An AI education practical examination method and system based on a cloud platform

Abstract

This invention discloses an AI education training and assessment method and system based on a cloud platform, relating to the field of artificial intelligence education technology. It includes: S1. Constructing a compliance essence model based on industry operating standards, wherein the compliance essence model includes multiple interconnected compliance nodes, each corresponding to action elements and speech / semantic elements, and establishing a causal binding relationship between the action elements and speech / semantic elements; S2. Configuring causal verification rules for the compliance nodes, and synchronizing the compliance essence model and causal verification rules to the training execution terminal. This invention, by constructing a compliance essence model containing the causal binding relationship between action elements and speech / semantic elements, and matching it with causal verification rules for corresponding compliance nodes, completes the training assessment judgment based on the business logic of industry operating standards, solving the problem of misjudgment of practical behaviors that conform to industry operating standards but differ from preset standard samples in timing or form.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence education technology, specifically to an AI education training and assessment method and system based on a cloud platform. Background Technology

[0002] AI-based education and training assessment is a core component of the vocational skills training system, bridging theoretical teaching and practical application. The cloud-based AI training and assessment model enables cross-regional sharing of training resources, unified management of assessment processes, and centralized analysis of practical data. It provides digital support for the standardization of vocational education and personalized teaching, and is a core application direction in the current digital upgrade of vocational education.

[0003] The invention patent with publication number CN120431512A discloses a multimodal intelligent detection method for practical training behavior. This solution completes the determination of the completion degree of practical training actions and the verification of audio compliance by directional triggering of audio and video devices and separate recognition of multimodal data, providing a feasible technical implementation path for intelligent detection of the practical training process.

[0004] However, in existing technologies, the AI ​​judgment logic for practical training assessments all use preset standard action sequences and standard terminology as the basis for judgment. It can only achieve similarity matching between action images and standard samples, and similarity comparison between audio semantics and standard terminology. It cannot establish a causal relationship verification between the execution logic of practical actions and the business logic of voice expression, nor can it dynamically adjust the reasonable boundaries of judgment based on the compliance nature of practical behavior. As a result, practical behaviors that comply with industry operating standards but have temporal or formal differences from preset standard samples are judged by the system as unqualified for assessment. It cannot achieve accurate judgment of practical training assessment based on the essence of business compliance, nor can it complete the essential optimization of judgment logic and synchronous implementation across all terminals through cloud platforms. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based education and training assessment method and system based on a cloud platform, which solves the problems of high misjudgment rate, poor adaptability of compliance judgment, and inconsistent assessment standards across terminals in existing training assessments.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI education and training assessment method based on a cloud platform, comprising: S1. Construct a compliance essence model based on industry operating standards. The compliance essence model includes multiple interconnected compliance nodes. Each compliance node corresponds to action elements and voice semantic elements. Establish a causal binding relationship between the action elements and the voice semantic elements. S2. Configure causal verification rules for the compliance node, and synchronize the compliance essence model and causal verification rules to the training execution terminal; S3. Collect audio and video data during the training process at the training execution terminal, align the time axis of the action screen sequence and the audio semantic sequence, and form a multimodal training data stream; S4. Extract action elements and speech semantic elements from the multimodal training data stream, and mark the temporal nodes corresponding to the action elements and speech semantic elements; S5. Verify the causal binding relationship between action elements and speech semantic elements according to the causal verification rules, complete the matching judgment of the compliance node and output the assessment result; S6. The training data is sent back to the cloud management platform. Based on the practical samples, the judgment boundary of the compliance node is optimized and the adaptation range of the causal verification rule is adjusted to complete the synchronous update of the training execution terminal.

[0007] Preferably, the construction of the compliance essence model includes: The industry operating procedures are broken down into multiple sub-tasks arranged in chronological order according to the operating process, with each sub-task corresponding to a compliance node; Define standard action feature vectors and standard speech semantic feature vectors for each compliant node, and establish the conditional triggering relationship between the two during the execution of the standard operation process. The conditional triggering relationship includes action-guided speech relationship, speech-accompanied action relationship, and action-feedback speech relationship.

[0008] Preferably, the timeline alignment of the completed action scene sequence with the audio semantic sequence includes: The video stream is subjected to frame rate normalization processing, and the audio stream is subjected to sampling rate resampling processing. The time offset between the video frame and the audio sampling block is calculated using a preset start synchronization pulse signal. The time offset is compensated by an interpolation algorithm to complete the one-to-one matching of the timestamps of the action frame and the corresponding audio segment.

