A digital management method for operational efficiency based on embodied intelligent data collection

By using wearable sensing devices and individualized musculoskeletal models, the system can collect and dynamically schedule operator muscle activation and joint movement data in real time, solving the problem of existing technologies being unable to accurately assess operator physiological data. This enables dynamic task scheduling from the individual to the production line, reducing fatigue risk and improving production efficiency.

CN122596604BActive Publication Date: 2026-11-10ZHONGQI LIANXIN TECHNOLOGY GROUP CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202611088803.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-22
Publication Date
2026-11-10
Estimated Expiration
2046-07-22

AI Technical Summary

Technical Problem

Existing technologies cannot acquire detailed physiological data such as the operator's muscle activation level and joint torque in real time, leading to the accumulation of fatigue operation risks and production line efficiency losses. They also cannot dynamically adjust task weights or production line rhythms based on the operator's real-time work efficiency status.

Method used

Wearable sensing devices are used to collect muscle activation and joint motion data. Combined with individualized electromyography-mechanical mapping relationships and musculoskeletal models, online adaptive calibration is performed to generate joint torque temporal characteristics. Dynamic task scheduling is then performed using an operational ergonomics evaluation model, including task scheduling optimization at the individual, workstation, and production line levels.

Benefits of technology

It achieves precise quantification of the operator's instantaneous load status, dynamically adjusts task scheduling, reduces fatigue risk, improves production line efficiency, and forms a closed-loop management system of perception-evaluation-scheduling-feedback-optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122596604B_ABST
    Figure CN122596604B_ABST
Patent Text Reader

Abstract

This invention relates to the field of intelligent manufacturing engineering management technology, and provides a digital management method for operational efficiency based on embodied intelligent data acquisition. The method includes: acquiring muscle activation state, joint motion state, and anthropometry parameters through wearable sensing devices to obtain individualized electromyography-mechanical mapping relationships; using these parameters to adaptively calibrate a musculoskeletal model online; calculating the net joint torque time-series characteristics based on kinematic parameters and signal confidence weights through inverse dynamic model fusion; inputting these into an operational efficiency evaluation model to output load scores and fatigue trend prediction values; and generating scheduling instructions at the individual, workstation, and production line levels according to preset conditions and issuing them to the manufacturing execution system, while incrementally updating the model using feedback signals. This invention achieves closed-loop management from precise perception of individual physiological load to dynamic scheduling of production line tasks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing engineering management technology, specifically to a digital management method for operational efficiency based on embodied intelligent data acquisition. Background Technology

[0002] Currently, in the on-site operation management of the manufacturing and logistics industries, the assessment of personnel operation efficiency and task scheduling usually rely on coarse-grained methods such as time statistics, output counting, or manual observation. However, in actual operation, there may still be problems such as the inability to obtain refined physiological data such as the operator's muscle activation level and joint torque in real time, and the difficulty in truly reflecting the instantaneous fatigue accumulation state. Moreover, existing task allocation and production line scheduling are mostly based on fixed rhythm and static workstation configuration, which cannot dynamically adjust task weight or production line rhythm according to the operator's real-time efficiency status, resulting in the accumulation of fatigue operation risks and production line efficiency loss. Therefore, there is an urgent need for a digital management method for operation efficiency that can realize the real-time perception of the operator's physiological load and make dynamic task scheduling accordingly. Summary of the Invention

[0003] To address the problem that existing technologies cannot dynamically schedule tasks based on the operator's real-time physiological load, this invention provides a digital management method for operational efficiency that uses embodied intelligent data acquisition to achieve closed-loop management from precise perception of individual physiological load to dynamic scheduling of production line tasks.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: Wearable sensing devices are used to collect muscle activation data, joint motion data, and anthropometric parameters of the operator. An individualized electromyography (EMG) mapping relationship is obtained through calibration movements. Using the anthropometric parameters and the individualized EMG mapping relationship, an online adaptive calibration of the musculoskeletal model is performed to obtain a calibrated musculoskeletal model. Based on the joint motion data, kinematic parameters of each joint of the operator are calculated, and the confidence weight of each signal source is dynamically calculated according to signal quality indicators. Based on the calibrated musculoskeletal model, the kinematic parameters, and the confidence weights, a fusion calculation is performed using an inverse dynamics model. During the fusion calculation, the contribution ratio of each signal source is allocated according to the confidence weights to obtain the net joint torque of each joint of the operator, generating a joint torque temporal feature. The joint... The torque timing characteristics are input into a pre-constructed operational efficiency evaluation model, which outputs the operator's operational load score and fatigue trend prediction value. In response to the operational load score or fatigue trend prediction value meeting preset scheduling conditions, corresponding task scheduling optimization instructions are generated according to the scheduling level, which includes individual-level scheduling, workstation-level scheduling, and production line-level scheduling. Specifically: individual-level scheduling includes adjusting the corresponding operator's task intensity or inserting micro-rests; workstation-level scheduling includes triggering workstation ergonomic redesign suggestions when multiple operators experience high loads consecutively; and production line-level scheduling includes triggering an overall reduction in production line takt time when multiple workstations experience high loads simultaneously. The task scheduling optimization instructions are sent to the manufacturing execution system for execution, and the actual feedback signal after execution is obtained. The operational efficiency evaluation model is incrementally updated using the actual feedback signal. As one implementation method, the net joint torque is smoothed and outlier removed to obtain the joint torque timing characteristics. This implementation method improves the data quality and stability of the input evaluation model by smoothing and processing the original net joint torque data.

[0005] As one implementation, the calibrated musculoskeletal model includes muscle origin and insertion points, muscle physiological cross-sectional area, and muscle force-length relationship parameters. The fusion calculation based on the calibrated musculoskeletal model, the kinematic parameters, and the confidence weights via an inverse dynamics model includes: determining the lever arm vector of each muscle group based on the muscle origin and insertion points, where the lever arm vector represents the geometric distance and direction of the muscle action line relative to the joint rotation center; calculating the instantaneous muscle force of each muscle group based on the muscle physiological cross-sectional area and the muscle force-length relationship parameters, combined with the surface electromyography signal amplitude in the muscle activation state data, using a muscle activation degree-force mapping function, where the muscle activation degree-force mapping function represents the nonlinear relationship between muscle activation level and voluntary force output; and calculating the muscle torque components generated by each muscle group at the corresponding joint based on the lever arm vector and the instantaneous muscle force, and vector-superimposing the muscle torque components of all relevant muscle groups to obtain the net joint torque, wherein the vector superposition process considers the co-contraction effect of antagonistic and voluntary muscle groups. This implementation significantly improves the physical accuracy and individual adaptability of muscle force and joint torque calculations by introducing biomechanical parameters such as muscle origin and insertion points, physiological cross-sectional area, and force-length relationship, and by combining the antagonistic muscle co-contraction effect.

