A virtual torque-driven motor disassembly and assembly intelligent evaluation system and method

Abstract

The present application relates to the field of intelligent technology evaluation, and discloses a virtual torque driven motor disassembly and assembly intelligent evaluation system and method, comprising: acquiring tool posture feature sequence and operation time sequence, extracting key point tracking completeness and time sequence missing rate; inputting neural network to estimate virtual torque and network output variance, and calculating virtual torque confidence; generating operation evidence flow based on the estimated value, confidence and features, extracting step identification and loading corresponding knowledge point sub-model; calculating step sequence, posture deviation and mechanical evidence item, and dynamically adjusting basic weight based on confidence to calculate multi-modal comprehensive score; updating the mastery probability by taking the score as a soft evidence interpolation coefficient, and generating augmented reality guidance information combined with the confidence. The present application gets rid of the dependence on physical torque sensors, uses confidence weight and soft evidence probability interpolation reasoning, overcomes data distortion caused by high noise features, and realizes high noise resistance and smooth training evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent technology evaluation, and in particular to an intelligent evaluation system and method for training the disassembly and assembly of a motor with virtual torque drive. Background Technology

[0002] In industrial manufacturing and maintenance training, the tightening torque of bolts or nuts is a core physical indicator determining assembly quality. Traditional motor disassembly and assembly training systems typically require numerous high-precision physical torque sensors installed on the tool or bench end to collect torque feedback during the assembly process. However, in vocational teaching and field operations, physical sensors are expensive and easily damaged or drifted under frequent, intense impacts, wear, and collisions, resulting in a heavy burden of physical maintenance and periodic offline calibration.

[0003] With the development of machine vision and edge computing, estimating "virtual torque" based on visual depth images and operational time series has become a trend to replace physical sensors. However, existing multimodal virtual estimation results have high uncertainty. When faced with situations such as student hand occlusion, ambient lighting fluctuations, or data communication loss, the output virtual torque will be drastically distorted. Directly substituting this noisy, continuous estimated value into the Bayesian Knowledge Tracking (BKT) model used to assess the probability of learners mastering skills presents two serious drawbacks: First, the standard BKT model is essentially a binary observation Hidden Markov system (i.e., learners' responses can only be judged as discrete correct or incorrect), and cannot directly consume high-dimensional continuous and noisy observation data such as virtual torque and attitude deviation. Forcibly implementing hard threshold binary truncation will lead to the loss of extremely rich mechanical gradient details. Second, when performing estimation or adaptive retraining of BKT parameters on a large sample in the cloud, due to the large gradient space and lack of hard constraints, "degeneration" phenomena (such as parameter overflow or semantic inversion of guessing rate, error rate, etc.) are very likely to occur, causing the assessment probability output to no longer conform to basic cognitive facts, resulting in assessment failure.

[0004] Furthermore, at distributed edge devices (such as industrial control computers or AR terminals), if each operation requires loading and fully reasoning through a massive, multi-level, interconnected knowledge graph, it will lead to severe computational latency, failing to meet the stringent low-latency requirement of millisecond-level augmented reality rendering guidance when critical deviations occur. Therefore, there is an urgent need for a highly feasible and noise-resistant intelligent evaluation solution that does not limit the edge processor hardware platform. Summary of the Invention

[0005] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides an intelligent evaluation system and method for training and evaluating the disassembly and assembly of motors driven by virtual torque.

[0006] The first technical solution provided by this invention is: A virtual torque-driven motor disassembly and assembly training intelligent evaluation method includes: Acquire the tool posture feature sequence and operation time sequence of motor disassembly and assembly operations, wherein the tool posture feature sequence includes the depth image key point tracking integrity rate; Extract the time series missing rate from the operation time series; The tool posture feature sequence and the operation time sequence are input into a pre-trained neural network to estimate the virtual torque and the network output variance. The virtual torque confidence level is calculated based on the network output variance, the depth image key point tracking completeness rate, and the time series missing rate. Based on the virtual torque estimate, the virtual torque confidence level, and the tool attitude feature sequence, an operation evidence stream is generated, and the current disassembly / assembly step identifier is extracted from the operation evidence stream. Based on the current disassembly / assembly step identifier, dynamically load the corresponding knowledge point sub-model containing the mastery probability from the edge cache; Based on the current disassembly / assembly step identifier, the tool posture characteristic sequence, and the virtual torque estimation value, calculate the step sequence deviation item, the tool posture deviation item, and the mechanical evidence item, respectively. Based on the virtual torque confidence level, the preset weights of various basic evidence items are dynamically adjusted to obtain the adjusted weights of various evidence items. The multimodal comprehensive score is then calculated by combining the step sequence deviation item, the tool posture deviation item, and the mechanical evidence item. The multimodal composite score is used as the soft evidence interpolation coefficient to update the mastery probability of the knowledge point sub-model; Augmented reality guidance information is generated based on the updated mastery probability and the virtual torque confidence level.

