Experimental intelligent evaluation method based on AI multi-modal and experimental rule combined evidence chain

Abstract

The present application relates to the technical field of online education, and more particularly to an experimental intelligent evaluation method based on AI multi-modal and experimental rule combination evidence chain output, the present application obtains the operation video uploaded by the student end and splits into step level video clips according to the experiment process, for each clip, the multi-modal AI model is driven according to the corresponding scoring rule to match and identify the scoring items and the deduction items, the video time interval of the hit item is backtracked and labeled synchronously, the evidence chain of each scoring result is formed, the confidence is calculated based on the evidence chain, the low confidence segment is marked as to be reviewed and pushed to the administrator end for manual review, and finally the comprehensive evaluation is obtained by summarizing. The scheme generates the evidence chain and targets the review, breaks the traditional AI scoring black box, makes each step of scoring traceable, realizes the risk control of AI evaluation, targets the review to avoid administrator traversal video, and improves the review efficiency. The method takes into account the automation, interpretability and high reliability of the experimental operation evaluation.

Description

Technical Field

[0001] This invention relates to the field of online education technology, and in particular to an intelligent evaluation method for experiments that generates a chain of evidence based on the combination of AI multimodal and experimental rules. Background Technology

[0002] In recent years, the deep integration of artificial intelligence and educational informatization is changing the way experimental teaching is evaluated. Automating the scoring of students' experimental operations using computer vision technology can reduce the repetitive burden of manual evaluation for teachers, while improving the standardization of assessment. Experimental operation scoring differs from general action recognition tasks; the system needs to accurately determine the start and end times of operation steps from a complete video, identify whether key actions are performed correctly, and correlate the scoring results with the specific operational evidence in the video. This places high demands on the system's semantic understanding capabilities and the interpretability of the results.

[0003] Current mainstream intelligent evaluation schemes for experimental operations generally follow a path of first manually annotating, then training a dedicated model, and finally scoring according to rules. Specifically, this requires collecting a large number of operation videos for each experimental course, with annotators marking the step boundaries and operation labels frame by frame or segment by segment. Then, a recognition model is trained based on this annotated data, and finally, the model output is scored using preset rules. When the experimental project, equipment, or scoring criteria are adjusted, samples need to be collected again, annotated again, and the model needs to be retrained. The deployment cycle for new experimental projects is long, and the reuse cost is high. At the same time, the model only outputs classification labels or total scores, and the scoring results lack evidence of correlation with specific screen locations and time segments in the video. When there is uncertainty in recognition due to factors such as screen occlusion or changes in lighting in the video, there is a lack of quantitative assessment of confidence and control measures for low-confidence results. It is also difficult for reviewers to quickly locate segments that need to be focused on for verification.

[0004] Therefore, how to establish a traceable link between the scoring results and the original operation records, and how to quantitatively evaluate and verify the confidence level of the scoring, so as to make the scoring process interpretable and controllable, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent evaluation method for experiments that combines AI multimodal analysis with experimental rules to generate a chain of evidence. This method aims to solve the technical problem of how to establish a traceable link between the scoring results and the original operation records in the prior art, and to quantitatively evaluate and verify the confidence level of the scores, so as to make the scoring process interpretable and controllable.

[0006] To achieve the above objectives, this invention provides an intelligent evaluation method for experiments that generates a chain of evidence based on the combination of AI multimodal analysis and experimental rules. The method includes the following steps: Obtain the operation video data uploaded by the user terminal, as well as the experimental process information corresponding to the operation video data; Based on the preset experimental procedure information, the experimental operation video is divided into video segments corresponding to each experimental step. For each video segment, a score is assigned according to the scoring rules associated with the corresponding experimental steps to obtain the score result of the video segment and the chain of evidence for the score result; Based on the chain of evidence for the scoring results, the confidence level of the scoring results is obtained; Video segments whose confidence level of the scoring result is lower than the preset confidence threshold are marked as segments to be reviewed, and the segments to be reviewed are pushed to the administrator terminal, which performs segment review and obtains the review scoring result. The comprehensive score result is obtained by summarizing the scoring results and the review scoring results.

[0007] Optionally, the step of scoring each video segment according to the scoring rules associated with the corresponding experimental step to obtain the scoring result of the video segment and the chain of evidence for the scoring result includes: Based on the scoring rules associated with the video clip and the experimental steps, a multimodal AI model is invoked to match and identify scoring items and deduction items to obtain scoring matching results. The scoring matching results include items that are scored, items that are deducted, and items that are not scored. Based on the scoring matching results, the video timestamp range, the description of the basis for the hit, and the snapshot address corresponding to each hit item are extracted and aggregated to obtain the evidence chain of the scoring results. Based on the matching items in the scoring results, preset weight values ​​are retrieved from the scoring rules associated with the experimental steps and summed to obtain the scoring result of the video segment.

[0008] Optionally, the step of calling a multimodal AI model to match and identify scoring items and deduction items according to the scoring rules associated with the video clip and the experimental steps, and obtaining the scoring matching result, includes: Based on the scoring rules associated with the experimental steps, the recognition features and hit rules corresponding to each scoring item are obtained. The scoring items include scoring items and deduction items. Based on the identification features, the corresponding video time interval is located from the video segment to obtain the video segment to be analyzed for each scoring item. The video segment to be analyzed and the corresponding hit rules are input into the multimodal AI model for hit analysis to obtain the preliminary recognition results of each evaluation item; Based on the preliminary identification results of all evaluation items, the matching scores, the deductions, and the missing items are summarized to obtain the scoring matching results.

[0009] Optionally, based on the scoring matching results, the step of extracting the video timestamp range, the description of the basis for the hit, and the snapshot address corresponding to each hit item, and aggregating them to obtain the scoring result evidence chain, includes: Based on the scoring matching results, extract the hit scoring items and the hit deduction items to obtain the hit item set; Extract the start and end times of each hit item in the video segment to obtain the timestamp interval of each hit item; Based on the timestamp interval of each hit item, the video segment is extracted and stored to obtain the snapshot address of each hit item; The image feature regions that match the hit items in the video intervals corresponding to each snapshot address are labeled to obtain the labeled content; The snapshot address and the labeled content are used to generate a description of the basis for each hit item; The evidence chain for the scoring result is obtained by associating and combining the timestamp interval, the description of the basis for the hit, and the snapshot address of each hit item.

[0010] Optionally, based on preset experimental procedure information, the experimental operation video is divided into video segments corresponding to each experimental step, including: Based on the pre-set experimental procedure information, the key visual recognition points and step duration features corresponding to each experimental step are obtained. Based on the key visual recognition points, the operation video data is matched with process steps to obtain the video time intervals corresponding to the successfully matched experimental steps. Based on the video time interval of the successfully matched experimental steps and the duration characteristics of the steps, the time interval of the missing experimental steps that failed to pass feature recognition is estimated by reverse calculation to obtain the video time interval of the missing experimental steps. The operation video data is segmented according to the video time interval of all steps to obtain video segments corresponding to each experimental step.

[0011] Optionally, obtaining the confidence level of the scoring result based on the chain of evidence for the scoring result includes: Based on the evidence chain of the scoring results, the snapshot address and annotation content corresponding to each hit item are extracted to obtain the associated evidence of each hit item; Based on the correlation evidence of each hit item, the degree of match between the labeled area and the preset identification target is evaluated, and the individual confidence score of each hit item is obtained. The confidence level of the video segment scoring result is obtained based on the individual confidence scores of all hit items and the influence weights of the missing items in the evidence chain.

[0012] Optionally, the step of marking video segments whose confidence level of the scoring result is lower than a preset confidence threshold as segments to be reviewed, and pushing the segments to be reviewed to the administrator, whereby the administrator performs segment review to obtain a review scoring result, includes: Based on the segment to be reviewed, obtain the video segment corresponding to the segment to be reviewed, the evidence chain of the scoring result, and the user terminal operation records of the time interval where the video segment is located; The administrator displays the video clip, the chain of evidence for the scoring results, and the user's operation records on the administrator's end. The administrator confirms or corrects the hit scoring items and deduction items in the segment to be reviewed based on the displayed content, and obtains the hit item review results. The non-hit items in the segment to be reviewed are re-evaluated, and the non-hit item review results are obtained. Based on the review results of the hit items and the review results of the missing items, a segment review score and a corresponding review description are generated. The review score of the segment is combined with the review description to obtain the review score result.