[0006] As one implementation, the pre-built operational ergonomics evaluation model includes a multi-head self-attention mechanism and a temporal position encoding encoder network. The multi-head self-attention mechanism is used to capture long-range dependencies in the temporal features of joint torques, and the temporal position encoding is used to preserve the temporal order information of the signals. The step of inputting the temporal features of joint torques into the pre-built operational ergonomics evaluation model and outputting the operator's operational load score and fatigue trend prediction value includes: embedding the temporal features of joint torques into positions using the temporal position encoding to obtain position-encoded features; inputting the position-encoded features into the encoder network containing the multi-head self-attention mechanism for feature interaction, and outputting the operational load score and the fatigue trend prediction value. This implementation utilizes the multi-head self-attention mechanism and temporal position encoding to effectively capture long-range temporal dependencies and order information in the joint torque sequence, improving the accuracy of fatigue trend prediction.

[0007] As one implementation, the step of generating corresponding task scheduling optimization instructions according to the scheduling level in response to the operation load score or the fatigue trend prediction value meeting the preset scheduling conditions includes: triggering a hierarchical scheduling event when the operation load score falls into the high load range or the fatigue trend prediction value exceeds the warning threshold; in response to the hierarchical scheduling event, executing the following according to the scheduling level: for the individual-level scheduling, generating task allocation adjustment instructions and micro-rest insertion instructions for the operator; for the workstation-level scheduling, outputting a workstation ergonomics evaluation report and redesign suggestion parameters; for the production line-level scheduling, using a multi-objective optimization algorithm to calculate the task allocation weight adjustment amount for each operator, the correction value of the current workstation production line takt time, and the optimization scheme of the shift sequence with the objective function of minimizing the fatigue difference of all employees and maximizing the production line efficiency; encapsulating the task allocation adjustment instructions, the micro-rest insertion instructions, the redesign suggestion parameters, the task allocation weight adjustment amount, the correction value, and the optimization scheme into the task scheduling optimization instructions, and writing them into the manufacturing execution system for execution through an application programming interface. This implementation method achieves collaborative decision-making from individual health protection to overall production line efficiency optimization through a hierarchical scheduling mechanism and multi-objective optimization algorithm. It also ensures the executability of scheduling instructions and system integration by linking with the manufacturing execution system through a standard interface.

[0008] As one implementation method, augmented reality smart glasses and multi-finger haptic feedback gloves are used to simultaneously collect the operator's eye-tracking gaze distribution data and fine motor skills data. The eye-tracking gaze distribution data includes gaze coordinates, gaze duration, and saccade path, while the fine motor skills data includes finger joint angles, fingertip tactile pressure values, and pinching force. This implementation method expands the dimensions of multimodal perception, providing a data foundation for assessing cognitive and fine motor skills workload.

[0009] As one implementation method, the eye-tracking fixation point distribution data and the fine hand operation data are synchronized and aligned on the time axis to form a multimodal temporal data pair for eye-hand coordination. This multimodal temporal data pair is input into an edge-side embodied intelligent model based on a lightweight Transformer. The edge-side embodied intelligent model includes an eye-tracking attention encoding branch and a hand posture encoding branch. These two branches interact through a cross-modal attention mechanism, outputting an operation intention category and an ergonomic risk index. When the ergonomic risk index exceeds a preset risk threshold, the current operation is determined to be in a high-risk state. In response to the high-risk state determination, visual guidance prompts are superimposed on the operator's field of vision through the augmented reality smart glasses. These visual guidance prompts include a highlighted target operation area and a virtual guide line to correct the movement trajectory. Simultaneously, a vibration warning signal is emitted from the corresponding finger part of the multi-finger haptic feedback glove. The vibration frequency of the vibration warning signal is positively correlated with the ergonomic risk index, and the vibration duration covers the duration of the high-risk state, thereby achieving real-time correction and guidance during the operation. This implementation method achieves operational intent understanding and risk prediction through end-side cross-modal modeling, and combines visual and tactile feedback for real-time correction, thereby improving operational safety and standardization.

[0010] As one implementation method, the operator's global physical load characteristics are extracted using the muscle activation state data as a baseline physiological signal; the operator's cognitive load characteristics are extracted using the eye movement fixation point distribution data, including fixation concentration, scanning rate, and pupil change rate; the operator's fine manipulation load characteristics are extracted using the hand fine manipulation data, including finger joint range of motion, pinch force change rate, and operation rhythm stability; the global physical load characteristics, cognitive load characteristics, and fine manipulation load characteristics are weighted and fused to generate a multi-dimensional ergonomic comprehensive score; the operational load score output by the operational ergonomics evaluation model is corrected and calibrated using the multi-dimensional ergonomic comprehensive score, wherein the correction and calibration is a real-time output compensation of the operational load score during the inference stage, without changing the model parameters of the operational ergonomics evaluation model. This implementation method, by fusing physical, cognitive, and fine manipulation three-dimensional load characteristics to correct and calibrate the basic operational load score, makes the ergonomics evaluation results more comprehensive and more in line with actual working conditions.

[0011] As one implementation method, multiple heterogeneous embodied acquisition agents are deployed on the production line for collaborative data acquisition. These agents include a depth camera array fixedly installed around the workstation and an autonomous mobile embodied agent carrying a robotic arm. The autonomous mobile embodied agent navigates autonomously within the production line environment and actively seeks target acquisition perspectives, collaborating with the depth camera array to acquire complete skeletal motion data and hand movement trajectories of the operator. The multi-view depth images acquired by the depth camera array are spatiotemporally registered and fused with the dynamic view depth images acquired by the autonomous mobile embodied agent to reconstruct the operator's 3D skeletal motion sequence. This includes: extracting the coordinates of key skeletal points from the multi-view depth images and the dynamic view depth images; synchronizing the key skeletal point coordinates based on timestamps and aligning the spatial coordinate system using an iterative nearest-point algorithm; and inputting the aligned key skeletal point coordinates into a graph neural network for topological relationship inference to reconstruct the 3D skeletal motion sequence containing complete joint connections. This implementation method, through collaborative acquisition by fixed and mobile agents, solves the problem of single-view occlusion and achieves high-precision 3D motion reconstruction without blind spots.

[0012] As one implementation method, a spatiotemporal graph convolutional network is used to perform real-time motion recognition and cycle segmentation on the 3D skeletal motion sequence, extracting the effective operation cycle duration, invalid delay duration, and operation rhythm features of each workstation. The effective operation cycle duration, invalid delay duration, and operation rhythm features of each workstation are mapped onto a production line layout diagram to generate a real-time production line efficiency heatmap. Workstations with efficiency below a first preset threshold or delay exceeding a second preset threshold are identified as bottleneck workstations based on the real-time production line efficiency heatmap, and bottleneck workstation identifiers and their corresponding bottleneck feature parameters are output. Minimizing the bottleneck feature parameters is the first optimization objective, and maximizing the overall throughput of the production line is the second optimization objective. A swarm intelligence optimization algorithm is used to search for the optimal reconstruction strategy in the solution space. The reconstruction strategy includes a task splitting and redistribution scheme among workstations, a personnel sharing scheduling scheme among workstations, and a mobile auxiliary robot intervention scheme. The feasibility of the searched optimal reconstruction strategy is verified, including verifying whether the operational load of each workstation after reconstruction exceeds the operator's capacity limit and verifying whether the logistics rhythm after reconstruction matches. The optimal reconstruction strategy that passes the feasibility verification is transmitted via OPC. The UA protocol is converted into workstation configuration instructions and line speed adjustment instructions recognizable by the Manufacturing Execution System (MES), and written into the MES to perform dynamic adjustments to production line workstation configurations and online adjustments to production line cycle time. This implementation method constructs a production line efficiency heatmap based on multi-agent collaborative data collection, accurately identifies bottleneck workstations, and generates executable reconfiguration strategies through swarm intelligence optimization and feasibility verification, achieving dynamic optimization and adaptive adjustment at the production line level. Beneficial effects

[0013] This invention uses wearable sensing devices to collect muscle activation and joint motion states in real time, and combines individualized calibrated musculoskeletal models with inverse dynamics fusion calculations to generate joint torque temporal characteristics that reflect real physical load. This overcomes the shortcomings of traditional ergonomic assessments that rely on coarse-grained statistics and cannot obtain refined physiological data, and achieves accurate quantification of the operator's instantaneous load state.