[0007] Furthermore, the neural network is a lightweight temporal convolutional network, and the step of inputting the tool pose feature sequence and the operation time sequence into the pre-trained neural network includes: The relative pose change features in the tool pose feature sequence are extracted by the two-dimensional convolutional branch of the lightweight temporal convolutional network. The action rhythm features in the operation time series are extracted by the one-dimensional temporal convolutional branch of the lightweight temporal convolutional network; the action rhythm features include tightening duration, pause interval and number of repeated tightenings. The relative pose change features and the motion rhythm features are input into the fusion regression layer, and the virtual torque estimate and the network output variance are output. After generating augmented reality guidance information, it also includes: The operation evidence stream is asynchronously sent to the cloud-based full-precision processing engine; The cloud-based full-precision processing engine performs Bayesian knowledge tracking parameter updates on the global model, generates and returns global parameter update results with model version number and evidence digest hash value. At the edge, by comparing the local hash value generated by the operational evidence stream with the returned evidence digest hash value, and by comparing the local version number with the model version number, sub-model replacement or incremental parameter merging in the edge cache is performed.

[0008] Furthermore, the virtual torque confidence level is calculated using the following formula: in, For virtual torque confidence; Output the variance for the network; Improve the accuracy of keypoint tracking in depth images; This represents the missing rate of the time series data. and These are the weighting coefficients for missing key points and missing time series data, respectively. and The sum is less than or equal to 1.

[0009] Furthermore, the loading strategy for dynamically loading the corresponding knowledge point sub-model containing the mastery probability from the edge cache includes: In the preset knowledge tree of motor disassembly and assembly, determine the master sub-model corresponding to the identifier of the current disassembly and assembly step; The master-child model, the parent node child model of the master-child model, the leaf node child model that has a prior repair dependency with the current disassembly / assembly step, and the adjacent subsequent knowledge point child model are all loaded together as the corresponding knowledge point model.

[0010] Furthermore, after generating the operational evidence stream based on the virtual torque estimate, the virtual torque confidence level, and the tool attitude feature sequence, a hybrid scheduling step is also included: The timing density of operation events within the sliding time window is calculated based on the operation evidence stream, and the current edge resource occupancy rate of the device is obtained. When the timing density of the operation events is lower than the density threshold, perform a batch update to merge the operation evidence stream; When the timing density of the operation event is not lower than the density threshold and the edge resource occupancy rate is lower than the resource occupancy limit, perform an immediate event update on the current operation evidence stream. When the interval between two consecutive operation events in the operation evidence stream is greater than the interval threshold, pulse interval preprocessing is performed. The edge resource occupancy rate is calculated using the following formula: in, For edge resource occupancy rate, For processor utilization, Graphics processor utilization, For edge cache utilization, , and The corresponding weights are 1, and the sum of the three weights is 1.

[0011] Furthermore, the mechanical evidence item The calculation formula is: in, For mechanical evidence, This is a virtual torque estimate. and These are the lower and upper limits of the target torque range, respectively. Tool attitude deviation term The calculation formula is: in, The angle between the center axis of the tightening tool and the bolt axis normal, determined based on the tool posture characteristic sequence; Step sequence deviation item Set as: If the current step conforms to the preset cross-tightening sequence, then ; If a student violates the order rule, a corresponding step order deviation value will be set. .

[0012] Furthermore, the adjustment logic for the dynamically adjusted preset weights of various basic pieces of evidence is as follows: Obtain the preset mechanical foundation weights, tool posture foundation weights, and step sequence foundation weights; Multiplying the mechanical basis weights by the virtual torque confidence level yields the adjusted mechanical evidence weights. The portion of the mechanical basis weight that is lost is compensated to the tool posture basis weight and the step sequence basis weight in an equal proportion, so as to obtain the adjusted tool posture weight and step sequence weight respectively. The adjustment calculation formula is as follows: In the formula, For virtual torque confidence, Weighting of evidence for the fundamentals of mechanics For tool pose basic weights and The basic weights for the order of steps; To adjust the weight of post-mechanical evidence, To adjust the tool pose weights and The weights are adjusted to reflect the order of steps.

[0013] Furthermore, the multimodal comprehensive score is calculated as follows: In the formula, For multimodal comprehensive score, For mechanical evidence, For tool attitude deviation, This refers to the deviation item in the order of steps; To adjust the weight of post-mechanical evidence, To adjust the tool pose weights and The weights are adjusted to reflect the order of steps.