[0009] Preferably, the extracted action elements include: The coordinates of key points of the human skeleton and the trajectory of hand movements in the training video are extracted by a convolutional neural network. The spatial displacement features of the operation object are identified, and the coordinates, trajectory and displacement features are encapsulated into motion feature vectors. The start and end timestamps of the motion feature vectors in the multimodal training data stream are recorded.

[0010] Preferably, the extraction of speech semantic elements includes: Automatic speech recognition technology is used to convert audio signals into text sequences. Natural language processing algorithms are used to extract business terminology keywords, tone pause features, and logical connectors from the text to construct speech semantic feature vectors. The audio duration intervals corresponding to the text sequences in the multimodal training data stream are recorded.

[0011] Preferably, the causal binding relationship between the verification action element and the speech semantic element includes: Retrieve the temporal nodes corresponding to the action elements and the speech semantic elements, and determine whether the temporal overlap and temporal order of the two meet the temporal threshold requirements in the causal verification rules. Determine whether the behavioral direction of the action element and the logical word direction in the speech semantic element are consistent with the standard operation process logic. If they meet the time sequence threshold requirements and are consistent, then the causal binding relationship is determined to be established.

[0012] Preferably, the determination boundary for optimizing the compliance node includes: Cluster analysis is performed on the returned practical samples that are deemed qualified. The deviations in action amplitude and voice expression of different practitioners when completing the same compliance node are extracted. The statistical distribution law of the deviations and differences is calculated. The similarity threshold and time tolerance value in the causal verification rule are adjusted to expand the judgment boundary to be compatible with practical behaviors that are compliant with the standard but have diverse forms.

[0013] Preferably, after outputting the assessment results, the following is also included: Based on the compliance nodes that fail the judgment, the corresponding normative guidance information is retrieved from the compliance essence model, and voice prompts and action demonstrations are displayed on the training execution terminal interface to guide trainees to perform corrective operations. The corrected practical data is recorded as a reference sample for the boundary optimization module.

[0014] Preferably, the synchronous update includes: When the cloud-based management platform detects a change in the causality verification rules, it generates an incremental update package and sends an update command to the local training execution terminal that is in a connected state through a two-way communication link. The training execution terminal silently downloads the incremental update package in the background and completes the parameter overwrite of the local operating environment during the idle period after the current training task ends.

[0015] This invention also provides an AI-based education, training, and assessment system based on a cloud platform, comprising: The cloud-based management platform includes a compliance essence modeling module, a causal rule configuration module, a boundary dynamic optimization module, a full data management module, and a synchronous distribution module; The compliance essence modeling module is used to construct a compliance essence model based on industry operating standards. The causality rule configuration module is used to configure causality verification rules for compliant nodes; The boundary dynamic optimization module is used to optimize the judgment boundary of compliance nodes based on the training data returned by the training execution terminal. The full data management module is used to store model and practical data; The synchronous distribution module is used to synchronize the updated model and rules to the training execution terminal in real time. The local training execution terminal cluster is connected to the cloud management and control platform, including a terminal adaptation module, a multimodal time series alignment module, a causal feature extraction module, a causal correlation verification module, a local assessment execution module, and a data feedback module. The terminal adaptation module is used to receive the distributed models and rules and update the model and rule parameters of the local operating environment; The multimodal temporal alignment module is used to generate multimodal training data streams; The causal feature extraction module is used to extract action elements, speech semantic elements, and temporal nodes; The causal correlation verification module is used to perform matching and determination based on causal verification rules; The local assessment execution module is used to output assessment results and guide practical training. The data feedback module is used to send the training data back to the cloud management platform.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention constructs a compliance essence model that includes the causal binding relationship between action elements and speech semantic elements, and provides corresponding causal verification rules for compliance nodes. It completes the practical training assessment judgment based on the business logic of industry operation standards, solving the problem of misjudgment of practical behaviors that comply with industry operation standards but have temporal or formal differences from preset standard samples. At the same time, it dynamically optimizes the boundary of compliance node judgment based on qualified practical samples returned by the training terminal, adjusts the adaptation range of causal verification rules, and is compatible with diverse compliant practical behaviors. Then, through the cloud incremental update mechanism, it completes the synchronous implementation of rules across all training terminals, realizes the unified iteration of assessment standards, ensures the consistency of judgment in cross-regional practical training assessments, adapts to the diverse practical scenarios of vocational skills training, and provides a stable and accurate judgment scheme for digital practical training assessment in vocational education. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 Flowchart for constructing the compliance essence model of this invention; Figure 3 This is a flowchart of the multimodal temporal alignment and feature extraction process of the present invention; Figure 4This is a system structure diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-3 This embodiment, combined with the application scenario of AI education training and assessment in vocational education that requires standardized operation and compliant voice expression, provides a detailed description of an AI education training and assessment method based on a cloud platform. The specific implementation includes the following steps: S1. Construct a compliance essence model based on industry operating standards. The compliance essence model includes multiple interconnected compliance nodes. Each compliance node corresponds to action elements and voice semantic elements. Establish a causal binding relationship between the action elements and the voice semantic elements. S2. Configure causal verification rules for the compliance node, and synchronize the compliance essence model and causal verification rules to the training execution terminal; S3. Collect audio and video data during the training process at the training execution terminal, align the time axis of the action screen sequence and the audio semantic sequence, and form a multimodal training data stream; S4. Extract action elements and speech semantic elements from the multimodal training data stream, and mark the temporal nodes corresponding to the action elements and speech semantic elements; S5. Verify the causal binding relationship between action elements and speech semantic elements according to the causal verification rules, complete the matching judgment of the compliance node and output the assessment result; S6. The training data is sent back to the cloud management platform. Based on the practical samples, the judgment boundary of the compliance node is optimized and the adaptation range of the causal verification rule is adjusted to complete the synchronous update of the training execution terminal.