[0013] Furthermore, to achieve the above objectives, this invention also proposes an intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules. The intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules includes: The video preprocessing module is used to acquire operation video data uploaded by the user terminal, as well as the experimental process information corresponding to the operation video data; The video preprocessing module is also used to split the experimental operation video into video segments corresponding to each experimental step based on preset experimental process information. The experimental operation evaluation module is used to score each video segment according to the scoring rules associated with the corresponding experimental steps, and to obtain the scoring result of the video segment and the evidence chain of the scoring result. The experimental operation evaluation module is also used to obtain the confidence level of the scoring result based on the evidence chain of the scoring result; The review processing module is used to mark video segments whose confidence level of the scoring result is lower than a preset confidence threshold as segments to be reviewed, and push the segments to be reviewed to the administrator terminal, whereby the administrator terminal performs segment review and obtains the review scoring result; The results summary module is used to summarize the scoring results and the review scoring results to obtain a comprehensive scoring result.

[0014] Furthermore, to achieve the above objectives, this invention also proposes an intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules. The intelligent experimental evaluation device includes: a memory, a processor, and an intelligent experimental evaluation program that generates an evidence chain based on the combination of AI multimodal and experimental rules, stored in the memory and executable on the processor. The intelligent experimental evaluation program is configured to implement the steps of the intelligent experimental evaluation method based on the combination of AI multimodal and experimental rules described above.

[0015] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing an experimental intelligent evaluation program that generates an evidence chain based on the combination of AI multimodal and experimental rules. When the experimental intelligent evaluation program that generates an evidence chain based on the combination of AI multimodal and experimental rules is executed by a processor, it implements the steps of the experimental intelligent evaluation method that generates an evidence chain based on the combination of AI multimodal and experimental rules as described above.

[0016] The one or more technical solutions proposed in this application have at least the following technical effects: By directly inputting structured scoring rules containing step descriptions, scoring points, and scoring rules into the operation recognition model, the model guides the temporal action localization and scoring of the video according to these rules. This means that when the experimental project or scoring criteria are adjusted, only the input rules need to be updated without re-collecting samples and retraining the model, reducing the cost and cycle of launching new experimental projects. Simultaneously, during the operation localization process according to the scoring rules, the model synchronously generates a record linking each scoring item to the corresponding spatiotemporal segment in the video, allowing the basis for deductions or scores to be directly traced back to the specific scene location and time range, improving the traceability and completeness of the scoring results. Furthermore, the model outputs a confidence assessment value for each scoring item, and low-confidence potential misjudgments can be identified and controlled before being included in the final score. Combined with the confidence value, a review label is generated, allowing manual review to focus only on the high-risk target segments, avoiding the need for reviewers to search the entire video for possible deductions, thus improving review efficiency and reducing the probability of omissions. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the intelligent experimental evaluation method based on the combination of AI multimodality and experimental rules to generate a chain of evidence according to the present invention. Figure 2 This is a flowchart illustrating the second embodiment of the experimental intelligent evaluation method based on the combination of AI multimodality and experimental rules to generate a chain of evidence according to the present invention. Figure 3 This is a flowchart illustrating the third embodiment of the experimental intelligent evaluation method based on the combination of AI multimodality and experimental rules to generate a chain of evidence according to the present invention. Figure 4 This is a timeline diagram of the entire process of the intelligent experimental evaluation method based on the combination of AI multimodality and experimental rules to generate an evidence chain, as presented in this invention. Figure 5 This is a structural block diagram of the first embodiment of the experimental intelligent evaluation device of the present invention, which generates an evidence chain based on the combination of AI multimodality and experimental rules. Figure 6 This is a schematic diagram of the structure of an experimental intelligent evaluation device that combines AI multimodal and experimental rules to generate an evidence chain, which is part of the hardware operating environment of the embodiment of the present invention.

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] The main solution of this application embodiment is as follows: Obtain operation video data uploaded by the user terminal, and experimental process information corresponding to the operation video data; based on the preset experimental process information, split the experimental operation video into video segments corresponding to each experimental step; for each video segment, score it according to the scoring rules associated with the corresponding experimental step to obtain the scoring result and the scoring result evidence chain of the video segment; obtain the confidence level of the scoring result based on the scoring result evidence chain; mark video segments whose confidence level is lower than a preset confidence threshold as segments to be reviewed, and push the segments to be reviewed to the administrator terminal, whereby the administrator terminal performs segment review to obtain the review scoring result; summarize the scoring result and the review scoring result to obtain a comprehensive scoring result.

[0024] Currently, mainstream intelligent evaluation schemes for experimental operations generally follow a path of first manually labeling, then training a dedicated model, and finally scoring according to rules. Specifically, this requires collecting a large number of operation videos for each experimental course, with labelers marking the step boundaries and operation labels frame by frame or segment by segment. A recognition model is then trained based on this labeled data, and finally, the model output is scored using preset rules. This approach requires re-collecting samples, re-labeling, and retraining the model when experimental projects, equipment, or scoring standards are adjusted. This results in long deployment cycles for new experimental projects and high reuse costs. Furthermore, the model only outputs classification labels or total scores, lacking evidence linking the scoring results to specific locations and time segments in the video. When recognition is uncertain due to factors such as occlusion or lighting changes, there is a lack of quantitative assessment of confidence and control measures for low-confidence results. Reviewers also find it difficult to quickly locate segments requiring focused verification. Therefore, establishing a traceable link between scoring results and original operation records, and quantifying and verifying the confidence of the scores to ensure interpretability and controllability of the scoring process, are urgent technical problems that need to be solved.

[0025] It should be noted that the executing entity of this invention can be an intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone. Alternatively, it can be a thermal management device capable of achieving the aforementioned functions, such as an intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules. This embodiment does not specifically limit the scope of the invention. The following description uses an intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules as an example to illustrate this embodiment and the subsequent embodiments.

[0026] Based on this, embodiments of this application provide an intelligent evaluation method for experiments that generates a chain of evidence based on the combination of AI multimodal analysis and experimental rules, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain, as described in this application.

[0027] In this embodiment, the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to produce an evidence chain includes steps S10 to S60: Step S10: Obtain the operation video data uploaded by the user terminal, and the experimental process information corresponding to the operation video data.

[0028] It should be noted that the primary entity executing this step is the system server, and its data source is the user terminal. The user terminal typically refers to the camera equipment installed in the experimental location or the mobile terminal held by the student participating in the experimental exam. Operational video data refers to multimedia files recorded in real-time by the user terminal's camera, documenting the entire experimental process performed by the experimenter; its content includes continuous footage from the start to the end of the experiment. Experimental procedure information refers to structured data related to the current experimental course, used to describe the standard operating sequence to be followed and the grading criteria. The specific content of the experimental procedure information typically includes the experiment name, experiment version, names of each step, and the logical relationships between the steps.

[0029] It is understandable that there is a clear correspondence between the operation video data and the experimental procedure information. After obtaining the two, the system first establishes a data association, that is, it clarifies which experimental course's operation video is being processed, thus providing a basis for subsequent visual recognition analysis according to the experimental steps. In practice, when users upload videos, they usually include the identification information of the experimental course, and the system automatically matches the corresponding experimental procedure configuration accordingly. The course configuration adopts a three-level structure of "major steps, sub-steps, and scoring points". Among them, major steps are used for teaching grouping and page organization and sorting; sub-steps are the smallest business units for AI analysis and scoring, and their fields include operation description, completion criteria, whether it is mandatory, AI supplementary explanation, etc.; scoring points are observable and judgmental check items, carrying specific scores. The total course score is automatically summed from the scores of each scoring point, thus avoiding the problem of inconsistent score configuration between the total course score and the step score. For example, when a user uploads a video of a chemical titration experiment, the system will simultaneously call up the scoring criteria for each sub-step of the experiment, which specifies the sequence of steps from cleaning the instrument, loading the solution, titrating to reading and recording, as well as the scoring points for each step.

[0030] It should be understood that the acquisition process in this step does not require the video data and process information to arrive at the same time; the system can support asynchronous reception and matching. It should also be noted that the experimental process information can be preset and modified by the administrator in the system backend to adapt to different experimental versions or adjustments to scoring criteria.

[0031] In one embodiment, the administrator adopts a three-level structure of "major steps, sub-steps, and scoring points" during course configuration. Major steps are used for teaching grouping and page organization; sub-steps are the smallest business units for AI analysis and scoring, with fields including operation description, completion criteria, whether it is mandatory, and AI supplementary explanations; scoring points are observable and verifiable check items, carrying specific scores. The total course score is automatically summed from the scores of each scoring point, thus avoiding inconsistencies between the total course score and the step-by-step score configuration.

[0032] Step S20: Based on the preset experimental procedure information, the experimental operation video is divided into video segments corresponding to each experimental step.

[0033] It should be noted that the preset experimental procedure information refers to the configuration data entered into the system beforehand, describing the standard operating sequence of the experiment. This data typically includes the name, sequence number, and expected start and end markers for each experimental step. A video segment refers to a continuous time period extracted from a complete operation video, containing only the footage corresponding to a specific experimental step. The splitting of this step uses this procedure information as a timeline template. By detecting action transition points or user operation nodes in the video footage, the start and end times of each step are located, thus dividing the entire video into several independent segments that are sequentially connected and correspond one-to-one with each step.

