Taking a three-level, multi product, and dual channel supply chain as an example, the inventory control problem in the supply chain under stochastic demand was explored. A dual channel supply chain simulation model was established based on the independent control, information sharing, and pre warehouse replenishment model of “single manufacturer-dual distributor-dual retailer-dual customer”. In node enterprises, the Pull/Push strategy was adopted for ordering decisions, and information entropy was used to measure the uncertainty of nodes. Finally, the whale optimization algorithm was used to adjust the inventory control parameters in the simulation model. The results show that in the case of interruption, the pre warehouse replenishment mode can increase customer satisfaction in the interrupted channel by 80%. The whale optimization algorithm can ensure customer satisfaction while controlling total costs and reducing uncertainty in the supply chain system.

In dynamic wireless environments, the distortion of transmission waveform is inevitably present, deteriorating the accuracy of identifying high altitude electromagnetic pulse (HEMP) parameters. To address this issue, an extreme learning machine parameter identification network (ELM-PInet)-based parameter identification method was investigated, which leverages the characteristics of HEMP waveform and considers the impact of wireless channels, thereby improving the accuracy of HEMP parameter identification. To demonstrate the nonlinear effects of wireless channels, the transmission model of HEMP waveform was first constructed based on wireless transmission theory. Subsequently, an ELM-PInet was developed to suppress waveform distortion and improve the identification accuracy of HEMP parameters. Finally, the proposed method was validated through field irradiation test on the experimental platform. Simulation results demonstrate that compared to classical HEMP parameter identification methods, the identification accuracy of HEMP parameters is enhanced by the proposed method. Furthermore, the ELM-PInet-based parameter identification method exhibits its robustness against the impacts of different parameters. Additionally, the effectiveness of the proposed method is further validated through field irradiation experiments.

For gearboxes, variations in fault characteristics under different operating conditions, susceptibility to noise interference in fault diagnosis, lead to poor generalization and low recognition accuracy of fault diagnosis models. An end-to-end convolutional block attention module-sparse temporal convolutional network with soft thresholding(CBAM-STCN) was proposed for gearbox fault diagnosis. Firstly, the Hilbert transform was employed to convert the gear fault vibration signal into an envelope spectrum signal. Then, this signal was input into the CBAM-STCN fault diagnosis model. The model integrates a hybrid attention mechanism module, the convolutional block attention module (CBAM), which adaptively learns the weights of channel and spatial attention to extract information sensitive to fault features. The embedded soft thresholding function minimizes the discrepancy between the model’s output and the original input. Finally, the proposed method was utilized to identify and classify various types of gear faults under two different conditions. The results indicate that the CBAM-STCN model achieves an average accuracy of 98.95% in intelligent gear fault diagnosis, demonstrating its potential value for gearbox fault diagnosis.

With the continuous promotion of the “dual carbon” strategic goals and the construction of new power systems, traditional distribution networks are gradually transforming into information-based, digital, and intelligent new distribution systems. To accurately characterize and analyze the characteristics of different types of loads in the distribution network, and support efficient operation and control of the distribution network, a data-driven classification method for typical load curves in the distribution network was proposed. Firstly, based on load data, various classification scenarios of typical loads in the distribution network were analyzed, and performance evaluation indicators for classification scenarios including error rate, accuracy, and confusion matrix were proposed. On this basis, a data-driven load classification method for distribution networks was proposed, which converts 24 dimensional daily load vectors into image data and uses convolutional neural networks to identify load curve images, achieving accurate classification of distribution network load curves. Finally, the accuracy and effectiveness of the proposed method were verified by combining actual distribution network load data, and analyzed and compared with existing methods. The results indicate that the proposed method for classifying typical load curves in power distribution networks has better classification speed and accuracy.

Aiming at the problems such as small and medium-sized obstacles on the road are prone to miss detection, small target obstacles are difficult to detect, and the number of model parameters is large in smart driving scenarios, the obstacle target detection algorithm with improved YOLOv8n was proposed. Distribution shifting convolution (DSConv) was used in the backbone network to replace floating point operation with integer operation, reducing the amount of redundant computation, and maintaining the accuracy by imitating the original convolution layer by quantization and distribution shifting. By adding small target detection layer, the feature information of small target can be captured better and the scale characteristics of small target can be adapted. Combined with SimAM parameterless attention mechanism, SPPF-SimAM module was introduced to improve the quality and diversity of feature representation, and the detection accuracy was improved without increasing the number of parameters. By combining ghost-shuffle convolution (GSConv) and VoV-GSCSP modules, the neck feature fusion network was lightweight, reducing the number of parameters and calculation of the model. The experimental results show that the accuracy, recall, and mean average precision of the improved model are improved by 1.6%, 8.0%, and 6.2%, respectively. The number of parameters is reduced by 6.7% compared with the original model, and the proposed algorithm effectively improves the detection accuracy of small and medium-sized obstacles in smart driving scenarios, and achieves a better balance between the detection performance and the model lightweighting.