[0020] In this embodiment, the method is based on a cloud-edge collaborative architecture. The cloud is responsible for model building, rule configuration and dynamic optimization, while the terminal is responsible for local training data collection, real-time verification and assessment execution. This avoids the network latency problem of pure cloud solutions. At the same time, by verifying the causal binding of actions and voice, it replaces the independent similarity comparison mode in the existing technology. It completes the assessment judgment based on the business essence of compliant operation, which solves the problem of misjudgment of practical behaviors that are compliant with the specifications but have different forms in the existing technology. At the same time, through dynamic optimization in the cloud and synchronization with the terminal, it realizes the unified iteration of assessment standards and full terminal implementation.

[0021] In one specific embodiment, the construction of the compliance essence model includes: The industry operating procedures are broken down into multiple sub-tasks arranged in chronological order according to the operating process, with each sub-task corresponding to a compliance node; Define standard action feature vectors and standard speech semantic feature vectors for each compliant node, and establish the conditional triggering relationship between the two during the execution of the standard operation process. The conditional triggering relationship includes action-guided speech relationship, speech-accompanied action relationship, and action-feedback speech relationship.

[0022] Specifically, the industry operation standards adopt the practical assessment standards for the corresponding job types in the national vocational skills standards. During the decomposition process, the irreversible links of the operation process are used as the basis for decomposition to ensure that each sub-task has independent business compliance judgment value. Each compliance node corresponds to a unique node number and time priority to avoid judgment failure caused by process disorder.

[0023] In this embodiment, the standard action feature vector Corresponding to the One compliant node, expressed as: ; in, The motion feature dimension includes the relative coordinates of key points on the human skeleton, the curvature of the hand movement trajectory, the displacement of the manipulated object, and the execution duration of the motion. The values ​​for each dimension are derived from the quantitative features of standard operating actions specified in the national vocational skills standards. All dimensions in the vector are normalized, and the value range is [range missing]. This ensures the computability of features across different dimensions.

[0024] Standard speech semantic feature vector Corresponding to the One compliant node, expressed as: ; in, The speech semantic feature dimension includes business term keyword matching degree, tone and pause feature matching degree, logical connector compliance, and speech duration. Each dimension is also normalized, and the value range is [value range missing]. .

[0025] For the For each compliant node, the causal binding relationship between action elements and speech semantic elements satisfies: ; or ; in, In the action elements Under the triggering conditions, the corresponding speech and semantic elements The conditional probability of occurrence In terms of speech and semantic elements Under the triggered conditions, the corresponding action element The conditional probability of occurrence The minimum conditional probability threshold is set in advance, and its value is determined based on the triggering logic of the standard process in the industry's operating specifications. For example, it can be set to 0.85 to ensure a strong correlation between causality.

