Colleges and universities are an important part of higher education, providing a large number of talents for social development. The study optimizes the way of student management in colleges and universities based on artificial intelligence technology. Firstly, the K-means algorithm in cluster analysis is used to classify students’ campus behavioral characteristics. Then use Apriori algorithm to correlate students’ behavioral characteristics with academic performance. Finally, colleges and universities can take differentiated management measures for different categories of students. The clustering analysis of 12,885 students’ consumption behavior, work and rest behavior, and study behavior in college Z, followed by the correlation analysis between the clustering results and academic performance, and a total of 10 correlation rules were found. Colleges and universities can formulate management rules based on the analysis results to improve management efficiency. In addition, the student management work of colleges and universities can be optimized and upgraded in several directions, including the awareness of student management work in colleges and universities, the information platform, the archive management work, the student management team, and the information security work.

Under the background of the current information age, the electronic and intelligent transformation of the bidding industry has become an inevitable trend, and the e-bidding model stands out and greatly improves people’s understanding of bidding. Aiming at the traditional e-bidding system, in order to solve the problem of the lack of the traditional e-bidding system that provides the bidding body with referable opinions, this study firstly constructs the e-bidding risk assessment indexes and realizes the optimization of the evaluation module of the system. Then the recommendation algorithm based on deep learning implements the optimization design of the e-header bidding system. This study constructs an optimized recommendation model by fusing knowledge graphs on the basis of deep learning. Then the e-tendering optimization system is designed according to the actual needs of e-tendering, combined with the recommendation model of this paper. The accuracy index ACC of this paper’s recommendation model is improved by about 3% on average compared with other best-performing recommendation models on each dataset, which verifies the excellent performance of this paper’s recommendation algorithm. This study constructs an optimized e-tendering system and proposes suggestions for the development and operation strategy of corporate e-tendering, contributing to the development of e-tendering transactions and the participation of social capital.

The flipped classroom relies on a smart platform to assist the implementation of English translation teaching, combining the smart platform with the students as the core to realize the efficient interaction of English translation teaching and enhance the students’ interest in English translation learning. This paper develops an easy-to-use interactive system for English translation teaching in flipped classroom based on Fine Report, and utilizes MySQL database to store the relevant data generated in the process of use. In this system, the BERT model trained by matrix masking strategy is used as the basis, and the neural machine translation model that assists teachers in English translation homework correction is established by combining the NMT model. Then the K-Means clustering algorithm is optimized by the adaptive K-value selection method, and the students’ learning data on the system is clustered by using the improved K-Means, and the student performance evaluation model is established by combining the CART decision tree. A pedagogical comparison experiment was carried out for the feasibility of the interactive system for teaching English translation in the flipped classroom. The BLUE value of machine translation using the BERT-NMT model was always above 30, and the average accuracy of student performance prediction of the K-Means-CART model could reach 84.85%. The English translation performance of the students in the experimental class was significantly improved after the teaching experiment, and the overall satisfaction of the students with the interactive system for teaching English translation was 4.038 points, which was between the satisfied~very satisfied level. Fully combining intelligent technology to assist teachers in teaching English translation under the flipped classroom can help to enhance the quality of cultivating English translation talents in colleges and universities.