The format and content of items such as product names and specifications in the detailed section of VAT invoices are highly flexible and complex, lacking complete gridlines to separate information fields. Existing methods for all-element structural recognition of VAT invoices face issues like low element recognition rates and high computational complexity. A structured recognition method for full face information based on computer morphology was proposed, which uses morphological operations to detect invoice table lines, cuts and recognizes text in different areas of the invoice. Then the implicit rules of the layout of the value-added tax invoice product details area was reused, combined with the text connected areas obtained through computer morphology operations, to construct a complete table structure. Finally, text detection and recognition were achieved using text detection neural network with differentiable binarization (DBNet) and convolutional recurrent neural networks (CRNN). The proposed method was tested on a dataset of 49 value-added tax invoices in three different formats, and the results show that the element recognition rates reached 99.9%, 97.4%, and 98.8%, respectively. The average running time per invoice is 0.90, 0.47, and 0.82 s, respectively. The structural recognition performance of the entire invoice exceeded multiple comparison table recognition models and literature methods.

In order to solve the problem of poor image denoising performance caused by the simple encoder-decoder structure of the convolutional neural network image denoising model, a residual dense image denoising network (RDIDNet) based on the residual dense network and attention mechanism was proposed. Firstly, the global residual block was used to enhance the nonlinear mapping ability of the network model. Secondly, the double-element convolutional attention module was introduced to realize the adaptive feature fusion in the decoding process of RDIDNet model. Finally, the RDIDNet denoising model was compared with 14 representative denoising methods, and ablation experiments were conducted to verify the effectiveness of using RDU Sub Network, DE-CAM, and PSNRLoss for network optimization on the benchmark model. The experimental results show that in the Set12 dataset and BSD68 dataset, RDIDNet improves the peak signal to noise ratio (PSNR) and structural similarity (SSIM) metrics by an average of 1.03 dB and 0.027 5, respectively, compared to the traditional classical method BM3D. Compared to SwinIR based on Vision Transformers architecture, the average improvement is 0.03 dB and 0.001 4, respectively. Compared to the latest CNN based denoising method NHNet, it has an average improvement of 0.22 dB and 0.008 9. The RDIDNet denoising network focuses more on low-frequency information and has more stable model training. It can effectively eliminate image noise while preserving image details and textures, and has good performance.

In order to solve the problem of automatic sorting of garbage, an intelligent sorting method for household garbage was designed based on the artificial intelligence computing platform Jetson NANO and YOLOv8 algorithm. Firstly, the lightweight YOLOv8 multi-target detection algorithm was used to classify garbage into four categories and achieve accurate recognition in the presence of occlusion. Secondly, the CoreXY coordinate transformation algorithm was used to determine the optimal angle for mechanical claw grasping and achieve precise garbage grasping. Finally, based on TensorRT optimization technology, the lightweight YOLOv8 multi-objective detection algorithm was deployed on the NANO hardware platform, completing the overall construction of the four major categories of garbage intelligent sorting system based on YOLOv8-NANO. The results show that in the self-made garbage test consisting of four major categories and a total of 13 sub categories, the sorting system can effectively achieve detection of single and multiple targets. The detection mAP0.5:0.95 value is 97.3%, and the target detection speed reaches 30.6 frames/s. The fastest garbage sorting speed could reach 6 pieces/min.

To solve the problem of deterioration of concrete performance caused by low temperature in cold regions. Based on the theory of nanomaterials to improve the properties of concrete, the effect of nano-silica on the properties of concrete was studied from the macro and micro scales. The results of the compressive strength test show that the compressive strength of ordinary concrete is attenuated by about 10% under low temperature curing. After being mixed with nano silica, the compressive strength of concrete is increased by about 20%, and the optimal dosage is 2%. The improvement mechanism of nano-silica on concrete properties was explored through microscopic test data such as mercury intrusion, X-ray diffraction and scanning electron microscopy. The results show that nano-silica can promote cement hydration at room temperature and low temperature curing, consume calcium hydroxide generated by hydration, and produce more hydrated calcium silicate and hydrated calcium aluminate gel, thereby reducing the porosity of concrete, optimizing the microstructure of concrete, and improving the performance of concrete. Compared with the room temperature environment, the improvement effect of nano-silica on concrete at low temperature is slightly reduced, but it completely overcomes the adverse effects of low temperatures on the performance of ordinary concrete.

Concrete arc beams in the support mold is often difficult to ensure the molding accuracy, the production is more difficult to high cost, and 3D printing technology has a construction speed, design freedom and high characteristics, so in order to solve the problems such as the complexity of concrete arc beam support, the effectiveness of 3D printing arc shell-cast-in-place beam construction was studied. According to the existing 3D printing concrete ratio and process parameters, three 3D printing curved beam mold shells were designed and printed, and the printing and molding accuracy was measured. The mold shells were equipped with reinforcing cages and cast-in-place concrete materials, and 3D printing concrete curved mold shells-cast-in-place beams were produced. The beam specimens were subjected to vertical loading tests to validate the effectiveness of the construction method. The results show that the 3D printed curved mold shell is basically the same size as the 3D model, with a maximum error of 4% in the middle, and the overall printing and molding quality is good. Under vertical loading, the damage patterns of the three 3D printed curved mold shell-cast-in-place beam specimens are similar. The cracking load and ultimate capacity of the beam specimen with reinforcement between the curved mold shell and the cast-in-place beam have been significantly improved, with an increase in the ultimate load of about 25%.