Aiming at the problem of white noise amplification of differential microphone array, a design method of parametric differential beamformer was proposed. Through theoretical derivation, it is proved that the delay summing beamformers can maximize the white noise gain and the superdirected beamformers can maximize the directional gain. The orthogonal eigenvector was obtained by using the unitary diagonalization method to deal with the pseudo-correlation matrix between the steering vector and the white noise gain, and the parameters of the beamformer machine were designed based on it. Through simulation experiments, the performance of parametric differential beamformers under different parameter settings was analyzed. Experimental results show that the proposed method can flexibly balance and adjust the white noise gain and directional gain by adjusting the parameters.

Multi-label learning is a common problem in real application scenarios. The construction of large-scale multi-label datasets often means high cost, so semi-supervised learning technology appears. At present, most semi-supervised learning is mainly used in the field of single label classification. Although semi-supervised learning in the field of multiple labels classification has made some progress, there is still much room for improvement in training time consumption, training effects and the use of potential relationships between labels. A multi-label semi-supervised curriculum learning model was proposed under the dual structure semi supervised course learning under dual structure(SSCD) to solve the above problems. Firstly, a curriculum learning scheme based on dual difference was designed, which greatly reduces the training time and improves the robustness of the model. Secondly, a single attention mechanism was designed to explore the potential relevance between labels. The performance of SSCD in the prediction task was evaluated on three open test datasets, and the results compared with four benchmark models show that the comprehensive indicators of SSCD are optimal in all aspects. Finally, the structure ablation experiment was carried out to prove the effectiveness of the proposed single attention mechanism.

In response to the problem of low accuracy in epilepsy detection and recognition using single-view networks, a multi-view convolutional network model with fused attention mechanism (FAM-MCNN) was proposed. Multiple view features were extracted from time domain, frequency domain, time-frequency domain and nonlinear domain to characterize electroencephalogram(EEG) signals comprehensively. Multi-scale convolution was used to capture different levels of detail information. In order to improve the ability to distinguish different types of EEG signals in epileptic patients, the attention mechanism was introduced to combine the features from view dimension and single feature vector dimension respectively. The results of the comparison experiments performed on the CHB-MIT epilepsy dataset show that the average accuracy, sensitivity, and specificity of the FAM-MCNN model are improved by 14.29%, 16.13%, and 12.54%, respectively, when compared to a single-view network. In addition, experiments under a small number of training samples (25%) show that its detection performance reaches the level of the comparison model with a large number of training samples (80%~90%).

Aiming at the problem of time-consuming and labor-intensive routing path design in the cable layout design of complex electromechanical products, an automatic routing technology for complex electromechanical product cables based on multi rules particle swarm algorithm was proposed. Firstly, the cable routing environment of electromechanical products was analyzed, and the routing path was abstracted into a sequence of points to complete the definition of cable routing space. Through pose transformation, the problem of difficult interference detection between wiring paths and parts in electromechanical products was solved. In order to make full use of the wiring space, the particle multiple rules were introduced into the particle swarm optimization algorithm. By using particle number, multi-scale collision detection, adjacent waypoint replacement method and fourth-order quasi-uniform B-spline curve method, the problem that the routing environment is complicated and the optimal solution cannot be obtained was solved, and the searching ability, solving speed and routing quality of the algorithm were improved. Through simulation analysis and comparison with other algorithms, the superiority of the algorithm is proved. The example proves that the proposed method can search feasible paths efficiently during routing. The generated routing paths do not interfere with parts in three-dimensional space, and there are no mutation points in the path fairing, which provides a new idea for the automatic routing of complex electromechanical products.

Compared to image instance segmentation in general scenes, instance segmentation in complex stacked scenes is affected by complex situations such as severe occlusion and stacking of similar objects, making instance segmentation more difficult. To solve the problem of garbage instance segmentation in complex stacking scenarios, an instance segmentation algorithm combining YOLOv8 and two-layer feature network strategy was proposed. Firstly, the feature data was layered in the data preprocessing part, and the two-branch feature fusion was realized through the graph convolutions network (GCN), which reduces the influence of stacking on the features of the occluded objects, thus solving the instance segmentation problem under complex stack occlusion. At the same time, in order to solve the problem that similar objects are easily confused, a soft threshold non-maximum suppression algorithm and a new intersection ratio algorithm were integrated. Finally, according to the complexity of application scenarios and data sets, the feature extraction module of the backbone network was optimized, and the multi-scale attention mechanism was introduced in the backbone network, which effectively improves the detection performance of the model. In the experiment, examples of occlusive garbage classification were used to segment the dataset. The experimental results show that this method outperforms other methods in terms of average accuracy, average accuracy when the intersection to union ratio threshold is 0.5 (AP₅₀), and average accuracy when the intersection to union ratio is 0.5~0.95 (AP_50~95). Compared with the original YOLOv8 algorithm, the detection AP₅₀is increased by 7.9% and the segmentation AP₅₀ is increased by 5.4%, which has better detection and segmentation effects.

