Assessment and evaluation method for virtual maintenance training process

Abstract

The invention relates to the technical field of intelligent assessment management, and discloses an assessment method for a virtual maintenance training process, and the method comprises the following steps: collecting real-time operation data in the maintenance training process through a virtual reality device, and the real-time operation data comprises a gesture track, a tool use sequence and an environment interaction log; performing noise filtering and format unification processing on the real-time operation data to generate a standardized operation data stream; key operation nodes in the standardized operation data flow are extracted based on a maintenance knowledge graph, and operation feature vectors are generated; according to the invention, systematic fusion and standardization processing are carried out on multi-source heterogeneous real-time operation data, and an objective evaluation benchmark is constructed based on a maintenance knowledge graph, so that comprehensive quantification and description of an operation process can be realized; the problems that the evaluation dimension is single and the real skill level is difficult to comprehensively reflect due to the dependence on subjective judgment and a single index are solved, and the objectivity and comprehensiveness of assessment are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent assessment and management technology, specifically to an assessment and evaluation method for a virtual maintenance training process. Background Technology

[0002] Virtual maintenance is a comprehensive application technology that combines virtual reality and computer technology. It simulates the maintenance process by driving digital prototypes and 3D human models in a computer-generated environment, realizing human-computer interaction. Its core objectives include disassembly and assembly operation guidance, status data extraction, and maintenance performance verification, breaking through the spatial and temporal limitations of equipment maintenance.

[0003] Currently, the assessment of virtual maintenance training relies on manual observation and simple rule-based judgment, which makes it impossible to deeply integrate and quantify the large amount of multi-dimensional real-time operational data generated during the trainees' operation. This results in a single assessment dimension, making it difficult to fully reflect the trainees' true skill level.

[0004] Therefore, a method for assessing and evaluating the virtual maintenance training process is proposed to address the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an assessment method for virtual maintenance training processes, which solves the problem mentioned in the background that the assessment dimensions are too singular and cannot fully reflect the trainees' true skill level.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for assessing and evaluating a virtual maintenance training process, the method comprising the following steps: S1. Collect real-time operation data during the maintenance training process using virtual reality equipment, including gesture trajectories, tool usage sequences, and environmental interaction logs; S2. Perform noise filtering and format unification processing on the real-time operation data to generate a standardized operation data stream; S3. Extract key operation nodes from the standardized operation data stream based on the maintenance knowledge graph, and generate operation feature vectors; S4. Utilize a deep learning model to evaluate the operation feature vector in multiple dimensions, including operation accuracy, efficiency, and security, and output the original evaluation score. S5. Combine the trainees' historical training records and the expert scoring database to dynamically weight and calibrate the original assessment scores, and generate a comprehensive assessment result. S6. Based on the comprehensive assessment results, construct a personalized feedback report, including analysis of weaknesses and training suggestions; S7. Based on the analysis of weaknesses in the personalized feedback report, the subsequent training content is dynamically optimized through the adaptive learning engine, and the evaluation threshold is updated according to the comprehensive assessment results, forming an adaptive adjustment loop of training content based on assessment result feedback.

[0007] Preferably, the real-time operation data acquisition in step S1 includes the following steps: S11. The inertial measurement unit and optical sensors built into the head-mounted virtual reality device collect data on the trainee's hand movements, tool gripping force, and gaze focus coordinates in the virtual maintenance scenario. S12. Use an environment simulator to record the interaction events between the trainee and the virtual device, including the disassembly sequence, the accuracy of parts installation, and the number of incorrect operations; S13. The hand movement data, tool gripping force, gaze focus coordinates and interaction events are synchronized and integrated according to timestamps to generate a time-seriesd real-time operation data matrix.