[0026] The three trigger relationships correspond to different practical scenarios. The action-guided voice relationship is suitable for scenarios where key points of the operation need to be verbally stated before the operation. The voice-accompanied action relationship is suitable for scenarios where the operation content needs to be broadcast synchronously during the operation. The action-feedback voice relationship is suitable for scenarios where the operation result and confirmation matters need to be verbally stated after the operation is completed. It covers most of the practical assessment requirements in vocational skills training, ensuring that the compliance essence model can fit the compliance logic of actual business, rather than simply an independent matching of actions and voice.

[0027] In one specific embodiment, the timeline alignment of the completed action scene sequence with the audio semantic sequence includes: The video stream is subjected to frame rate normalization processing, and the audio stream is subjected to sampling rate resampling processing. The time offset between the video frame and the audio sampling block is calculated using a preset start synchronization pulse signal. The time offset is compensated by an interpolation algorithm to complete the one-to-one matching of the timestamps of the action frame and the corresponding audio segment.

[0028] Specifically, the video stream's frame rate is standardized to 25 frames per second to ensure a unified time reference for video data from different acquisition devices. The audio stream's sampling rate is resampled to 48kHz, with mono quantization to eliminate sampling differences between different audio acquisition devices. The initial synchronization pulse signal is the visual and audio prompts output synchronously by the training execution terminal before the training begins. The visual prompts are full-screen solid color transitions, and the audio prompts are fixed-frequency monotone pulses. Both are triggered synchronously, serving as the reference zero point for audio-video time axis alignment.

[0029] Time offset The calculation expression is: ; in, The timestamp for detecting a synchronization pulse signal in the audio stream. The timestamp for the detected synchronization pulse in the video stream. A positive value indicates that the audio signal is ahead of the video image, while a negative value indicates that the audio signal is behind the video image.

[0030] The interpolation algorithm uses linear interpolation to compensate the time axis of the audio stream, and the timestamps of the compensated audio sample points are... The calculation expression is: ; in, The original timestamps of the audio sampling points before compensation are used. For the video stream, a frame interpolation algorithm is used to calibrate the timestamps of the frame rate-normalized frames, ensuring that the timestamp of each video frame perfectly matches the timestamp of the corresponding audio segment, with a matching error not exceeding 40ms. This meets the synchronous perception threshold of human ear and vision, avoiding the failure of causality verification due to audio-video timing misalignment. After timeline alignment, the resulting multimodal training data stream has a unified time reference, providing a reliable foundation for subsequent timing node matching of actions and speech elements.

[0031] In one specific embodiment, the extraction of action elements includes: The coordinates of key points of the human skeleton and the trajectory of hand movements in the training video are extracted by a convolutional neural network. The spatial displacement features of the operation object are identified, and the coordinates, trajectory and displacement features are encapsulated into motion feature vectors. The start and end timestamps of the motion feature vectors in the multimodal training data stream are recorded.

[0032] Specifically, the convolutional neural network employs a pre-trained OpenPose network model, optimized through transfer learning for the practical training scenario. During optimization, real-world action samples from the corresponding job type are used as the training set to ensure the accuracy of skeletal keypoint extraction matches the needs of the training scenario. Human skeletal keypoints include trunk keypoints such as the shoulder, elbow, wrist, hip, knee, and ankle, as well as hand keypoints such as finger joints and metacarpophalangeal joints, covering all core action execution parts in the practical movements. The hand movement trajectory is generated by fitting the coordinate sequence of hand keypoints in consecutive frames, and trajectory features include trajectory length, curvature change, movement speed, and acceleration.

[0033] The spatial displacement features of the manipulated object are identified using a target detection algorithm to determine the center coordinates of the bounding box. The displacement amount and direction of the center coordinates are calculated across consecutive frames to distinguish between valid and invalid limb movements. After extraction, all action features are normalized and encapsulated into action feature vectors with the same dimensions as the standard action feature vectors in the compliance model. Simultaneously, the start and end timestamps corresponding to these action feature vectors are recorded. The timestamp accuracy is consistent with the time base of the multimodal training data stream, providing precise timing data for subsequent time-series verification.

[0034] In one specific embodiment, the extraction of speech semantic elements includes: Automatic speech recognition technology is used to convert audio signals into text sequences. Natural language processing algorithms are used to extract business terminology keywords, tone pause features, and logical connectors from the text to construct speech semantic feature vectors. The audio duration intervals corresponding to the text sequences in the multimodal training data stream are recorded.

[0035] Specifically, the automatic speech recognition technology employs an ASR model optimized for specialized training scenarios such as industrial and medical settings. During model training, a corpus of professional business terminology for the corresponding job is incorporated to reduce the error rate in recognizing specialized terms. Before converting the audio signal into a text sequence, a noise reduction algorithm is used to preprocess the audio signal, filtering out background noise in the training environment and improving the accuracy of speech recognition.

