黑龙江科技大学

To address the problem of high-precision prediction of mine airflow in mine ventilation systems,an Improved Sparrow Search Algorithm(ISSA)-based BP neural network model(ISSA-BP)was developed.The ISSA,incorporating Cauchy mutation and opposition-based learning mechanisms,was used to globally optimize the initial weights and thresholds of the BP neural network,thereby reducing the interference of randomly initialized parameters with the training results and improving the convergence stability,optimization accuracy,and nonlinear fitting capability of the model when handling complex airflow time-series data.Comparative experimental results showed that the ISSA-BP model achieved a Mean Absolute Error(MAE),Mean Absolute Percentage Error(MAPE),Mean Squared Error(MSE),and Root Mean Square Error(RMSE)of 0.75 m3/s,6.2％,0.98 m6/s2,and 0.98 m3/s,respectively,with a coefficient of determination(R2)of 0.98,outperforming the Long Short-Term Memory(LSTM),BP,Particle Swarm Optimization(PSO)-BP,and SSA-BP models.The ablation experimental results showed that,after the opposition-based learning mechanism was introduced into the SSA-BP model,the MAE,MAPE,MSE,and RMSE decreased from 2.08 m3/s,8.9％,10.37 m6/s2,and 3.22 m3/s to 1.96 m3/s,8.2％,9.85 m6/s2,and 3.08 m3/s,respectively.After the Cauchy mutation mechanism was introduced,the corresponding metrics further decreased to 1.48 m3/s,7.3％,5.95 m6/s2,and 2.44 m3/s,respectively.When the two mechanisms were integrated,the prediction errors were further reduced,indicating that their synergistic effect further improved the prediction accuracy and stability of the model.

In complex environments commonly encountered during routine inspection of power equipment,such as low illumination,occlusion,and dense distributions of small targets,infrared thermal imaging often suffers from poor clarity and blurred edges.To address these issues,an improved CA-ESRGAN-based infrared image enhancement algorithm is proposed.Firstly,LDConv is used to replace the standard convolution,dynamically adjusting the convolution kernel sampling positions to enhance the model's adaptability to complex inspection conditions.Secondly,SCSA is adopted to replace the original channel attention mechanism,strengthening the capture of both local details and global contextual information.Finally,R-RaGAN is introduced to optimize the adversarial training process,improving image quality and training stability.Experiments conducted on the FLIR public dataset and an infrared dataset collected from a substation demonstrate that well performance on both datasets the is achieved by the proposed algorithm.In particular,on the substation dataset,PSNR,SSIM,and LPIPS are improved by 1.84 dB,0.038 1,and 0.013 1,respectively,compared to the best-performing baseline,further confirming its effectiveness in enhancing infrared image quality under complex real-world conditions.

Existing mine ventilation network solution methods do not consider the distribution of error weights in the ventilation network during the solution process,which leads to deviations in the airflow of each branch and limits the accuracy of the results.To address this issue,a ventilation network solution method integrating rough set attribute reduction and Weighted Least Squares(WLS)method was proposed.Based on rough set theory,a forward greedy algorithm was used to construct equivalence classes and calculate dependency degree,thereby achieving attribute reduction,and the key attributes highly correlated with the airflow distribution were identified as airflow,resistance,and cross-sectional area.According to the weights of the key attributes,Prim's algorithm with a priority queue was employed to construct the minimum spanning tree of the ventilation network,and the co-tree branch set was further obtained.The WLS co-tree branch determination method was introduced to iteratively optimize the weights of each branch,reduce computational errors,and improve prediction accuracy.The results of engineering application analysis showed that,after introducing rough set attribute reduction,the solution time of the network was effectively reduced,and the efficiency and accuracy of ventilation network solution were improved.The WLS co-tree branch determination method controlled the airflow calculation deviations of 16 key co-tree branches within 0.1％-2.0％,demonstrating higher computational accuracy and stability.

To address the challenge of extracting fault features from vibration signals in rolling bearing fault diagnosis under strong noise interference,an improved feature extraction method based on symplectic geometry modal decomposition(SGMD)is proposed.In this method,the density of samples around core points is calculated,and then the radius based on density ratios is adaptively adjusted to perform cluster analysis on the initial components obtained through SGMD.And this approach is used to resolve the parameter sensitivity issue encountered during initial component recombination in traditional SGMD.Com-parative simulation experiments demonstrate that the proposed improved SGMD does not require param-eter selection and achieves optimal denoising performance with a signal-to-noise ratio of 22.9 dB.By applying the improved SGMD and short-time Fourier transform time-frequency analysis to vibration sig-nals of normal bearings,inner race faults,and outer race faults,the method is validated to effectively extract the rolling bearing characteristic features.Combining with the AlexNet algorithm,precise fault diagnosis with the highest accuracy rate reaching 98.53％is achieved.

To investigate the mesoscopic crack evolution behavior and the dominant mechanisms during the failure process of sandstone,an integrated approach combining acoustic emission moment tensor inversion with RA-AF parameter analysis was employed to quantitatively characterize the types,spatiotemporal distribution,and stress response characteristics of microcracks.Based on moment tensor theory and incorporating sensor coupling coefficients calibrated through pencil-lead break experiments,microcracks were classified into five categories—shear cracks,tensile-shear mixed-mode cracks,compressive-shear mixed-mode cracks,tensile cracks,and compressive cracks—using the crack tensile angle criterion.Furthermore,an RA-AF empirical model was established to support the analysis.The results indicate the following:(1)Under various loading paths,microcracks resulting from sandstone failure are predominantly shear cracks.The number of each of the five microcrack types exhibits a positive correlation with stress level,with shear cracks showing the most significant increase.(2)As stress increases,microcracks initiate,propagate,and gradually coalesce,forming a fracture zone that corresponds to the macroscopic failure surface.(3)RA-AF analysis reveals that shear cracks account for more than 50%of all microcracks in sandstone,which aligns with findings from moment tensor inversion.(4)Waveforms generated by tensile cracks exhibit abrupt characteristics,with concentrated signal energy in the frequency domain,whereas waveforms associated with shear cracks display oscillatory behavior,featuring dispersed frequency-domain energy and higher amplitude.This distinction provides a physical mechanism that explains the heterogeneity observed in RA-AF parameters.(5)Moment tensor inversion is well-suited for theory-driven,detailed analysis of crack mechanisms,while RA-AF analysis is more appropriate for rapid identification of crack types in engineering practice.This study elucidates the dominant micromechanical mechanism of shear failure in sandstone and the co-evolutionary behavior of multiple crack types,thereby providing a theoretical foundation for rock fracture prediction.