An AI-assisted experimental teaching system

Abstract

The application discloses an AI-assisted experimental teaching system, which organically integrates a teaching process with artificial intelligence technology, constructs a whole-process intelligentized auxiliary teaching of 'pre-class preparation-in-class operation-post-class review', and preliminarily constructs a knowledge framework for students through man-machine intelligent interaction and virtual two-dimensional simulation experiment operation, and completes intelligent preparation; the system realizes auxiliary guidance and error avoidance in the operation process through real-time monitoring and intelligent judgment; the system provides post-class homework correction assistance for teachers through a photographing automatic evaluation system, and reduces single repetitive work; the system provides accurate weakness evaluation and improvement suggestions for students through dynamic review question generation and individualized evaluation reports, and is characterized by AI assistance as a core, full-fledged teaching, teaching individualization, high efficiency and intelligence.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent teaching, and in particular to the technical field of an AI-assisted experimental teaching system. Background Technology

[0002] The application of artificial intelligence in physics experiments is increasingly demonstrating its important value. Through advanced machine learning algorithms, AI can automatically analyze large amounts of experimental data, identify subtle signals, and discover patterns that are difficult to detect by traditional methods, thereby significantly improving data processing efficiency. In experimental design, AI can also optimize parameter settings and provide more targeted experimental solutions. In addition, AI's real-time monitoring and control capabilities enable physical experiments to be dynamically adjusted, improving experimental accuracy.

[0003] Chinese patent application number CN202310155413.6 discloses a digital teaching system. This system includes an application layer for teachers to use in teaching; a data transmission layer for the transmission and sharing of various teaching resources; and a foundation layer for data storage and resource classification throughout the system. This system effectively merges and classifies mathematical teaching resources, enabling digital resource sharing across multiple campuses and improving teaching quality. However, the system still has shortcomings. While it achieves resource sharing, it lacks intelligent editing capabilities based on multimodal analysis, making it unable to dynamically adapt to students' learning situations. Furthermore, the transmission module lacks a hash verification and Q-value comprehensive scoring mechanism, resulting in insufficient reliability of resource transmission.

[0004] Chinese patent application number CN2022111330902 discloses an artificial intelligence adaptive interactive teaching system. It achieves interactive teaching by playing teaching videos with dynamic dialogues, tracks students' learning status in real time, pushes training questions based on students' learning status, and detects students' learning obstacles through training questions. Based on the learning obstacles, it assigns corresponding learning tasks. However, the system still has shortcomings. After assigning learning tasks, it does not analyze and compare the subsequent effects of students' learning tasks, nor does it track and analyze the learning process, thus failing to guarantee students' learning effectiveness.

[0005] Chinese patent application number CN201810991964.5 discloses a virtual reality-based 3D animation assembly teaching method and device, including step one: importing AI models of all parts of the product to be assembled and forming a virtual assembly scene; step two: planning the initial positions of all parts mentioned in step one; step three: establishing a mathematical AI model for each part based on the adaptive algorithm of the running path and running speed of the feature positions; step four: according to steps two and three, forming a 3D assembly teaching animation of all parts mentioned in step one from the initial position to the final position. By constructing a virtual assembly scene and scientifically planning the running path and running speed of the parts, a high-precision AI model is formed for 3D animation assembly teaching, which effectively reduces teaching costs, improves students' learning efficiency, and achieves the effect of quickly becoming familiar with each part and the assembly process. However, the method still has shortcomings, namely, the lack of teaching interactivity, which will have a negative impact on teaching feedback, teaching quality, students' learning mood and enthusiasm.

[0006] In the process of integrating AI into experimental teaching, there are still many problems and challenges, which are manifested as follows: (1) Difficulty in data processing and analysis: The experiment involves the detection of various components, which will generate a large amount of heterogeneous data. AI needs to have efficient data fusion, feature extraction, anomaly detection and multimodal analysis capabilities to support accurate analysis and inference; (2) Cost of data acquisition and labeling: AI training usually requires a large amount of experimental datasets. Data collection, labeling and quality control often require a lot of human resources as support; (3) Impact of real-time guidance on inquiry ability: Although the immediate feedback in the experimental operation guidance improves safety and efficiency, over-reliance may weaken students' active inquiry ability and hinder the formation of a complete experimental thinking chain of "discovery-analysis-solution"; (4) Potential weakening of physical meaning understanding by data processing: If the automated data processing and report generation do not emphasize the principles and process steps, students may not fully understand the underlying physical meaning and ignore key links such as outlier handling and fitting principles. Summary of the Invention

[0007] The purpose of this invention is to solve the problems in the prior art and propose an AI-assisted experimental teaching system that can systematically improve data governance, AI model transparency, evaluation and feedback mechanisms, and alignment with teaching objectives, thereby improving the quality of experimental teaching.

