Multimodal ai-driven vocational skill adaptive training platform

Abstract

This invention discloses a multimodal AI-driven adaptive vocational skills training platform, primarily relating to the fields of vocational skills training and adaptive learning. It includes: a multimodal data acquisition unit for collecting multimodal data from trainees; a skills profile construction unit for generating proficiency profiles of trainees across various skills dimensions based on a multimodal fusion model; an adaptive path generation unit for generating personalized training paths based on the skills profiles using reinforcement learning algorithms; a training interaction unit for executing the training paths and collecting interaction data; and a dynamic evaluation and adjustment unit for evaluating learning effectiveness and dynamically adjusting the training paths. The beneficial effects of this invention are: it achieves accurate profiling of trainees' skill levels and real-time dynamic adjustment of training paths, significantly improving the personalization and intelligence of vocational skills training through multimodal data fusion and reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of vocational skills training and adaptive learning technology, specifically a multimodal AI-driven adaptive vocational skills training platform. Background Technology

[0002] Currently, the vocational skills training sector is undergoing a critical transformation from traditional offline models to digital and intelligent approaches. Existing online training platforms primarily rely on pre-set course systems, providing learning resources to students through video recordings, online question banks, and other means. While these platforms overcome time and space limitations, their core logic remains the same: a uniform, one-size-fits-all approach to course delivery, lacking precise identification and dynamic adaptation to individual student differences.

[0003] First, in terms of student ability assessment, existing technologies mostly use single-modal test data, such as periodic test scores or answer accuracy rates. This assessment method is difficult to fully capture the student's true state. Research shows that multimodal data such as students' speech expression, facial expressions, and eye movement trajectories during the learning process contain rich information and can reflect their attention level, emotional state, and depth of understanding. However, how to effectively integrate these heterogeneous data to build accurate student profiles remains a challenge for current technologies. Secondly, in terms of training path planning, mainstream platforms use static knowledge graphs or rule-based recommendation systems. These methods cannot dynamically adjust subsequent content based on students' real-time learning performance. When students encounter difficulties in understanding or show signs of fatigue, the system often continues to proceed according to the original plan, resulting in poor learning outcomes. Although adaptive algorithms such as reinforcement learning have been explored in the field of education, a mature solution has not yet been formed for how to combine multimodal feedback for real-time path optimization. In addition, existing assessment mechanisms are mostly ex post facto, that is, judging learning effectiveness through unit tests or final exams, lacking real-time assessment and intervention during the learning process. This delayed feedback mechanism makes it impossible for students to receive targeted guidance in a timely manner, missing the best opportunity to adjust teaching.

[0004] Therefore, developing an intelligent training platform that can integrate multimodal data, achieve accurate profiling, dynamically plan paths, and evaluate and adjust in real time has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal AI-driven adaptive vocational skills training platform, which realizes accurate profiling of trainees' skill levels and real-time dynamic adjustment of training paths. Through multimodal data fusion and reinforcement learning, it significantly improves the personalization and intelligence of vocational skills training.

[0006] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a multimodal AI-driven adaptive vocational skills training platform, comprising: The multimodal data acquisition unit is used to collect input data from trainees in multiple modalities during the training process; The skill profile construction unit is used to generate a student's skill profile based on the multimodal data through a multimodal fusion model. The skill profile represents the student's current proficiency in each professional skill dimension. An adaptive path generation unit is used to generate a personalized training path based on the skill profile using a reinforcement learning algorithm. The training path consists of a series of learning units. The training interaction unit is used to execute the training path, present learning content to trainees, and collect trainees' interaction data. The dynamic evaluation and adjustment unit is used to evaluate the learners' learning outcomes based on the interactive data and dynamically adjust the training path based on the evaluation results.

[0007] Preferably, the multimodal data includes at least two modalities among text, image, speech, and video; the multimodal fusion model adopts a multimodal fusion network based on an attention mechanism. This network first extracts features of each modality through their respective encoders, then aligns the features of different modalities through a cross-modal attention module, and finally performs weighted fusion of the aligned features through a fusion layer to obtain multimodal fusion features.