[0034] Understandably, when splitting a video into segments, the system locates the start and end frames of each individual step defined in the experimental procedure information, rather than simply cutting the video evenly according to duration. For example, in an experimental video about preparing temporary plant cell slides, the procedure specifies five steps: material collection, flattening, coverslipping, staining, and observation. The system needs to identify when the operator puts down the tweezers to flatten the material and when they pick up the coverslip to prepare for covering, thus accurately identifying the actual start point of the coverslip operation and encapsulating the continuous footage within that start and end time period into a video segment for the "coverslip" step.

[0035] It should be understood that this splitting function does not change the storage state of the original video; it only copies the video segments used for analysis. Furthermore, when a step in the process information is not detected as being executed in the video, the system will mark the corresponding video segment as empty or missing, rather than forcibly splitting it into a content-free segment, thus ensuring the accuracy of subsequent scoring and judgment processes.

[0036] Step S30: For each video segment, score it according to the scoring rules associated with the corresponding experimental step to obtain the scoring result of the video segment and the chain of evidence for the scoring result.

[0037] It should be noted that the purpose of this step is to perform independent automated scoring for the experimental steps corresponding to each segment. Scoring rules refer to pre-configured evaluation criteria tied to specific experimental steps, typically including the visual conditions that the step must meet, the required operational sequence, or whether key actions have occurred. The scoring result refers to the qualitative or quantitative conclusion given by the system after analysis of the step's execution. The chain of evidence for the scoring result refers to a set of evidentiary data associated with the scoring result, used to trace the original basis for the system's judgment.

[0038] Understandably, the granularity of the scoring rules corresponds one-to-one with each sub-step, and the scoring rules for different sub-steps are independent and do not interfere with each other. When processing a video segment, the system only calls the scoring rules bound to that step for targeted analysis, without mixing in the evaluation criteria of other steps. When the scoring rules include multiple dimensions, the system will analyze each dimension separately and finally generate a comprehensive score for that step. For example, in the "covering with a coverslip" step, the scoring rules may simultaneously include two dimensions: "no air bubbles generated during the operation" and "operation speed not exceeding two seconds." The system will determine whether air bubbles exist in the image, calculate the operation time, and then give the total score for the step and the sub-item score for each dimension based on the achievement of the two dimensions.

[0039] It should be understood that, in order to ensure that the chain of evidence for the scoring results makes the scoring conclusions traceable and verifiable, this chain of evidence must include at least the visual features that triggered the scoring result, the confidence score of the AI ​​model, and a description of the judgment logic. By establishing a correlation between the scoring results and the chain of evidence, the system can support teachers in quickly verifying abnormal scoring in subsequent stages. Teachers can directly jump to the corresponding segment in the video for manual confirmation based on the timestamp in the chain of evidence.

[0040] Step S40: Obtain the confidence level of the scoring result based on the chain of evidence for the scoring result.

[0041] It should be noted that the scoring result evidence chain refers to the associated record generated synchronously during the process of the model scoring the experimental operation video item by item according to the structured scoring rules. This record binds each score or deduction point to the corresponding frame position and time segment in the video. This evidence chain not only includes the conclusion of whether the operation was judged to be compliant, but also includes video frames or video snapshots that support this judgment, so that specific video evidence can be found for each scoring item.

[0042] Understandably, the confidence level of a rating result refers to a quantitative assessment of the reliability of the rating conclusion based on the completeness of information and the clarity of features reflected in the aforementioned chain of evidence. For example, when the chain of evidence shows that a key operational step is precisely located within a clear, continuous time period, and that this segment is unobstructed and free from sudden changes in lighting, and the model's feature response value for this action is stable and clearly distinct from other action categories, then the confidence level of the rating is high. Conversely, if the start and end boundaries of the operation recorded in the chain of evidence are ambiguous, or if there are mixed responses from multiple competing action features, then the confidence level of the corresponding rating item will decrease.

[0043] It should be understood that by extracting clues such as spatiotemporal consistency and feature discrimination strength from the chain of evidence to obtain confidence levels, it is possible to identify low-certainty scores caused by environmental interference or motion blur. This approach transforms the reliability of the scoring conclusion into a quantifiable indicator, providing a direct screening basis for subsequent targeted review of high-risk segments, reducing the workload of reviewers in searching the entire video, and preventing low-confidence misjudgments from directly entering the final score.

[0044] In a feasible embodiment, obtaining the confidence level of the scoring result based on the chain of evidence for the scoring result includes steps A10 to A30: Step A10: Based on the evidence chain of the scoring results, extract the snapshot address and annotation content corresponding to each hit item to obtain the associated evidence for each hit item.

[0045] It should be noted that the snapshot address is the storage path of the video screenshot that records the moment the hit occurred, while the annotation content consists of the key areas marked on the image and text descriptions. Correlated evidence refers to the original data that can be independently traced back after combining the snapshot address and the annotation content. For example, for the hit item "left hand holding a test tube," the system saves a screenshot marked with a red rectangle, along with its path and annotation coordinates. Administrators can directly open the screenshot to verify its accuracy.

[0046] Step A20: Based on the correlation evidence of each hit item, evaluate the degree of match between the labeled area and the preset identification target, and obtain the individual confidence score of each hit item.

[0047] It's important to note that the single-item confidence score refers to the confidence level calculated based on the degree of similarity between the labeled area and the preset recognition target for a single hit item. The labeled area is the range of target pixels within the image that is bounded, while the preset recognition target is the ideal operational form defined in the scoring criteria, such as "the test tube tilt angle should be between 30 and 45 degrees." The degree of similarity is assessed by comparing the geometric features, color distribution, or spatial relationships of the labeled area with the preset standard. For example, if the system determines "test tube tilt" as a hit item, and the labeled area shows that the actual tilt angle of the test tube is 40 degrees, deviating from the preset standard of 50 degrees, then the single-item confidence score for this hit item may be low, prompting the administrator to conduct a thorough review.

[0048] Step A30: Based on the individual confidence scores of all hit items and the influence weights of the missing items in the evidence chain, obtain the confidence level of the video segment scoring result.

[0049] It's important to note that the confidence level of the scoring result is a comprehensive measure of the reliability of the entire video segment's scoring result. It requires integrating the individual confidence scores of all hit items and additionally considering the impact weight of missed items. Missed items refer to key actions that the automated system failed to identify but actually exist; their impact weight reflects the importance of that action to the overall score. For example, in a segment with three scoring items and two deduction items, where two scoring items have high individual confidence scores and one has a low score, and the system misses a crucial deduction item—"extinguishing the alcohol lamp"—this missed item has a high weight, leading to a significant decrease in the overall scoring confidence. The final calculated confidence level will guide the administrator in deciding whether to conduct a more in-depth manual review of the segment or directly adopt the score.

[0050] Step S50: Mark video segments whose confidence level of the scoring result is lower than the preset confidence threshold as segments to be reviewed, and push the segments to be reviewed to the administrator terminal, so that the administrator terminal can perform segment review and obtain the review scoring result.

[0051] It should be noted that the confidence level of the scoring result refers to the quantified value of how confident the system's internal algorithm is in its judgment when making a scoring decision. The preset confidence threshold is a baseline set during the system configuration phase to determine whether the current scoring result is sufficiently reliable. Segments awaiting review refer to video segments whose automatically scored confidence level is below this threshold. These segments are not suitable as final scores due to the inherent uncertainties in their analysis process. The administrator terminal is the operating terminal used by personnel with review permissions, typically equipped with an interactive interface for video playback, segment annotation, and score modification.

[0052] Understandably, the core value of this step lies in filtering out low-quality judgments generated during the automatic scoring process and submitting them to human reviewers for secondary confirmation. This design acknowledges the possibility of blind spots or misjudgments in the automatic scoring system, thus effectively compensating for the limitations of the model's capabilities through human review.

[0053] It should be understood that when administrators perform segment review, they can directly view the original video segment and the evidence chain of the automatically generated scoring results, including the specific basis for the system's judgment and screenshots of the scene at the time. Administrators can either fully adopt the original system score or modify the score based on manual observation to generate a review score result. The review score result will replace the automatic score result as the final result of this step. At the same time, the system will record the personnel information, review time, and reasons for modification of the review operation, forming a complete review audit record for subsequent data analysis and score quality traceability. This process of machine initial evaluation plus manual review effectively balances scoring efficiency and scoring accuracy.

[0054] Step S60: Summarize the scoring results and the review scoring results to obtain a comprehensive scoring result.