[0014] Furthermore, updating the mastery probability of the knowledge point sub-model includes: The posterior value of skill mastery probability and transition probability parameter of the previous time step are obtained from the knowledge point sub-model, and the predicted prior value of the current step is calculated. Calculate the first provisional posterior probability under the assumption that the observations are completely correct, and the second provisional posterior probability under the assumption that the observations are completely wrong. Using the multimodal composite score as interpolation coefficients, linear interpolation is performed between the first temporary posterior probability and the second temporary posterior probability to obtain the updated mastery probability.

[0015] Based on the first technical solution, the second technical solution is proposed as follows: A virtual torque-driven motor disassembly and assembly training intelligent evaluation system, executing the aforementioned virtual torque-driven motor disassembly and assembly training intelligent evaluation method, includes: The data acquisition module is used to acquire the tool posture feature sequence and operation time sequence of motor disassembly and assembly operations, and extract the key point tracking completeness rate and time sequence missing rate respectively; The virtual torque estimation module is used to input the tool posture feature sequence and the operation time sequence into a pre-trained neural network to estimate the virtual torque estimate and the network output variance, and to calculate the virtual torque confidence based on the network output variance, the key point tracking completeness rate and the time sequence missing rate. The evidence stream generation and model loading module is used to generate an operation evidence stream based on the virtual torque estimate, the virtual torque confidence level, and the tool posture feature sequence, extract the current disassembly and assembly step identifier from the operation evidence stream, and dynamically load the corresponding knowledge point sub-model containing the mastery probability according to the current disassembly and assembly step identifier. The edge knowledge tracking and reasoning module is used to calculate the step sequence deviation, tool posture deviation, and mechanical evidence item based on the current disassembly / assembly step identifier, the tool posture feature sequence, and the virtual torque estimation value, respectively; dynamically adjust the preset basic evidence weights based on the virtual torque confidence level to obtain the adjusted evidence weights; combine the step sequence deviation, tool posture deviation, and mechanical evidence item to calculate the multimodal comprehensive score; and use the multimodal comprehensive score as a soft evidence interpolation coefficient to update the mastery probability of the knowledge point sub-model. An augmented reality guidance module is used to generate augmented reality guidance information based on the updated mastery probability and the virtual torque confidence level.

[0016] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: Instead of using expensive and easily damaged physical high-precision torque sensors, a low-cost depth camera and machine vision are used to estimate "virtual torque," solving the problems of sensor drift and periodic offline calibration burden caused by frequent knocking and bumping in industrial training sites.

[0017] A "virtual torque confidence" quantitative evaluation mechanism has been introduced. When faced with unavoidable student hand occlusion or wireless packet loss leading to severe distortion in visual estimation, the system can "self-detect" and automatically reduce the mechanical weights and smoothly compensate for them by adjusting the posture and sequence weights, preventing poor noise from contaminating the overall evaluation results.

[0018] It breaks through the limitations of traditional Bayesian Knowledge Tracking (BKT) models, which can only handle binary observations of "right / wrong". By using the dynamically adjusted multimodal comprehensive score as the full probability interpolation coefficient of "soft evidence", it avoids the loss of mechanical details caused by hard threshold truncation, and the probability flow of the evaluation results is smoother, which is consistent with objective cognitive facts.

[0019] Based on a hybrid scheduling model that considers both "operation event timing density" and "resource utilization," the massive global knowledge graph is dynamically loaded in fragmented form. This successfully frees up the computing power of AR terminals or industrial control computers, ensuring millisecond-level augmented reality rendering feedback in the event of dangerous actions or operational deviations.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart of an intelligent evaluation method for virtual torque-driven motor disassembly and assembly training in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0024] like Figure 1 The diagram illustrates the main process of an intelligent evaluation method for virtual torque-driven motor disassembly and assembly training according to the present invention.

[0025] Example 1 A virtual torque-driven motor disassembly and assembly training intelligent evaluation method, the specific steps of which include: S101, a depth camera acquires a sequence of images depicting the disassembly and assembly of the motor. This sequence of images covers at least the areas containing the trainee's hand, tools, fasteners, motor end caps, rotor, bearings, and terminals. The bench edge computing module performs noise reduction, hand keypoint localization, tool endpoint localization, and fastener region localization on the continuous depth images, obtaining the tool rotational angular velocity, wrist rotation amplitude, tool axis offset, and relative position of the fasteners. Simultaneously, it outputs the depth image keypoint tracking completeness rate, defined as the ratio of the number of successfully located hand keypoints to the total number of keypoints that should be located in the current frame. The aforementioned tool rotational angular velocity, wrist rotation amplitude, tool axis offset, relative position of the fasteners, and keypoint tracking completeness rate constitute the tool posture feature sequence, which is used by the virtual torque estimation module to perform virtual torque estimation.