[0036] The natural language processing algorithm employs a pre-trained BERT model, fine-tuned to meet the semantic recognition requirements of practical training assessments. During fine-tuning, a corpus of standard operational terms is used as the training set, enabling accurate extraction of business terminology keywords from the text, assessment of keyword matching and completeness, and identification of intonation and pause features, including pause location and duration, to determine compliance with standard operational terminology. Logical connectors are also extracted to verify the logical order of expressions within the industry. After extraction, all semantic features are normalized to construct a speech semantic feature vector with dimensions consistent with the standard speech semantic feature vector in the compliance essence model. The audio duration interval corresponding to the text sequence, i.e., the start and end timestamps, is recorded to maintain consistency with the time base of the multimodal training data stream, ensuring direct comparison with the temporal nodes of action elements.

[0037] In one specific embodiment, the causal binding relationship between the verification action element and the speech semantic element includes: Retrieve the temporal nodes corresponding to the action elements and the speech semantic elements, and determine whether the temporal overlap and temporal order of the two meet the temporal threshold requirements in the causal verification rules. Determine whether the behavioral direction of the action element and the logical word direction in the speech semantic element are consistent with the standard operation process logic. If they meet the time sequence threshold requirements and are consistent, then the causal binding relationship is determined to be established.

[0038] Specifically, the temporal nodes include the start and end timestamps of action elements and speech semantic elements. The temporal overlap is used to determine the degree of temporal matching between action and speech. The temporal sequence is used to determine whether it conforms to the triggering logic of the corresponding causal binding relationship. For example, the action-guided speech relationship requires the start timestamp of the speech to be earlier than the start timestamp of the action. The speech-accompanied action relationship requires the temporal overlap between the action and the speech to reach a preset threshold. The action-feedback speech relationship requires the start timestamp of the speech to be later than the end timestamp of the action.

[0039] Temporal overlap The calculation expression is: ; in, This refers to the duration of the intersection between the time intervals of action elements and the time intervals of speech and semantic elements. The duration is the union of the time intervals of action elements and the time intervals of speech and semantic elements. The range of values ​​is The closer the value is to 1, the higher the temporal overlap.

[0040] The timing thresholds are set according to different causal binding relationship types. For example, the timing overlap threshold for voice-accompanied action relationship can be set to 0.75, the maximum duration threshold for voice lag behind action in action-guided voice relationship can be set to 3s, and the maximum duration threshold for voice lag behind action end in action feedback voice relationship can be set to 5s. The threshold settings are determined based on the practical assessment specifications for the corresponding job.

[0041] The consistency judgment between the behavioral orientation of action elements and the logical word orientation of speech semantic elements is achieved by matching the operation behavior type corresponding to the action feature vector with the business operation content corresponding to the speech semantic feature vector. This determines whether they correspond to the standard operation content of the same compliance node. For example, if the action element is a device power-off operation and the speech semantic element is a terminology related to device power-off, then they are considered consistent. Conversely, if the action element is a device power-off operation and the speech semantic element is a statement related to device power-on, then they are considered inconsistent. Only when both the time sequence check and the orientation consistency check pass is the causal binding relationship of the compliance node considered valid, and the node passes the assessment. Otherwise, it is considered unsuccessful. This approach completes the compliance judgment from the essence of business logic, avoiding the misjudgment problem caused by simply comparing similarity in existing technologies.

[0042] In one specific embodiment, the optimization of the determination boundary of the compliance node includes: Cluster analysis is performed on the returned practical samples that are deemed qualified. The deviations in action amplitude and voice expression of different practitioners when completing the same compliance node are extracted. The statistical distribution law of the deviations and differences is calculated. The similarity threshold and time tolerance value in the causal verification rule are adjusted to expand the judgment boundary to be compatible with practical behaviors that are compliant with the standard but have diverse forms.

[0043] Specifically, the cluster analysis employs the K-Means clustering algorithm, clustering the action feature vectors and speech semantic feature vectors of qualified practical samples under the same compliance node. The number of clusters is determined based on the distribution characteristics of the samples, ensuring that samples within the same cluster all meet the requirements of industry operating standards. Action amplitude deviation includes deviations in action execution speed, action amplitude, and operation trajectory; speech expression differences include differences in word order of terminology, differences in the duration of pauses, and differences in the synonymous expression of professional terms.