[0014] This invention inputs the timing characteristics of joint torques into a pre-built operational efficiency evaluation model, outputs operational load scores and fatigue trend prediction values, and dynamically generates task scheduling optimization instructions at multiple levels, including individual, workstation, and production line levels. This solves the problem that existing task scheduling, which is based on fixed beats and static configurations, cannot respond to changes in individual states, and achieves a leap from passive post-event statistics to proactive predictive efficiency management.

[0015] This invention sends task scheduling optimization instructions to the manufacturing execution system for execution and incrementally updates the operation efficiency evaluation model with actual feedback signals, forming a complete closed loop of "perception-evaluation-scheduling-feedback-optimization". This ensures seamless integration of scheduling instructions with the existing production system and ensures that the evaluation model can continuously adapt to individual differences and changes in working conditions, thereby improving the long-term applicability and robustness of the system. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a digital management method for operational efficiency of embodied intelligent data collection, according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, 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.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] Embodiment 1 of the present invention is as follows: like Figure 1 As shown, this embodiment provides a digital management method for operational efficiency based on embodied intelligent data collection. This method constructs a complete closed-loop framework from physiological perception to production line scheduling. Specifically, the method includes the following steps.

[0020] Step S100 involves collecting muscle activation data, joint motion data, and anthropometric parameters of the operator using wearable sensing devices, and obtaining the operator's individualized electromyography-mechanical mapping relationship through calibration movements. Specifically, wearable sensing devices can take various forms, such as flexible fabric sensors, inertial measurement units, or patch electrodes, as long as they can conform to the human body surface and acquire physiological and motion signals in real time. Muscle activation data reflects the electrophysiological activity level of muscle fibers, joint motion data characterizes the changes in posture and angular velocity of the limb in three-dimensional space, and anthropometric parameters include anatomical dimensions such as limb length and circumference. Calibration movements refer to guiding the operator to perform standardized isometric contractions or movements within a specific range before or during formal work, thereby establishing a correspondence between the operator's unique electromyographic signal amplitude and the muscle's active force output. Because there are significant differences in subcutaneous fat thickness, muscle fiber type, and skin impedance among individuals, general population models often fail to accurately reflect the true mechanical output of individuals. Therefore, obtaining a specific mapping relationship through individualized calibration can eliminate measurement bias caused by individual physiological differences at the source, laying the foundation for subsequent accurate assessment.

[0021] Step S200 involves online adaptive calibration of the musculoskeletal model using anthropometry parameters and individualized electromyography-mechanical mapping relationships, resulting in a calibrated musculoskeletal model. Specifically, general musculoskeletal models are typically built based on standard human databases, making it difficult to directly adapt to all operators. This step involves substituting measured anthropometry parameters into the model and adjusting geometric parameters such as bone length, joint center position, and muscle attachment points to match the model's topology with the operator's actual body shape. Simultaneously, individualized electromyography-mechanical mapping relationships are used to correct the muscle biomechanical properties in the model. Crucially, online adaptive calibration is a continuous or periodically triggered process because during prolonged work, sensors may experience slight displacement, skin perspiration leading to impedance changes, or the muscles themselves may undergo fatigue-related changes, all of which can cause the initial calibration parameters to become invalid. By monitoring signal quality and model prediction residuals in real time and dynamically updating calibration parameters, it is possible to ensure that the musculoskeletal model maintains a high-fidelity representation of the operator's current physiological state throughout the entire work cycle, avoiding the accumulation of evaluation errors due to model mismatch.

[0022] Step S300 involves calculating the kinematic parameters of each joint of the operator based on joint motion state data, and dynamically calculating the confidence weight of each signal source according to signal quality indicators. Specifically, kinematic parameters include joint angles, angular velocities, and angular accelerations, which are necessary inputs for inverse dynamics calculations. Signal quality indicators may include the signal-to-noise ratio of electromyography (EMG) signals, baseline drift, data integrity of inertial sensors, or occlusion probability of visual capture. The confidence weights are not fixed values ​​but are dynamically adjusted according to real-time changes in the working environment and sensor status. For example, if a large swing of the operator's arm causes poor contact of an EMG sensor, the confidence weight of that channel will automatically decrease; conversely, if a signal is stable and clear, its weight will increase accordingly. This dynamic weighting mechanism enables the system to automatically identify and rely on high-quality data sources and suppress low-quality noise interference in complex and ever-changing industrial environments, thereby ensuring the robustness of subsequent fusion calculations.

[0023] Step S400 involves fusion calculation based on the calibrated musculoskeletal model, kinematic parameters, and confidence weights, using an inverse dynamics model. During the fusion calculation, the contribution ratio of each signal source is allocated according to its confidence weight to obtain the net joint torque of each joint and generate temporal characteristics of the joint torque. Specifically, the human musculoskeletal system is a typical redundant drive system, meaning multiple muscles jointly drive a joint, and the force distribution of each muscle cannot be uniquely determined solely by kinematic data. Introducing muscle activation state data as a constraint can effectively solve this static indeterminate problem. In the fusion calculation, the confidence weights act as a dynamic regulator: high-confidence electromyographic signals will dominate in muscle force estimation, while low-confidence signals are weakened or used only as auxiliary references. The final output net joint torque eliminates the influence of gravity, inertial forces, and external loads, purely reflecting the biomechanical cost borne by the operator to maintain the current movement. Arranging the net joint torques in a time series generates a joint torque time series characteristic that can characterize the dynamic evolution of the load. This is more revealing of the intrinsic laws of fatigue accumulation than a single instantaneous torque value.

[0024] In step S500, the joint torque temporal features are input into a pre-constructed operational ergonomics evaluation model, which outputs the operator's operational load score and fatigue trend prediction value. Specifically, the operational ergonomics evaluation model is a deep learning or machine learning model trained offline using a large amount of labeled data, and in this embodiment, it is only used as an inference engine. This model can extract implicit patterns that are highly correlated with subjective fatigue perception and objective performance decline from high-dimensional temporal features. The operational load score quantifies the current instantaneous workload, while the fatigue trend prediction value predicts the fatigue development trend in the future based on historical data windows. This dual output mechanism of "current state + future trend" enables the management system not only to respond to current overload risks but also to have a forward-looking early warning capability, thereby gaining a valuable time window for proactive intervention.