[0008] Preferably, the skill profile construction unit includes: The feature extraction subunit is used to extract feature vectors for each modality of data using a pre-trained neural network, wherein the pre-trained neural network includes BERT for text, ResNet for images, and Wav2Vec for speech. Cross-modal alignment subunits are used to map features from different modalities to the same semantic space through contrastive learning, so that features from different modalities with similar semantics are close in distance in space; The fusion subunit is used to weighted fuse the aligned features to obtain a comprehensive feature vector, where the weights are adaptively learned through an attention mechanism. The skill label prediction subunit is used to input the comprehensive feature vector into the multilayer perceptron classifier and output the probability distribution of the learner's proficiency in each preset skill dimension.

[0009] Preferably, the cross-modal alignment subunit is trained using a contrastive learning loss function, the loss function being: ; in, and These represent feature vectors of two different modalities of the same student, and together they form a positive sample pair; The negative sample feature vector representing other students or different semantics; The cosine similarity function; This is a temperature parameter used to control the smoothness of the distribution.

[0010] Preferably, the adaptive path generation unit includes: The knowledge graph storage module is used to store knowledge points in the field of professional skills and their prerequisites and relationships, forming a directed graph structure. The state representation module is used to use the student's skill profile as the current state, where the skill profile is a proficiency vector for each skill dimension; The path planning module is used to generate the optimal learning path based on the current state and knowledge graph using a reinforcement learning algorithm. The action space consists of recommended learning units, and the reward function is set according to the performance evaluation value of the learner after completing the learning unit. The reinforcement learning algorithm employs a deep Q-network, and its Q-value update formula is as follows: ; in, This is the current state. The action to be performed, i.e., the recommended learning unit. For the instant reward, The new state after the action is performed. As a discount factor, This is the learning rate.

[0011] Preferably, the reward function It consists of learning performance evaluation values, specifically represented as follows: ; in, The improvement in skill proficiency after completing this learning unit is provided by the dynamic assessment and adjustment unit; This represents the student's learning progress percentage within this learning unit. This is the ratio of the time spent learning to the expected time. , , These are the weighting coefficients.

[0012] Preferably, the training interaction unit includes: A multimodal interaction interface is used to present learning content in the form of text, images, voice, and virtual reality, and to collect interactive data such as students' voice feedback, facial expressions, eye movements, and gestures in real time. The learning behavior recording module is used to preprocess and store the collected interaction data, including the accuracy of answers, answering time, attention level, and emotional state indicators.

[0013] Preferably, the dynamic evaluation and adjustment unit includes: The learning effectiveness evaluation model is used to calculate learners' mastery of each knowledge point based on their answer records and behavioral data from the interaction data, employing a cognitive diagnostic model: ; in, For students capability parameters For knowledge points Difficulty parameters, For students In the knowledge points The answer results are displayed as 1 for correct and 0 for incorrect; updates are made in real time using maximum likelihood estimation. and This allows us to obtain real-time feedback on students' understanding of the material. The path correction module is used to trigger the adaptive path generation unit to replan the path when a student's mastery of a certain knowledge point is lower than a preset threshold, thereby increasing the review of that knowledge point or the learning of related prerequisite knowledge.

[0014] Preferably, the learning performance evaluation model also incorporates multimodal behavioral data, uses multimodal fusion to estimate learners' engagement, and integrates engagement as a correction factor into the cognitive diagnostic model. ; in, For students The level of attention in the current learning unit is predicted by a neural network using facial expression and eye movement data. For adjustment coefficients, This is the sigmoid function.

[0015] Preferably, it also includes a data preprocessing unit for cleaning, normalizing, and data augmentation operations on the collected multimodal data, specifically including: segmenting text data and removing stop words, normalizing the size of images and videos and performing data augmentation, and denoising and extracting features from speech.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention collects multimodal data such as text, images, voice, and video from learners, and uses a multimodal fusion network based on attention mechanism and contrastive learning alignment technology to deeply explore learners' explicit and implicit learning states, and constructs an accurate profile covering proficiency in various skill dimensions. Compared with traditional single-modal assessment, this invention significantly improves the accuracy and comprehensiveness of learner ability judgment, laying a solid foundation for personalized teaching. 2. Based on knowledge graphs and reinforcement learning algorithms, this invention can intelligently generate the optimal learning path according to the real-time profile of the learner, and can dynamically adjust the subsequent content according to the effect feedback during the learning process. This path planning avoids the inefficient repetition of fixed courses, significantly improves learning efficiency and interest, and realizes the true implementation of personalized teaching. 3. This invention uses a cognitive diagnostic model that incorporates multimodal behavioral data to assess students’ mastery of each knowledge point in real time. It also adjusts the assessment results by combining indicators of engagement such as attention level. Once the mastery is found to be below the threshold, path correction is immediately triggered to push review or prior knowledge, so as to realize the immediate discovery and precise intervention of learning problems and effectively prevent the accumulation of knowledge gaps. 4. This invention supports multiple forms such as text, voice, and virtual reality through a multimodal interactive interface to create an immersive learning experience. At the same time, it collects rich interactive data (such as facial expressions, eye movements, gestures, etc.) to form a complete closed loop from data collection to evaluation and then to path optimization. This closed loop enables the system to have the ability to continuously self-optimize and continuously improve the effectiveness of training strategies. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0018] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0019] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0020] Example: Cardiopulmonary resuscitation (CPR) is a crucial first aid skill that requires operators to master the correct compression positions, depths, and frequencies, as well as artificial respiration techniques. Traditional training typically uses a combination of theoretical instruction and mannequin practice, but it lacks detailed feedback and personalized guidance for trainees' operational processes. This embodiment will take CPR first aid skills training as an example to provide a more detailed explanation of this platform.