When a laser beam passes through a solid physical material with a nonlinear refractive index, it can produce an optical nonlinear effect, which depends on the refractive index that changes with the light intensity. Based on an analysis of the linear principle of nonlinear optics, the article describes the coupled wave equations under the nonlinear optical phenomenon. It introduces the phase-matching method of frequency conversion and the theoretical basis of optical waveguide. Starting from the classical Maxwell’s equations, the nonlinear optical transmission equations and the optical effect model are established, and then the finite element method (FEM) simulation model is constructed based on the FEM model to analyze the nonlinear optical phenomena of solid-state physical silicon materials. To verify the validity of the FEM model, the optical bistability effect and four-wave mixing spectrum of the nonlinear optical phenomena are simulated and analyzed, and the homochiral spinning effect and transmission spectrum are investigated. When the solid-state physical silicon material is rotating, the laser power required to observe the optical bistability is up to 9.51 W when the rotation rate is increased from 12 kHz to 24 kHz, and the four-wave mixing intensity decreases from 0.115 to about 0.028 when the oscillator frequency of the solid-state physical silicon material is increased from 15 MHz to 30 MHz. The plasma resonance absorption wavelength of the solid physical silicon material is at 791 nm, and the effective refractive index obtained from the simulation is 0.61 in the real part, which is only 1.64% lower than the actual refractive index. The trend of nonlinear optical phenomena in solid-state physics can be effectively obtained by using the FEM model, which provides a new idea for the application expansion of the optical force system.

With the development of the informationization era, it has become the norm for teachers of Civics and Political Science courses in colleges and universities to assist classroom teaching through network resources. In order to further utilize network resources to make them better serve the classroom teaching of Civics and Politics courses in colleges and universities, this paper optimizes the teaching resources recommendation technology based on deep neural network. Defining the network teaching resources data as a ternary group , we put forward the research hypothesis and LSTM model, and establish the G-LSTM recommendation model for recommending the teaching resources of ideological network. The overall framework of G-LSTM model is described, and the recommendation based on G-LSTM is applied to the ideological network teaching resources recommendation. Adopt AUC, MRR and NDCG as evaluation indexes to check the performance indexes of G-LSTM model. Combined with the actual teaching of ideologic theory class, the practical effect of G-LSTM recommendation model is analyzed. 67.81% of students and 39.71% of teachers recognize each recommended online teaching resources. It shows that the improved LSTM model in this paper can further screen the ideological and political network teaching resources, and the teaching resources recommended by the model are more suitable for the teaching of ideological and political theory.

Machine learning-based learning analytics can fully use the learner learning behavior interaction data recorded by online English teaching systems, providing support for observing students’ learning process from the perspective of learning behavior. In this paper, we construct a framework for recognizing college students’ English learning behavior patterns, propose an SGT-based feature extraction algorithm for learning sequences, and use Gaussian mixture models to identify the extracted learning characteristic sequences. Subsequently, a K-means clustering algorithm is used for sequence clustering and lag sequence analysis. At the same time, the English personalized teaching method is designed by combining the proposed personalized knowledge point recommendation method of multi-knowledge fusion in-depth knowledge tracking and group feature collaborative filtering. The results show that college students’ English learning behaviors are classified as active, passive, and passive, and the behavioral sequences of students in different modes are differentiated, in which the sequence residual value of active learners is greater than 1.96. There is a significant difference between the personalized teaching mode and the ordinary teaching mode in terms of the learning mode and the learning effect (P<0.05), and it can achieve a better English teaching effect.

The development of media technology profoundly affects the presentation mode, dissemination rate and scale of news information, which in turn reshapes the business chain and business landscape of the entire news media industry. Based on the analysis of the shortcomings of the LDA model, this paper proposes an improved LDA model with binomial distribution, and applies it to the analysis of the evolution of news topics. The model introduces binomial distribution to enhance the discriminative ability of lexical items, and parallelises it to improve the classification effect of news topics. In order to effectively obtain the relevant features of cultural communication in news documents, this paper introduces BERT to obtain word embedding and word vector matrix, and then realises the generation of theme word structure and theme words. The performance of the improved LDA model is verified through the THUCNews dataset, and the news topic morphology is visualised and analysed with the example data, and its morphological evolution, as well as the degree of contribution to cultural communication, is studied. The theme consistency score of the improved LDA model is -13.39 when the word generation probability is 1, which is 19.14% higher than that of the traditional LDA model. The intensity of the ‘cultural policy’ news format theme increases 14.44 times from 2010 to 2023, and the mean value of the ‘cultural governance’ news format theme’s contribution to cultural dissemination reaches 0.091. Based on the innovation and evolution of news forms, we can empower more communication channels for culture and spirit, so as to enhance people’s cultural self-confidence and national cultural soft power.

