To addressing the critical issues faced by traditional heating systems, such as uneven temperature distribution, significant energy wastage, and inefficiencies associated with manual control, an advanced IoT-based intelligent monitoring system utilizing NB-IoT(Narrow Band Internet of Things) technology is designed.The primary objective is to enhance the efficiency, reliability, and user satisfaction of community heating networks through modern technological integration. The proposed system integrates contemporary communication technologies, sensor data acquisition, cloud-based AI analysis, and edge computing. It consists of portable temperature measurement nodes, NB-IoT communication modules, an AI co-processor for local intelligence, and a cloud server for centralized management. Specifically, up to 16 temperature-measuring nodes and two base stations are deployed within the network. These terminal nodes collect indoor heating temperatures alongside user-defined settings, transmitting gathered data via NB-IoT to a cloud server where it undergoes intelligent processing facilitated by an edge gateway. This processed data is then visually displayed on a user interface,enabling historical reference, anomaly detection, and predictive analysis for temperature patterns. Testing results demonstrate substantial improvements in system performance, including a notable reduction in power  consumption, and the overall system current in sleep mode is≤8μA,which extends battery life expectancy beyond two years. The system significantly enhances user satisfaction by providing more consistent and comfortable indoor temperatures. With its proven capability to reduce energy waste and improve operational efficiencies, the system not only promises considerable economic benefits but also establishes itself as an effective solution to the inherent limitations of conventional heating systems. Thus, this work contributes to the advancement of smart community infrastructure, promoting sustainable and efficient heating solutions.

In recent years, fisheye cameras have been widely used in various fields due to their ultra-wide field of view, but they are prone to image distortion that affects application, requiring calibration. Traditional calibration methods are time-consuming and cumbersome when capturing multiple images continuously, and the conventional KB(Kannala-Branolt) model fails to project severely distorted feature points of fisheye cameras with a field of view exceeding 180°. To address these issues, a stereo calibration box for large-wide-angle fisheye lenses devised and an adaptive Christopher Mei model is used for one-shot calibration. The calibration box is designed, then a centroid positioning method is used for precise feature point extraction, and finally MEI calibration is enhanced with the particle swarm optimization algorithm. Experiments show that this method is stable, accurate, and robust, reducing time consumption and improving accuracy compared to traditional methods.

The learner model is an important foundation for effectively recommending personalized learning resources. Although it has widely integrated multiple dimensions of learner characteristics, it has not incorporated online learning self-efficacy, a key indicator for measuring learners‘ psychology and learning motivation.Therefore, self-efficacy of online learning is introduced and a comprehensive evaluation method that combines static and dynamic approaches is proposed, using scales and a series of designed behavioral indicators to achieve a comprehensive assessment of online learning self-efficacy. Diversified features such as learning interests,current levels, and learning styles is integrated to construct a multi-dimensional learner model and an improved neural collaborative filtering method that incorporates the learner model is designed. Experimental results show that the multi-dimensional learner model has significantly better performance in recommendation methods than the comparison methods that do not consider this factor, providing a new perspective for the construction of learner models and enabling more effective personalized and precise learning resource recommendations for learners.

Aiming at the problem of difficult vehicle localization in underground coal mines, a real-time NLOS(Non-Line-of-Sight) error suppression positioning algorithm with a tight combination of UWB(Ultra-Wideband) and IMU( Inertial Measurement Unit) is proposed. First, a recursive model of UWB ranging is constructed to dynamically identify the NLOS error by combining the historical filtering information, and a moving average model is adopted to correct the ultra-wideband measurement. Second, a UWB / IMU tight combination model based on the ESKF(Error State Kalman Filter) is designed to adaptively regulate the process noise covariance matrix through the introduction of a forgetting factor enhancing the correction effect of the measurement information on the state estimation. Simulation experiments show that the proposed scheme improves the maximum 3D localization accuracy by 30. 68% and the average accuracy by 28. 61% compared to the scheme in the least squares support vector machine-based correction method. Compared to the scheme in relative to the residual cliscrimination-based method, the 3D localization accuracy is improved by 35. 86% and the average accuracy is improved by 27. 88% . This research provides a high-precision and low-latency solution for vehicle positioning in the underground complex environment, which is of great engineering significance for promoting the construction of intelligent transportation system in coal mines.

Addressing the challenge of teachers' difficulty in effectively utilizing educational big data to Implement differentiated teaching decisions, methods for differentiated teaching decision-making and implementation is studied based on student learning behavior data. Firstly, an input indicator set that can fully describe students' learning behavior is constructed, and data on each student's learning behavior during the online course teaching process is collected. Based on the collected learning behavior data, training samples are constructed and students' online learning effectiveness is manually evaluated to obtain the corresponding label values for the training samples. Then, using the constructed training samples and the evaluated label values, the convolutional neural network is trained to approximate the mapping relationship between student behavior data and evaluation results. A well-trained network can automatically provide differentiated assessment results based on different students' learning behaviors. Based on the evaluation results of each student and combined with their teaching experience, teachers can develop differentiated intervention strategies for different students. Finally, the implementation effect of the intervention strategy is examined in detail, and dynamic adjustments are made to the intervention strategy based on the actual situation during the implementation process. Empirical research results have shown that compared to traditional teaching decision-making methods, teaching decision-making methods based on online learning behavior data are more significant in improving students' academic performance. The findings reveal that the implementation of differentiated teaching decisions based on student learning behavior data is effective and feasible. The research provides support for teachers in analyzing learning behavior data and adjusting corresponding teaching decisions.