[0021] like Figure 1 As shown, this embodiment provides a multimodal AI-driven adaptive vocational skills training platform, including: a multimodal data acquisition unit, a skills profile construction unit, an adaptive path generation unit, a training interaction unit, and a dynamic evaluation and adjustment unit. The platform is equipped with the following hardware devices: Virtual Reality Headset (HTC Vive Pro): Used to present immersive training scenarios, including first aid scene simulations, knowledge explanation videos, etc., and also has a built-in eye-tracking module to collect the trainees' eye movement trajectories; High-definition camera (Logitech Brio): Placed in front of the trainee to capture the trainee's facial expressions, upper body posture, and hand movements; Microphone array: Used to capture students' voice questions, answers, and conversations with virtual tutors; Force Feedback Manikin (Laerdal Resusci Anne QCPR): Built-in pressure sensor and accelerometer to measure in real time the depth, frequency, rebound of the student's compressions, as well as the volume and flow rate of air. Computing server: Equipped with a high-performance GPU (NVIDIA RTX 3090) for running multimodal fusion models, reinforcement learning algorithms, and cognitive diagnostic models.

[0022] After the trainees begin training, the multimodal data acquisition unit continuously collects the following data: Text data: Text answers and notes entered by students in the system, as well as theoretical test questions provided by the system; Image data: Facial expression frames of trainees are captured via camera to analyze emotional states (such as focus, confusion, fatigue); hand posture images are captured to assess operational standardization. Voice data: Recordings of conversations between students and virtual tutors are transcribed into text through speech recognition, and voice features (pitch, speech rate) are extracted to analyze emotions. Video data: First-person perspective videos of trainees recorded by virtual reality headsets, including eye-tracking fixation point overlays, used to analyze the areas of trainees' attention; Sensor data: Simulated human output of compression depth (mm), frequency (times / minute), rebound speed (mm / s), and ventilation volume (ml); All data is timestamped and transmitted synchronously to the data preprocessing unit.

[0023] The data preprocessing unit cleans and standardizes the raw data: Text: Perform word segmentation, stop word removal, and spell correction; Images: Facial images were cropped to 224×224 pixels, and brightness normalization and random horizontal flipping were performed for data augmentation; key point coordinates were extracted from hand images. Speech: Noise reduction processing, extraction of Mel spectrum features, and normalization; Video: Extract keyframes and align them with eye-tracking data; Sensor data: Filtered to remove noise, converted to standard units; The preprocessed data is stored in a temporary buffer for use by subsequent units.

