Music therapy is the treatment of college students’ psychology through various techniques and methods of music, and this paper focuses on researching and analyzing the improvement effect of music therapy on college students’ mental health in the context of cultural education. Students’ physiological data are collected and denoised, and machine learning models are used to realize the multimodal fusion of all kinds of physiological signal features to obtain the objective psychological state assessment values of college students. The subjective assessment results of the mental health assessment scales were then combined to analyze the improvement effect of mental health of college students in the music therapy intervention. The analysis of the psychological health status of the students before and after the intervention experiment revealed that the objective assessment values of the psychological state of the college students in the intervention group gradually tended to be positive with the music therapy, and the subjective assessment results of the psychological health scales of the students in the intervention group were signiϐicantly better than those of the nonintervention group after the experiment (P<0.05). Music therapy has a signiϐicant role in intervening in the mental health of college students and resolving their psychological malaise, which is of great practical and guiding signiϐicance in improving the psychological tolerance and health of college students.

For national grid power line infrastructure construction construction, quality management and control can ensure improved safety standards, long-term reliability and cost savings through avoiding rework. In this paper, a high-definition image of a transmission line is collected from multiple viewpoints by a UAV, and a model for recognizing surface defects on infrastructure lines is proposed to reduce the computational complexity to improve the YOLOv8 algorithm. The model uses ResNet50 as the feature extraction backbone network and fuses convolution and attention mechanisms to enhance global and local feature extraction. A multi-scale feature aggregation diffusion module is added to the neck network of the model to enhance the detection of small targets on infrastructure lines. Finally, the classification loss function combined with the PIOU bounding box loss function is introduced to further enhance the recognition accuracy of infrastructure line surface defects. The experimental results show that the mAP of the infrastructure line surface defect recognition model is up to 0.935, which is 2.41% higher than that of the baseline model, and the performance is significantly better than that of some of the current mainstream defect recognition models. Therefore, from the computational complexity, combined with the target detection YOLOv8 algorithm can realize the accurate recognition of surface defects on infrastructure lines, and provide reliable data support for improving the timely repair of grid infrastructure lines.

In recent years, with the development of science and technology, image enhancement has become a very important topic in scientific research, become an indispensable part of machine vision, and has a wide range of applications in various fields of computer vision. In this paper, the image gradient enhancement algorithm is first improved based on the image gradient field, and its enhancement effect on low quality (low resolution) images is found to be poor through experiments. For this reason, the study constructs a multi-scale feature image enhancement model (LIEN-MFC) by convolutional neural network to further optimize the image enhancement effect. By comparing with different algorithms, the average PSNR of the model is 21.80 and the average SSIM is 0.8767, and it outperforms other compared algorithms in both PSNR and SSIM. In addition, the ablation experiments demonstrate that the enhancement effect of the LIEN-MFC model is further improved on the basis of the improved image gradient enhancement algorithm. The results show that the image enhancement model algorithm with multi-scale features proposed in this paper has a significant image enhancement effect and the improved image gradient enhancement in image enhancement of convolutional neural networks improves the model performance to some extent.

The expansion of English vocabulary is the foundation of college students’ English learning and the key to improve English learning. This project centers on the quantitative analysis of college English vocabulary learning efϐiciency improvement, through the questionnaire survey to understand the use of English vocabulary learning strategies of students. The inϐluencing factors of English vocabulary learning efϐiciency improvement are selected, correlation analysis is carried out, and then multiple regression model is used to explore the role of each variable on the improvement of English vocabulary learning efϐiciency. The results show that students most often use the metacognitive strategy of preplanning (3.674), and that students who are good at learning are more inclined to adopt the metacognitive strategy to control vocabulary learning from a macro perspective. Multiple selfelements and environmental elements together positively affect the improvement of English vocabulary learning efϐiciency (p < 0.01), with the most signiϐicant effects of learning strategies (0.482), teaching methods (0.457) and learning strategies (0.416). It is recommended to promote the efϐiciency of English vocabulary learning through innovative teaching methods, combining word class memorization, expanding the scope of reading, and vocabulary association learning.

In order to better realize the automatic classification and change detection of remote sensing images, this paper proposes an automatic remote sensing image classification model based on CNN and migration learning, and constructs a remote sensing image change detection model by combining CNN and Transformer network. In the remote sensing image classification model, DenseNet network and Inception network are used as the backbone network, combined with the new channel attention module to mine the image features of remote sensing images, and then realize the accurate classification of remote sensing images. In the remote sensing image change detection model, the convolution operation of CNN with different sizes of void rate and expansion rate is utilized to better guide the feature map to focus on local information. Combined with the dynamic deformable Transformer to provide more accurate remote sensing image location information and detail information, to reduce the impact of background interference on remote sensing image change detection, and to improve the model’s ability to recognize the target of remote sensing images. The parameter count and floating-point computation of the remote sensing image classification model are 7.69MB and 1.89GB, respectively, which are smaller than the parameter count and floating-point computation of the single network model. The RSICD models mF1 and mIoU are 1.66% and 0.58% higher than the optimal ones. Through the effective integration of convolutional neural networks and many different types of deep learning techniques, automated classification and change detection of remote sensing images can be realized.