Without considering the feature differences between large-scale data, using a single feature as the query basis can result in significant query errors. Therefore, a parallel query algorithm for large-scale data based on PAT(Pump Algebra Tutor) algebra is proposed. Using PAT algebra to optimize the semantics and logic of parallel data, setting initial sequence blocks for large-scale parallel data, obtaining data block density, and implementing low weight key filtering in a directed graph by adjusting node density according to data block density, the effective filtering is achieved. On this basis, the strategy of minimizing the product of subqueries is used to determine the sequence points where the target data is located. Greedy rules are used to search for clause sets that meet the conditions in the neighborhood set, establish query connections, and achieve efficient parallel data queries. The experimental results show that the proposed method has high data transmission and query volume, indicating that it can achieve accurate queries for large-scale data and has certain practical value.

Currently, low illumination enhancement algorithms applied to crude oil storage stations are prone to color distortion and excessive sharpening in the enhanced images due to the low illumination and contrast of on-site images. Therefore a low brightness image enhancement method is proposed based on LogRetinex-Net. First, a logarithmic transformation layer is introduced to the Retinex-Net to enhance the overall grayscale of the images, reducing the impact of low illumination and contrast in crude oil storage station images. Then, a channel attention mechanism is utilized to increase the network’s focus on color channels, thereby mitigating the issue of color distortion. Finally, the model is trained and validated on a crude oil storage station dataset. Experimental results show that the LogRetinex-Net network improves artifacts and over-sharpening phenomena, significantly enhancing image quality.

Due to the large number of user groups and resources in the employment system for vocational college graduates, the weight of employment resources varies significantly, making it difficult to generate recommendation labels uniformly, resulting in the inability of the graduate employment system to complete on-demand resource recommendations. Therefore, a multi type resource on-demand recommendation algorithm is designed for the employment system of vocational college graduates. By extracting multidimensional features of user information from historical data and utilizing long short-term memory neural networks to fuse multiple sources of data, effective labels are extracted to establish a user label library for the graduate employment system, forming user profiles. Based on the general situation of user profiles, combined with the Ebbinghaus forgetting curve, the label matrix of multiple types of employment resources is evaluated, a content topic model is established, and spectral clustering algorithm is used for graph segmentation. Different weight values are assigned to different employment resources based on similarity, and normalization is implemented to generate secondary labels, completing the labeling process of employment resources. A regional preference base for graduates' employment is constructed, associating and matching user profiles with employment resource labels in designated geographical locations, and score the recommendation results through expert ranking weighting. Experiments are conducted on the above design, and the results show that the hit rate of the algorithm's recommended results is greater than 0. 9, indicating high accuracy.

Multi source all media information comes from different platforms, channels, and formats, including text, images, videos, audio, and other forms. The data has high dimensions and complexity, making information representation difficult. Therefore, a multi-source all media interaction information similarity search algorithm based on double-layer attention is proposed. A dual layer attention model is applied to extract features from all media information, effectively capturing key features of information at different levels. The fuzzy C-means algorithm is used to cluster the entire media information database and classify similar media information together.Using similarity search algorithms, by calculating the similarity between sample information and other information in the entire media information database. The most similar information content can be quickly and accurately searched, providing users with a better search experience. Experimental results have shown that the proposed method can achieve similarity search for all media information with accurate search results.

When volleyball players are blocking and spikling, their upper limbs show the characteristics of turbulent movement. The velocity at the end of the arms generates a non-steady-state fluid trajectory similar to the water vortex ring in the diving phenomenon, making it difficult to capture the position of the end of the volleyball players' arms and resulting in a large error in tracking the end position. To this end, a real-time tracking algorithm for the end trajectories of both arms in volleyball blocking and spikstroke actions is designed. By using the boundary area information reconstruction method, the start and end frames of the volleyball blocking spike action in the action video are determined. By using the LM(Levenberg-Marquardt algorithm) algorithm, the Euler Angle between the skeletal points at the ends of the arms in the start and end frame images is taken as the optimization variable to capture the positions at the ends of the arms of volleyball players. Based on the capture results of the end positions of both arms, the Mean Shift algorithm is selected. By measuring the similarity between the target model for tracking the end trajectories of both arms and the candidate models, the end trajectories of both arms for volleyball blocking spiks are tracked in real time. The experimental results show that this method can accurately track the end trajectories of both arms of the volleyball blocking spikstroke action. The end position error is less than 3 cm, and the tracking speed is higher than 30 f / s, which can meet the requirements of volleyball action analysis in a high-speed motion environment.

The power business involves a wide range of complex domain knowledge, usually existing in an unstructured form, including power system operation, energy management, power grid planning, etc. , which reduces the efficiency of intelligent services for power business. In order to improve service quality, a semantic retrieval algorithm for power business AI(Artificial Intelligence) Q&A based on the subject language model is proposed. A knowledge base for power business is established using a knowledge graph, providing rich semantic information and knowledge storage. Using TF-IDF(Term Frequency Inverse Document Frequency) to match the semantics of power business AI Q&A, a RWT BERT(Retrieval Augmented WaveNet Transformers) model is established based on semantic matching. This model is used to achieve more accurate semantic retrieval function for power business AI Q&A. The experimental results show that the proposed method has a recall and accuracy rate of over 96% , with an MRR(Mean Reciprocal Rank) value of up to 94% , indicating high retrieval accuracy and efficiency.