[0008] Preferably, the processing of real-time operation data in step S2 includes the following steps: S21. The gesture trajectory data in the real-time operation data matrix is ​​smoothed by using a wavelet transform algorithm to remove high-frequency noise caused by device vibration. S22. Using the Z-score standardization method, tools with different dimensions are converted into a uniform numerical range using sequence data and environmental interaction logs, thus eliminating differences in magnitude. S23. The processed data stream is segmented and recombined using time window sliding technology to generate a standardized operation data stream based on operation cycles. Each data segment contains a complete maintenance action unit.

[0009] Preferably, extracting key operation nodes in S3 includes the following steps: S31. Construct a maintenance operation knowledge graph, where nodes represent maintenance steps, edges represent logical dependencies between steps, and standard operation time and common error types are labeled. S32. Use a graph matching algorithm to align the standardized operation data stream with the maintenance operation knowledge graph to identify key nodes in the trainee's operation, including the initial action, core disassembly steps and final inspection points; S33. Extract feature parameters based on key nodes, including operation path length, tool switching frequency, and error recovery time, and encode them into fixed-dimensional operation feature vectors.

[0010] Preferably, the multi-dimensional evaluation using a deep learning model in step S4 includes the following steps: S41. Design a multi-task deep learning model in which the convolutional neural network branch processes spatial operation features and the long short-term memory network branch processes temporal operation features. S42. Input the operation feature vector into the multi-task deep learning model, and calculate the operation accuracy score, efficiency score and security score respectively. S43. The multi-task outputs are fused through a fully connected layer and normalized to the original evaluation scores in the 0-1 range using the Sigmoid function.

[0011] Preferably, designing a multi-task deep learning model in step S41 includes the following sub-steps: S411. An adversarial training technique is adopted, which simulates the generation of abnormal operation data through the generator of the generative adversarial network, and uses a discriminator to distinguish between real operation data and generated abnormal operation data, and updates the parameters of the deep learning model in an iterative manner. S412. Introduce a multi-scale feature extraction mechanism, use dilated convolutional layers to capture long-range dependencies in the operation sequence, and combine local attention networks to calculate the weight distribution of detailed operation features. S413. Conduct model interpretability analysis and generate heat map data corresponding to key decision-making areas through gradient weighted class activation mapping technology; S414. Establish a closed loop for model performance monitoring, detect the consistency deviation between the evaluation score and the manual review result in real time, and call the model retraining module to update the network parameters when the deviation exceeds the preset threshold.

[0012] Preferably, the dynamic weighted calibration in S5 includes the following steps: S51. Establish an expert rating database to store expert rating records for typical operation modes under different maintenance scenarios, and associate them with operation complexity coefficients. S52. Retrieve trainees’ historical training records and extract recent assessment trends and progress as dynamic weighting factors; S53. Use a Bayesian inference algorithm to perform probability fusion of the original evaluation scores and the expert scoring database, and combine the dynamic weighting factors to calculate the comprehensive assessment result after confidence adjustment.

[0013] Preferably, the probability fusion and calculation using the Bayesian inference algorithm in step S53 includes the following sub-steps: S531. Construct a time decay weight function so that the influence of recent training records on the weighting factor is higher than that of earlier records, where the decay coefficient is dynamically adjusted based on the student's learning curve. S532. Use a fuzzy logic system to handle subjective differences in expert scoring, transform qualitative evaluation into quantitative weights, and eliminate abnormal scores through a consistency test algorithm. S533 integrates a real-time feedback mechanism, introduces student self-assessment data during the calibration process, and uses a Bayesian update algorithm to dynamically correct weighted parameters; S534. Implement source tracing for calibration results, record the logical path and influencing factors of each weight adjustment, and support the auditing and review of evaluation results.

[0014] Preferably, the process of building a personalized feedback report in step S6 includes the following steps: S61. Based on the comprehensive assessment results, a decision tree model is used to identify the weak points in the trainees' operations, including incorrect tool selection, reversed order, and omission of key steps. S62. Generate hierarchical feedback content based on weak links, including error location, cause analysis, and demonstration video links; S63. Use natural language generation technology to convert feedback content into personalized text reports and add visual charts to display score distribution.