[0025] Step S600: In response to the operation load score or fatigue trend prediction value meeting the preset scheduling conditions, corresponding task scheduling optimization instructions are generated according to the scheduling level, which includes individual-level scheduling, workstation-level scheduling, and production line-level scheduling. Specifically, the preset scheduling conditions may be that the load score exceeds the safety threshold or the fatigue trend slope is abnormally steep. Individual-level scheduling targets individual operators, including adjusting their current task intensity, switching to a low-load process, or inserting a short micro-rest, aiming to alleviate immediate physiological stress. Workstation-level scheduling is triggered when multiple operators experience high loads at the same workstation, outputting ergonomic redesign suggestions, such as adjusting tooling height, optimizing material placement, or replacing auxiliary tools, aiming to eliminate systemic fatigue-causing factors from an environmental perspective. Production line-level scheduling is triggered when multiple workstations experience high loads simultaneously or the overall fatigue trend rises synchronously, including reducing the overall production line takt time or rebalancing the task allocation of each workstation, aiming to ensure the safety and sustainability of the entire production line. This hierarchical scheduling strategy avoids crude management and achieves precise matching from micro-level individual protection to macro-level system optimization.

[0026] Step S700 involves sending the task scheduling optimization instruction to the Manufacturing Execution System (MES) for execution and obtaining the actual feedback signals after execution. Specifically, the MES is the core control platform in industrial production, responsible for task assignment, equipment control, and personnel scheduling. Writing scheduling instructions into this system through a standard interface ensures that efficiency management decisions are seamlessly translated into actual production actions, rather than remaining merely theoretical suggestions. The actual feedback signals include objective data such as changes in production capacity, quality fluctuations, and subsequent improvements in operators' physiological indicators after instruction execution. These feedback signals constitute key evidence for verifying the effectiveness of scheduling and serve as a bridge connecting digital decision-making and physical execution.

[0027] Step S800 involves incrementally updating the operational efficiency evaluation model using actual feedback signals. Specifically, the personnel composition, process flow, and equipment status in industrial settings are constantly changing, and offline-trained models inevitably experience performance degradation over time. The incremental update mechanism uses the latest feedback data to fine-tune the model parameters, enabling it to continuously learn new operating conditions and individual adaptive changes without retraining from scratch. This not only reduces computational resource consumption but, more importantly, endows the system with the ability to continuously evolve, ensuring that efficiency evaluation and scheduling decisions maintain high accuracy and applicability throughout long-term operation, truly achieving data-driven closed-loop self-optimizing management.

[0028] Embodiment 2 of the present invention is as follows: Building upon Example 1, this example further refines the process of generating the temporal characteristics of joint torques and the internal mechanism of inverse dynamic fusion calculation. As one implementation method, the net joint torque is smoothed and outlier removed to obtain the temporal characteristics of the joint torques. Specifically, the originally calculated net joint torques often contain high-frequency noise from sensors, impedance fluctuations caused by skin sweat, or numerical jumps caused by momentary occlusion during motion capture. If directly used as input to the evaluation model, it can easily lead to false peaks or oscillations in fatigue trend prediction. Smoothing can employ methods such as sliding window averaging filtering, Savitzky-Golay filtering, or low-pass Butterworth filtering, aiming to preserve the overall trend of torque changes while suppressing high-frequency spikes. Outlier removal can employ statistical methods such as the 3σ principle, interquartile range method, or threshold judgment method based on physical constraints. For example, when the calculated joint torque instantaneously exceeds the anatomical limit of the joint or the rate of change at the previous moment exceeds the physiologically impossible range, it is marked as an invalid point and interpolated for repair. This preprocessing mechanism ensures that the data input to the subsequent evaluation model has a high degree of temporal continuity and physical rationality, avoiding the triggering of erroneous scheduling instructions due to signal quality fluctuations.

[0029] As one implementation method, the calibrated musculoskeletal model includes muscle origin and insertion points, muscle physiological cross-sectional area, and muscle force-length relationship parameters. Specifically, these three parameters are the cornerstone of constructing a high-fidelity individualized biomechanical model. The muscle origin and insertion points determine the geometric lever efficiency of the muscle lever arm. This can be obtained by scaling coordinate points in standard anatomical atlases using affine transformations based on the operator's measured limb length, or by personalized calibration through medical image segmentation or touch-based localization of surface landmarks. The muscle physiological cross-sectional area directly characterizes the muscle's maximum isometric contractile force potential. Since it is difficult to measure directly on the production line, it is usually estimated using multiple regression equations based on anthropometric parameters such as height, weight, and limb circumference, or by using portable ultrasound equipment to collect thickness data of key muscle groups for estimation. The muscle force-length relationship parameter describes the change in the muscle's ability to generate active force under different stretching states. It is usually characterized using the classic Hill-type muscle model or higher-order polynomial curves. This parameter allows the model to distinguish the actual force differences caused by different joint angles at the same electromyographic level, thereby significantly improving the dynamic accuracy of torque estimation.

[0030] Furthermore, based on the calibrated musculoskeletal model, kinematic parameters, and confidence weights, the specific process of fusion calculation through the inverse dynamics model includes: determining the lever arm vector of each muscle group according to the origin and insertion points of the muscles; the lever arm vector represents the geometric distance and direction of the muscle action line relative to the joint rotation center; based on the physiological cross-sectional area of ​​the muscles and the force-length relationship parameters of the muscles, combined with the amplitude of surface electromyography signals in the muscle activation state data, calculating the instantaneous muscle force of each muscle group through the muscle activation degree-force mapping function; the muscle activation degree-force mapping function represents the nonlinear relationship between the muscle activation level and the active force output; based on the lever arm vector and the instantaneous muscle force, calculating the muscle torque components generated by each muscle group at the corresponding joint, and vector superimposing the muscle torque components of all related muscle groups to obtain the net joint torque. The vector superposition process considers the co-contraction effect of antagonistic and agonist muscle groups. Specifically, the muscle activation degree-force mapping function is not a simple linear proportional relationship, but a complex nonlinear function modulated by multiple factors such as the current muscle length, contraction speed, and fatigue state. Only by introducing this nonlinear mapping can the force characteristics of human muscles under different postures be truly reflected. More importantly, when the human body performs delicate operations or heavy-duty tasks, the nervous system often simultaneously activates agonist and antagonist muscles to increase joint stiffness and maintain postural stability; this phenomenon is known as co-contraction. If only the torque generated by the agonist muscles is calculated during vector superposition while ignoring the reverse torque of the antagonist muscles, although the calculated net joint torque may be close to the external load requirements in numerical terms, it will severely underestimate the contact pressure inside the joint and the actual metabolic consumption of the muscles, leading to misjudgments of the operator's fatigue level and risk of injury. This embodiment, by explicitly modeling and superimposing the torque components of the antagonist muscle group in the fusion calculation, restores the true biomechanical cost of human movement, making the generated joint torque temporal characteristics not only reflect "how much work was done" but also more accurately represent "how much physiological cost was incurred," providing a physically realistic data foundation for subsequent ergonomics evaluation.