[0008] To achieve the above objectives, this invention proposes an AI-assisted experimental teaching system, comprising: Pre-class preparation guidance module: Construct a dedicated intelligent agent that contains only physical theory knowledge and physical experiment-related knowledge, and is oriented towards knowledge retrieval, reasoning and human-computer interaction. Call the Doubao API to encapsulate it into a mini program. In-class operation guidance module: Construct a real-time image recognition monitoring module for students' experimental process, capture video streams in real time, and perform detection, analysis and judgment process frame by frame based on multi-model inference, spatial relationship calculation and connection legality judgment. When students make incorrect connections during the experiment, timely voice reminders are given. After the students obtain the experimental fitting image, it is judged whether the image is correct. If it is incorrect, the reason for the error is given automatically, and the student's mastery of the experiment is scored. The post-class analysis and consolidation module includes an image upload module and an intelligent question generation module. The image upload module collects student assignments and compares them with correct images in the database to score them. The intelligent question generation module generates different post-class questions based on different assignment submissions and classroom experiments, and scores the students' answers.

[0009] Preferably, the dedicated intelligent agent core module in the pre-class preparation guidance module includes a knowledge base and retrieval module, a reasoning and response generation module, and an interaction module, and its operation process is as follows: Input phase: Users ask questions through text input, a dedicated intelligent agent recognizes the user's intent, and the knowledge base and retrieval module combine the knowledge base content with the retrieval results of the Doubao API to form a preliminary semantic context; Reasoning and Response Stage: The reasoning and response generation module analyzes the search results, derives formulas or explains experimental principles, generates verifiable answers or guiding questions, and includes a description of the chain of evidence. Output phase: The reasoning and response generation module presents the results through images and text output, and provides feedback to the user in text form through the output module, providing learning paths, relevant experimental examples and extended resources; Interaction Phase: The interaction module is based on multiple connection games and uses Pygame as the basic framework to design the movement of simulation experimental components and wire connection interaction. It judges whether the circuit connection is correct through three steps: rule definition, connection diagram construction, and multi-dimensional verification. It designs multi-dimensional status prompts and UI interaction, including real-time status prompts and interaction enhancement. The overall image is rendered in a visual hierarchy.

[0010] As a preferred embodiment, the operation procedure of the in-class operation guidance module is as follows: Data collection and annotation: Shoot videos of the experimental process, extract images from the videos at fixed time intervals as the dataset, annotate the experimental components in the dataset, generate label files, and generate training, validation and test sets; Data preprocessing: Preprocess the experimental data, check for format errors and missing values ​​in the data files, and correct and remove them; normalize different data to the same units and orders of magnitude; convert the cleaned and normalized data into a format suitable for AI model input; AI Model Construction: A convolutional neural network (CNN) from deep learning is used as the core AI model. Residual connections and attention mechanisms are combined to construct the AI ​​model. Multiple convolutional layers are stacked. In each convolutional layer, convolution operations are performed on the input data to extract features at different scales and directions. Padding is used to effectively extract edge features. Max pooling is used to select the maximum value in a local region for data downsampling. Residual blocks are added between the convolutional layers of the AI ​​model. Each residual block contains two convolutional layers and an identity mapping. The output of the residual block is added to the input to form a residual connection. After the convolutional layers of the AI ​​model, a channel attention module is added. Global average pooling is used to compress the feature map into a channel-dimensional vector. Then, it passes through two fully connected layers and a sigmoid activation function to generate attention weights for each channel. The attention weights are multiplied by the original feature map to perform feature map weighting. The input data is flattened into a one-dimensional vector, and then multiple fully connected layers are used for feature extraction and transformation. A softmax activation function is used to map the output to different categories to obtain the probability distribution of each category. AI Model Training: The training set is used for training the AI ​​model, the validation set is used to evaluate the performance of the AI ​​model and select the optimal AI model parameters, the loss function is selected to measure the difference between the AI ​​model output and the true label, the optimizer is selected to update the AI ​​model parameters to minimize the value of the loss function, the training data is input into the AI ​​model in batches, the output and loss function values ​​of the AI ​​model are calculated, the gradient of the AI ​​model parameters is calculated through the backpropagation algorithm, and the AI ​​model parameters are updated using the optimizer. This process is repeated until the loss function value of the AI ​​model on the training set no longer decreases or reaches the preset number of training rounds. In each training round, the loss value and accuracy of the AI ​​model on the training set and validation set are recorded. During the training process, after each batch of data, the loss value and accuracy of the AI ​​model are calculated, all files that are incorrectly judged in the AI ​​model are picked out, and the number of files belonging to the test set is found from the files. The misclassification rate is obtained by dividing the misclassification rate by the total number of files in the test set. AI Model Inference and Parameter Estimation: After the AI ​​model is trained, it is used to perform inference and parameter estimation on new experimental data. For the experimental data to be tested, the same data preprocessing operations as in the training phase are performed first. The preprocessed data is then input into the AI ​​model, which calculates the probability distribution of each category through forward propagation and selects the category with the highest probability as the prediction result of the AI ​​model. Based on the determined categories of the experimental data, parameter estimation is performed for each category. The inference results and parameter estimates of the AI ​​model are then output to the experimenters for judgment and analysis, using a graphical interface and user interaction.

[0011] Preferably, the after-class analysis and consolidation module collects a large number of standard correct images and homework image datasets with different error levels from previous years to construct a convolutional neural network AI model. The dataset is divided into multiple categories according to the scoring criteria, and the convolutional neural network AI model is used to learn and train on the dataset to obtain the basic data support for AI model training.