[0055] Understandably, for video segments that meet the confidence level, the system directly adopts their automatic scoring result as the final score for that step; while for segments marked as requiring review, the review score generated by the administrator replaces the original automatic score. The comprehensive score result refers to the final complete score conclusion representing the operational level of each experimental step and even the entire experiment, obtained after merging the two types of score data according to preset rules.

[0056] This invention first acquires the operation video and corresponding experimental process information uploaded by the user; then, based on the process information, the complete video is divided into video segments corresponding to each experimental step; next, for each segment, the scoring rules associated with its step are called to automatically score it, and an evidence chain of scoring results carrying the judgment criteria is generated; the confidence level of each scoring result is calculated based on the evidence chain, and segments with confidence levels lower than a set threshold are marked as segments to be reviewed and pushed to the administrator for manual review; finally, the automatic scoring results and the manual review results are summarized to obtain the comprehensive scoring result of each step.

[0057] Based on the above solution, this invention achieves the following beneficial effects: By breaking down and scoring videos step by step, scoring efficiency is significantly improved, reducing the burden of manual grading. By generating a chain of evidence and calculating confidence levels, the system can automatically select reliable scores for direct adoption, while triggering manual review for segments with high uncertainty, effectively balancing the accuracy and reliability of the scoring. Furthermore, the chain of evidence provides precise location and reference for manual review, facilitating rapid error correction by administrators, thereby improving the overall objectivity and traceability of experimental assessment scoring.

[0058] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 In the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain, the step of determining the conflicting grid by judging semantic conflicts based on the semantic probability distribution of each grid after the update includes steps S301 to S303: Step S301: Based on the video clip and the scoring rules associated with the experimental steps, call the multimodal AI model to perform matching and identification of scoring items and deduction items, and obtain the scoring matching result.

[0059] It should be noted that multimodal AI models refer to artificial intelligence analysis models capable of simultaneously processing multiple information types such as video images, audio, and text. The scoring items and deduction items are specific descriptions of the bonus and penalty behaviors stipulated in the scoring rules, respectively.

[0060] Understandably, the core operation of this step involves calling a multimodal AI model to compare and identify video segments item by item according to scoring rules. The model analyzes the actions, temporal relationships, and audio content in the video, and determines whether there are scoring or deducting behaviors in the segment. Finally, it outputs a complete scoring and matching result, which includes matching scoring items, matching deducting items, and missing items.

[0061] It should be understood that the scoring matching results not only include the match status of each scoring item and deduction item, but also retain the video timestamp or image features corresponding to each matching criterion. This matching recognition output data will serve as the basic material for generating the evidence chain of subsequent scoring results, ensuring the traceability and interpretability of the automatic scoring process.

[0062] In a feasible embodiment, the step of calling a multimodal AI model to perform matching and identification of scoring items and deduction items based on the scoring rules associated with the video clip and the experimental steps, and obtaining the scoring matching result, includes steps B10 to B40: Step B10: Based on the scoring rules associated with the experimental steps, obtain the recognition features and hit rules corresponding to each scoring item. The scoring items include scoring items and deduction items.

[0063] It should be noted that the scoring rules associated with the experimental steps refer to a set of pre-defined evaluation criteria for specific operational steps. Scoring items are the basic units constituting these rules, specifically divided into points-earning items and points-deducting items. Point-earning items refer to the bonus points earned for conforming to the standard operating procedures, while points-deducting items refer to the points deducted for violations or non-standard operating procedures. Identification features refer to specific visual or auditory markers that can be extracted from video or audio data to determine whether a scoring item is valid.

[0064] Understandably, the core task of this step is to clarify the corresponding identification features and hit rules based on the meaning of each scoring item. The hit rules refer to the specific logical conditions for determining the validity of a scoring item, such as angle range, temporal sequence, or the presence or absence of an object. For example, for the scoring item "test tube opening facing the direction of no one," the corresponding identification feature might be the angle parameter of the test tube opening, and the hit rules stipulate that the angle deviation must be less than a certain allowable range to be considered a hit. For the deducted item "not wearing goggles," the system first identifies the object feature of goggles. When it detects that the facial area in the image lacks goggles, it locates that time interval as the segment to be analyzed for subsequent scoring.

[0065] Step B20: Based on the identification features, locate the corresponding video time interval from the video segment to obtain the video segment to be analyzed corresponding to each scoring item.

[0066] It should be noted that a video segment refers to a complete operational image obtained by dividing the original experimental video into steps. A video time interval refers to the start and end time periods within a video segment that are highly correlated with a certain recognition feature. The video segment to be analyzed refers to a short-term image that is located and extracted from the original video and is highly correlated with a specific scoring item.

[0067] Understandably, this fine-grained stratification stems from the administrator's scoring configuration logic. The system first divides the original recordings into multiple video segments according to the experimental steps, each segment corresponding to a standard operating procedure. These video segments, as the first level, address the need for step-by-step grouping, providing clear step boundaries for subsequent analysis and avoiding cross-step interference. Then, the system needs to identify the start and end times of the appearance or disappearance of specific recognition features within a video segment, thus forming a video time interval. This level addresses the question of "within which time period it occurred," such as pinpointing the moment a student was not wearing goggles. However, this interval may contain irrelevant transitional segments, requiring further trimming. Finally, the video segments to be analyzed, as the third level, are short images precisely extracted from the video time interval, containing only the most relevant key actions. This level addresses the problem of "simplifying input data." For example, for the scoring item "test tube orientation," only the key frames during the test tube rotation process are extracted, excluding the preceding and following waiting time. Through a three-tiered process, the system can progressively compress massive amounts of raw video footage into extremely small data units that are compatible with AI model processing. This not only improves analysis efficiency but also ensures that the identification features of each scoring item can be independently and accurately identified.

[0068] It should be understood that each scoring item may correspond to multiple discrete video time intervals. The system will generate independent segments to be analyzed for each interval, thereby ensuring that each suspected scoring or deduction behavior can be examined and judged individually and accurately, providing spatiotemporally accurate input material for the final scoring matching and recognition.

[0069] Step B30: Input the video segment to be analyzed and the corresponding hit rules into the multimodal AI model for hit analysis to obtain the preliminary recognition results of each evaluation item.

[0070] It should be noted that the hit rule is the specific judgment logic condition set for this scoring item, such as angle threshold or overlap relationship. A multimodal AI model refers to an intelligent analysis model capable of comprehensively processing image, sound, and text information. The preliminary recognition result refers to the model's first-round judgment conclusion on whether each segment to be analyzed meets the hit rule.

[0071] Understandably, this step involves taking each video segment to be analyzed and its corresponding matching rule as a set of inputs, which is then fed into a multimodal AI model for frame-by-frame or segment-by-segment matching analysis. The model combines the position of objects in the scene, the movement trajectory of people, and possible audio information to check whether all the conditions listed in the matching rule are met. For example, for the scoring item "heating the test tube, cooling it first, and then shaking it," the model will analyze whether there is a complete cooling action in the segment to be analyzed and whether the shaking action occurs after cooling is completed. If this condition is met, it is initially identified as a match.

[0072] It should be understood that the preliminary identification results usually include "hit" or "miss" labels and corresponding confidence scores. This result is still in the preliminary stage, and its accuracy may be limited by video quality or model recognition capabilities. Further verification or fusion steps are needed to generate the final scoring configuration items, thereby ensuring the reliability and rigor of the entire automatic scoring system.

[0073] Step B40: Based on the preliminary identification results of all evaluation items, summarize and determine the matching points, the deduction points, and the missing items to obtain the scoring matching results.

[0074] It's important to clarify that "hit" in scoring refers to experimental operations that the model determines fully meet the scoring rules. For example, if the scoring item "checking instruments before the experiment" is identified as a hit, it will receive a positive score. "Hit" in deduction refers to operations that the model determines involve violations. For example, if "not wearing gloves" is identified as a hit, it will receive a negative score. "Miss" refers to situations where no evidence conforming to the rules is detected in the corresponding video clip. This neither affects the score nor triggers a deduction. Generally, this is due to the model's limited video recognition capabilities, which may result in the model missing important steps or segments, making it impossible to form a definitive basis for scoring or deduction.

[0075] Step S302: Based on the scoring matching results, extract the video timestamp interval, the description of the basis for the hit, and the snapshot address corresponding to each hit item, and aggregate them to obtain the scoring result evidence chain.

[0076] In a feasible embodiment, the step of extracting the video timestamp interval, the description of the basis for the hit, and the snapshot address corresponding to each hit item based on the scoring matching result, and aggregating them to obtain the scoring result evidence chain, includes steps C10 to C60: Step C10: Based on the scoring matching results, extract the hit scoring items and the hit deduction items to obtain the hit item set.

[0077] It should be noted that this step is the first round of screening from the score matching results. Hit items refer to the evaluation items that were determined to meet the scoring conditions in the previous steps. Hit items are the evaluation items that were determined to be violations. Together, they constitute the set of hit items, which only contains entries judged as "hit" by the model; non-hit items are filtered out and do not enter the subsequent evidence chain construction process.