[0031] Embodiment 3 of the present invention is as follows: Building upon Example 1, this example further refines the internal architecture and operational mechanism of the operational ergonomics evaluation model in step S500. As one implementation method, the pre-constructed operational ergonomics evaluation model includes a multi-head self-attention mechanism and a temporal position encoding encoder network. Specifically, this model is not a simple fully connected neural network or a traditional convolutional neural network, but rather employs an encoder structure based on the Transformer architecture. The core reason for choosing this architecture is that human fatigue accumulation is a typical long-term temporal dependent process. The current physiological state depends not only on the instantaneous intensity of the action but also on the nonlinear modulation of the load history over the past tens of minutes or even hours. Traditional recurrent neural networks are prone to the gradient vanishing problem when processing long sequences, making it difficult to capture the contribution of early load signals to the current fatigue state. While convolutional neural networks can extract local features, their receptive field is limited, often requiring extremely deep network layers to cover the entire work cycle, increasing inference latency. In contrast, the encoder network used in this example can directly establish the correlation between any two time points in the sequence through a self-attention mechanism, regardless of their distance on the time axis, thereby accurately representing the dynamic process of fatigue evolution.

[0032] Furthermore, multi-head self-attention mechanisms are used to capture long-range dependencies in the temporal features of joint torques, while temporal position encoding preserves the temporal order information of the signals. Specifically, because the computation of self-attention mechanisms is permutation-invariant (i.e., it does not naturally perceive the order of data), and joint torque data, as a time series, contains crucial causal logic in its sequence—for example, "the high-intensity torque peak appears first, followed by a decrease in muscle activation"—is a typical fatigue compensation pattern. Disrupting this order completely loses its physical meaning. Therefore, temporal position encoding must inject absolute or relative temporal position information into the feature vector, enabling the model to distinguish between "the current torque" and "the torque ten minutes ago." Multi-head self-attention mechanisms map input features to multiple different subspaces, where each "head" can focus on learning dependencies at different time scales or semantic dimensions. For example, one attention head might focus on capturing instantaneous impact loads at the second level, another head on tracking continuous static loads at the minute level, and yet another head might specifically identify periodic rest and recovery patterns. This multi-perspective feature extraction capability enables the model to comprehensively consider multiple fatigue-causing factors, avoiding evaluation bias caused by a single feature dimension.

[0033] At the execution level, the process of inputting the temporal features of joint torques into a pre-built operational ergonomics evaluation model and outputting the operator's operational load score and fatigue trend prediction value includes: embedding the temporal features of joint torques into positions through temporal position encoding to obtain position-encoded features; inputting the position-encoded features into an encoder network containing a multi-head self-attention mechanism for feature interaction, and outputting the operational load score and fatigue trend prediction value. Specifically, position embedding can employ various methods such as sine / cosine position encoding, learnable embedding vectors, or rotational position encoding, as long as it provides an effective temporal index for the model. Feature interaction refers to the process in the multi-layered stacked structure of the encoder network where each layer updates its feature representation through self-attention calculation and combines it with a feedforward neural network for nonlinear transformation, gradually abstracting higher-order fatigue semantic features from the original torque values. The final output operational load score reflects the comprehensive physiological cost at the current moment, while the fatigue trend prediction value is an estimate of the fatigue development slope over a future period based on current and historical features. It should be understood that this embodiment emphasizes the inference application process of the model. The model is trained and converged in the offline stage using a large amount of labeled historical ergonomic data and then deployed to the production environment. In actual operation, it is only used as a deterministic mapping function and does not involve online training or parameter adjustment.

[0034] To more clearly illustrate the technical advantages of this embodiment, a comparison is made below using two typical work modes as examples. Assume operator A is in a "continuous moderate load" state, and operator B is in an "intermittent extremely high load" state. Their average joint torque amplitudes per unit time may be very similar. If only time-domain statistical characteristics (such as mean and variance) are used as the evaluation basis, the system would easily classify their fatigue risk as the same level. However, in actual physiological mechanisms, operator B's intermittent extremely high load leads to faster accumulation of metabolic products and a decline in neuromuscular control, resulting in a significantly higher rate of fatigue accumulation than operator A. This embodiment preserves the precise temporal structure of load fluctuations through temporal position encoding and captures the specific fatigue-causing pattern of "peak-short interval" using a multi-head self-attention mechanism. This allows for accurate differentiation between these two seemingly similar but actually very different work states, significantly improving the physical realism and early warning accuracy of fatigue trend prediction.

[0035] Embodiment 4 of the present invention is as follows: Building upon Example 1, this example further elaborates on the generation logic of the task scheduling optimization instruction in step S600, the multi-objective optimization mechanism, and the interaction details with the manufacturing execution system. As one implementation method, in response to the operation load score or fatigue trend prediction value meeting preset scheduling conditions, corresponding task scheduling optimization instructions are generated according to the scheduling level. Specifically, this process is not a simple threshold comparison, but a complete decision-making chain encompassing status monitoring, event triggering, strategy calculation, and instruction encapsulation.

[0036] First, regarding the setting of triggering conditions, a hierarchical scheduling event is triggered when the operational load score falls into the high load range or the fatigue trend prediction value exceeds the warning threshold. Specifically, the high load range is usually defined based on the statistical distribution of historical work efficiency data or industry safety standards. For example, after normalizing the operational load score, 0.7 to 0.9 is defined as the high load range, and above 0.9 is defined as the extremely high risk range. The fatigue trend prediction value reflects the rate of fatigue accumulation over a future period. The warning threshold is set to identify potential risk states where fatigue is increasing too rapidly even though the current load is acceptable. Using "OR" logic as the triggering condition allows the system to cope with instantaneous physiological overload and proactively intervene in chronic fatigue accumulation, avoiding missed or false judgments caused by a single indicator. It should be understood that the high load range and warning threshold can be statically configured fixed values ​​or adaptive parameters dynamically adjusted based on the operator's individual baseline data; this invention does not limit this.

[0037] Secondly, the system generates significantly differentiated and targeted instructions for different scheduling levels. In response to hierarchical scheduling events, the system performs the following operations according to the scheduling level. For individual-level scheduling, it generates task allocation adjustment instructions and micro-rest insertion instructions for the operator. Specifically, task allocation adjustment instructions can manifest as reducing the task weight of the operator's subsequent tasks, reducing the maximum single-transfer mass, or temporarily switching high-precision assembly tasks to low-cognitive-load auxiliary tasks. Micro-rest insertion instructions, without affecting the overall cycle time, forcibly insert a short recovery window, such as a 15-60 second active relaxation period, during the current process interval. This immediate intervention aims to quickly alleviate the individual's acute physiological stress and prevent injury.

[0038] For workstation-level scheduling, the system outputs a workstation ergonomics assessment report and redesign recommendations. Specifically, when multiple operators consistently exhibit high workload characteristics at the same workstation, it indicates a systemic ergonomic design flaw in the workstation itself, rather than individual differences. The generated assessment report includes a historical load heatmap of the workstation, high-frequency fatigue-inducing action segments, and environmental factor analysis. The redesign recommendations provide specific quantitative indicators for engineering improvements, such as suggesting lowering the material rack height by 5 cm, adjusting the tooling fixture tilt angle by 10 degrees, or increasing the auxiliary torque setting of the robotic arm. These parameters can be directly fed back to the process engineering department as a basis for production line modifications, eliminating potential fatigue hazards at their source.