[0012] Preferably, the knowledge base in the knowledge base and retrieval module is organized in a way that combines structured knowledge and vectorized representation. The structured part covers the key points of physical laws, formula derivation, experimental principles and operation steps; the vectorized part includes semantic retrieval and similarity matching, as well as supporting fast cross-document retrieval and context completion.

[0013] Preferably, the core of the Doubao API includes: generating a relevance ranking of multi-source texts based on query intent, summarizing and extracting key points from the results to form a semantic context that can be directly used for reasoning; aligning, detecting conflicts, and evaluating consistency of knowledge from different sources, supplemented by a brief description of the chain of evidence to improve the interpretability of the answer.

[0014] Preferably, the pre-class preparation guidance module constructs a two-dimensional physical simulation experiment model with controllable object placement and line connection positions, and encapsulates the two-dimensional physical simulation experiment model into a mini-program.

[0015] The beneficial effects of this invention are as follows: This invention constructs a full-process intelligent teaching system that, while breaking through the limitations of time, space, manpower, and efficiency in traditional experimental teaching, organically integrates the teaching process with artificial intelligence technology. It constructs a full-process intelligent assisted teaching system encompassing "pre-class preparation - in-class practice - post-class review." The system guides students to initially build a knowledge framework and complete intelligent pre-class preparation through intelligent human-computer interaction and virtual two-dimensional simulation experiments; it provides auxiliary guidance and error avoidance during the operation process through real-time monitoring and intelligent judgment; it provides teachers with assistance in grading homework through an automatic photo evaluation system, reducing repetitive work; and it provides students with precise weakness assessments and improvement suggestions through dynamic review question generation and personalized evaluation reports. With AI assistance as its core, the system provides comprehensive... This system integrates location-based learning into teaching, promoting personalized, efficient, and intelligent instruction. It employs convolutional neural networks and YOLOv8m to identify errors during experiments, enabling real-time monitoring of the experimental process, extraction of features from volt-ampere characteristic curves, and curve classification. It can intelligently analyze experimental data, accurately identify component types and their potential error types, providing immediate feedback to help experimenters quickly identify and correct problems, reduce trial-and-error costs, and improve experimental efficiency. By introducing computer vision and intelligent recognition technologies, the AI ​​system can automatically extract key information from student work images and perform refined analysis and quantitative evaluation of experimental results and reports. Using a large database of student work, it can objectively and accurately grade assignments. The evaluation and feedback function significantly improves grading efficiency and consistency, reduces teachers' workload, and provides students with immediate and objective feedback on their learning outcomes. By automatically collecting and evaluating students' behavior throughout the entire process of "pre-study-operation-testing," a dynamic digital profile of each student is constructed. This not only accurately identifies each student's weaknesses and cognitive biases in areas such as knowledge comprehension, instrument operation, and data analysis, but also automatically generates highly targeted personalized analysis reports and improvement plans. This provides a scientific basis and data support for personalized teaching assessment. In experimental teaching, AI technology, leveraging its advantage of full-process intervention, provides comprehensive support, specifically in the three core stages of pre-class, in-class, and post-class learning. In the pre-class preparation stage, AI can assist students in learning... Students complete the learning of relevant theoretical knowledge and simulation experiments. Through systematic theoretical knowledge delivery and an interactive virtual simulation environment, students can build an experimental principle framework and clarify the core logic and key operations of volt-ampere characteristic testing. In a virtual scenario without the risk of physical equipment damage, students repeatedly practice the experimental process and anticipate operational difficulties such as device placement and circuit connection, laying the foundation for efficient classroom practice. During the course implementation phase, the real-time intervention function of AI can effectively improve the efficiency of experimental teaching: on the one hand, for experimental operation links, AI can identify operational errors in real time and provide feedback on the cause of errors and correction suggestions, which not only reduces equipment damage and safety risks, but also greatly reduces the time cost for teachers to troubleshoot one by one, allowing teachers to focus more on explaining the principles.On the other hand, AI can automatically process experimental data and generate reports, guiding students to focus on in-depth data analysis and the derivation of experimental conclusions. During the post-class consolidation phase, AI's precise services can enhance teaching effectiveness. Based on the key points of classroom experiments, AI can intelligently push targeted exercises to help students identify and fill knowledge gaps and deepen their understanding. Simultaneously, relying on an automated scoring system, the timely and objective grading of exercises allows students to quickly receive feedback and identify their weaknesses.