[0026] S102, the time sequencer records the operation time sequence. The operation time sequence includes the operation start time, operation end time, tool pause interval, number of repeated tightening operations, and step switching time. The operation time sequence and the depth image sequence are aligned using the same bench clock or synchronization timestamp.

[0027] S103, the calibration torque sampling interface receives real torque calibration samples. These samples come from calibration torque wrenches, electronic torque screwdrivers, bench torque sampling units, or calibrated teaching torque tools. Each real torque calibration sample includes the calibration tool identifier, fastener location, calibration torque value, sampling timestamp, and calibration status marker. Samples that fail the calibration status check are not added to the training sample library. These real torque calibration samples are used for offline training of lightweight temporal convolutional networks and for global parameter correction of the cloud-based full-precision processing engine.

[0028] S104, the virtual torque estimation module performs virtual torque estimation. This module receives a tool posture feature sequence and an operation time sequence. The lightweight temporal convolutional network includes a two-dimensional convolutional branch, a one-dimensional temporal convolutional branch, and a fusion regression layer. The two-dimensional convolutional branch takes the tool posture feature sequence as input and extracts the relative pose change features of the hand, tool, and fastener. The one-dimensional temporal convolutional branch takes the operation time sequence as input and extracts rhythmic features such as tightening duration, pause interval, and number of repeated tightening operations. The fusion regression layer outputs the virtual torque estimate and virtual torque confidence score for the current operation of the fastener. The lightweight temporal convolutional network is trained offline using collected real torque calibration samples, with the calibration torque values ​​serving as supervision labels.

[0029] The virtual torque estimation module calculates the current virtual torque confidence level using the following formula. : in, The virtual torque confidence level can sensitively and quantitatively characterize the degree of interference of current environmental noise (such as hand obstruction and wireless transmission packet loss) on virtual mechanical estimation.

[0030] This represents the output variance of a lightweight temporal convolutional network within the current operating window. To improve the accuracy of keypoint tracking in depth images, This is the missing rate of the operation time series (i.e., the ratio of the duration of missing or invalid records in the time series to the total operation duration). and These are the weighting coefficients for missing key points and missing time series data, respectively. and The sum of is less than or equal to 1. and Determined by calibration set cross-validation or bench configuration.

[0031] The physical meaning of the virtual torque confidence calculation formula is: when all key points are tracked ( =1), No missing time series data ( =0) and the network output variance approaches zero ( When →0), the confidence level The confidence level approaches 1; however, it decreases as key points are lost, time series data is missing, or output variance increases.

[0032] S105, the operation evidence stream generation module generates the operation evidence stream. For each disassembly / assembly action, the operation evidence stream generation module writes the virtual torque estimate and virtual torque confidence level, the current disassembly / assembly step identifier, tool posture features, key point integrity markers, model version number, and operation timestamp into the same record, and calculates the evidence summary hash value. The operation evidence stream enables virtual torque evidence, visual posture evidence, and step sequence evidence to be consumed by the same knowledge point sub-model.

[0033] S106, a cloud-based full-precision processing engine, maintains the motor disassembly and assembly knowledge point tree and the Bayesian knowledge tracing global model. The motor disassembly and assembly knowledge point tree can include parent nodes such as disassembly preparation, end cover removal, rotor extraction, bearing inspection and resetting, terminal fixing, fastener reinstallation, and safe finishing. Each parent node includes leaf nodes such as cross-loosening bolts, target torque range control, tool axis alignment, end cover positioning sequence, and terminal insulation inspection.

[0034] The S107 cloud-based full-precision processing engine divides the global Bayesian knowledge tracing model into multiple knowledge point sub-models according to the parent-child topology and prior dependencies of the knowledge point tree. Each knowledge point sub-model contains at least mastery probability parameters, transition probability parameters, guessing rate parameters, error rate parameters, and evidence weight parameters. The evidence weight parameters are used to limit the contribution of mechanical evidence, step sequence evidence, and tool posture evidence to the mastery probability update. The set of segmented knowledge point sub-models is distributed to the edge cache of the bench edge computing module for storage, and is loaded by the step sub-model dynamic loading module.

[0035] In step S108, the sub-model dynamic loading module determines the current task knowledge point label. The edge terminal searches for the corresponding knowledge point label in the motor disassembly and assembly knowledge point tree based on the current disassembly / assembly step identifier, tool type, fastener location, and operation timestamp. For example, when the current step identifier is end cover reassembly and tightening, the tool posture feature corresponds to a tightening tool, and the fastener location is at the four corners of the end cover, the current task knowledge point label is determined to be cross-tightening and target torque range control. The current disassembly / assembly step identifier originates from the current disassembly / assembly step identifier field in the operation evidence stream; the tool posture feature is obtained from the acquisition process in step S101; and the fastener location is derived from the fastener area positioning result.