[0044] The statistical distribution pattern is fitted using a normal distribution, and the mean of the deviation and difference is calculated. with standard deviation The adjusted similarity threshold and time tolerance value cover The distribution range is defined to ensure that the vast majority of compliant operational behaviors are included in the pass / fail judgment range, while excluding non-compliant operational behaviors that fall outside this range. After the adjustment, the adaptation range of the causality verification rules is updated simultaneously to ensure that the judgment boundary can adapt to the operating habits of different operators, while not deviating from the essential requirements of industry operating standards. This solves the problem in existing technologies where the judgment boundary is fixed and cannot accommodate compliant but diverse operational behaviors.

[0045] In one specific embodiment, the output of the assessment results further includes: Based on the compliance nodes that fail the judgment, the corresponding normative guidance information is retrieved from the compliance essence model, and voice prompts and action demonstrations are displayed on the training execution terminal interface to guide trainees to perform corrective operations. The corrected practical data is recorded as a reference sample for the boundary optimization module.

[0046] Specifically, the guidelines include the standard operating procedures, action specifications, standard terminology, and common error prompts for the compliance node. The voice prompts use standardized human voices to clearly state the errors and key points for correction. The action demonstrations use 3D animations or standard practical videos to show the correct actions step by step. Trainees can control the playback progress of the demonstrations through the terminal interface to facilitate comparison and correction.

[0047] During the correction process, the training terminal synchronously collects audio and video data, repeatedly executes the processes of timing alignment, feature extraction, and causal verification, and completes the compliance judgment of the correction operation. If the correction operation is judged to be qualified, the practical data is recorded as a reference sample for the boundary optimization module to further enrich the coverage of the compliance sample. If the correction operation still fails, the corresponding standard guidance information is pushed again to guide the trainees to complete the second correction. At the same time, the operation data of multiple corrections are recorded for subsequent personalized teaching analysis to help trainees improve their practical skills in a targeted manner.

[0048] In one specific embodiment, the synchronous update includes: When the cloud-based management platform detects a change in the causality verification rules, it generates an incremental update package and sends an update command to the local training execution terminal that is in a connected state through a two-way communication link. The training execution terminal silently downloads the incremental update package in the background and completes the parameter overwrite of the local operating environment during the idle period after the current training task ends.

[0049] Specifically, the incremental update package only contains the changed causal verification rule parameters and the decision boundary parameters of compliant nodes, without including the complete model file, significantly reducing the size of the update package and the bandwidth and time required for terminal download. The bidirectional communication link uses 5G or wired Ethernet communication, supporting the TCP / IP transmission protocol to ensure the stability and security of data transmission. The update package is encrypted during transmission to prevent parameter tampering.

[0050] After receiving an update command, the training terminal silently downloads the incremental update package in the background, without affecting the currently executing training and assessment task, thus avoiding training interruption caused by the update operation. Once the current training task is completed and the terminal is idle, it automatically overwrites the parameters of the local operating environment. After overwriting, it automatically verifies the integrity and validity of the parameters. If the verification passes, the updated parameters are enabled; if the verification fails, it automatically rolls back to the original parameters, ensuring that the terminal's training and assessment functions are not disabled due to update failure. Through incremental updates and silent downloads, the assessment standards of the cloud management platform and all training terminals are synchronized, ensuring a unified compliance judgment standard for cross-regional training and assessment execution, while not affecting the normal use of the terminals.

[0051] Please see Figure 4 This embodiment also provides a cloud-based AI education and training assessment system to implement the above-mentioned cloud-based AI education and training assessment method. The system includes: The cloud-based management platform includes a compliance essence modeling module, a causal rule configuration module, a boundary dynamic optimization module, a full data management module, and a synchronous distribution module; The compliance essence modeling module is used to construct a compliance essence model based on industry operating standards. The causality rule configuration module is used to configure causality verification rules for compliant nodes; The boundary dynamic optimization module is used to optimize the judgment boundary of compliance nodes based on the training data returned by the training execution terminal. The full data management module is used to store model and practical data; The synchronous distribution module is used to synchronize the updated model and rules to the training execution terminal in real time. The local training execution terminal cluster is connected to the cloud management and control platform, including a terminal adaptation module, a multimodal time series alignment module, a causal feature extraction module, a causal correlation verification module, a local assessment execution module, and a data feedback module. The terminal adaptation module is used to receive the distributed models and rules and update the model and rule parameters of the local operating environment; The multimodal temporal alignment module is used to generate multimodal training data streams; The causal feature extraction module is used to extract action elements, speech semantic elements, and temporal nodes; The causal correlation verification module is used to perform matching and determination based on causal verification rules; The local assessment execution module is used to output assessment results and guide practical training. The data feedback module is used to send the training data back to the cloud management platform.