[0039] For production line-level scheduling, with the objective functions of minimizing overall employee fatigue variation and maximizing production line efficiency, a multi-objective optimization algorithm is used to calculate the task allocation weight adjustment for each operator, the correction value of the current workstation's production line takt time, and the optimized shift rotation sequence. Specifically, this is a typical multi-objective constrained optimization problem. "Minimizing overall employee fatigue variation" in the objective function reflects consideration of worker health fairness, avoiding situations where some employees are overworked while others are idle; "maximizing production line efficiency" ensures the economy of production management. These two objectives often conflict, therefore, a set of non-dominated solutions needs to be obtained through Pareto optimality, and then the final execution plan is selected based on the current production priority. Constraints include, but are not limited to, the skill qualification matrix of each operator, the maximum operating speed of equipment, the work-in-process buffer capacity, and statutory working hour limits. In addition, the constraints include job switching cost constraints, specifically: imposing an upper limit on the frequency of job switching between adjacent time periods in the job rotation sequence, to avoid operators experiencing decreased proficiency, increased operational error rates, and fluctuations in output per workstation due to the learning curve effect caused by frequent job changes. This ensures that the optimization results balance fatigue distribution without compromising the overall efficiency of the production line and the stability of product quality. Through this global optimization, the system can automatically balance the load distribution of each workstation while ensuring that overall output targets are met, or, when necessary, reduce the overall production line takt time to maintain the continuous working capacity of personnel.

[0040] Finally, at the instruction issuance and execution level, task allocation adjustment instructions, micro-rest insertion instructions, redesign suggestion parameters, task allocation weight adjustment amounts, correction values, and optimization schemes are encapsulated into task scheduling optimization instructions, and written into the Manufacturing Execution System (MES) for execution via an application programming interface (API). Specifically, the encapsulation process converts the aforementioned heterogeneous decision results into a structured data format conforming to the MES interface specifications, such as JSON objects, XML documents, or standardized OPC UA information models. Writing via the API ensures seamless integration and automated closed-loop between the ergonomics management module and the production control platform, eliminating the need for manual transcription or secondary confirmation. Upon receiving the instructions, the MES automatically updates the scheduling queue, adjusts conveyor line speeds, or pushes notifications to field terminals, and sends back feedback signals such as actual capacity changes and personnel response status after execution. This is used to verify the effectiveness of the scheduling strategy and support subsequent incremental model updates. This standardized interaction mechanism not only enhances the system's industrial applicability but also provides clear and visible data interface evidence for potential future technical rights protection.

[0041] Embodiment 5 of the present invention is as follows: Building upon Example 1, this example further constructs a real-time correction and model adjustment sub-loop based on multimodal perception to compensate for the shortcomings of single body kinematic perception in cognitive state and fine manipulation dimensions. As one implementation method, augmented reality smart glasses and multi-finger haptic feedback gloves simultaneously collect the operator's eye-tracking gaze distribution data and fine hand manipulation data. Specifically, the augmented reality smart glasses integrate an eye-tracking module, capable of recording the operator's visual search behavior during the task at a high sampling rate. The eye-tracking gaze distribution data includes gaze coordinates, gaze duration, and saccade path. The gaze coordinates reflect the operator's current visual focus area, the gaze duration characterizes the depth of information processing for a specific target, and the saccade path reveals the operator's attention shifting strategy and search efficiency among multiple work objects. Simultaneously, the multi-finger haptic feedback gloves incorporate flexible bending sensors and a pressure sensor array to capture fine hand manipulation data, including finger joint angles, fingertip tactile pressure values, and pinching force. The angles of each finger joint precisely describe the hand's grasping posture and dexterity, the tactile pressure values ​​at the fingertips reflect the mechanical interaction state of the contact interface, and the magnitude of the pinching force is directly related to the operator's control stability over the tool or workpiece. This "eye-hand" synchronous acquisition mechanism enables the system to simultaneously acquire the operator's cognitive intention input and physical execution output, providing a complete data foundation for subsequent understanding of complex work behaviors.

[0042] As one implementation method, eye-tracking gaze distribution data and fine motor hand operation data are synchronized and aligned on the time axis to form a multimodal temporal data pair for eye-hand coordination. Specifically, since eye-tracking signals and hand movement signals are typically acquired by different hardware, their sampling rates and transmission delays may differ. Therefore, microsecond-level temporal alignment is required using a unified timestamp or hardware trigger signal. Only by analyzing the aligned data pair can the "eye-hand coordination" pattern or the cognitive-executive separation phenomenon of "eye looking but hand not moving" be accurately captured. The aligned multimodal temporal data pair is input into a lightweight Transformer-based edge embodied intelligence model. This model includes an eye-tracking attention encoding branch and a hand posture encoding branch. The two branches interact through a cross-modal attention mechanism, outputting the operation intention category and ergonomic risk index. Specifically, the lightweight Transformer architecture is adopted to meet the stringent real-time requirements of industrial environments, ensuring inference latency is controlled within milliseconds, thereby supporting immediate safety intervention. The eye-tracking attention encoding branch focuses on extracting cognitive load features from visual search patterns, while the hand posture encoding branch extracts biomechanical risk features during action execution. The cross-modal attention mechanism allows the model to dynamically establish the association weights between visual focus and hand movements. For example, when it detects that the operator's gaze has left the current workstation but the hand is still applying force, the model can identify this potential risk of misoperation through the cross-attention layer. The output operation intention category can be a semantic label such as "normal assembly," "searching for parts," "adjusting posture," or "abnormal pause," while the ergonomic risk index is a continuous value that quantifies the degree to which the current operation deviates from standard safety specifications.

[0043] Furthermore, when the ergonomic risk index exceeds a preset risk threshold, the current operation is determined to be in a high-risk state. In response to this high-risk determination, visual guidance prompts are overlaid in the operator's field of vision via augmented reality smart glasses. These prompts include a highlighted target operation area and a virtual guide line to correct the movement trajectory. Simultaneously, vibration warning signals are emitted from the corresponding finger parts of the multi-finger haptic feedback glove. The vibration frequency of these signals is positively correlated with the ergonomic risk index, and the vibration duration covers the duration of the high-risk state, enabling real-time correction and guidance of the operation process. Specifically, this dual feedback mechanism follows the cognitive engineering principles of human-computer interaction. Visual guidance provides positive error correction information, indicating "how to do it," while the highlighted display quickly captures the operator's peripheral visual attention, redirecting their gaze back to the correct area. The virtual guide line presents the standard movement path in an intuitive spatial trajectory, reducing the operator's cognitive reconstruction burden. The haptic vibration warning provides immediate negative feedback, indicating "what was done wrong." Compared to auditory or visual warnings, the haptic channel has higher anti-interference capabilities and wakefulness in noisy or visually overloaded industrial environments. More importantly, the design, which shows a positive correlation between vibration frequency and risk index, means that the feedback signal itself carries information about the risk level: minor risk deviations trigger only low-frequency, gentle vibrations, serving as a reminder without disrupting the workflow; while serious safety hazards trigger high-frequency, rapid vibrations, forcibly alerting the operator. This tiered feedback strategy avoids the desensitization problem caused by the "boy who cried wolf" effect, ensuring a balance between the effectiveness of real-time correction and user experience.