[0024] The skills profile building unit is responsible for transforming multimodal data into trainees’ proficiency in various dimensions of CPR skills. In this embodiment, the CPR skills dimensions include: theoretical mastery, compression accuracy, ventilation accuracy, emergency response speed, and operational standardization. The feature extraction subunit extracts features using a pre-trained model for different modalities. (1) Text: 768-dimensional feature vectors were extracted using the Chinese BERT model; (2) Facial images: 2048-dimensional features were extracted using ResNet-50 and then reduced to 512-dimensionality through a fully connected layer; (3) Speech: Use Wav2Vec 2.0 to extract 1024-dimensional features and reduce the dimensionality to 512-dimensional; (4) Eye-tracking video: Spatiotemporal features were extracted using 3D CNN (C3D) to obtain a 512-dimensional vector; (5) Sensor data: Construct time series data and use LSTM to extract 128-dimensional time series features; To map features from different modalities to the same semantic space, the cross-modal alignment subunit uses contrastive learning for alignment. The specific steps are as follows: (1) Construct positive sample pairs: Different modal features of the same student within the same time period (e.g., a 5-second window) are considered as positive sample pairs (e.g., facial expressions and vocal emotion expression). (2) Construct negative sample pairs: randomly sample features from other trainees or different time periods as negative samples; (3) Training the contrastive loss function This makes the feature vectors of positive sample pairs closer together in the embedding space, and negative sample pairs farther apart; After alignment, the features of all modalities are mapped to a 128-dimensional common space, denoted as . , , , , ; The fusion subunit employs an attention-based fusion layer, first calculating the attention weights for each modality feature: ; ; in, Represents mode, , , , The parameters are then learned, and the fused features are calculated: ; The skill tag prediction subunit will integrate features. The input is a multilayer perceptron containing two hidden layers (256-dimensional and 128-dimensional), and the output layer has 5 neurons, corresponding to 5 skill dimensions. The sigmoid activation function is used to obtain the proficiency of each dimension (between 0 and 1). For example, the output [0.85, 0.62, 0.71, 0.93, 0.58] indicates that the student's theoretical mastery is 0.85, compression accuracy is 0.62, ventilation accuracy is 0.71, emergency response speed is 0.93, and operational standardization is 0.58. The model is trained using a labeled dataset, which is obtained by experts scoring the trainees' performance.

[0025] The adaptive path generation unit dynamically plans the learner's next learning content based on the current skill profile: The knowledge graph storage module constructs a CPR skills knowledge graph, with nodes representing knowledge points such as "cardiac arrest recognition," "chest compression location," "compression depth requirements," "artificial respiration methods," and "AED usage," totaling 50 nodes. Edges represent prerequisite relationships (e.g., "chest compression location" must be mastered before "compression depth requirements" can be learned) and association relationships. The knowledge graph is stored in the Neo4j graph database in the form of a directed graph. The status representation module is used to represent the current status of the student. Student's current status It consists of two parts: a skill profile vector (5-dimensional proficiency) and a learned knowledge point vector (50-dimensional, 0 / 1 indicating whether it has been learned). Therefore, the state space is 55-dimensional. The path planning module uses a Deep Q-Network (DQN) for path planning. The action space consists of 30 recommended learning units, each corresponding to a single knowledge point or a comprehensive practice of a set of knowledge points. The reward function... The design is as follows: ; in: This indicates the improvement in relevant dimensions of the learner's skills profile after completing the learning unit (provided by the dynamic assessment and adjustment unit). If the unit corresponds to multiple knowledge points, the average value is taken. This indicates the percentage of learning activities completed by trainees in this unit, such as the progress of watching videos and the accuracy rate of answering questions, and is statistically analyzed by the training interaction unit. This represents the ratio of actual learning time to expected learning time; a penalty is applied if the ratio exceeds 1. (Weight) , , ; The training process of DQN is as follows: use - Greedy strategy for selecting actions, experience replay pool storage and transfer. The target network is updated every 100 steps. The Q network structure is a three-layer fully connected network (input 55, hidden layers 256 and 128, output 30), and it is trained until convergence. In actual reasoning, regarding the current student state DQN outputs the Q-value of each action and selects the learning unit corresponding to the largest Q-value as the next recommendation.

[0026] The training interaction unit is responsible for presenting learning content and collecting interaction data. Multimodal interaction interfaces are used for: Theoretical teaching: 360° emergency rescue scene videos are played through VR headsets, supplemented by text explanations and voice narration. Students can ask questions by voice, and the system uses speech recognition and natural language understanding to answer them. Skills practice: Trainees perform pressing and breathing operations on a force feedback mannequin. The VR headset displays real-time feedback (such as the pressure depth waveform and prompts such as "Press harder"). The camera captures hand posture, and if the pressing position is found to be off, the system provides visual prompts. Assessment Test: The system randomly selects theoretical questions (answered via text or voice) and practical questions, and records the accuracy rate, reaction time, etc. The learning behavior recording module records the following data in real time: Answer record: Question ID, Student's answer, Correct or incorrect, Time taken to answer; Operation log: Depth, frequency, and rebound of each press; volume of each ventilation; operation timestamp; Behavioral data: eye-tracking gaze heatmap, facial expression classification (using a pre-trained emotion recognition model to output probabilities of focus, confusion, fatigue, etc.), and head movement trajectory; Interaction log: Voice interaction text between students and the system, and system feedback content; All data is stored in a MongoDB database and synchronized to the dynamic evaluation and adjustment unit.