[0016] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is a system framework diagram of an AI-assisted experimental teaching system according to the present invention; Figure 2 This is a flowchart of the real-time image recognition process of an AI-assisted experimental teaching system according to the present invention; Figure 3 This is a simulation experiment logic diagram of an AI-assisted experimental teaching system according to the present invention; Figure 4 This is a real-time image recognition logic diagram of an AI-assisted experimental teaching system according to the present invention; Figure 5 This is a diagram illustrating the in-class operation logic of an AI-assisted experimental teaching system according to the present invention. Figure 6 This is a convolutional neural network recommendation logic diagram for an AI-assisted experimental teaching system according to the present invention. Detailed Implementation

[0018] See Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 The present invention provides an AI-assisted experimental teaching system, comprising: Pre-class preparation guidance module: Construct a dedicated intelligent agent that contains only physical theory knowledge and physical experiment-related knowledge, and is oriented towards knowledge retrieval, reasoning and human-computer interaction. Call the Doubao API to encapsulate it into a mini program. In-class operation guidance module: Construct a real-time image recognition monitoring module for students' experimental process, capture video streams in real time, and perform detection, analysis and judgment process frame by frame based on multi-model inference, spatial relationship calculation and connection legality judgment. When students make incorrect connections during the experiment, timely voice reminders are given. After the students obtain the experimental fitting image, it is judged whether the image is correct. If it is incorrect, the reason for the error is given automatically, and the student's mastery of the experiment is scored. The post-class analysis and consolidation module includes an image upload module and an intelligent question generation module. The image upload module collects student assignments and compares them with correct images in the database to score them. The intelligent question generation module generates different post-class questions based on different assignment submissions and classroom experiments, and scores the students' answers.