[0015] Preferably, optimizing the training content and updating the evaluation threshold in S7 includes the following steps: S71. Utilize reinforcement learning algorithms to build an adaptive learning engine, dynamically adjusting the difficulty and focus of subsequent virtual training tasks based on the learner's weaknesses. S72. Set up an evaluation threshold update mechanism to automatically recalibrate the parameter thresholds of the scoring model when the distribution of assessment scores among trainees changes. S73. Push the updated evaluation threshold and training content to the virtual reality device, and perform the next round of training and evaluation based on the new threshold and content to form a data feedback loop.

[0016] Compared with existing technologies, the present invention provides an assessment method for a virtual maintenance training process, which has the following beneficial effects: 1. In this invention, by systematically integrating and standardizing real-time operational data from multiple heterogeneous sources, and constructing an objective evaluation benchmark based on a maintenance knowledge graph, it is possible to achieve a comprehensive quantification and characterization of the operational process. This solves the problem of a single evaluation dimension and difficulty in fully reflecting the true skill level caused by reliance on subjective judgment and single indicators, thereby improving the objectivity and comprehensiveness of the assessment.

[0017] 2. In this invention, by introducing a dynamic weighted calibration mechanism, the evaluation results are personalized by combining individual historical performance and expert knowledge, and subsequent training content and evaluation standards are adaptively optimized in real time based on feedback. This enables the evaluation to dynamically adapt to the learning curves and skill levels of different learners, solving the problems of rigid evaluation standards and inability to achieve personalized guidance, and enhancing the pertinence of training and teaching effectiveness.

[0018] 3. In this invention, a multi-dimensional evaluation score is generated by a deep learning model to generate operational feature vectors, and a personalized feedback report containing analysis of weaknesses and training suggestions is constructed based on the comprehensive assessment results. This can accurately locate the root cause of operational errors and provide intuitive improvement guidance, solving the problem that errors are difficult to correct due to untimely or vague feedback, and improving training efficiency and the scientificity and credibility of assessment results. Attached Figure Description

[0019] Figure 1 This is a flowchart of an assessment method for a virtual maintenance training process according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] For specific implementation examples, please refer to: Figure 1 An assessment method for a virtual maintenance training process, comprising the following steps: S1. Collect real-time operation data during the maintenance training process using virtual reality equipment, including gesture trajectories, tool usage sequences, and environmental interaction logs; S2. Perform noise filtering and format unification processing on real-time operation data to generate a standardized operation data stream; S3. Extract key operation nodes from the standardized operation data stream based on the maintenance knowledge graph and generate operation feature vectors; S4. Utilize a deep learning model to evaluate the operation feature vector in multiple dimensions, including operation accuracy, efficiency, and security, and output the raw evaluation score. S5. Combine trainees’ historical training records and expert scoring database to dynamically weight and calibrate the original assessment scores to generate comprehensive assessment results; S6. Based on the comprehensive assessment results, construct a personalized feedback report, including analysis of weaknesses and training suggestions; S7. Based on the analysis of weaknesses in the personalized feedback report, the subsequent training content is dynamically optimized through the adaptive learning engine, and the evaluation threshold is updated according to the comprehensive assessment results, forming an adaptive adjustment loop of training content based on assessment result feedback.

[0022] The process of collecting real-time operation data in S1 includes the following steps: S11. The inertial measurement unit and optical sensors built into the head-mounted virtual reality device collect data on the trainee's hand movements, tool gripping force, and gaze focus coordinates in the virtual maintenance scenario. S12. Use an environment simulator to record the interaction events between the trainee and the virtual device, including the disassembly sequence, the accuracy of parts installation, and the number of incorrect operations; S13. Synchronize and integrate hand movement data, tool gripping force, gaze focus coordinates, and interaction events according to timestamps to generate a time-seriesd real-time operation data matrix.