[0044] In addition to real-time correction, multimodal perception data is also used to improve the long-term accuracy of the main evaluation model. As one implementation method, muscle activation state data is used as the baseline physiological signal to extract the operator's global physical load characteristics. Specifically, muscle activation state data directly reflects the metabolic consumption level of muscle fibers and is the most objective physiological gold standard for measuring physical exertion. The extracted global physical load characteristics form the foundation of ergonomics evaluation. Eye movement fixation point distribution data is used to extract the operator's cognitive load characteristics, including fixation concentration, scan rate, and pupillary change rate. Specifically, lower fixation concentration and higher scan rate usually indicate that the operator is experiencing difficulty in information retrieval or decision-making uncertainty; pupillary change rate is a sensitive indicator of autonomic nervous system activity, reflecting the degree of mental tension and the utilization of mental resources. Fine manipulation data is used to extract the operator's fine manipulation load characteristics, including finger joint range of motion, pinch force change rate, and operational rhythm stability. Specifically, an excessive range of motion in the finger joints may indicate compensatory manipulation under unnatural postures, drastic fluctuations in the rate of change of pinching force reflect a decline in force control, and loss of operational rhythm stability is often an early behavioral sign of neuromuscular fatigue.

[0045] Based on this, the global physical load characteristics, cognitive load characteristics, and fine operational load characteristics are weighted and fused to generate a multi-dimensional ergonomic comprehensive score. Specifically, the weighted fusion is not a simple linear superposition, but rather the weights are dynamically adjusted according to the current job type. For example, in precision assembly processes, the weights of cognitive load and fine operational load are increased; while in heavy object handling processes, the weight of global physical load dominates. This adaptive fusion mechanism ensures that the comprehensive score can fully and evenly represent the operator's true load state under different working conditions. Finally, the operational load score output by the operational ergonomics evaluation model is corrected and calibrated using the multi-dimensional ergonomic comprehensive score. The correction and calibration is a real-time output compensation of the operational load score during the inference stage, without changing the model parameters of the operational ergonomics evaluation model. Specifically, the operational load score generated based on the joint torque time sequence characteristics in Example 1 mainly reflects the macroscopic biomechanical cost, but there may be perceptual blind spots in some scenarios. For example, although the operator's joint torque is not large, they may feel extremely fatigued due to high mental tension, or the overall torque may not exceed the standard due to local muscle strain in the hand, but the local area may be close to the damage threshold. This embodiment effectively fills these blind spots by introducing supplementary information from the cognitive and fine operational dimensions. The correction and calibration process can involve superimposing the multi-dimensional ergonomic comprehensive score as a residual term onto the original score, or using it as a gating coefficient to nonlinearly modulate the original score. This hierarchical architecture of "basic assessment + multi-modal correction" retains the physical interpretability of the dynamic assessment in Embodiment 1 while enhancing the system's robustness to complex human factors, making the final ergonomic management decisions more accurate and human-centered.

[0046] Embodiment 6 of the present invention is as follows: Building upon Example 1, this example further expands the scope of ergonomics perception from individual wearable devices to production line-level environmental perception, and achieves dynamic reconfiguration of the production line through multi-agent collaboration and swarm intelligence optimization. As one implementation method, multiple heterogeneous embodied acquisition agents are deployed on the production line for collaborative data acquisition. These agents include a depth camera array fixedly installed around the workstations and an autonomous mobile embodied agent carrying a robotic arm. Specifically, this heterogeneous combination aims to solve the occlusion and blind spot problems inherent in a single fixed perspective in complex industrial environments. The fixed depth camera array provides all-weather, high-frame-rate basic coverage, stably capturing operational actions in the core areas of the workstations; while the autonomous mobile embodied agent possesses spatial mobility, with its robotic arm end-effector integrated with a depth sensor, allowing it to autonomously navigate to specific locations based on real-time task requirements, actively seeking and locking onto the optimal acquisition perspective. For example, when a workstation's view is obstructed due to temporary increases in material stacking, the mobile agent can automatically adjust its posture to fill the gap from the side and above. This collaborative model of "fixed benchmark + mobile enhancement" ensures the integrity and robustness of data acquisition, laying a physical foundation for subsequent accurate analysis.

[0047] Furthermore, the autonomous mobile embodied agent navigates autonomously in the pipeline environment and actively seeks target acquisition perspectives, collaborating with a depth camera array to acquire complete skeletal motion data and hand operation trajectories of the operator. Specifically, this "active search" differs from traditional preset path navigation; it means the agent can plan its movement trajectory in real time to maximize observational information gain based on the confidence feedback of human keypoint detection in the current scene. When the system detects that the loss rate of a certain operator's hand or torso keypoints exceeds a threshold, the mobile agent's perspective optimization behavior is triggered. Collaborative acquisition requires not only spatial complementarity but also strict temporal synchronization. All acquisition devices achieve microsecond-level clock alignment through the PTP protocol or hardware trigger signals to ensure the temporal consistency of multi-source data.

[0048] After acquiring multi-source data, it is necessary to perform spatiotemporal registration and fusion of multi-view depth images acquired by a depth camera array and dynamic view depth images acquired by an autonomous mobile embodied agent to reconstruct the operator's 3D skeletal action sequence. Specifically, this process includes three key sub-steps. First, the coordinates of skeletal keypoints are extracted from the multi-view depth images and dynamic view depth images; these coordinates constitute a sparse representation of human posture. Second, the skeletal keypoint coordinates are time-synchronized based on timestamps, and the spatial coordinate system is aligned using an iterative nearest-point algorithm. Since the mobile agent itself is in motion, the coordinate system of its acquired data is dynamically changing. Therefore, it is necessary to use SLAM technology to estimate its own pose, combined with the iterative nearest-point algorithm, to accurately transform the point cloud from the dynamic view to the global coordinate system of the fixed camera, eliminating geometric deviations caused by relative motion. Finally, the aligned skeletal keypoint coordinates are input into a graph neural network for topological relationship inference to reconstruct a 3D skeletal action sequence containing complete joint connections. Specifically, a simple point cloud or keypoint set lacks prior knowledge of human anatomy and is prone to limb misconnections or breaks. Graph neural networks model the human body as a graph structure, where nodes represent joints and edges represent skeletal connections. By learning the topological constraints of a large number of normal human movements, they can infer a complete skeleton that conforms to biomechanical laws from noisy or partially missing data. This reconstruction method based on deep learning topological reasoning significantly improves the adaptability to occlusion, self-similar poses, and rapid movements compared to traditional pure geometric matching algorithms, ensuring the accuracy of production line-level ergonomic analysis data.

[0049] Based on the reconstructed high-quality 3D skeletal motion sequence, this embodiment further realizes the automatic identification and dynamic optimization of production line bottlenecks. As one implementation method, a spatiotemporal graph convolutional network is used to perform real-time motion recognition and cycle segmentation of the 3D skeletal motion sequence, extracting the effective operation cycle duration, ineffective delay duration, and operation rhythm features for each workstation. Specifically, the spatiotemporal graph convolutional network performs convolution operations simultaneously in the spatial dimension (human joint topology) and the temporal dimension (motion evolution), enabling precise differentiation of fine-grained motion semantics such as "assembly," "transportation," "waiting," and "searching." The effective operation cycle duration reflects the time spent on operations that directly create value, while the ineffective delay duration quantifies the non-value-adding time caused by material shortages, tool malfunctions, or fatigue-induced pauses. The operation rhythm features describe the smoothness and stability of motion execution. These features constitute atomic indicators for evaluating workstation efficiency.

