This study introduces and implements a local, lightweight, intelligent teaching-assistant system. Using the IPEX-LLM (Intel PyTorch extention for large language model) acceleration library, the system can efficiently deploy and execute large language models that are fine-tuned using the QLoRA (quantum-logic optimized resource allocation) framework on devices with limited computational resources. Combining this with enhanced retrieval techniques, the system provides flexible course customization through four major functional modules: intelligent Q&A, automated question generation, syllabus creation, and course PPT generation. This system is intended to assist educators in improving the quality and efficiency of lesson preparation and delivery, safeguarding data privacy, supporting personalized student learning, and offering real-time feedback. Performance tests exemplified by the optimized Chatglm3-6B model show the rapid inference capability of the system via the processing of a 64-token output task within 4.08 s in a resource-constrained environment. A practical case study comparing the functionality of the system with native Chatglm-6B and ChatGPT 4.0 further validates its superior accuracy and practicality.
Building an intelligent education platform is an important process in the promotion of the intelligence of education. However, the artificial intelligence model on which intelligent education platforms rely consumes a large amount of electricity and energy during its training process, therefore, it is of great significance to carry out a short-term power load prediction for building an intelligent education platform. However, the major issue is the weak correlations between some attributes and power load data when considering multiple attributes during short-term power load forecasting, and the Transformer cannot capture the temporal correlation of power load data, which leads to a lack of accuracy in power load forecasting. Therefore, a short-term power load forecasting model, SF-Transformer is proposed that is based on the SR (Székely and Rizzo) distance correlation coefficient, fusion temporal localization coding and Transformer. The SF-Transformer filters the attributes that affect the power load data by using the SR distance correlation coefficient and selects the attributes that have higher SR distance correlation coefficients with the power load data. The SF-Transformer adopts fusion time localization coding, thereby combining global time coding and local position coding, which helps the model to comprehensively obtain time and localization information regarding power load data. Experiments conducted on the dataset show that SF-Transformer has a lower RMSE (root mean square error) and MAE (mean absolute error), compared with those of other power load forecasting models over two-time durations.
Recently, with the development of artificial intelligence, visual recognition, and edge intelligent computing, intelligent patrol or online monitoring technology based on visual recognition has found important applications in conventional campus security, laboratory safety monitoring, and industrial production operation and maintenance monitoring. Campus security and laboratory safety monitoring aim to protect the personal safety of students and teachers and avoid incidents such as campus bullying or laboratory safety accidents. Industrial production operation and maintenance monitoring is the identification and early warning of hidden dangers and defects in equipment or operational behaviors in industrial scenarios to avoid huge losses caused by faults and hazards. In security and production operation monitoring tasks, using manual methods for real-time detection can be labor-intensive and inefficient, and human negligence that leads to undetected dangers could occur. Therefore, based on the needs of campus security and safety or industrial production operation and maintenance monitoring, this study designs and implements an intelligent patrol system based on microservices and the operation and maintenance monitoring of industrial substations. The system does not require excessive manual participation and can automatically conduct patrols, identify dangers, and provide early warnings. Subsequently, the system adopts an advanced scheduling system that takes only 3–5 min to perform one patrol, considerably improving the efficiency of hazard detection during patrols. The system can be applied to intelligent patrols of campus and industrial substations, security, and safety.
中文核心期刊