[0023] Processing real-time operational data in S2 includes the following steps: S21. The gesture trajectory data in the real-time operation data matrix is ​​smoothed using a wavelet transform algorithm to remove high-frequency noise caused by device vibration, including the following steps: S211. Select the Daubechies wavelet basis function to perform multi-resolution decomposition on the gesture trajectory data, set the decomposition level to 5 levels, and extract high-frequency detail coefficients and low-frequency approximation coefficients. The wavelet decomposition formula is: ; in Represents wavelet coefficients, For scale parameters, For translation parameters, This is the original gesture trajectory data. These are the Daubechies wavelet basis functions; S212. A soft threshold denoising method is used for high-frequency detail coefficients, with the threshold set to 0.1 to suppress coefficient fluctuations caused by noise. The soft thresholding formula is: ; in For high-frequency detail coefficients, For the threshold, These are the denoised coefficients; S22. Using the Z-score standardization method, tools with different dimensions are converted into a uniform numerical range using sequence data and environmental interaction logs to eliminate magnitude differences, including the following steps: S221. The calculation tool uses the mean μ and standard deviation σ of the sequence data and the environmental interaction log, where μ is the arithmetic mean of the data sample and σ is the sample standard deviation. S222, For each data point Applying the standardization formula, the data is mapped to a distribution with a mean of 0 and a standard deviation of 1: ; in For the original data points, For standardized data points, The mean, Standard deviation; S223. Perform range pruning on the standardized data, limiting the values ​​to the range of [-3,3] to prevent extreme values ​​from affecting subsequent processing; S23. Using time window sliding technology, the processed data stream is segmented and reassembled to generate a standardized operation data stream based on operation cycles. Each data segment contains a complete maintenance action unit, including the following steps: S231. Set the sliding window length to 5 seconds and the step size to 1 second, and divide the processed data stream into windows; S232. Extract data segments within each window and calculate the statistical characteristics of the data within each segment, including mean, variance, and maximum value. S233. Based on statistical characteristics, perform overlap verification on data segments to ensure data consistency in overlapping parts between adjacent segments and generate a standardized data stream in units of operation cycles.

[0024] Extracting key operation nodes in S3 includes the following steps: S31. Construct a maintenance operation knowledge graph, where nodes represent maintenance steps, edges represent logical dependencies between steps, and standard operation time and common error types are labeled. S32. Use a graph matching algorithm to align the standardized operation data flow with the maintenance operation knowledge graph to identify key nodes in the trainee's operation, including the initial action, core disassembly steps, and final inspection points. Specifically, this includes the following steps: S321. Construct a node similarity matrix between the standardized operation data flow and the maintenance operation knowledge graph. Calculate the node feature matching degree using cosine similarity. The cosine similarity formula is: ; in and For node feature vectors, Describing the vector norm, This represents the similarity value. S322. Apply the Hungarian algorithm to find the optimal matching pairs in the similarity matrix and maximize the overall matching score. S323. Post-process the matching results to remove node pairs with a similarity of less than 0.7 to ensure the accuracy of key node alignment; S33. Extract feature parameters based on key nodes, including operation path length, tool switching frequency, and error recovery time, and encode them into fixed-dimensional operation feature vectors.

[0025] Multidimensional evaluation using deep learning models in S4 includes the following steps: S41. Design a multi-task deep learning model in which the convolutional neural network branch processes spatial operation features and the long short-term memory network branch processes temporal operation features. S42. Input the operation feature vector into the multi-task deep learning model, and calculate the operation accuracy score, efficiency score and safety score respectively. S43. The multi-task outputs are fused through a fully connected layer and normalized to the original evaluation scores in the 0-1 range using the Sigmoid function.