Dance drama is a comprehensive art with dramatic conflicts and plots based on the use of dance’s own language system, which plays an important role in cultural dissemination and aesthetic experience. The article designs a resource library of classic dance drama works in the way of WEB site, establishes a data dynamic distribution strategy to deal with structured data, and combines the consistent hash algorithm to optimize the load balancing of structured data in the resource library. Then, a graph convolutional neural network model and a sample-weighted aesthetic classification model are combined to establish an aesthetic assessment model for images of classical dance drama works, and a regularized matching module is designed. For the application effectiveness of the structured data processing strategy, the structured data processing of the classic dance and drama works resource library is verified, and the hyperparameters of the model, evaluation results and ablation experiments are also analyzed. Combined with the data in the resource library of classic dance drama works, the aesthetic experience of the audience was analyzed using a questionnaire. After using the dynamic distribution strategy to process the structured data, its write and query times were shortened by 40.05% and 17.89% compared to before use, and the response time under different index query load balance degrees did not exceed 55ms.The accuracy of the aesthetic assessment model for classical dance and drama works was 48.85%, and the accuracy improvement of the G-AANet model compared to BoTNet ranged from 0.93% ~ between 6.12%. The resource base of classical dance drama works established through structured data processing helps to enhance the audience’s aesthetic experience of dance drama works and helps them to appreciate the spiritual connotation of dance drama works.

This paper is based on the digital image processing technology, using the undamaged image information to restore and protect the frescoes. The discrete binary wavelet change is used to decompose and denoise the image signal. And decompose and filter the high-frequency component and low-frequency component of the image, choose different components, respectively, carry out coefficient transformation, and solve the OMP least-paradigm for different random matrices. The color space is selected, and the mural color space is channel decomposed according to the grayscale mode and restored separately. Establish an assumed datum for each independent face of the mural, establish a spatial coordinate system for it, realize the transformation of spatial coordinates, and realize the super-resolution three-dimensional reconstruction of the mural based on the generative adversarial network and the self-attention mechanism. Objective evaluation indexes and subjective evaluation indexes are established to compare the protection effect of different algorithms on murals. Compared with the traditional algorithm CDD, this paper’s algorithm improves the restoration time by 9.545~15.625 s, and the peak signal-to-noise ratio index improves by 1.35~4.769 db. In the results of the image extraction and processing, the calculated values of discrete curvature of the mural segments AB, CD, and EF ranges from -0.00945 to -0.00478, and the difference of standard deviation of the curvature from the target curvature is 6.477%. The approximate target curvature is obtained, and the algorithm has strong adaptive ability.

Electricity theft management is closely related to the economy of electric power enterprises. This paper proposes a power theft estimation method based on semi-supervised learning and time series analysis prediction. The electricity consumption data of power theft users are extracted as time series data, and in order to achieve multi-step prediction, MMD is utilized to improve the LSSVR semi-supervised learning algorithm. In addition, a perturbation term is introduced to optimize the convergence effect of the artificial bee colony algorithm, and a time series prediction algorithm based on improved artificial bee colony is established. Bringing in the power theft monitoring process to identify whether the user has power theft behavior, using the real power consumption dataset as the experimental validation data, comparing the identification accuracy of the prediction model. Predict the potential power theft of each user, solve the optimization model with the goal of optimal economic efficiency, and determine the actual ranking order of power theft users. The improved time series prediction algorithm proposed in this paper has a global error of 0.0003 and 0.0027 in dataset 1 and dataset 2, respectively, with the lowest global error and the highest overall accuracy of PSE prediction. And the algorithm predicts the list of users to be scheduled is basically the same as the list of users determined by the real PSE, which can achieve the maximum economic benefits.