[0036] S109, the sub-model dynamic loading module loads a set of knowledge point sub-models from the edge cache. The loading strategy is as follows: load the master-sub-model corresponding to the current task's knowledge point, the parent node sub-model of the master-sub-model, the leaf node sub-model that has a pre-repair dependency with the current disassembly / assembly step, and adjacent subsequent knowledge point sub-models. When the edge resource occupancy rate is not lower than the resource occupancy limit, only the master-sub-model and the knowledge point sub-model that recently experienced an operational deviation are retained. The edge resource occupancy rate is calculated by the hybrid update scheduling module. The adjacent subsequent knowledge point sub-model refers to the sub-model corresponding to the next leaf node of the current task's knowledge point label in the topological sequence of the knowledge point tree.

[0037] S110, the hybrid update scheduling module calculates the timing density of operation events. The operation events originate from the operation evidence stream generated by the operation evidence stream generation module.

[0038] Operation event timing density: in, For the timing density of operation events, This refers to the number of valid operation events within the sliding time window (i.e., the number of records in the operation evidence stream with a virtual torque confidence level not lower than the minimum confidence threshold). The window length of the sliding time window.

[0039] S111, the hybrid update scheduling module calculates the edge resource occupancy rate.

[0040] Edge resource utilization rate: in, For edge resource occupancy rate, For processor utilization, Graphics processor utilization, For edge cache utilization, , and The corresponding weights are 1, and the sum of the three weights is 1.

[0041] S112, Edge Knowledge Tracking Inference Update Scheduling. When the temporal density of operation events is lower than the density threshold, the system performs low-frequency batch updates, merging multiple operation evidence stream records to update the sub-model parameters; when the temporal density of operation events is not lower than the density threshold and the edge resource occupancy rate is lower than the resource occupancy limit, the system performs immediate event updates; when the interval between two consecutive operation events is greater than the interval threshold, the system performs pulsed intermittent preprocessing, including data cleaning, feature caching, evidence digest hash value calculation, and packaging of events to be corrected in the cloud. The update scheduling mode determines the triggering timing and batch size of edge knowledge tracking inference.

[0042] S113, the edge knowledge tracking and reasoning module performs local reasoning. The edge knowledge tracking and reasoning module takes the operation evidence flow as input to the loaded knowledge point sub-model set, takes the deviation between the virtual torque estimate and the target torque range as the mechanical evidence item, takes the correctness of the step sequence as the step sequence evidence item, takes the tool axis offset and tool attitude stability as the tool attitude evidence item, and adjusts the evidence weight of the mechanical evidence item according to the virtual torque confidence level.

[0043] Mechanical evidence item Calculation: in, For mechanical evidence, This is a virtual torque estimate. and These are the lower and upper limits of the target torque range, respectively; when When within the target torque range, The value is 0, and it increases with the degree of deviation when the target torque range is exceeded.

[0044] Tool attitude deviation term Calculation: in, The angle between the center axis of the tightening tool and the bolt axis normal is determined based on the tool posture characteristic sequence.

[0045] Step sequence deviation item The setting method is as follows: If the current step conforms to the preset cross-tightening sequence, then If a trainee violates the sequence rule by continuously tightening adjacent bolts on the same side, a corresponding deviation value for the sequence of steps will be set. .

[0046] S114, Edge Reasoning Judgment. The edge knowledge tracking and reasoning module performs local reasoning on the mastery probability of the current task knowledge point based on mechanical evidence, step sequence evidence, and tool posture evidence, and outputs the mastery probability of the current task knowledge point and the operational deviation level. The operational deviation levels include no deviation, slight deviation, correction deviation, and safety risk deviation.

[0047] S115, the augmented reality guidance module generates augmented reality guidance information. When the mastery probability is lower than the mastery probability threshold and the operation deviation level reaches the correction level, the augmented reality display terminal outputs a correction level prompt; when the virtual torque confidence level is lower than the confidence threshold and the operation deviation level reaches the safety risk level, the augmented reality display terminal outputs a pause level prompt, requiring the trainee to stop the current action, re-align the tool, or wait for the teacher's confirmation.

[0048] S116, Edge-Cloud Asynchronous Communication and Global Parameter Update. The edge-cloud asynchronous communication module asynchronously sends operational evidence streams, low-confidence events, and real torque calibration samples to the cloud-based full-precision processing engine. The cloud-based full-precision processing engine performs Bayesian knowledge tracing global parameter updates based on complete historical records and calibration samples, generates personalized training sequences, and returns the versioned global parameter update results to the edge.