[0026] Designing a multi-task deep learning model in S41 includes the following sub-steps: S411. Using adversarial training techniques, the generator of a generative adversarial network simulates the generation of anomalous operation data, and a discriminator distinguishes between real operation data and generated anomalous operation data. The parameters of the deep learning model are updated iteratively. The specific steps for updating the model parameters using adversarial training techniques include: S4111. Initialize the generator and discriminator of the generative adversarial network. The generator consists of fully connected layers and deconvolutional layers, and the discriminator consists of convolutional layers and pooling layers. S4112. In the training iteration, the generator generates simulated abnormal operation data based on random noise vectors, and the discriminator performs a binary classification task on the real data and the generated data. S4113. Update the network parameters by minimizing the adversarial loss function between the generator and the discriminator. The loss function is the standard minimum-maximum loss of the generative adversarial network. S412. Introduce a multi-scale feature extraction mechanism, using dilated convolutional layers to capture long-range dependencies in the operation sequence, and combine it with a local attention network to calculate the weight distribution of detailed operation features. Specifically, this includes the following steps: S4121. Set multiple sets of dilation rate parameters in the dilated convolutional layer, including convolutional kernels with dilation rates of 1, 2, and 4, to capture operation sequence features at different scales respectively; S4122. The local attention network calculates the weight distribution at each location in the feature map and normalizes the attention score using the Softmax function, with the following formula: ; in For the first Attention weights for each position, For input features, For the number of features, , For location index; S4123. Multi-scale features and attention-weighted features are concatenated by channels and fused through a 1x1 convolutional layer to output enhanced operational features. S413. Perform model interpretability analysis, and generate heat map data corresponding to key decision-making areas through gradient weighted class activation mapping technology. The core calculations include the following steps: S4131. Obtain the output feature map of the last convolutional layer of the deep learning model, denoted as... ,in Indicates the feature map channel index; S4132. Calculate the score of the target category relative to the feature map. The gradient is calculated, and the weights for each feature channel are obtained through global average pooling. : ; in For the first Each feature map corresponds to the target category. Importance weight, For the model to class The score, For the first Each feature map in spatial location activation value, The spatial dimensions of the feature map. , For spatial location index, For feature map channel index; S4133, Using the calculated weights For feature maps We perform weighted summation and use the ReLU activation function to highlight feature regions that contribute positively to classification, generating a coarse-grained heatmap. : ; in For the generated coarse-grained heatmap data, As importance weight, For the first Each feature map For feature map channel index; S4134. Upsample the coarse-grained heatmap to make its size consistent with the original input image, and generate the final heatmap data that can be used for visualization interpretation. S414. Establish a closed loop for model performance monitoring, detect the consistency deviation between the evaluation score and the manual review result in real time, and call the model retraining module to update the network parameters when the deviation exceeds the preset threshold.

[0027] Dynamic weighted calibration in S5 includes the following steps: S51. Establish an expert rating database to store expert rating records for typical operation modes under different maintenance scenarios, and associate them with operation complexity coefficients. S52. Retrieve trainees’ historical training records and extract recent assessment trends and progress as dynamic weighting factors; S53. Use Bayesian inference algorithm to probabilistically fuse the original evaluation scores with the expert scoring database, and combine dynamic weighting factors to calculate the comprehensive assessment result after confidence adjustment. S53 uses Bayesian inference algorithms for probability fusion and computation, which includes the following sub-steps: S531. Construct a time decay weight function so that the influence of recent training records on the weighting factor is higher than that of earlier records, where the decay coefficient is dynamically adjusted based on the student's learning curve. S532. Using a fuzzy logic system to handle subjective differences in expert scoring, transforming qualitative evaluations into quantitative weights, and eliminating outlier scores through a consistency check algorithm, including the following steps: S5321. Define the input variable as expert scoring bias and the output variable as weight adjustment amount, and set the fuzzy set as "low", "medium" and "high"; S5322. Design a fuzzy rule base, including IF-THEN rules, including "when the deviation is low, the adjustment amount is small"; S5323. The fuzzy output is converted into precise weight values ​​using a defuzzification method, and the final quantitative weights are calculated using the centroid method. The calculation formula is as follows: ; in The final quantitative weight value obtained from the calculation, Indicates the first The values ​​of each data point For the first Data points The corresponding weights The total number of discrete points; S533 integrates a real-time feedback mechanism, introduces student self-assessment data during the calibration process, and uses a Bayesian update algorithm to dynamically correct weighted parameters; S534. Implement source tracing for calibration results, record the logical path and influencing factors of each weight adjustment, and support the auditing and review of evaluation results.