[0078] Understandably, the set of hit items serves as the starting point for all subsequent evidence extraction. For example, in a physics experiment assessment, the set of hit items might include a point for "correctly connecting the circuit" and a point for "not using insulating gloves." By extracting these two types of hit items separately, a timestamp, snapshot, and description can be generated for each item, avoiding confusion between the logical relationship between points and deductions and ensuring a clear classification of the evidence chain.

[0079] Step C20: Extract the start and end time points of each hit item in the video segment to obtain the timestamp interval of each hit item.

[0080] It should be noted that the video clips refer to the complete operational footage divided according to the experimental steps mentioned above. The timestamp interval refers to the instant the hit action begins and ends, represented by the absolute time position of the video frame. Extracting the timestamp interval relies on the video time intervals generated in the previous steps, but this step requires more precise location of the actual boundary where the hit action occurred. For example, the deduction item "test tube mouth to person" might only last one frame, so its start and end times might be the same. By obtaining an independent timestamp interval for each hit action, it can be ensured that subsequent snapshots only include the most relevant footage, avoiding the inclusion of irrelevant transitional segments.

[0081] It should be understood that timestamp intervals are the most basic data units in the chain of evidence. They are used to guide the start and end points of video captures and to inform users in the evidence description which moment in the experiment the behavior occurred. Multiple timestamp intervals for different hits may overlap; the system needs to record them independently so that they can be arranged chronologically or by rating category in the final chain of evidence, facilitating rapid location and verification.

[0082] Step C30: Based on the timestamp interval of each hit item, the video segment is extracted and stored to obtain the snapshot address of each hit item.

[0083] It should be noted that the snapshot address refers to the file path of the extracted video clip on the storage device. This step involves cropping a short video segment or several keyframes within the corresponding time range from the original video clip, based on the timestamp intervals obtained in the preceding steps. Each matched item corresponds to an independent snapshot, and the snapshot format can be a common video format, such as a frame-by-frame image sequence. During storage, the file name must be associated with the unique identifier of the matched item.

[0084] Understandably, snapshot URLs serve as visual evidence in the chain of evidence. For example, for the penalty of "not wearing goggles," the snapshot should capture the entire process from the student's face being exposed to the completion of the action. The snapshot duration should not be too long, typically covering only a few seconds before and after the offending action to ensure reasonable storage space. The snapshot URL, as a link, can be directly played or displayed within the chain of evidence for reviewers to access and view.

[0085] It should be understood that the quality of the snapshot directly affects the validity of the evidence. If the snapshot range is too narrow, key frames for judgment may be missed; if it is too wide, redundant information will be included. The system should combine the identification features of the hit items and set a reasonable snapshot buffer, such as adding a few seconds of tolerance before and after the start and end times, to ensure that the snapshot can fully present the context in which the behavior occurred, thereby enhancing the credibility of the evidence.

[0086] Step C40: Mark the image feature regions that match the hit items in the video intervals corresponding to each snapshot address to obtain the marked content.

[0087] It's important to note that the image feature region refers to a local area in the snapshot video directly related to the hit judgment. For example, for the deduction item "test tube opening to person," the feature region is the spatial location of the test tube opening and the human face. The annotation operation in this step involves marking these regions on the snapshot image with rectangles, markers, or highlight colors, and generating corresponding coordinate information or text descriptions. The annotation content refers to this annotated image identification data, used to visually demonstrate the basis of the model's judgment.

[0088] Understandably, the purpose of annotation is to help non-technical personnel understand why the AI ​​model determined the hit to be true. For example, in the snapshot of the score item "correctly shaking the reagent bottle," the annotation would circle the student's hands holding the bottle and the shaking effect of the liquid inside. The annotation content can be stored as a vector graphic overlaid on the video frame or saved separately as an annotation coordinate file. In this way, when viewing the snapshot in the chain of evidence, the key areas are immediately apparent, avoiding blind searching.

[0089] It should be understood that the labeled content needs to be associated with the snapshot address to ensure synchronous display within the same chain of evidence. The accuracy and completeness of the labeling directly reflect the interpretive ability of the AI ​​model. Discrepancies in labeling may lead to misjudgments or disputes. Therefore, the labeling algorithm in this step should be based on the same-layer feature mapping of the multimodal AI model from the preceding steps, ensuring that the labeled location is consistent with the feature definition in the hit rules, thereby improving the confidence of the chain of evidence.

[0090] Step C50: Combine the snapshot address and the annotation content to generate a hit basis description for each hit item.

[0091] It should be noted that the "hit evidence description" refers to the textual explanation of why a particular item was judged as a hit, and it is usually automatically generated by combining the snapshot address and the annotation content. For example, for the deduction item "the alcohol lamp was not covered after use," the evidence description could be "Within the timestamp interval, the video snapshot shows that the alcohol lamp flame was not extinguished, and the lamp cap was not placed in the lamp opening; the annotation area marks the location of the uncovered lamp opening." The description generated in this step integrates visual evidence and judgment logic into readable text.

[0092] Understandably, the description of the hit evidence is the core user-facing output of the evidence chain, replacing the original model output probabilities or feature vectors, making the scoring results more comprehensible. For example, the description can directly quote coordinates or region descriptions from the annotations, such as "Annotation box A shows the test tube opening pointing towards the experimenter's head." This allows both teachers and students to quickly understand the specific reasons behind the scoring conclusion based on the description.

[0093] Step C60: Combine the timestamp interval, hit basis description and snapshot address of each hit item to obtain the evidence chain of the scoring result.

[0094] It should be noted that association and combination refers to packaging the three data items corresponding to each hit item—timestamp interval, hit evidence description, and snapshot address—into a structured record according to a one-to-one correspondence. For example, an evidence chain unit for a hit item includes: start and end time "at frame YY of second XX in the video", description text "the student was not wearing gloves", and snapshot address " / evidence / 0723.png". All evidence chain units for hit items are arranged in order or chronological order according to the scoring matching results, ultimately forming a complete scoring result evidence chain.

[0095] Understandably, the chain of evidence can be used for internal archiving as well as for user access. For example, the chain of evidence for a chemistry experiment's scoring results might contain three pieces of evidence for scoring points and two pieces of evidence for deducting points, each clearly labeled with the time, reason, and screenshot. This structured approach facilitates subsequent statistical analysis or batch export, while also supporting manual review of each item.

[0096] It should be understood that the completeness of the evidence chain directly determines the traceability of the scoring system. The absence of any single element may make the scoring conclusion difficult to verify. Therefore, this step requires completing the association of all matching items without omission, and using a verification mechanism to ensure that the three data items in each evidence chain unit are not empty and point to the correct values. The final generated evidence chain should be a closed data set, usable independently of the original video, serving as strong evidence for the scoring results.

[0097] Step S303: Based on the matching items in the scoring results, retrieve the preset weight values ​​from the scoring rules associated with the experimental steps and sum them to obtain the scoring result of the video segment.

[0098] It should be noted that the scoring matching result in this step refers to the list of hit items output in the preceding steps, including the hit scoring items and the hit deduction items. A hit item refers to an operation judged as correct or incorrect. The scoring rules associated with the experimental step include a preset weight value for each scoring item or deduction item. The weight value can be understood as a positive or negative score, used to quantify the value of a correct operation or the severity of the deduction.

[0099] Understandably, the summation operation accumulates the preset weight values ​​of all hit items. For example, if there are two hit items under the step "Connecting the Circuit," with "Correctly Connecting the Wire" having a weight of +1 and "Short Circuit Not Disconnected in Time" having a weight of -2, then the summation yields -1 as the local score for that step. The score result is the numerical value obtained after weighting the video segment.

[0100] It should be understood that the final video segment score is only a part of the total score. The scores for all experimental steps must be aggregated to obtain the final score for the entire experiment. The weighting of values ​​must be based on the experimental teaching requirements; for example, the deduction weight for critical safety matters should be higher than that for general operational errors. Through this weighted summation method, the score reflects both the standardization of the operation and the differences in importance between different evaluation items.

[0101] This embodiment acquires experimental video clips and their associated scoring rules, then uses a multimodal AI model to match and identify scoring and deduction items in the video, obtaining a scoring matching result that includes both hit and miss items. Subsequently, it extracts the video timestamp, the description of the basis for the hit, and the snapshot address for each hit item, aggregating them into a traceable chain of scoring result evidence. Finally, it sums the hit items according to preset weight values ​​to obtain the scoring result for the video clip. This scheme achieves a fully automated closed loop from video input to scoring output.