[0049] S117, the edge caching and version management module corrects the edge cache based on the results returned from the cloud. The edge cache stores the model version number, local update time, cloud correction time, and evidence digest hash value for each knowledge point sub-model. When the model version number returned from the cloud is higher than the edge cache version number and the evidence digest hash value matches, sub-model replacement or incremental parameter merging is performed; when the version number is lower, the hash is inconsistent, or communication times out, the previous valid sub-model is retained and recorded in the merging queue. The corrected edge cache is used by the next round of sub-model dynamic loading module.

[0050] In the above embodiments, the virtual torque estimation value is not an isolated detection result, but is first written into the operation evidence stream (S104→S105); the step labels, confidence levels, and resource status in the operation evidence stream determine the dynamic loading range of the sub-model (S105→S108 / S109); the loaded knowledge point sub-model outputs the mastery probability and operation deviation level (S109→S113 / S114); the temporal density of operation events and edge resource occupancy determine the update scheduling mode (S110 / S111→S112); the augmented reality guidance module outputs prompts based on the mastery probability, deviation level, and confidence level (S114 / S104→S115); the full-precision processing result in the cloud is then written back to the edge cache according to the version (S116→S117→S109). This forms a continuous technology chain from visual temporal input, virtual torque estimation, knowledge tracking and reasoning, edge scheduling, augmented reality feedback to cloud correction.

[0051] Example 2 1. Confidence-guided dynamic adjustment scheme for evidence weights After quantifying the deviation in independent dimensions, the system does not perform a rigid, one-size-fits-all approach to physically deviating. The inference module bases its calculations on the current virtual torque confidence level. Weighting of pre-defined mechanical basis evidence Tool pose basic weights and step order basic weight Dynamic adaptive fine-tuning is performed to prevent poor-quality virtual torque noise from contaminating overall cognitive reasoning when confidence is extremely low (e.g., due to large-area hand occlusion). The adjusted weights of the three pieces of evidence are denoted as: To adjust the weight of post-mechanical evidence, To adjust the tool pose weights and The weights are adjusted for the order of steps.

[0052] The adjusted formulas for calculating the weights of the three pieces of evidence are as follows: like and If the sum is zero, then Assigned value , Assigned value .

[0053] After weight reorganization, the system dynamically weights and merges these three deviation items to calculate a multimodal composite score that reflects the true quality of the trainee's operational behavior. : Multimodal composite score The closer it is to 1, the higher the probability of the student's operation being correct (excellent in both mechanics and posture, and no violation of steps). The closer it is to 0, the more likely it is that a serious non-compliance has occurred with a high degree of confidence.

[0054] 2. BKT's probabilistic smoothing update scheme The edge knowledge tracking and reasoning module will combine the above scores. This serves as soft evidence in Bayesian updates. The parameters of the currently loaded knowledge point sub-model include: error rate parameter. Guessing rate parameter Transition probability The posterior probability of the student mastering the skill at the previous moment is... First, calculate the predicted prior value for the current step. : Subsequently, calculations were performed assuming the observations were completely correct ( ) and the assumption is completely wrong ( Temporary posterior probabilities under two discrete states: Finally, the multimodal composite score is used. As the full probability interpolation coefficients, the true posterior probability of mastery after fusing multimodal soft evidence is calculated. : Updated posterior mastery probability As the latest The cache is located at the edge. This mathematical derivation mechanism allows continuous operational performance deviations to flow smoothly within the Markov chain of BKT, avoiding drastic fluctuations in evaluation results caused by forced binarization.

[0055] Example 3 This embodiment provides a virtual torque-driven intelligent evaluation system for motor assembly and disassembly training, suitable for vocational schools or enterprise training benches. The training object can be a three-phase asynchronous motor, a servo motor, or a modular motor for teaching. The bench includes a motor fixing fixture, a tool recognition area, a depth camera, a time-series recorder, a calibration torque sampling interface, a bench edge computing module, a cloud-based full-precision processing engine, an augmented reality display terminal, and an edge-cloud asynchronous communication module.