[0019] 1.1 Pre-class preparation This invention designs an intelligent agent filled with physics experiments and related physics knowledge, and encapsulates it into a mini-program using the Doubao API. This enables intelligent question-and-answer interaction via voice, allowing students to easily learn this invention. Students unfamiliar with an experiment may easily find videos online that don't match the experimental setup if they blindly search for them online. Therefore, the mini-program includes precautions and operation videos for this experiment, and designs a two-dimensional simulation of the experiment. Students can control the placement of items and wiring connections in the two-dimensional physics simulation model for pre-class preparation, allowing them to understand the experimental methods before class. If the wiring or item placement in the two-dimensional physics simulation model is incorrect, the mini-program will automatically report the error and provide correct operation prompts. These features effectively solve the problem of students' difficulty in pre-class preparation. 1.2 In-class Experiment Guidance During the experimental operation, some students may not know how to connect the instruments or set the experimental conditions. Therefore, this invention designs a real-time monitoring system for the students' experimental process. When students make incorrect connections during the experiment, the system will promptly remind them and determine whether the experimental fitting image is correct after the students obtain it. If it is incorrect, the system will automatically give the reason for the error. 1.3 Post-lesson evaluation and consolidation At the end of the course, teachers usually assign drawing assignments to students to test their learning outcomes. This invention adds an image upload function to the mini-program to collect student assignments, compares them with correct images in the collected database, and scores the images. In addition, an intelligent question generation module is added to the mini-program to give different after-class questions based on different assignment submissions and classroom experiments. 2. Experimental Apparatus Regulated power supply, slide wire rheostat, voltmeter, camera, milliammeter, metal film resistor under test, crystal diode, wire and electronic display screen; 3. AI-assisted experimental process 3.1 Pre-study section 3.1.1 Intelligent Question Answering (1) Technical Design 1. System Objectives and Overall Architecture This invention is designed for physics teaching scenarios and constructs a dedicated intelligent agent for knowledge retrieval, reasoning, and human-computer interaction. The agent's knowledge base contains only physics theory knowledge and physics experiment-related knowledge to ensure the controllability and verifiability of the knowledge source. The system adopts a modular architecture, with core modules including a knowledge base and retrieval module, a reasoning and response generation module, and a voice perception / interaction module. By introducing external APIs, it achieves efficient cross-source information integration and natural language voice interaction capabilities. 2. Knowledge Base Design and Representation The knowledge base is organized in a way that combines structured knowledge and vectorized representation. The structured part covers key points such as physical laws, formula derivation, experimental principles and operating procedures, which facilitates rule-based reasoning and procedural explanation. The vectorized part is used for semantic retrieval and similarity matching, supporting fast cross-document retrieval and context completion. Knowledge acquisition channels include textbook-level theoretical entries, standard experimental operating procedures, common misconceptions and typical examples, etc., and the content is version controlled and traceable to ensure repeatability and verifiability. 3. API Introduction To improve the coverage, timeliness, and robustness of content retrieval, the system introduces the Doubao API to enhance the ability to retrieve, integrate, and present knowledge. The core functions of the Doubao API include: Content retrieval and aggregation: Based on query intent, generate relevance ranking of multi-source texts, and summarize and extract key points from the results to form a semantic context that can be directly used for reasoning; Content integration and consistency verification: Alignment, conflict detection, and consistency assessment of knowledge from different sources, supplemented by concise chain of evidence descriptions, to improve the interpretability of the answers; 4. Workflow and Interaction Logic Input phase: Users ask questions in text form. The NLU module identifies the user's intent, and the retrieval module combines the knowledge base content with the retrieval results from the Doubao API to form a preliminary semantic context. Reasoning and Response Stage: The reasoning module analyzes the search results, performs formula derivation or step-by-step explanation of experimental principles when necessary, generates verifiable answers or guiding questions, and includes a description of the chain of evidence. Output phase: The response generation module presents the results through images and text, and provides feedback to the user in text form through the output module; if needed, the system can provide further learning paths, related experimental examples, or extended resources; 3.1.2 Simulation Experiment (1) Design concept The core design features game design, which is incorporated into the simulation experiment. Based on the Pygame framework, two main functions are designed: component drag-and-drop functionality and wire connection status storage. Based on multiple interconnected games, the basic functions of moving components and connecting wires are implemented. Adding circuit judgment function: Based on the obtained circuit connection status, this invention realizes the judgment of whether the circuit is connected correctly through three steps: rule definition, connection diagram construction, and multi-dimensional verification. Visual Experience Enhancement: To make the model easier for users to use, multi-dimensional status prompts and UI interactions were designed, mainly including real-time status prompts and enhanced interactions; layered rendering was also designed for the overall image to ensure correct visual hierarchy; 3.2 In-class Experiment Process 3.2.1 Visual monitoring of the process YOLOv8 is a version of the YOLO (You Only Look Once) series of models that can be used for instance segmentation tasks. The YOLO series of models aims to achieve fast and efficient object detection. Unlike the R-CNN series, it adopts the One-Stage method. Its simple principle is to first divide the input image into a grid, extract image features through N, and each grid is responsible for predicting multiple bounding boxes and their corresponding class probabilities. Then, object detection is directly achieved through forward propagation of the prediction results. Design route The main approach involves using YOLOv8m to capture video in real time, then using the YOLO model to detect components and wires, analyzing the connection relationships between components, determining whether the connections are correct, and uploading the experimental operation results. Data acquisition and annotation: The circuit connection process was filmed, and one image was extracted every 30 frames as the dataset. The circuit components and wires in the dataset were labeled using labelme, and labelme format label files were generated. Subsequently, the labelme2yolo library was used to convert them to YOLO format, and training set, validation set and test set were generated in a 7:2:1 ratio. Model training: Based on the YOLO framework, the "circuit element detection model" and the "wire color detection model" are trained respectively, and the confidence level of each instrument can be obtained. Real-time detection and connection relationship analysis: The system captures video streams in real time through a camera and performs detection, analysis, and judgment processes frame by frame; this is mainly achieved through multi-model reasoning, spatial relationship calculation, and connection legality determination. 3.2.2 Intelligent Diagnosis of Results 3.2.2.1 Design Process (1) Data preprocessing Before training the AI ​​model, the experimental data needs to be preprocessed. The experimental data comes from the current-voltage characteristic curves of current-carrying components. This data is usually stored in text files and contains measured values ​​of current and voltage. In order for the data to be effectively processed by the AI ​​model, the following preprocessing steps are required: 1. Data cleaning: Check data files for formatting errors, missing values, etc., and correct or remove them accordingly; 2. Data Normalization: This invention uses Z-score standardization: ( μ The mean of the training set data. σ To normalize the current and voltage data to the same units and orders of magnitude (for the standard deviation of the training set), the model can better learn the features and patterns in the data. 3. Data transformation: The cleaned and normalized data is transformed into a format suitable for model input. In this invention, the data is transformed into tensor form so that it can be accepted by the deep learning model. (2) Model Construction This invention employs a convolutional neural network (CNN) from deep learning as the core model, combined with residual connections and attention mechanisms, to improve the model's ability to extract features from current-voltage characteristic curve data and the accuracy of curve classification. The model construction process is as follows: 1. Convolutional layer: The convolutional layer is the basic part of the model and is used to extract local features from the data. By stacking multiple convolutional layers, the model can gradually learn the complex features in the current-voltage characteristic curve. In each convolutional layer, the input data is convolved using a convolutional kernel to extract features of different scales and directions. Padding is used to effectively extract edge features and keep the input and output formats the same or similar. 