[0102] In summary, this technical solution improves the accuracy of scoring by leveraging the multi-dimensional recognition capabilities of a multimodal AI model; enhances the traceability and interpretability of scoring results by generating a complete chain of evidence including timestamps, snapshots, and textual descriptions, facilitating manual review and teaching feedback; and reasonably reflects the differences in importance among evaluation items through a weighted summation mechanism, thereby reducing the workload of manual scoring and improving the objectivity and efficiency of experimental teaching evaluation.

[0103] Based on the first embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 In the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to produce an evidence chain, step S50 includes steps S501 to S505: Step S501: Based on the segment to be reviewed, obtain the video segment corresponding to the segment to be reviewed, the evidence chain of the scoring result, and the user terminal operation records of the time interval where the video segment is located.

[0104] It should be noted that the segments to be reviewed refer to video sections that, after automatic scoring, are marked by the system or manually as requiring further review. The chain of evidence for the scoring results is a set of structured data generated by the aforementioned steps, including the timestamp range for each hit item, a description of the basis for the hit, and the corresponding snapshot address.

[0105] Understandably, user-side operation logs refer to the sequence of events automatically collected and reported by the client during the examination, reflecting the candidate's terminal behavior status. This includes violations such as window losing focus, network disconnection, exiting full-screen mode, and periodic heartbeat signals. For example, if the automatic scoring has doubts about the "extinguishing the lamp cap" action in the "using the alcohol lamp" step, the system will capture that video segment, extract the corresponding evidence chain (such as snapshots and descriptions), and retrieve operation logs from that time period to see if the candidate's actions were affected by screen switching or network disconnection.

[0106] Step S502: Display the video clip, the scoring result evidence chain, and the user terminal operation record on the administrator's end.

[0107] It should be noted that the administrator interface refers to the user interface specifically for reviewers, typically presented as a standalone system via a webpage or client. The video clip is the original exam recording corresponding to the segment to be reviewed, supporting playback, pause, and frame-by-frame skipping functions. Displaying these three types of information together provides administrators with a comprehensive review view. The video footage allows administrators to directly observe the student's actual actions, the evidence chain demonstrates how the automated scoring system judges each action, and the operation log reveals whether there were any abnormalities or violations on the student's terminal. These three elements complement each other. For example, when the automated scoring determines "incorrect reagent addition order," the administrator can simultaneously view the student's actual operation sequence in the video, screenshots and descriptions in the evidence chain, and whether there were any submission delays due to network latency in the operation log, thereby determining the accuracy of the automated scoring.

[0108] Step S503: The administrator confirms or corrects the hit scoring items and deduction items in the segment to be reviewed item by item according to the displayed content, and obtains the hit item review result. The non-hit items in the segment to be reviewed are re-evaluated and the non-hit item review result is obtained.

[0109] Understandably, segments awaiting review are video sections marked in the automatic scoring system due to insufficient confidence or rule disputes. Hit items (points awarded and points deducted) refer to specific action points successfully identified and judged by the automatic scoring system within that segment. The review result for hit items refers to the final confirmation or correction conclusion made by the administrator after item-by-item verification for each hit record. Missed items refer to key actions that the automatic scoring system failed to identify within that segment but which were actually performed by the examinee. The review result for missed items is a supplementary judgment made by the administrator after re-evaluation based on the video segment, evidence chain, and operation records.

[0110] It should be understood that the administrator's workflow involves two parallel tasks. First, the administrator reviews each correct score and deduction: They view the description, snapshot, and timestamp of each correct score in the interface, and assess the accuracy of the automatic judgment by reviewing the video playback. For example, if the automatic scoring deducts points based on "the student pointing the test tube opening towards themselves," but the administrator finds after playing the video that the student was actually facing the wall, the deduction can be corrected, and the reason for the correction recorded. Second, the administrator re-evaluates incorrect scores: They actively scan all action points in the video clip, compare them to the scoring criteria, and add any key actions missed by the automatic system. For example, if the system fails to recognize the step of "using filter paper," the administrator can add a score item after confirming it through the video and record its timestamp and basis.

[0111] Step S504: Based on the review results of the hit items and the review results of the missing items, generate a segment review score and a corresponding review description.

[0112] It should be noted that the "hit item review result" refers to the final set of judgments after the administrator confirms or corrects each of the scoring and deduction items identified in the automatic scoring. The "missed item review result" is the set of judgments supplemented by the administrator after re-evaluating key actions missed by the automatic system. The segment review score refers to the comprehensive score calculated for the segment to be reviewed after adding the scoring items and subtracting the deduction items from the above two types of results. The review description is a textual explanation of all corrections and additions to the judgment records during the segment scoring process, including the timestamp of each change, the original automatic result, the administrator's correction result, and the basis for the correction. For example, in a chemistry experiment segment, the administrator corrects the "unextinguished flame" which was automatically judged as a deduction to not deducting points and notes the reason "the video shows that the student extinguished the flame, but the angle caused the automatic recognition to miss it." At the same time, a score item "using a glass rod to guide the flow" that was automatically missed is added, and the final score for the segment is calculated as full marks minus one actual violation deduction.

[0113] Understandably, the process of generating segment review scores and review explanations essentially transforms the administrator's item-by-item review results into structured scoring data and a traceable decision log. Segment review scores should not only include the final score but also reflect the extent of change in the segment relative to the original automatic score. The review explanations should record each change in detail so that the scoring basis can be clearly understood during subsequent audits or inquiries.

[0114] Step S505: Combine the segment review score with the review description to obtain the review score result.

[0115] Understandably, this combined operation represents the final output packaging stage of the manual review process. It transforms the fragmented, item-by-item review operations into a structured, deliverable conclusion. The review scoring results are not only used to replace or correct the original automatic scores but also stored as high-quality labeled samples in the training database to optimize subsequent automatic recognition models. For example, if the description field in multiple review scoring results frequently mentions "angle occlusion causing recognition omissions," the system can adjust the image recognition algorithm's judgment strategy for occluded scenes accordingly. Through this combination, the score and description corroborate each other. When any subsequent reviewer consults the result, they can quickly obtain the scoring conclusion and simultaneously understand the logical context of the scoring decision, effectively avoiding controversial issues that cannot be explained by the score alone.

[0116] In this embodiment, the video segment corresponding to the segment to be reviewed, the evidence chain of the scoring results, and the user's operation records are automatically acquired and centrally displayed on the administrator's end, providing reviewers with a panoramic review view. Subsequently, the administrator confirms or corrects each of the hit scoring and deduction items based on the displayed content, generating a review result for the hit items. Simultaneously, items missed by the automatic system are re-evaluated, generating a review result for the missing items. Next, the scoring items are added together and the deduction items are subtracted based on the two types of review results to generate a segment review score. The original result, correction basis, and timestamp for each change are recorded as a review explanation. Finally, the segment review score and review explanation are combined into a structured review score result, which is used both to replace or correct the original automatic score and as a high-quality labeled sample stored in the training database.

[0117] In summary, this embodiment unifies the presentation of video footage, evidence chains, and operation logs, enabling administrators to quickly locate points of contention and make accurate judgments, significantly reducing the risk of misjudgments due to fragmented information. The administrator's item-by-item correction and supplementation of both hit and miss items not only improves the scoring accuracy of individual segments but also fully preserves the decision-making logic through review explanations, ensuring sufficient traceability of the scoring results. Furthermore, the generated review scoring results, as high-quality labeled data, are fed back to the automatic scoring model, guiding the algorithm to identify common misjudgment scenarios, such as angle obstruction or abnormal operation logs, thereby continuously improving the overall reliability of the system's automatic scoring and forming a virtuous cycle of mutual optimization between human and artificial intelligence.

[0118] like Figure 4 As shown, Figure 4 This is a timeline diagram of the entire process of the intelligent evaluation method for producing an evidence chain based on the combination of AI multimodality and experimental rules, as described in this invention.

[0119] This system consists of four interactive entities: teacher / administrator end, student end, backend system, and AI analysis service. The overall evaluation process is divided into four stages: course and rule configuration, exam execution and monitoring, AI analysis and adjudication, and scoring and review. Each stage achieves data flow and status synchronization through standardized interfaces.

[0120] Phase 1: Course and rule configuration is initiated by the teacher / administrator. The teacher first creates an experimental course and configures the course structure on the management side. Then, the experimental process is broken down into a three-level structure of major steps, sub-steps, and observable scoring points. For each sub-step, the completion standard, AI supplementary explanation, and scoring point score are configured. After the configuration is completed, the course version is published. The backend system persists the course rules and generates a version identifier.

[0121] Phase Two: Exam Execution and Monitoring. Teachers create classroom exercises and distribute exam information and rules. Students obtain an ExamToken, initialize their cameras, enter standardized recording mode, and continuously report heartbeats and abnormal events. During the exam, teachers monitor students' status in real-time via SSE. After students complete the experiment, recording ends, the client verifies the MP4 file's validity, streams it to OSS, and finally submits the exam record. During the exam process (looped execution), students report their running status at preset intervals. The backend system only updates the database and records monitoring events when the status changes, reducing database write frequency.