[0027] The dynamic assessment and adjustment unit evaluates students' mastery of each knowledge point in real time and triggers path correction: The learning performance evaluation model uses an IRT (Intuitive Therapy) model to estimate learners' abilities. and the difficulty of knowledge points For each knowledge point Collect students' answer records on questions related to this knowledge point. The model formula is expressed as follows: ; Real-time updates using Bayesian estimation and The initial value is set to 0. After each new question is answered, the parameters are updated using maximum a posteriori estimation. Mastery is defined as... The threshold is set to 0.7; anything below the threshold is considered not mastered. Simultaneously, by combining multimodal behavioral data to estimate learner engagement, an attention level prediction model is trained. The inputs are facial expression features, eye movement features, and speech features, and the output is an attention value. This value is then incorporated as a correction factor into the IRT model: ; in , The sigmoid function is used, and the corrected probability is taken as the student's real-time mastery level. The path correction module is used to calculate the mastery of all knowledge points after each learning unit is completed. If the mastery of a knowledge point is lower than 0.7 and the knowledge point has already been learned, a review path is triggered: the knowledge point and its prerequisite knowledge are added to the learning list and their weight in the reinforcement learning reward is increased. The adaptive path generation unit re-plans the path according to the new status to ensure that students can consolidate their weak points in a timely manner.

[0028] Taking student A as an example, describe the complete process of their first use of the platform: 1. Registration and Initial Assessment: After registration, A will first receive an initial assessment consisting of 10 theoretical questions and a simple practical exercise. The multimodal data acquisition unit will record the answers and operation data, and the skills profile building unit will generate an initial profile (theory 0.4, compression 0.2, ventilation 0.1, reaction 0.5, standardization 0.3). 2. Path planning: Based on the initial profile, the adaptive path generation unit recommends the first unit, "Principles of Cardiac Arrest Recognition and First Aid." A watches a video through VR, and eye-tracking data shows that he focuses on key information points and his facial expression is attentive. 3. Learning and Interaction: After watching, the system conducts a short test. If A answers correctly, the training interaction unit is 100% complete, and the time penalty is 0.8 (for early completion). The dynamic evaluation and adjustment unit updates the mastery level: the theoretical level is improved to 0.55. 4. Route Adjustment: Rewards for the Route Planning Module Update the Q network; the next unit will recommend "chest compression positions". 5. Skill Practice: A practices pressing on a mannequin. The camera captures the position of A's hand and detects that the pressing point is off to the right. The system provides real-time feedback. The sensors record an average pressing depth of 45mm (standard 50-60mm) at a frequency of 110 times / minute. After completing the unit, the skill profile is updated: pressing accuracy 0.35, standard accuracy 0.4. The evaluation model finds that the pressing accuracy is below the threshold, triggering a review. 6. Review and reinforcement: The system redesigned the path and added a "refined practice on pressing position" unit. A practiced again to get more feedback and finally the pressing depth reached 52mm, which is 0.7 standard. 7. Continuous iteration: After repeating this process for 10 units, A's proficiency in all skill dimensions exceeded 0.8, and he passed the graduation assessment.

[0029] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A multimodal AI-driven adaptive vocational skills training platform, characterized in that, include: The multimodal data acquisition unit is used to collect input data from trainees in multiple modalities during the training process; The skill profile construction unit is used to generate a student's skill profile based on the multimodal data through a multimodal fusion model. The skill profile represents the student's current proficiency in each professional skill dimension. An adaptive path generation unit is used to generate a personalized training path based on the skill profile using a reinforcement learning algorithm. The training path consists of a series of learning units. The training interaction unit is used to execute the training path, present learning content to trainees, and collect trainees' interaction data. The dynamic evaluation and adjustment unit is used to evaluate the learners' learning outcomes based on the interactive data and dynamically adjust the training path based on the evaluation results.

2. The multimodal AI-driven adaptive vocational skills training platform according to claim 1, characterized in that, The multimodal data includes at least two modalities among text, image, speech, and video; the multimodal fusion model adopts a multimodal fusion network based on an attention mechanism. This network first extracts features of each modality through their respective encoders, then aligns the features of different modalities through a cross-modal attention module, and finally weights and fuses the aligned features through a fusion layer to obtain multimodal fusion features.