[0028] Building a personalized feedback report in S6 involves the following steps: S61. Based on the comprehensive assessment results, a decision tree model is used to identify weaknesses in the trainees' operations, including incorrect tool selection, reversed order, and omission of key steps, including the following steps: S611. Construct a C4.5 decision tree model, using operational features as input and weak link type as output, and set the information gain ratio as the splitting criterion. The formula for the information gain ratio is: ; in Information gain ratio, For information gain, For split information, For the sample set, As a characteristic; S612. Prune the decision tree using a pessimistic error pruning method to reduce the risk of overfitting. S613. Extract rule paths from the decision tree to generate a decision logic chain for weak links, supporting error location analysis; S62. Generate hierarchical feedback content based on weak links, including error location, cause analysis, and demonstration video links; S63. Use natural language generation technology to convert feedback content into personalized text reports and add visual charts to display score distribution.

[0029] Optimizing training content and updating evaluation thresholds in S7 includes the following steps: S71. Construct an adaptive learning engine using reinforcement learning algorithms to dynamically adjust the difficulty and focus of subsequent virtual training tasks based on the learner's weaknesses, including the following steps: S711. Define the state space as the trainee's historical operation sequence, the action space as the training task difficulty adjustment instructions, and the reward function as the degree of improvement based on the assessment results. S712. Train the agent using the Q-learning algorithm, setting the learning rate to 0.1 and the discount factor to 0.9. Update the Q-value table. The Q-value update formula is: ; in For state-action value function, Here, is the learning rate, is the reward value, and γ=0.9 is the discount factor. For state, For action; S713. Optimized training content is dynamically generated by balancing exploration and utilization through an ε-greedy strategy. S72. Set up an evaluation threshold update mechanism to automatically recalibrate the parameter thresholds of the scoring model when the distribution of assessment scores among trainees changes. S73. Push the updated evaluation threshold and training content to the virtual reality device, and perform the next round of training and evaluation based on the new threshold and content to form a data feedback loop.

[0030] The operational steps of this virtual maintenance training process assessment method are as follows: Step 1: Multi-source operational data acquisition and standardization processing This method first collects real-time operational data during maintenance training using virtual reality equipment. Specifically, the inertial measurement unit and optical sensors built into the head-mounted device acquire the trainee's hand movement data, tool gripping force, and gaze focus coordinates. Simultaneously, an environmental simulator records the trainee's interaction events with the virtual device, including disassembly sequence, parts installation accuracy, and the number of erroneous operations. Subsequently, the collected multi-source heterogeneous data undergoes noise filtering and format standardization. Wavelet transform algorithms are used to smooth and denoise the gesture trajectory data, and Z-score standardization converts data with different dimensions into a unified numerical range, ultimately generating a standardized operational data stream based on operation cycles, providing a high-quality, standardized data foundation for subsequent analysis.

[0031] Step 2: Key Operation Node Extraction and Feature Generation Based on Knowledge Graph After obtaining the standardized operation data flow, this method extracts key operation nodes based on a pre-constructed maintenance operation knowledge graph. This knowledge graph uses nodes to represent maintenance steps and edges to represent logical dependencies between steps. A graph matching algorithm aligns the trainee's actual operation data flow with the standard flow of the knowledge graph, intelligently identifying key nodes such as the initial action, core disassembly steps, and final inspection points. Subsequently, multi-dimensional feature parameters, including operation path length, tool switching frequency, and error recovery time, are extracted based on these key nodes and encoded into fixed-dimensional operation feature vectors, thereby transforming continuous operation sequences into structured features that can be processed by the model.