[0122] Phase 3: AI analysis and adjudication are asynchronously triggered by the backend system after exam submission. The backend first reads the exam record, course version, and scoring rules to generate a time-limited video URL accessible to AI. The AI ​​analysis service calls a multimodal model to perform video semantic detection, step visibility / semantic verification, and scoring point hit adjudication, outputting the start and end time, completion status, hit scoring points, risk items, confidence level, and evidence fragments for each step. After receiving the AI ​​analysis results, the backend performs deterministic scoring based on the scoring point values ​​and persists the step scores, evidence chains, and scoring operation records to the database.

[0123] Phase Four: Scoring and Review. In this phase, the backend system distributes review requests based on the adjudication status, visibility check results, and confidence level. Steps with sufficient confidence and complete visibility are automatically approved; steps marked as uncertain, partial / blocked, or failing semantic checks are marked as pending review. Teachers view the grade list and details drawer, locate video segments based on timestamp intervals and snapshots in the evidence chain, confirm or modify the steps pending review or automatically approved steps, and submit review explanations. The backend verifies the reviewVersion to prevent concurrent overwriting, re-summarizes the total score, and records the review version. Finally, the system generates exam reports and statistical analysis, supporting grade export and historical tracking.

[0124] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the experimental intelligent evaluation method of this application based on the combination of AI multimodal and experimental rules to produce evidence chains. Any simple modifications based on this technical concept are within the protection scope of this application.

[0125] This application also provides an intelligent experimental evaluation device that generates a chain of evidence based on the combination of AI multimodal and experimental rules. Please refer to [reference needed]. Figure 5 The experimental intelligent evaluation device based on the combination of AI multimodal and experimental rules to generate a chain of evidence includes: The video preprocessing module 10 is used to acquire operation video data uploaded by the user terminal, as well as the experimental process information corresponding to the operation video data; The video preprocessing module 10 is also used to split the experimental operation video into video segments corresponding to each experimental step based on preset experimental process information. The experimental operation evaluation module 20 is used to score each video segment according to the scoring rules associated with the corresponding experimental steps, and obtain the scoring result of the video segment and the evidence chain of the scoring result. The experimental operation evaluation module 20 is also used to obtain the confidence level of the scoring result based on the evidence chain of the scoring result; The review processing module 30 is used to mark video segments whose confidence level of the scoring result is lower than a preset confidence threshold as segments to be reviewed, and push the segments to be reviewed to the administrator terminal, so that the administrator terminal can perform segment review and obtain review scoring results; The result summary module 40 is used to summarize the scoring results and the review scoring results to obtain a comprehensive scoring result.

[0126] In one embodiment, the experimental operation evaluation module 20 is further configured to: invoke a multimodal AI model to perform matching and identification of scoring items and deduction items according to the scoring rules associated with the video segment and the experimental steps, and obtain a scoring matching result, wherein the scoring matching result includes hit scoring items, hit deduction items, and no hit items; extract the video timestamp interval, hit basis description, and snapshot address corresponding to each hit item according to the scoring matching result, and aggregate them to obtain a scoring result evidence chain; and retrieve preset weight values ​​from the scoring rules associated with the experimental steps according to the hit items in the scoring matching result and sum them to obtain the scoring result of the video segment.

[0127] In one embodiment, the experimental operation evaluation module 20 is further configured to obtain the recognition features and hit rules corresponding to each scoring item according to the scoring rules associated with the experimental steps, wherein the scoring items include scoring items and deduction items; locate the corresponding video time interval from the video segment according to the recognition features to obtain the video segment to be analyzed corresponding to each scoring item; input the video segment to be analyzed and the corresponding hit rules into the multimodal AI model for hit analysis to obtain the preliminary recognition results of each evaluation item; and summarize and determine the hit scoring items, hit deduction items and non-hit items according to the preliminary recognition results of all evaluation items to obtain the scoring matching results.

[0128] In one embodiment, the experimental operation evaluation module 20 is further configured to: extract the scoring items and deducting items from the matched scores to obtain a set of matched items; extract the start and end times of each matched item in the video segment to obtain the timestamp interval of each matched item; cut and store the video segment according to the timestamp interval of each matched item to obtain a snapshot address of each matched item; annotate the image feature regions in the video segment corresponding to each snapshot address that match the matched items to obtain annotation content; generate a matching basis description for each matched item by combining the snapshot address and the annotation content; and associate and combine the timestamp interval, matching basis description, and snapshot address of each matched item to obtain the evidence chain of the scoring result.

[0129] In one embodiment, the video preprocessing module 10 is further configured to: obtain key visual recognition points and step duration features corresponding to each experimental step according to preset experimental process information; perform process step matching on the operation video data according to the key visual recognition points to obtain the video time interval corresponding to the successfully matched experimental step; perform reverse estimation of the time interval for the missing experimental step that failed feature recognition according to the video time interval of the successfully matched experimental step and the step duration features to obtain the video time interval of the missing experimental step; and segment the operation video data according to the video time interval of all steps to obtain video segments corresponding to each experimental step.

[0130] In one embodiment, the experimental operation evaluation module 20 is further configured to extract the snapshot address and annotation content corresponding to each hit item according to the evidence chain of the scoring results, and obtain the associated evidence of each hit item; evaluate the degree of matching between the annotation area and the preset recognition target according to the associated evidence of each hit item, and obtain the individual confidence score of each hit item; and obtain the confidence level of the scoring result of the video segment scoring result according to the individual confidence scores of all hit items and the influence weight of the non-hit items in the evidence chain.

[0131] In one embodiment, the review processing module 30 is further configured to: obtain, based on the segment to be reviewed, the video segment corresponding to the segment to be reviewed, the scoring result evidence chain, and the user terminal operation records for the time interval of the video segment; display the video segment, the scoring result evidence chain, and the user terminal operation records on the administrator terminal; have the administrator terminal confirm or correct the hit scoring items and deduction items in the segment to be reviewed item by item according to the displayed content, to obtain the hit item review results; re-evaluate the miss items in the segment to be reviewed item, to obtain the miss item review results; generate a segment review score and a corresponding review description based on the hit item review results and the miss item review results; and combine the segment review score and the review description to obtain the review score result.

[0132] The experimental intelligent evaluation device based on the combination of AI multimodal and experimental rules to generate an evidence chain provided in this application adopts the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain in the above embodiments. It can solve the technical problem of how to establish a traceable association between the scoring results and the original operation records, and quantitatively evaluate and verify the confidence of the scoring, so as to make the scoring process interpretable and controllable. Compared with the prior art, the beneficial effects of the experimental intelligent evaluation device based on the combination of AI multimodal and experimental rules to generate an evidence chain provided in this application are the same as the beneficial effects of the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain provided in the above embodiments. Moreover, other technical features in the experimental intelligent evaluation device based on the combination of AI multimodal and experimental rules to generate an evidence chain are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0133] This application provides an intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules. The intelligent experimental evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules includes: at least one processor; and a memory communicatively connected to at least one processor; wherein the memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to execute the intelligent experimental evaluation method that generates an evidence chain based on the combination of AI multimodal and experimental rules in the above embodiment 1.

[0134] The following is for reference. Figure 6 This document illustrates a structural schematic diagram of an intelligent experimental evaluation device suitable for implementing embodiments of this application, which generates an evidence chain based on the combination of AI multimodal and experimental rules. The intelligent experimental evaluation device in this application, which generates an evidence chain based on the combination of AI multimodal and experimental rules, can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 6 The experimental intelligent evaluation device shown is merely an example of the evidence chain generated by combining AI multimodal and experimental rules, and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0135] like Figure 6As shown, the experimental intelligent evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1002 or the program loaded from the storage device 1003 into the random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the experimental intelligent evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the experimental intelligent evaluation device, which generates an evidence chain based on the combination of AI multimodal and experimental rules, to exchange data with other devices wirelessly or via wired means. Although the figure shows an experimental intelligent evaluation device with various systems that generate an evidence chain based on the combination of AI multimodal and experimental rules, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0136] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0137] The experimental intelligent evaluation device based on the combination of AI multimodal and experimental rules to generate an evidence chain provided in this application adopts the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain in the above embodiments. It can solve the technical problem of how to establish a traceable association between the scoring results and the original operation records, and quantitatively evaluate and verify the confidence of the scoring, so as to make the scoring process interpretable and controllable. Compared with the prior art, the beneficial effects of the experimental intelligent evaluation device based on the combination of AI multimodal and experimental rules to generate an evidence chain provided in this application are the same as the beneficial effects of the experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain provided in the above embodiments. Moreover, other technical features in the experimental intelligent evaluation device based on the combination of AI multimodal and experimental rules to generate an evidence chain are the same as the features disclosed in the previous embodiment method, and will not be repeated here.

