlusion and preserves essential global context information. Additionally, we have incorporated a specialized detection head to enhance the detection performance for small objects. To evaluate the effectiveness of YOLOX-SNH, we conducted experiments on the VisDrone2021 dataset, comparing it with state-of-the-art object detection methods such as ViT-YOLO. The results demonstrate that YOLOX-SNH outperforms these existing methods, achieving an impressive interpretability in drone capture scenarios. Specifically, when applied to the VisDrone2021 dataset, YOLOX-SNH achieved an average accuracy of 67.00%, surpassing the ViT-YOLO method by 1.11%.

Aiming at the difficulty of hyper-parameter tuning commonly found in traditional machine learning algorithms. Combine the Slime Mould Algorithm with traditional machine learning algorithms. Use XGBoost algorithm to construct a concrete compressive strength prediction model, use different algorithms to optimize the parameters, the experimental results show that the SMA algorithm performs well in optimizing the XGBoost parameters, the established concrete strength prediction model is highly efficient, the prediction accuracy is improved significantly, and the complex problem of concrete compressive strength prediction is solved.

Image aesthetic assessment is a complex and subjective task intended for computers to mimic the process of human visual perception to score the aesthetics of an image. Current research in image aesthetic assessment focuses on learning the aesthetic features of an image through convolutional networks, and then aesthetically scoring the image based on the learned features, and researchers are committed to improving the aesthetic assessment method by considering a more comprehensive set of aesthetic features. From aesthetic dichotomy to aesthetic regression to aesthetic distribution, images can provide more and more rich aesthetic information, but the interpretability is still weak. In recent studies, researchers have found that thematic information of images can effectively improve the interpretability of aesthetic evaluation models. As a research theme that has emerged only in the past two years, the aesthetic assessment of multi-theme images is still in the early stage of research. There are still fewer papers in the research direction of multi-theme image aesthetic assessment, and this paper provides an overview of the multi-theme image aesthetic assessment model at this stage. Firstly, it analyzes the research status and development trend of image aesthetics assessment; then it gives an overview of the multi-theme image aesthetics assessment model at the present stage, and details the research status and future development trend of the multi-theme image aesthetics assessment model at the present stage.