[0032] Step 3: Multi-dimensional intelligent assessment and dynamic weighted calibration The generated operational feature vectors are input into a multi-task deep learning model for multi-dimensional evaluation. This model processes spatial operational features through convolutional neural network branches and temporal operational features through long short-term memory network branches, calculating scores for operational accuracy, efficiency, and safety respectively. The raw evaluation scores are then fused through fully connected layers. To enhance the personalization and fairness of the evaluation results, a dynamic weighted calibration mechanism is introduced. This method combines trainees' historical training records and an expert scoring database, uses a Bayesian inference algorithm to probabilistically fuse the raw scores, and utilizes a fuzzy logic system to handle subjective scoring differences, ultimately generating a comprehensive assessment result. This ensures that the evaluation criteria can adapt to individual differences and long-term progress trends.

[0033] Step 4: Personalized Feedback Generation and System Adaptive Optimization Finally, personalized feedback reports are constructed based on the comprehensive assessment results. A decision tree model is used to identify specific weaknesses in the trainees' operations, including incorrect tool selection and reversed tool order, and hierarchical feedback content is generated, including error localization, root cause analysis, and links to demonstration videos. Furthermore, natural language generation technology is used to convert the analysis results into easily understandable text reports. To achieve continuous optimization of training, this method dynamically adjusts the difficulty and focus of subsequent training tasks based on the analysis of weaknesses in the feedback reports, using an adaptive learning engine built with reinforcement learning algorithms. The evaluation threshold is updated based on the assessment results, thus forming an "assessment-feedback-optimization" data feedback loop, continuously improving the systematic nature and effectiveness of virtual maintenance training.

[0034] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing and evaluating a virtual maintenance training process, characterized in that: The method includes the following steps: S1. Collect real-time operation data during the maintenance training process using virtual reality equipment, including gesture trajectories, tool usage sequences, and environmental interaction logs; S2. Perform noise filtering and format unification processing on the real-time operation data to generate a standardized operation data stream; S3. Extract key operation nodes from the standardized operation data stream based on the maintenance knowledge graph, and generate operation feature vectors; S4. Utilize a deep learning model to evaluate the operation feature vector in multiple dimensions, including operation accuracy, efficiency, and security, and output the original evaluation score. S5. Combine the trainees' historical training records and the expert scoring database to dynamically weight and calibrate the original assessment scores, and generate a comprehensive assessment result. S6. Based on the comprehensive assessment results, construct a personalized feedback report, including analysis of weaknesses and training suggestions; S7. Based on the analysis of weaknesses in the personalized feedback report, the subsequent training content is dynamically optimized through the adaptive learning engine, and the evaluation threshold is updated according to the comprehensive assessment results, forming an adaptive adjustment loop of training content based on assessment result feedback.

2. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: The process of collecting real-time operation data in S1 includes the following steps: S11. The inertial measurement unit and optical sensors built into the head-mounted virtual reality device collect data on the trainee's hand movements, tool gripping force, and gaze focus coordinates in the virtual maintenance scenario. S12. Use an environment simulator to record the interaction events between the trainee and the virtual device, including the disassembly sequence, the accuracy of parts installation, and the number of incorrect operations; S13. The hand movement data, tool gripping force, gaze focus coordinates and interaction events are synchronized and integrated according to timestamps to generate a time-seriesd real-time operation data matrix.

3. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: The processing of real-time operation data in S2 includes the following steps: S21. The gesture trajectory data in the real-time operation data matrix is ​​smoothed by using a wavelet transform algorithm to remove high-frequency noise caused by device vibration. S22. Using the Z-score standardization method, tools with different dimensions are converted into a uniform numerical range using sequence data and environmental interaction logs, thus eliminating differences in magnitude. S23. The processed data stream is segmented and recombined using time window sliding technology to generate a standardized operation data stream based on operation cycles. Each data segment contains a complete maintenance action unit.

4. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: Extracting key operation nodes in S3 includes the following steps: S31. Construct a maintenance operation knowledge graph, where nodes represent maintenance steps, edges represent logical dependencies between steps, and standard operation time and common error types are labeled. S32. Use a graph matching algorithm to align the standardized operation data stream with the maintenance operation knowledge graph to identify key nodes in the trainee's operation, including the initial action, core disassembly steps and final inspection points; S33. Extract feature parameters based on key nodes, including operation path length, tool switching frequency, and error recovery time, and encode them into fixed-dimensional operation feature vectors.

5. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: The multi-dimensional evaluation using a deep learning model in S4 includes the following steps: S41. Design a multi-task deep learning model in which the convolutional neural network branch processes spatial operation features and the long short-term memory network branch processes temporal operation features. S42. Input the operation feature vector into the multi-task deep learning model, and calculate the operation accuracy score, efficiency score and security score respectively. S43. The multi-task outputs are fused through a fully connected layer and normalized to the original evaluation scores in the 0-1 range using the Sigmoid function.

6. The assessment method for a virtual maintenance training process according to claim 5, characterized in that: The design of the multi-task deep learning model in S41 includes the following sub-steps: S411. An adversarial training technique is adopted, which simulates the generation of abnormal operation data through the generator of the generative adversarial network, and uses a discriminator to distinguish between real operation data and generated abnormal operation data, and updates the parameters of the deep learning model in an iterative manner. S412. Introduce a multi-scale feature extraction mechanism, use dilated convolutional layers to capture long-range dependencies in the operation sequence, and combine local attention networks to calculate the weight distribution of detailed operation features. S413. Conduct model interpretability analysis and generate heat map data corresponding to key decision-making areas through gradient weighted class activation mapping technology; S414. Establish a closed loop for model performance monitoring, detect the consistency deviation between the evaluation score and the manual review result in real time, and call the model retraining module to update the network parameters when the deviation exceeds the preset threshold.

7. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: The dynamic weighted calibration in S5 includes the following steps: S51. Establish an expert rating database to store expert rating records for typical operation modes under different maintenance scenarios, and associate them with operation complexity coefficients. S52. Retrieve trainees’ historical training records and extract recent assessment trends and progress as dynamic weighting factors; S53. Use a Bayesian inference algorithm to perform probability fusion of the original evaluation scores and the expert scoring database, and combine the dynamic weighting factors to calculate the comprehensive assessment result after confidence adjustment.

8. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: The probability fusion and calculation using the Bayesian inference algorithm in step S53 includes the following sub-steps: S531. Construct a time decay weight function so that the influence of recent training records on the weighting factor is higher than that of earlier records, where the decay coefficient is dynamically adjusted based on the student's learning curve. S532. Use a fuzzy logic system to handle subjective differences in expert scoring, transform qualitative evaluation into quantitative weights, and eliminate abnormal scores through a consistency test algorithm. S533 integrates a real-time feedback mechanism, introduces student self-assessment data during the calibration process, and uses a Bayesian update algorithm to dynamically correct weighted parameters; S534. Implement source tracing for calibration results, record the logical path and influencing factors of each weight adjustment, and support the auditing and review of evaluation results.

9. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: The process of building a personalized feedback report in S6 includes the following steps: S61. Based on the comprehensive assessment results, a decision tree model is used to identify the weak points in the trainees' operations, including incorrect tool selection, reversed order, and omission of key steps. S62. Generate hierarchical feedback content based on weak links, including error location, cause analysis, and demonstration video links; S63. Use natural language generation technology to convert feedback content into personalized text reports and add visual charts to display score distribution.

10. The assessment method for a virtual maintenance training process according to claim 1, characterized in that: The optimization of training content and updating of evaluation thresholds in S7 includes the following steps: S71. Utilize reinforcement learning algorithms to build an adaptive learning engine, dynamically adjusting the difficulty and focus of subsequent virtual training tasks based on the learner's weaknesses. S72. Set up an evaluation threshold update mechanism to automatically recalibrate the parameter thresholds of the scoring model when the distribution of assessment scores among trainees changes. S73. Push the updated evaluation threshold and training content to the virtual reality device, and perform the next round of training and evaluation based on the new threshold and content to form a data feedback loop.

Images