[0138] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0139] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0140] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the experimental intelligent evaluation method based on the combination of AI multimodality and experimental rules to produce a chain of evidence in the above embodiments.

[0141] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0142] The aforementioned computer-readable storage medium may be included in an experimental intelligent evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules; or it may exist independently and not be assembled into an experimental intelligent evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules.

[0143] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by an experimental intelligent evaluation device that generates an evidence chain based on the combination of AI multimodal and experimental rules, the experimental intelligent evaluation device generates an evidence chain based on the combination of AI multimodal and experimental rules: acquires operation video data uploaded by the user terminal, and experimental process information corresponding to the operation video data; based on the preset experimental process information, splits the experimental operation video into video segments corresponding to each experimental step; for each video segment, scores it according to the scoring rules associated with the corresponding experimental step, obtaining a scoring result and a scoring result evidence chain for the video segment; obtains the confidence level of the scoring result based on the scoring result evidence chain; marks video segments whose confidence level is lower than a preset confidence threshold as segments to be reviewed, and pushes the segments to be reviewed to the administrator terminal, whereby the administrator terminal performs segment review to obtain a review scoring result; and summarizes the scoring result and the review scoring result to obtain a comprehensive scoring result.

[0144] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0145] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0146] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0147] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the aforementioned intelligent experimental evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain. This solves the technical problem of how to establish a traceable correlation between scoring results and original operation records, and how to quantitatively evaluate and verify the confidence level of the scores, thereby making the scoring process interpretable and controllable. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the intelligent experimental evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain provided in the above embodiments, and will not be repeated here.

[0148] The computer program product provided in this application can solve the technical problem of intelligent experimental evaluation based on the combination of AI multimodal and experimental rules to generate evidence chains. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the intelligent experimental evaluation method based on the combination of AI multimodal and experimental rules to generate evidence chains provided in the above embodiments, and will not be repeated here.

[0149] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. An intelligent evaluation method for experiments based on the combination of AI multimodal and experimental rules to generate a chain of evidence, characterized in that, The method includes: Obtain the operation video data uploaded by the user terminal, as well as the experimental process information corresponding to the operation video data; Based on the preset experimental procedure information, the experimental operation video is divided into video segments corresponding to each experimental step. For each video segment, a score is assigned according to the scoring rules associated with the corresponding experimental steps to obtain the score result of the video segment and the chain of evidence for the score result; Based on the chain of evidence for the scoring results, the confidence level of the scoring results is obtained; Video segments whose confidence level of the scoring result is lower than the preset confidence threshold are marked as segments to be reviewed, and the segments to be reviewed are pushed to the administrator terminal, which performs segment review and obtains the review scoring result. The comprehensive score result is obtained by summarizing the scoring results and the review scoring results.

2. The experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain as described in claim 1, characterized in that, For each video segment, a score is assigned according to the scoring rules associated with the corresponding experimental step to obtain the score result and the chain of evidence for the score result, including: Based on the scoring rules associated with the video clip and the experimental steps, a multimodal AI model is invoked to match and identify scoring items and deduction items to obtain scoring matching results. The scoring matching results include items that are scored, items that are deducted, and items that are not scored. Based on the scoring matching results, the video timestamp range, the description of the basis for the hit, and the snapshot address corresponding to each hit item are extracted and aggregated to obtain the evidence chain of the scoring results. Based on the matching items in the scoring results, preset weight values ​​are retrieved from the scoring rules associated with the experimental steps and summed to obtain the scoring result of the video segment.

3. The experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain as described in claim 2, characterized in that, The step of calling a multimodal AI model to match and identify scoring items and deduction items based on the scoring rules associated with the video clip and the experimental steps, and obtaining the scoring matching result, includes: Based on the scoring rules associated with the experimental steps, the recognition features and hit rules corresponding to each scoring item are obtained. The scoring items include scoring items and deduction items. Based on the identification features, the corresponding video time interval is located from the video segment to obtain the video segment to be analyzed for each scoring item. The video segment to be analyzed and the corresponding hit rules are input into the multimodal AI model for hit analysis to obtain the preliminary recognition results of each evaluation item; Based on the preliminary identification results of all evaluation items, the matching scores, the deductions, and the missing items are summarized to obtain the scoring matching results.

4. The experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain as described in claim 2, characterized in that, Based on the scoring matching results, the video timestamp range, the description of the basis for the match, and the snapshot address corresponding to each hit item are extracted and aggregated to obtain the scoring result evidence chain, including: Based on the scoring matching results, extract the hit scoring items and the hit deduction items to obtain the hit item set; Extract the start and end times of each hit item in the video segment to obtain the timestamp interval of each hit item; Based on the timestamp interval of each hit item, the video segment is extracted and stored to obtain the snapshot address of each hit item; The image feature regions that match the hit items in the video intervals corresponding to each snapshot address are labeled to obtain the labeled content; The snapshot address and the labeled content are used to generate a description of the basis for each hit item; The evidence chain for the scoring result is obtained by associating and combining the timestamp interval, the description of the basis for the hit, and the snapshot address of each hit item.

5. The experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain according to claim 1, characterized in that, Based on the preset experimental procedure information, the experimental operation video is divided into video segments corresponding to each experimental step, including: Based on the pre-set experimental procedure information, the key visual recognition points and step duration features corresponding to each experimental step are obtained. Based on the key visual recognition points, the operation video data is matched with process steps to obtain the video time intervals corresponding to the successfully matched experimental steps. Based on the video time interval of the successfully matched experimental steps and the duration characteristics of the steps, the time interval of the missing experimental steps that failed to pass feature recognition is estimated by reverse calculation to obtain the video time interval of the missing experimental steps. The operation video data is segmented according to the video time interval of all steps to obtain video segments corresponding to each experimental step.

6. The experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain according to claim 1, characterized in that, The step of obtaining the confidence level of the scoring result based on the evidence chain of the scoring result includes: Based on the evidence chain of the scoring results, the snapshot address and annotation content corresponding to each hit item are extracted to obtain the associated evidence of each hit item; Based on the correlation evidence of each hit item, the degree of match between the labeled area and the preset identification target is evaluated, and the individual confidence score of each hit item is obtained. The confidence level of the video segment scoring result is obtained based on the individual confidence scores of all hit items and the influence weights of the missing items in the evidence chain.

7. The experimental intelligent evaluation method based on the combination of AI multimodal and experimental rules to generate an evidence chain according to claim 1, characterized in that, The process involves marking video segments whose confidence level is lower than a preset confidence threshold as segments requiring review, and pushing these segments to the administrator for review. The administrator then performs the review to obtain the review score, including: Based on the segment to be reviewed, obtain the video segment corresponding to the segment to be reviewed, the evidence chain of the scoring result, and the user terminal operation records of the time interval where the video segment is located; The administrator displays the video clip, the chain of evidence for the scoring results, and the user's operation records on the administrator's end. The administrator confirms or corrects the hit scoring items and deduction items in the segment to be reviewed based on the displayed content, and obtains the hit item review results. The non-hit items in the segment to be reviewed are re-evaluated, and the non-hit item review results are obtained. Based on the review results of the hit items and the review results of the missing items, a segment review score and a corresponding review description are generated. The review score of the segment is combined with the review description to obtain the review score result.

8. An intelligent experimental evaluation device that generates a chain of evidence based on the combination of AI multimodal and experimental rules, characterized in that, The device includes: The video preprocessing module is used to acquire operation video data uploaded by the user terminal, as well as the experimental process information corresponding to the operation video data; The video preprocessing module is also used to split the experimental operation video into video segments corresponding to each experimental step based on preset experimental process information. The experimental operation evaluation module is used to score each video segment according to the scoring rules associated with the corresponding experimental steps, and to obtain the scoring result of the video segment and the evidence chain of the scoring result. The experimental operation evaluation module is also used to obtain the confidence level of the scoring result based on the evidence chain of the scoring result; The review processing module is used to mark video segments whose confidence level of the scoring result is lower than a preset confidence threshold as segments to be reviewed, and push the segments to be reviewed to the administrator terminal, whereby the administrator terminal performs segment review and obtains the review scoring result; The results summary module is used to summarize the scoring results and the review scoring results to obtain a comprehensive scoring result.

9. An intelligent experimental evaluation device that generates a chain of evidence based on the combination of AI multimodal and experimental rules, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the experimental intelligent evaluation method for generating a chain of evidence based on the combination of AI multimodal and experimental rules as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, they implement the experimental intelligent evaluation method as described in any one of claims 1 to 7, which generates an evidence chain based on the combination of AI multimodal and experimental rules.

Images