2. Pooling layer: The pooling layer is used to reduce the spatial dimensionality of the data, reduce the computational cost and number of parameters of the model, while retaining the main features of the data. In this model, max pooling is used to downsample the data by selecting the maximum value in a local region. 3. Residual Connections: To alleviate the vanishing and exploding gradient problems during deep network training and improve model training efficiency and performance, residual connections are introduced. Residual connections allow signals in the network to bypass certain layers and propagate directly, thereby enhancing information transmission and model stability. Residual blocks are added between the convolutional layers of the model. Each residual block contains two convolutional layers and an identity mapping. Residual connections are achieved by adding the output of the residual block to the input. Its core mathematical representation is: (1); in: x For input features; F(x H( represents the nonlinear transformation for residual block learning; H( x The final output is the input; this design allows the network to learn only the residual between the input and output. This significantly reduces the optimization difficulty of deep networks; 4. Attention Mechanism: To further enhance the model's attention to key features of the current-voltage characteristic curve, an attention mechanism is introduced. This mechanism automatically learns the importance weights of different parts of the data, enabling the model to focus more on feature regions that play a crucial role in classification and parameter estimation. A channel attention module is added after the model's convolutional layers. The feature map is compressed into a channel-dimensional vector through global average pooling. Then, after passing through two fully connected layers and a sigmoid activation function, attention weights for each channel are generated. These weights are multiplied by the original feature map, which weights the feature map and thus enhance the model's attention to key features. 5. Fully Connected Layer: The fully connected layer is used to integrate and classify the features extracted by the convolutional layer and pooling layer. In the fully connected layer, the input data is flattened into a one-dimensional vector, and then the features are further extracted and transformed through multiple fully connected layers. Finally, the softmax activation function is used to map the output to different categories to obtain the probability distribution of each category. (3) Model Training Model training is the process of optimizing a model using labeled experimental data. The goal is to enable the model to learn the features and patterns in the data, thereby achieving accurate classification of voltage-current characteristics and estimation of related parameters. The training process is as follows: 1. Dataset partitioning: The collected experimental data is divided into training set and validation set. 70% of the data is used as training set and 30% as validation set. The training set is used for model training and the validation set is used to evaluate model performance and select optimal model parameters. 2. Loss Function Definition: A suitable loss function is chosen to measure the difference between the model output and the true label. In this invention, the Cross-Entropy Loss function is used, which effectively measures the difference between the probability distribution predicted by the model and the true distribution, thereby guiding the model's training direction. (2); 3. Optimizer selection: An optimizer is selected to update the model parameters in order to minimize the value of the loss function. In this invention, the Adam optimizer is used. It is an adaptive learning rate optimization algorithm that can automatically adjust the learning rate according to the gradient information of the model, thereby accelerating the convergence speed of the model and improving training efficiency. 4. Training Process: During training, training data is input into the model in batches. The model's output and loss function are calculated. Then, the gradient of the model parameters is calculated using the backpropagation algorithm, and the model parameters are updated using the optimizer. This process is repeated until the model's loss function value on the training set no longer decreases or reaches the preset number of training epochs. In each training epoch, the model's loss value and accuracy on the training and validation sets are recorded to monitor the model's training process and performance changes. During training, after each batch of data, the model's loss value and accuracy are calculated and printed out to observe the model's training status. (4) Model Inference and Parameter Estimation After the model training is completed, the trained model is used to perform inference and parameter estimation on new experimental data. The specific process is as follows: Data preprocessing: For the current-voltage characteristic data of the current-voltage components to be tested, the same data preprocessing operations as in the training phase are first performed, including data cleaning, normalization and transformation, so that it can be accepted by the model; Model inference: The preprocessed data is input into the trained model. The model calculates the probability distribution of each category through forward propagation and selects the category with the highest probability as the model's prediction result, thereby realizing the classification of the current-voltage characteristics of current components. Parameter estimation: Based on the determination of the type of current-voltage characteristic, the relevant parameters are further estimated. For its current-voltage characteristic curve, the parameter estimation method is adopted. For example, for the current-voltage characteristic of diodes, the parameters are estimated according to the characteristics of its forward conduction region. Based on the current and voltage data during forward conduction, parameters such as threshold voltage are calculated. Results output: The inference results and parameter estimates of the model are output to the experimenters so that they can make further judgments and analyses. The output results include the predicted category of the current-voltage characteristic, confidence level, estimated values ​​of relevant parameters, etc. (5) GUI Graphical Programming GUI (Graphics User Interface) is a graphical user interface programming language. This invention can use Python to provide a rich set of widgets to enable user interaction through a graphical interface. 3.2.3 Experimental Procedure During the experiment, the system will provide voice prompts to students who make mistakes, record the errors, and score the students' understanding of the experiment. 1. Instrument selection and placement First, after receiving the instrument, students need to select the instrument to use, place the instrument under the camera, and the program will recognize it. 2. Wire connection If an error occurs while connecting wires, the computer will issue a voice prompt and record the reason for the error, laying the foundation for students' after-class summary and grading; 3.2.4 Experimental Data Measurement and Analysis Unified preprocessing of experimental data Based on the experimental data imported with one click, this invention preprocesses the data. The data volume is based on 220 pairs; for fewer than 220 pairs, linear interpolation is used; for more than 220 pairs, random point removal is performed. After processing, the data is converted into tensors to facilitate model prediction. After model prediction, eighteen categories are output, and further processing is performed based on the output categories: (1) Fixed resistor: For a fixed resistor, the present invention uses a linear fitting method to calculate the slope and calculate the resistance value of the resistor itself. (2) Small light bulb: For the small light bulb, the present invention uses a segmented method to extract relevant data points, thereby calculating the resistance value of the small light bulb at the low, medium and high segments; (3) Light-emitting diode: The threshold voltage of the light-emitting diode is obtained by taking the intercept of the linear extension of the current rising part on the voltage coordinate axis, and then substituting it into the relevant formula to calculate the wavelength of the light-emitting diode. 3.2.5 Performance Indicators (1) Precision and uncertainty This invention calculates the precision and uncertainty using the following formulas:

[0020]

[0021] (3); (4); (5); (6); (7); In this invention, the precision and uncertainty can be directly calculated in the experiment. Since a single experiment is accidental, this invention needs to complete multiple experiments. Different numbers of experiments correspond to different coefficients. This invention only applies to experimental data within ten experiments. 1. Fixed resistor: Measure the current and voltage values ​​of the fixed resistor and calculate the resistance accordingly. After multiple tests, determine the accuracy and uncertainty of R. 2. Light bulb: Since the current-voltage characteristic curve of the light bulb is non-linear, it is known that the resistance of the light bulb is constantly changing. Therefore, this invention takes the resistance corresponding to the voltage at the beginning, middle and end stages for accuracy and uncertainty analysis. 3. Light Emitting Diode (LED): The main invention for LEDs is their wavelength. According to formula (8), their threshold voltage can be obtained, and the accuracy and uncertainty of the wavelength can be calculated: (8); (2) Category identification and response time This invention can currently determine a total of 18 types of circuit connection states. The number of types that can be determined will increase as the dataset grows. Currently, it can determine whether a light bulb is correctly connected, or whether the ammeter, voltmeter, or other circuits are reversed, etc. A detailed explanation follows: 1. Resistance This invention can identify circuit connection problems such as: correct resistor connection, reversed ammeter connection, reversed voltmeter connection, and both reversed ammeter and voltmeter connection. In terms of response time, the timing program designed in this invention has an average resistor response time of 193.33 milliseconds. Including data processing, category judgment, chart generation, and report presentation, the category judgment averages around 40ms. 2. Small light bulb This invention can identify circuit connection problems such as: the light bulb is connected correctly, the ammeter is reversed, the voltmeter is reversed, and both the ammeter and voltmeter are reversed. It is difficult to calculate the resistance of the light bulb three times, so the response time is about 217.4ms. 3. Light Emitting Diode This invention can identify: LED color, and circuit connection problems such as LEDs of different colors being reversed, ammeter being reversed, voltmeter being reversed, and both ammeter and voltmeter being reversed. Because LED calculations are relatively complex, the response time is around 215.1ms. (3) Accuracy To calculate the accuracy, first, all files incorrectly classified in the model are identified. Then, the number of files belonging to the test set is found within that set. Dividing this number by the total number of files in the test set yields the false positive rate, which gives an accuracy of 98.15%. (9); 3.3 Post-lesson assessment 3.3.1 Homework Correction (1) Technical Design This invention uses AI image recognition technology as its core and builds a convolutional neural network (CNN) model based on a dataset of student assignments from previous years to achieve automatic assignment scoring. After users upload assignment images, the model can automatically analyze them and output the scoring results. In practice, a large-scale dataset of assignment images needs to be collected first, covering various standard correct images and assignment images with different error levels, and divided into 6 categories according to the scoring criteria (100 points in category 1, 80 points in category 2, 60 points in category 2, and 0 points in category 1). Then, the CNN model is used to learn and train on the dataset. Currently, nearly 500 valid images have been selected from students' assignments from previous years as the basic data support for model training. (2) Operation demonstration First, you enter the assignment upload interface, which has many requirements for the assignment. After submitting the assignment, the system will automatically score it, and the score will be combined with the score of the intelligent question generation part. 3.3.2 Intelligent Question Generation Analyzing students' pre-study activities, the AI ​​autonomously extracts multiple-choice, fill-in-the-blank, and short-answer questions from the intelligent agent and scores the students' answers.

[0022] This invention proposes an AI-assisted experimental teaching system, encompassing three main parts: experimental preparation, operation, and testing. It ultimately provides an evaluation report on student learning progress. The experimental result testing utilizes a large dataset to overcome the shortcomings of traditional computer-generated plots that rely on outlier data. It autonomously identifies and removes outliers, reducing the impact of errors. Furthermore, it measures the internal resistance of the voltmeter, further improving experimental accuracy. This system helps students better understand and master experimental methods, while also stimulating their interest in applied innovation. It offers significant inspiration for students in designing experiments and applying knowledge to new situations.

[0023] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. An AI-assisted experimental teaching system, characterized in that: include: Pre-class preparation guidance module: Construct a dedicated intelligent agent that contains only physical theory knowledge and physical experiment-related knowledge, and is oriented towards knowledge retrieval, reasoning and human-computer interaction. Call the Doubao API to encapsulate it into a mini program. In-class operation guidance module: Construct a real-time image recognition monitoring module for students' experimental process, capture video streams in real time, and perform detection, analysis and judgment process frame by frame based on multi-model inference, spatial relationship calculation and connection legality judgment. When students make incorrect connections during the experiment, timely voice reminders are given. After the students obtain the experimental fitting image, it is judged whether the image is correct. If it is incorrect, the reason for the error is given automatically, and the student's mastery of the experiment is scored. The post-class analysis and consolidation module includes an image upload module and an intelligent question generation module. The image upload module collects student assignments and compares them with correct images in the database to score them. The intelligent question generation module generates different post-class questions based on different assignment submissions and classroom experiments, and scores the students' answers.