In image inpainting, it is crucial that the identification and inpainting of local detail features and the preservation of global features. The models based on fractional-order partial differential equations were characterized by rich evolutionary behaviors, which allow image details to be effectively understood and a certain sharpening effect to be exhibited in image inpainting. However, issues such as inaccurate identification of large-scale features and over-sharpening are prone to be encountered. An optimal control model was proposed and the objective function was defined by the total variation energy of image global features and the constraint was formulated by a spatial fractional-order vector-valued Cahn-Hilliard equation, aiming to achieve a balanced effect between local detail restoration and preservation of global features. L₂ gradient flow, H^-1 gradient flow, and convex splitting were applied to design a numerical scheme for non-convex constraint conditions. And then the split bregman method was used to optimize the objective function with a dynamic grayscale adjustment strategy was introduced to maintain grayscale discrimination capability while enhancing computational efficiency. The numerical experiments demonstrate that the new model achieves an improvement on peak signal to noise ratio(PSNR) ranging from 0.371 8 dB to 9.935 2 dB compared to other methods, exhibiting strong competitiveness in terms of structural similarity(SSIM) and greater effectiveness on images with fragmental damages. Moreover, compared to traditional fractional-order equation models, the computational time is reduced by a factor of 49.50% to 52.91%.

Weld defects present within pipelines constitute a considerable threat for leakage and rupture accidents. To elevate the detection precision of these defects, X-ray inspection was employed as a means to identify and locate them with greater accuracy. However, the diverse types, small sizes, and complex backgrounds of weld defects posed challenges for accurate detection. To address the limitations of current deep learning-based models, such as inadequate adaptability to complex backgrounds and lighting variations, as well as poor performance in detecting small targets, an improved faster region convolutional neural networks(Faster R-CNN) network model was investigated. This model incorporated a channel attention mechanism into the backbone network, modified the residual block structure, and employed ROI Align to replace the traditional ROI Pooling. The results show that compared to the original algorithm, the improved Faster R-CNN model achieves significant improvements in mean average precision (mAP) and F₁, with respective increases of 15.82% and 16.44%. It is concluded that this improved model can meet the high-precision requirements for weld defect detection and holds significant theoretical importance as well as promising prospects for engineering applications.

Aiming at the blindness of the current coal mine roadway surrounding rock support programme and its parameter design, in order to improve the effect of roadway surrounding rock support and meet the requirements of safe and efficient production of mines, taking the Jiaoping mining area as the engineering background, the roof strength, coal gang strength, bottom plate strength, the basic top comes to press the equivalent, mining disturbance, roadway buried depth, roadway protection coal pillar width, span height ratio, top height ratio and the maximum horizontal principal stress were selected as roadway stability master control indicators, and 16 typical roadways and chambers were selected as samples, and the weights of 10 classification indicators were determined based on the analytic hierarchy process. On this basis, the stability of the sample roadway was clustered and analyzed, and the optimal classification number was selected according to the F-statistic method to divide the sample roadway into five categories: very stable, stable, basically stable, unstable and extremely unstable, and then the cluster center of the stability of the mine roadway was constructed. Finally, based on the above theory, the stability of the surrounding rock in the 2407 return wind channel of Yuhua Mine was predicted, and the targeted support measures and parameters were proposed. The results show that the classification results of 2407 return wind channel roadway stability are in line with the actual field engineering, and the deformation control effect of surrounding rock is good, which provides a strong guarantee for the safe and efficient production of the working face.

In September 5, 2022, a M6.8 earthquake occurred in Luding County Sichuan Province. Quite a lot of store-front type buildings damaged or collapsed. Five representative buildings representing both positive and negative aspects were selected to analyze the earthquake damage mechanism through theoretical basis and model experiment. The results show that the earthquake damage is mainly concentrated in the bottom layer, which is composed of concrete column and masonry wall. The masonry wall with no lateral openings restrains the transverse and torsional deformation of the structure, and the floor only transports along the longitudinal direction, and the seismic shear force shared by each member is proportional to its longitudinal lateral stiffness. The rigid and brittle members with large stiffness will appear “internal force condensation”, and then reach “deformation saturation”, and lose the load-bearing capacity in the way of brittle failure, and the gravity of the upper layers will be borne by the transverse wall. If the earthquake does not stop at this time, the transverse wall will lose the role of “buttress” in the direction of exit plane, and the whole structure will collapse along the longitudinal contact with the ground. On the contrary, if the structure of the bottom layer avoids large differences in stiffness, it can significantly improve the seismic resistance.

The drainage consolidation characteristics of municipal sludge are closely related to its water occurrence form. However, there is insufficient understanding of the water transformation law of municipal sludge after chemical conditioning and consolidation. Based on the theory of soil science, the soil-water potential curves of different types of municipal sludge were tested by centrifuge method, and the water forms of municipal sludge were divided into bound water, capillary water and gravitational water according to the range of soil-water potential. On this basis, the different forms of water content in the original sludge, consolidated samples and conditioned sludge samples were compared to reveal the water transformation law of municipal sludge under the action of consolidation and chemical conditioning. The results show that the bound water content of the sludge is reduced by 70%~80% and the free water content is doubled after the sludge is modified by ionic salt. Under the consolidation pressure of 3.1 kPa, only a part of gravity water is discharged from different types of municipal sludge, and the contents of capillary water and bound water are basically unchanged. Under the consolidation pressure of 100 kPa, the gravity water is completely discharged, the capillary water is significantly reduced, and the bound water is slightly reduced.