[0052] Specifically, the cloud-based management platform adopts a distributed cloud architecture, supporting multi-tenant management. Different schools and training institutions can manage their own training models, rules, and student data separately, while also supporting cross-regional sharing of training resources. The compliance essence modeling module provides a visual model editing interface, allowing users to independently decompose compliance nodes based on industry operating standards, define action and voice elements, and configure causal binding relationships, lowering the threshold for model building. The causal rule configuration module supports configuring differentiated causal verification rules for different compliance nodes and different assessment levels, adapting to the assessment needs of different professional skill levels such as beginner, intermediate, and advanced. The boundary dynamic optimization module supports automatic execution of sample clustering analysis and parameter optimization, while providing a manual review interface. Optimized parameters must be manually reviewed and confirmed before being distributed, ensuring that the judgment rules do not deviate from the essential requirements of industry operating standards. The full data management module adopts a distributed storage architecture to store the compliance essence model, causal verification rules, full training audio and video data, and assessment result data, supporting data retrieval, backtracking, and statistical analysis, providing data support for teaching quality assessment. The synchronous distribution module supports both full synchronization and incremental synchronization modes, with incremental synchronization mode as the default to ensure update efficiency.

[0053] The local training execution terminal cluster comprises multiple training terminals deployed in different training sites. Each terminal is equipped with a high-definition camera, microphone, and interactive display device, supporting local offline operation. Even if the network connection to the cloud management platform is interrupted, local training assessment tasks can still be completed, and training data will be automatically transmitted back after the network is restored. The terminal adaptation module supports terminal devices with different hardware configurations and operating systems, ensuring that model and rule parameters can adapt to different terminal operating environments. The multimodal time series alignment module, causal feature extraction module, and causal association verification module all use lightweight AI models, supporting real-time local operation on the terminal without relying on cloud computing resources, reducing the impact of network latency on real-time assessment. The local assessment execution module supports real-time output of assessment results, synchronously displaying the pass status of compliance nodes, and pushing standard guidance information to unqualified nodes to guide trainees to complete corrective actions. The data transmission module supports breakpoint resumption, ensuring that training data can be completely transmitted back to the cloud management platform without data loss due to network interruption.

[0054] In summary, this invention constructs a compliance essence model that includes the causal binding relationship between action elements and speech semantic elements, and provides corresponding causal verification rules for compliance nodes. It uses the business logic of industry operating standards as a benchmark to complete the practical training assessment, solving the problem of misjudgment of practical behaviors that conform to industry operating standards but differ from preset standard samples in terms of timing or form. Simultaneously, based on qualified practical samples returned from the training terminals, it dynamically optimizes the boundary of compliance node judgment, adjusts the adaptability range of causal verification rules to accommodate diverse compliant practical behaviors, and completes the synchronous implementation of rules across all training terminals through a cloud-based incremental update mechanism. This achieves unified iteration of assessment standards, ensures consistency in judgment for cross-regional practical training assessments, adapts to diverse practical scenarios in vocational skills training, and provides a stable and accurate judgment solution for digital practical training assessments in vocational education.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cloud-based AI education and training assessment method, characterized in that, include: S1. Construct a compliance essence model based on industry operating standards. The compliance essence model includes multiple interconnected compliance nodes. Each compliance node corresponds to action elements and voice semantic elements. Establish a causal binding relationship between the action elements and the voice semantic elements. S2. Configure causal verification rules for the compliance node, and synchronize the compliance essence model and causal verification rules to the training execution terminal; S3. Collect audio and video data during the training process at the training execution terminal, align the time axis of the action screen sequence and the audio semantic sequence, and form a multimodal training data stream; S4. Extract action elements and speech semantic elements from the multimodal training data stream, and mark the temporal nodes corresponding to the action elements and speech semantic elements; S5. Verify the causal binding relationship between action elements and speech semantic elements according to the causal verification rules, complete the matching judgment of the compliance node and output the assessment result; S6. The training data is sent back to the cloud management platform. Based on the practical samples, the judgment boundary of the compliance node is optimized and the adaptation range of the causal verification rule is adjusted to complete the synchronous update of the training execution terminal.

2. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The aforementioned model for constructing the essential compliance framework includes: The industry operating procedures are broken down into multiple sub-tasks arranged in chronological order according to the operating process, with each sub-task corresponding to a compliance node; Define standard action feature vectors and standard speech semantic feature vectors for each compliant node, and establish the conditional triggering relationship between the two during the execution of the standard operation process. The conditional triggering relationship includes action-guided speech relationship, speech-accompanied action relationship, and action-feedback speech relationship.

3. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The timeline alignment of the completed action scene sequence with the audio semantic sequence includes: The video stream is subjected to frame rate normalization processing, and the audio stream is subjected to sampling rate resampling processing. The time offset between the video frame and the audio sampling block is calculated using a preset start synchronization pulse signal. The time offset is compensated by an interpolation algorithm to complete the one-to-one matching of the timestamps of the action frame and the corresponding audio segment.

4. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The extracted action elements include: The coordinates of key points of the human skeleton and the trajectory of hand movements in the training video are extracted by a convolutional neural network. The spatial displacement features of the operation object are identified, and the coordinates, trajectory and displacement features are encapsulated into motion feature vectors. The start and end timestamps of the motion feature vectors in the multimodal training data stream are recorded.

5. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The extracted speech semantic elements include: Automatic speech recognition technology is used to convert audio signals into text sequences. Natural language processing algorithms are used to extract business terminology keywords, tone pause features, and logical connectors from the text to construct speech semantic feature vectors. The audio duration intervals corresponding to the text sequences in the multimodal training data stream are recorded.

6. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The causal binding relationship between the verification action elements and the speech semantic elements includes: Retrieve the temporal nodes corresponding to the action elements and the speech semantic elements, and determine whether the temporal overlap and temporal order of the two meet the temporal threshold requirements in the causal verification rules. Determine whether the behavioral direction of the action element and the logical word direction in the speech semantic element are consistent with the standard operation process logic. If they meet the time sequence threshold requirements and are consistent, then the causal binding relationship is determined to be established.

7. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The decision boundary for optimizing the compliance node includes: Cluster analysis is performed on the returned practical samples that are deemed qualified. The deviations in action amplitude and voice expression of different practitioners when completing the same compliance node are extracted. The statistical distribution law of the deviations and differences is calculated. The similarity threshold and time tolerance value in the causal verification rule are adjusted to expand the judgment boundary to be compatible with practical behaviors that are compliant with the standard but have diverse forms.

8. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The output of the assessment results also includes: Based on the compliance nodes that fail the judgment, the corresponding normative guidance information is retrieved from the compliance essence model, and voice prompts and action demonstrations are displayed on the training execution terminal interface to guide trainees to perform corrective operations. The corrected practical data is recorded as a reference sample for the boundary optimization module.

9. The AI ​​education and training assessment method based on a cloud platform according to claim 1, characterized in that, The synchronous update includes: When the cloud-based management platform detects a change in the causality verification rules, it generates an incremental update package and sends an update command to the local training execution terminal that is in a connected state through a two-way communication link. The training execution terminal silently downloads the incremental update package in the background and completes the parameter overwrite of the local operating environment during the idle period after the current training task ends.

10. A cloud-based AI education and training assessment system, applied to the cloud-based AI education and training assessment method according to any one of claims 1 to 9, characterized in that, include: The cloud-based management platform includes a compliance essence modeling module, a causal rule configuration module, a boundary dynamic optimization module, a full data management module, and a synchronous distribution module; The compliance essence modeling module is used to construct a compliance essence model based on industry operating standards. The causality rule configuration module is used to configure causality verification rules for compliant nodes; The boundary dynamic optimization module is used to optimize the judgment boundary of compliance nodes based on the training data returned by the training execution terminal. The full data management module is used to store model and practical data; The synchronous distribution module is used to synchronize the updated model and rules to the training execution terminal in real time. The local training execution terminal cluster is connected to the cloud management and control platform, including a terminal adaptation module, a multimodal time series alignment module, a causal feature extraction module, a causal correlation verification module, a local assessment execution module, and a data feedback module. The terminal adaptation module is used to receive the distributed models and rules and update the model and rule parameters of the local operating environment; The multimodal temporal alignment module is used to generate multimodal training data streams; The causal feature extraction module is used to extract action elements, speech semantic elements, and temporal nodes; The causal correlation verification module is used to perform matching and determination based on causal verification rules; The local assessment execution module is used to output assessment results and guide practical training. The data feedback module is used to send the training data back to the cloud management platform.

Images