[0056] The aforementioned intelligent evaluation system for training and testing the disassembly and assembly of motors driven by virtual torque is used to implement the intelligent evaluation method for training and testing the disassembly and assembly of motors driven by virtual torque. Its functional modules include: The data acquisition module is used to acquire the tool posture feature sequence and operation time sequence of motor disassembly and assembly operations, and extract the key point tracking completeness rate and time sequence missing rate respectively; The virtual torque estimation module is used to input the tool posture feature sequence and the operation time sequence into a pre-trained neural network to estimate the virtual torque estimate and the network output variance, and to calculate the virtual torque confidence based on the network output variance, the key point tracking completeness rate and the time sequence missing rate. The evidence stream generation and model loading module is used to generate an operation evidence stream based on the virtual torque estimate, the virtual torque confidence level, and the tool posture feature sequence, extract the current disassembly and assembly step identifier from the operation evidence stream, and dynamically load the corresponding knowledge point sub-model containing the mastery probability according to the current disassembly and assembly step identifier. The edge knowledge tracking and reasoning module is used to calculate the step sequence deviation, tool posture deviation, and mechanical evidence item based on the current disassembly / assembly step identifier, the tool posture feature sequence, and the virtual torque estimation value, respectively; dynamically adjust the preset basic evidence weights based on the virtual torque confidence level to obtain the adjusted evidence weights; combine the step sequence deviation, tool posture deviation, and mechanical evidence item to calculate the multimodal comprehensive score; and use the multimodal comprehensive score as a soft evidence interpolation coefficient to update the mastery probability of the knowledge point sub-model. An augmented reality guidance module is used to generate augmented reality guidance information based on the updated mastery probability and the virtual torque confidence level.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A virtual torque-driven motor disassembly and assembly training intelligent evaluation method, characterized in that, include: Acquire the tool posture feature sequence and operation time sequence of motor disassembly and assembly operations, wherein the tool posture feature sequence includes the depth image key point tracking integrity rate; Extract the time series missing rate from the operation time series; The tool posture feature sequence and the operation time sequence are input into a pre-trained neural network to estimate the virtual torque and the network output variance. The virtual torque confidence level is calculated based on the network output variance, the depth image key point tracking completeness rate, and the time series missing rate. Based on the virtual torque estimate, the virtual torque confidence level, and the tool attitude feature sequence, an operation evidence stream is generated, and the current disassembly / assembly step identifier is extracted from the operation evidence stream. Based on the current disassembly / assembly step identifier, dynamically load the corresponding knowledge point sub-model containing the mastery probability from the edge cache; Based on the current disassembly / assembly step identifier, the tool posture characteristic sequence, and the virtual torque estimation value, calculate the step sequence deviation item, the tool posture deviation item, and the mechanical evidence item, respectively. Based on the virtual torque confidence level, the preset weights of various basic evidence items are dynamically adjusted to obtain the adjusted weights of various evidence items. The multimodal comprehensive score is then calculated by combining the step sequence deviation item, the tool posture deviation item, and the mechanical evidence item. The multimodal composite score is used as the soft evidence interpolation coefficient to update the mastery probability of the knowledge point sub-model; Augmented reality guidance information is generated based on the updated mastery probability and the virtual torque confidence level; The virtual torque confidence level is calculated using the following formula: in, For virtual torque confidence; Output the variance for the network; Improve the accuracy of keypoint tracking in depth images; This represents the missing rate of the time series data. and These are the weighting coefficients for missing key points and missing time series data, respectively. and The sum of is less than or equal to 1; The mechanical evidence item The calculation formula is: in, For mechanical evidence, This is a virtual torque estimate. and These are the lower and upper limits of the target torque range, respectively. Tool attitude deviation term The calculation formula is: in, The angle between the center axis of the tightening tool and the bolt axis normal, determined based on the tool posture characteristic sequence; Step sequence deviation item Set as: If the current step conforms to the preset cross-tightening sequence, then ; If a student violates the order rule, a corresponding step order deviation value will be set. ; The dynamic adjustment of the preset weights of various basic pieces of evidence follows the following logic: Obtain the preset mechanical foundation weights, tool posture foundation weights, and step sequence foundation weights; Multiplying the mechanical basis weights by the virtual torque confidence level yields the adjusted mechanical evidence weights. The portion of the mechanical foundation weight that is lost is compensated to the tool posture foundation weight and the step sequence foundation weight in an equal proportion, so as to obtain the adjusted tool posture weight and step sequence weight respectively. The adjustment calculation formula is as follows: In the formula, For virtual torque confidence, Weighting of evidence for the fundamentals of mechanics For tool pose basic weights and The basic weights for the order of steps; To adjust the weight of post-mechanical evidence, To adjust the tool pose weights and The weights of the adjusted steps; The multimodal comprehensive score is calculated as follows: In the formula, For multimodal comprehensive score, For mechanical evidence, For tool attitude deviation, This refers to the deviation item in the order of steps; To adjust the weight of post-mechanical evidence, To adjust the tool pose weights and The weights are adjusted to reflect the order of steps.