2. The AI-assisted experimental teaching system as described in claim 1, characterized in that: The dedicated intelligent agent core module in the pre-class preparation guidance module includes a knowledge base and retrieval module, a reasoning and response generation module, and an interaction module. Its operation process is as follows: Input phase: Users ask questions through text input, a dedicated intelligent agent recognizes the user's intent, and the knowledge base and retrieval module combine the knowledge base content with the retrieval results of the Doubao API to form a preliminary semantic context; Reasoning and Response Stage: The reasoning and response generation module analyzes the search results, derives formulas or explains experimental principles, generates verifiable answers or guiding questions, and includes a description of the chain of evidence. Output phase: The reasoning and response generation module presents the results through images and text output, and provides feedback to the user in text form through the output module, providing learning paths, relevant experimental examples and extended resources; Interaction Phase: The interaction module is based on multiple connection games and uses Pygame as the basic framework to design the movement of simulation experimental components and wire connection interaction. It judges whether the circuit connection is correct through three steps: rule definition, connection diagram construction, and multi-dimensional verification. It designs multi-dimensional status prompts and UI interaction, including real-time status prompts and interaction enhancement. The overall image is rendered in a visual hierarchy.

3. The AI-assisted experimental teaching system as described in claim 1, characterized in that: The operation procedure of the in-class operation guidance module is as follows: Data collection and annotation: Shoot videos of the experimental process, extract images from the videos at fixed time intervals as the dataset, annotate the experimental components in the dataset, generate label files, and generate training, validation and test sets; Data preprocessing: Preprocess the experimental data, check for format errors and missing values ​​in the data files, and correct and remove them; normalize different data to the same units and orders of magnitude; convert the cleaned and normalized data into a format suitable for AI model input; AI Model Construction: The convolutional neural network in deep learning is used as the core AI model. It is combined with residual connections and attention mechanisms to construct the AI ​​model. Multiple convolutional layers are stacked. In each convolutional layer, the input data is convolved to extract features at different scales and directions. Padding is used to effectively extract edge features. Max pooling is used to downsample the data by selecting the maximum value in a local region. Residual blocks are added between the convolutional layers of the AI ​​model. Each residual block contains two convolutional layers and an identity mapping. The output of the residual block is added to the input to form a residual concatenation. After the convolutional layers of the AI ​​model, a channel attention module is added. The feature map is compressed into a channel-dimensional vector through global average pooling. Then, it is passed through two fully connected layers and a sigmoid activation function to generate attention weights for each channel. The attention weights are multiplied with the original feature map to perform feature map weighting. The input data is flattened into a one-dimensional vector, and then features are extracted and transformed through multiple fully connected layers. The softmax activation function is used to map the output to different categories to obtain the probability distribution of each category. AI Model Training: The training set is used for training the AI ​​model, the validation set is used to evaluate the performance of the AI ​​model and select the optimal AI model parameters, the loss function is selected to measure the difference between the AI ​​model output and the true label, the optimizer is selected to update the AI ​​model parameters to minimize the value of the loss function, the training data is input into the AI ​​model in batches, the output and loss function values ​​of the AI ​​model are calculated, the gradient of the AI ​​model parameters is calculated through the backpropagation algorithm, and the AI ​​model parameters are updated using the optimizer. This process is repeated until the loss function value of the AI ​​model on the training set no longer decreases or reaches the preset number of training rounds. In each training round, the loss value and accuracy of the AI ​​model on the training set and validation set are recorded. During the training process, after each batch of data, the loss value and accuracy of the AI ​​model are calculated, all files that are incorrectly judged in the AI ​​model are picked out, and the number of files belonging to the test set is found from the files. The misclassification rate is obtained by dividing the misclassification rate by the total number of files in the test set. AI Model Inference and Parameter Estimation: After the AI ​​model is trained, it is used to perform inference and parameter estimation on new experimental data. For the experimental data to be tested, the same data preprocessing operations as in the training phase are performed first. The preprocessed data is then input into the AI ​​model, which calculates the probability distribution of each category through forward propagation and selects the category with the highest probability as the prediction result of the AI ​​model. Based on the determined categories of the experimental data, parameter estimation is performed for each category. The inference results and parameter estimates of the AI ​​model are then output to the experimenters for judgment and analysis, using a graphical interface and user interaction.

4. The AI-assisted experimental teaching system as described in claim 1, characterized in that: The post-class analysis and consolidation module collects a large number of standard correct images and homework image datasets with different error levels from previous years to construct a convolutional neural network AI model. The dataset is divided into multiple categories according to the scoring criteria, and the convolutional neural network AI model is used to learn and train on the dataset to obtain the basic data support for AI model training.

5. The AI-assisted experimental teaching system as described in claim 2, characterized in that: The knowledge base in the knowledge base and retrieval module is organized in a way that combines structured knowledge and vectorized representation. The structured part covers the key points of physical laws, formula derivation, experimental principles and operation steps; the vectorized part includes semantic retrieval and similarity matching, as well as supporting fast cross-document retrieval and context completion.

6. The AI-assisted experimental teaching system as described in claim 1, characterized in that: The core of the Doubao API includes: generating a relevance ranking of multi-source text based on query intent, summarizing and extracting key points from the results to form a semantic context that can be directly used for reasoning; aligning, conflict detection, and consistency assessment of knowledge from different sources, supplemented by a brief description of the chain of evidence to improve the interpretability of the answer.

7. The AI-assisted experimental teaching system as described in claim 1, characterized in that: The pre-class preparation guidance module constructs a two-dimensional physical simulation experiment model with controllable object placement and circuit connection positions, and encapsulates the two-dimensional physical simulation experiment model into a mini-program.

Images