3. The multimodal AI-driven adaptive vocational skills training platform according to claim 2, characterized in that, The skill profile construction unit includes: The feature extraction subunit is used to extract feature vectors for each modality of data using a pre-trained neural network, wherein the pre-trained neural network includes BERT for text, ResNet for images, and Wav2Vec for speech. Cross-modal alignment subunits are used to map features from different modalities to the same semantic space through contrastive learning, so that features from different modalities with similar semantics are close in distance in space; The fusion subunit is used to weighted fuse the aligned features to obtain a comprehensive feature vector, where the weights are adaptively learned through an attention mechanism. The skill label prediction subunit is used to input the comprehensive feature vector into the multilayer perceptron classifier and output the probability distribution of the learner's proficiency in each preset skill dimension.

4. The multimodal AI-driven adaptive vocational skills training platform according to claim 3, characterized in that, The cross-modal alignment subunit is trained using a contrastive learning loss function, which is: ； in, and These represent feature vectors of two different modalities of the same student, and together they form a positive sample pair; The negative sample feature vector representing other students or different semantics; The cosine similarity function; This is a temperature parameter used to control the smoothness of the distribution.

5. The multimodal AI-driven adaptive vocational skills training platform according to claim 1, characterized in that, The adaptive path generation unit includes: The knowledge graph storage module is used to store knowledge points in the field of professional skills and their prerequisites and relationships, forming a directed graph structure. The state representation module is used to use the student's skill profile as the current state, where the skill profile is a proficiency vector for each skill dimension; The path planning module is used to generate the optimal learning path based on the current state and knowledge graph using a reinforcement learning algorithm. The action space consists of recommended learning units, and the reward function is set according to the performance evaluation value of the learner after completing the learning unit. The reinforcement learning algorithm employs a deep Q-network, and its Q-value update formula is as follows: ； in, This is the current state. The action to be performed, i.e., the recommended learning unit. For the instant reward, The new state after the action is performed. As a discount factor, This is the learning rate.

6. The multimodal AI-driven adaptive vocational skills training platform according to claim 5, characterized in that, The reward function It consists of learning performance evaluation values, specifically represented as follows: ； in, The improvement in skill proficiency after completing this learning unit is provided by the dynamic assessment and adjustment unit; This represents the student's learning progress percentage within this learning unit. This is the ratio of the time spent learning to the expected time. , , These are the weighting coefficients.

7. The multimodal AI-driven adaptive vocational skills training platform according to claim 1, characterized in that, The training interaction unit includes: A multimodal interaction interface is used to present learning content in the form of text, images, voice, and virtual reality, and to collect interactive data such as students' voice feedback, facial expressions, eye movements, and gestures in real time. The learning behavior recording module is used to preprocess and store the collected interaction data, including the accuracy of answers, answering time, attention level, and emotional state indicators.

8. The multimodal AI-driven adaptive vocational skills training platform according to claim 1, characterized in that, The dynamic evaluation and adjustment unit includes: The learning effectiveness evaluation model is used to calculate learners' mastery of each knowledge point based on their answer records and behavioral data from the interaction data, employing a cognitive diagnostic model: ； in, For students capability parameters For knowledge points Difficulty parameters, For students In the knowledge points The answer results are displayed as 1 for correct and 0 for incorrect; updates are made in real time using maximum likelihood estimation. and This allows us to obtain real-time feedback on students' understanding of the material. The path correction module is used to trigger the adaptive path generation unit to replan the path when a student's mastery of a certain knowledge point is lower than a preset threshold, thereby increasing the review of that knowledge point or the learning of related prerequisite knowledge.

9. A multimodal AI-driven adaptive vocational skills training platform according to claim 8, characterized in that, The learning effectiveness evaluation model also incorporates multimodal behavioral data, using a multimodal fusion approach to estimate learners' engagement, and integrates engagement as a correction factor into the cognitive diagnostic model. ； in, For students The level of attention in the current learning unit is predicted by a neural network using facial expression and eye movement data. For adjustment coefficients, This is the sigmoid function.

10. A multimodal AI-driven adaptive vocational skills training platform according to claim 1, characterized in that, It also includes a data preprocessing unit for cleaning, normalizing, and data augmentation of the collected multimodal data. Specifically, it includes: segmenting text data and removing stop words, normalizing the size of images and videos and augmenting the data, and denoising and extracting features from speech.

Images