2. The intelligent evaluation method for virtual torque-driven motor disassembly and assembly training according to claim 1, characterized in that, The neural network is a lightweight temporal convolutional network. The step of inputting the tool pose feature sequence and the operation time sequence into the pre-trained neural network includes: The relative pose change features in the tool pose feature sequence are extracted by the two-dimensional convolutional branch of the lightweight temporal convolutional network. The action rhythm features in the operation time series are extracted by the one-dimensional temporal convolutional branch of the lightweight temporal convolutional network; the action rhythm features include tightening duration, pause interval and number of repeated tightenings. The relative pose change features and the motion rhythm features are input into the fusion regression layer, and the virtual torque estimate and the network output variance are output. After generating augmented reality guidance information, it also includes: The operation evidence stream is asynchronously sent to the cloud-based full-precision processing engine; The cloud-based full-precision processing engine performs Bayesian knowledge tracking parameter updates on the global model, generates and returns global parameter update results with model version number and evidence digest hash value. At the edge, by comparing the local hash value generated by the operational evidence stream with the returned evidence digest hash value, and by comparing the local version number with the model version number, sub-model replacement or incremental parameter merging in the edge cache is performed.

3. The intelligent evaluation method for virtual torque-driven motor disassembly and assembly training according to claim 1, characterized in that, The loading strategy for dynamically loading the corresponding knowledge point sub-model containing the mastery probability from the edge cache includes: In the preset knowledge tree of motor disassembly and assembly, determine the master sub-model corresponding to the identifier of the current disassembly and assembly step; The master sub-model, the parent node sub-model of the master sub-model, the leaf node sub-model that has a prior repair dependency with the current disassembly / assembly step, and the adjacent subsequent knowledge point sub-model are all loaded together as the corresponding knowledge point model containing the mastery probability.

4. The intelligent evaluation method for virtual torque-driven motor disassembly and assembly training according to claim 1, characterized in that, After generating the operational evidence stream based on the virtual torque estimate, the virtual torque confidence level, and the tool attitude feature sequence, a hybrid scheduling step is also included: The timing density of operation events within the sliding time window is calculated based on the operation evidence stream, and the current edge resource occupancy rate of the device is obtained. When the timing density of the operation events is lower than the density threshold, perform a batch update to merge the operation evidence stream; When the timing density of the operation event is not lower than the density threshold and the edge resource occupancy rate is lower than the resource occupancy limit, perform an immediate event update on the current operation evidence stream. When the interval between two consecutive operation events in the operation evidence stream is greater than the interval threshold, pulse interval preprocessing is performed. The edge resource occupancy rate is calculated using the following formula: in, For edge resource occupancy rate, For processor utilization, Graphics processor utilization, For edge cache utilization, , and The corresponding weights are 1, and the sum of the three weights is 1.

5. The intelligent evaluation method for virtual torque-driven motor disassembly and assembly training according to claim 1, characterized in that, The update of the mastery probability of the knowledge point sub-model includes: The posterior value of skill mastery probability and transition probability parameter of the previous time step are obtained from the knowledge point sub-model, and the predicted prior value of the current step is calculated. Calculate the first provisional posterior probability under the assumption that the observations are completely correct, and the second provisional posterior probability under the assumption that the observations are completely wrong. Using the multimodal composite score as interpolation coefficients, linear interpolation is performed between the first temporary posterior probability and the second temporary posterior probability to obtain the updated mastery probability.

6. A virtual torque-driven motor disassembly and assembly training intelligent evaluation system, executing the virtual torque-driven motor disassembly and assembly training intelligent evaluation method as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire the tool posture feature sequence and operation time sequence of motor disassembly and assembly operations, and extract the key point tracking completeness rate and time sequence missing rate respectively; The virtual torque estimation module is used to input the tool posture feature sequence and the operation time sequence into a pre-trained neural network to estimate the virtual torque estimate and the network output variance, and to calculate the virtual torque confidence based on the network output variance, the key point tracking completeness rate and the time sequence missing rate. The evidence stream generation and model loading module is used to generate an operation evidence stream based on the virtual torque estimate, the virtual torque confidence level, and the tool posture feature sequence, extract the current disassembly and assembly step identifier from the operation evidence stream, and dynamically load the corresponding knowledge point sub-model containing the mastery probability according to the current disassembly and assembly step identifier. The edge knowledge tracking and reasoning module is used to calculate the step sequence deviation, tool posture deviation, and mechanical evidence item based on the current disassembly / assembly step identifier, the tool posture feature sequence, and the virtual torque estimation value, respectively; dynamically adjust the preset basic evidence weights based on the virtual torque confidence level to obtain the adjusted evidence weights; combine the step sequence deviation, tool posture deviation, and mechanical evidence item to calculate the multimodal comprehensive score; and use the multimodal comprehensive score as a soft evidence interpolation coefficient to update the mastery probability of the knowledge point sub-model. An augmented reality guidance module is used to generate augmented reality guidance information based on the updated mastery probability and the virtual torque confidence level.

Images