The development of globalization has contributed to the increasing demand for cross-language communication, and machine translation, as an effective language conversion tool, has improved the quality and efficiency of English translation. The article discusses the syntactic optimization and semantic reconstruction strategies for English translation based on machine learning. The machine translation model of English syntax optimization and semantic reconstruction based on EM algorithm is constructed by using key technologies such as EM algorithm and multi-head attention mechanism. The model adopts a joint learning method, combining the Transformer model with the EM algorithm. The dependency between any two words in the input sequence is captured using the multi-head attention mechanism, and the new translation corpus is generated by multi-task joint training algorithm. The training phase of this paper’s model has good translation effect, and the model of this paper gets the highest BLEU score of 32.86 when the number of multi-head attention layers is 1. The distribution of semantic features of translation reconstruction under this paper’s method is basically consistent with the simulation results, and the error elimination rate of semantic reconstruction is 99.64% when the number of samples is 500. The method in this paper is more effective in syntactic structure optimization, with the highest BLEU scores on “Chinese to English” and “English to Chinese”, and the syntactic correctness rate on English long sentences of different topics reaches 88.69%~96.57%.

This topic is centered around temperature and stress, and describes the theory of electric power thermal characteristics. There are usually two methods for thermal coupling analysis, for direct coupling and sequential coupling. Considering that the stress field of the cable does not have much influence on the temperature field, it is proposed to use the sequential coupling method for the calculation of the thermal characteristics of the cable. The calculated and solved cable temperature and stress distribution values are put into the Lap-ML-ELM algorithm for training. When the contact coefficient k=1, 4, 7, 10, 13 and 15, the cable joints and surfaces produce a monotonically increasing law of temperature, and the stress exhibits the same situation.During the training of the model on the thermal characteristics of the cables, it is found that the accuracy curve of the thermal characteristics detection of the Lap-ML-ELM algorithm is higher than that of both the RNN network and the CNN network, which shows that in the detection of the thermal characteristics of cables, the Laplace Multilayer Extreme Learning Machine fusion algorithm performs more obviously.

Students have the problems of insufficient self-control, insufficient learning motivation and unplanned and unsystematic for independent learning of university French. In order to solve this problem effectively, this study proposes the reform of French blended education model guided by POA theory. In this paper, we design a hybrid intelligent teaching mode of university French guided by the output-oriented approach, improve it based on the mutation operation in the genetic algorithm, propose the adaptive mutation genetic algorithm, and optimize the BP neural network with this algorithm. The GA-BP neural network is trained through simulation experiments to verify the performance of the algorithm. Using SEM structural equation modeling, the measurement model of six dimensions, namely, learning effect, teaching effect, learning input, objective learning conditions, subjective learning factors and learning ability, is established, integrating factor analysis and path analysis, and relevant research hypotheses are proposed. The feasibility of the hypotheses is verified one by one through empirical research. The path coefficients between each variable in the model and the path coefficients of the factor loadings are all at the significant level of 0.000, and all of them are positive, the path coefficients’ validity is within the acceptable range, and the hypotheses proposed in this paper are all supported. Compared with the default path, 69.78% of the students in the recommended path for learning French think that the knowledge of the recommended learning path is easy to understand, and the learning path constructed on the basis of the educational resources of the output-oriented method can better satisfy the learning needs of the students compared with the default learning path.

The reasonable division of power supply grid plays an important role in the feasibility and stability of power grid operation. This paper mainly explores the feasible methods of power supply grid division under the dynamic change of grid load. The grid load prediction model is constructed by the improved long and short-term memory network algorithm (ILSTM) based on expert rules to visualize the dynamic changes of the grid load. Based on the study of hierarchical architecture of power supply grid, the objective function is constructed using hierarchical recursive method, and the power supply grid division model is constructed with adjacent connection as the basic constraint. The power consumption information of JH urban area is selected as the data source of this paper, and the method of this paper is used to forecast the grid load of JH urban area and perform the power supply grid division. The power supply network in JH city can effectively meet the objective function and constraints set in the model, and the average number of faults in the power supply network decreases by 94.8% compared with that before the grid demarcation, which fully ensures the safety and reliability of the power supply network operation.

In this paper, for the influence of non-metallic inclusions on the contact fatigue performance of steel, based on the finite element method and rolling contact fatigue theory, the contact fatigue model of U26Mn2Si2CrNiMo bainitic austenitic steel containing non-metallic inclusions is established. The characteristics of non-metallic inclusions and U26Mn2Si2CrNiMo bainitic austenitic steel are analyzed. To investigate the changes in the composition, density and size of each inclusions during the production steps of U26Mn2Si2CrNiMo bainitic austenitic steel by using the inclusions detection technique in steel, the stress and strain response algorithm and the thermodynamic calculations (deoxidization equilibrium calculations of the steel liquid). To analyze the range of fatigue damage concentration caused by non-metallic inclusions by characterizing the distribution of subsurface fatigue damage in the RCF process of U26Mn2Si2CrNiMo bainitic austenitic steel. Explore the effect of the distribution depth of individual non-metallic inclusions on the contact fatigue life of U26Mn2Si2CrNiMo bainitic austenitic steel, and the role of the angle of arrangement of dual nonmetallic inclusions on the properties of U26Mn2Si2CrNiMo bainitic austenitic steel. When circular alumina inclusions with a radius of 5 m are located at different depths of the bainitic austenitic steel, the von Mises stress reaches a maximum value of 770.0 MPa at a depth of 0.53 mm (0.67 Hb ) of inclusions, which is increased by 18.5% compared to the case without inclusions (650 MPa). When the spacing of the two inclusions is 2.5 r (12.5 m ) and the depth is 0.5 mm, the arrangement of the nonmetallic inclusions affects the predicted fatigue life, and the two inclusions reduce the predicted fatigue life around them to different degrees.