[0050] Subsequently, the valid operation cycle duration, invalid delay duration and operation rhythm characteristics of each workstation are mapped to the assembly line layout diagram to generate a real-time production line work efficiency heat map. Specifically, the heat map visually presents the efficiency status of each workstation through color coding, for example, red indicates high delay and low efficiency, green indicates normal operation, and yellow indicates potential risks. This visual expression enables managers to grasp the operation status of the entire production line at a glance. Workstations with efficiency lower than a first preset threshold or delay exceeding a second preset threshold are identified as bottleneck workstations according to the real-time production line work efficiency heat map, and the identifiers of the bottleneck workstations and their corresponding bottleneck characteristic parameters are output. The bottleneck characteristic parameters not only include the degree of the bottleneck, but also include the type of leading factor causing the bottleneck (such as "redundant motion type", "waiting type" or "overload type"), which provides a basis for subsequent targeted optimization.

[0051] For the identified bottlenecks, the system does not simply give an alarm, but automatically generates an optimization scheme. As an implementation mode, with the minimization of the bottleneck characteristic parameters as the first optimization objective and the maximization of the overall throughput of the production line as the second optimization objective, a swarm intelligence optimization algorithm is used to search for the optimal reconfiguration strategy in the solution space. Specifically, this is a typical bi-objective combinatorial optimization problem. Swarm intelligence optimization algorithms (such as particle swarm optimization, ant colony algorithm or genetic algorithm) can efficiently explore in a huge solution space, and simulate the cooperation and evolution mechanism of natural groups to find the Pareto optimal solution set. The reconfiguration strategy includes the splitting and redistribution scheme of tasks among workstations, the personnel sharing scheduling scheme among workstations, and the intervention operation scheme of mobile auxiliary robots. For example, if a certain workstation is overloaded for a long time due to excessively complex assembly processes, the algorithm may suggest splitting the "pre-assembly" subtask in this process to the previous idle workstation; or when the loads of two adjacent workstations are uneven, arrange a multi-skilled worker to move between the two for support; or schedule a mobile robot to undertake the material delivery link of this workstation to reduce manual handling time. This diverse strategy library enables the system to flexibly cope with different types of bottleneck problems.

[0052] Crucially, any optimization strategy must undergo rigorous safety and feasibility verification before implementation. The optimal reconfiguration strategy obtained through the search is then validated for feasibility. This validation includes verifying whether the operational load at each workstation exceeds the operator's capacity limit after reconfiguration and whether the reconfigured logistics cycle time matches the requirements. Specifically, efficiency improvements must never come at the expense of personnel health or disruption of production order. Load verification utilizes the individualized work efficiency assessment model established in Examples 1 to 5 to predict whether the operator's physiological load will exceed the limit under the new task allocation, preventing new fatigue risks caused by task reconfiguration. Logistics cycle time verification verifies whether the new work cycle is coordinated with the upstream and downstream conveyor line speeds and AGV delivery frequency, avoiding work-in-process accumulation or material shortages due to localized speed increases. Only strategies that pass both verifications will be adopted, reflecting the "people-oriented, safety first" technical philosophy of this invention.

[0053] Finally, the optimal reconfiguration strategy, verified for feasibility, is converted into workstation configuration instructions and line speed adjustment instructions recognizable by the Manufacturing Execution System (MES) via the OPC UA protocol. These instructions are then written into the MES to dynamically adjust the workstation configuration and online adjust the production line cycle time. Specifically, OPC UA, as a universal standard protocol for industrial interconnection, shields the communication differences between underlying heterogeneous devices, ensuring seamless integration between the efficiency optimization module and various mainstream MES systems. The workstation configuration instructions include information such as new work instruction version numbers, updated personnel schedules, and changes to robot task queues; the line speed adjustment instructions directly act on the frequency converter or main control PLC, smoothly adjusting the conveyor line's operating speed. The entire closed loop, from environmental perception, bottleneck identification, strategy generation to final execution, is fully automated, requiring no manual intervention, truly achieving adaptive intelligent management at the production line level. It should be understood that although this embodiment uses the OPC UA protocol as an example, other implementations may also use MQTT, RESTful API, or other industrial communication protocols, as long as data interaction with the MES can be achieved.

[0054] Embodiment 7 of the present invention is as follows: To more clearly illustrate the application value of the technical solution of this invention in actual industrial production, this embodiment takes the interior assembly process in an automobile assembly line as an example, demonstrating how the dispersed technical features in the aforementioned embodiments work together to form a complete closed loop from individual physiological perception to global optimization of the production line. It should be understood that although this embodiment uses automobile interior assembly as a specific scenario, the method of this invention is also applicable to other manufacturing or operational scenarios involving manual operation, such as electronic product assembly, heavy machinery assembly, and logistics sorting.

[0055] In step S701, at the dashboard installation station on the automotive interior assembly line, the operator wears a flexible exoskeleton integrated with fabric electromyography sensors and an inertial measurement unit, along with augmented reality smart glasses and multi-finger haptic feedback gloves. Simultaneously, a fixed depth camera array deployed above the station continuously collects basic motion data, while an autonomous mobile avatar carrying a robotic arm navigates the assembly line environment autonomously, actively seeking and locking onto the optimal viewing angle for clearly observing the operator's fine hand movements and torso posture. It collaborates with the fixed camera array to collect complete skeletal motion data and hand operation trajectories. Specifically, the dashboard installation process requires operators to frequently perform complex movements such as tilting their heads back and raising their arms, internal wrist rotation, and delicate wiring connections, which can easily lead to shoulder and wrist muscle fatigue and joint damage. Through this multimodal, multi-view collaborative acquisition mechanism, the system can comprehensively acquire raw data reflecting the operator's true physiological load and operational status without interfering with the normal production cycle, laying a solid foundation for subsequent accurate evaluation.

[0056] In step S702, during the operation, the edge-side embodied intelligent model receives and aligns eye-tracking gaze point distribution data and fine hand operation data in real time. It then interacts with features through a cross-modal attention mechanism, outputting the operation intention category and ergonomic risk index. When the system detects that the operator's biceps are under high load and there is a risk of excessive wrist deflection, causing the ergonomic risk index to exceed a preset risk threshold, the current operation is determined to be in a high-risk state. In response to this high-risk state determination, the augmented reality smart glasses immediately highlight the target area of ​​the dashboard wiring harness connection in the operator's field of vision and overlay a virtual guide line to correct the wrist angle. Simultaneously, the multi-finger haptic feedback glove emits high-frequency vibration warning signals at the corresponding thumb and index finger positions, with the vibration frequency positively correlated with the current ergonomic risk index. Specifically, this immediate dual feedback mechanism can transform abstract risk quantification indicators into sensory guidance that the operator can intuitively perceive within milliseconds, allowing them to subconsciously adjust their force application posture and operation path without interrupting the operation, thereby effectively preventing acute musculoskeletal injuries while ensuring assembly quality. Compared to traditional post-training or regular medical checkups, this embedded real-time correction significantly reduces human error rates and occupational disease incidence.

[0057] In step S703, while real-time correction is performed at the end-user side, the cloud-based operational efficiency evaluation model continuously outputs operational load scores and fatigue trend prediction values ​​based on joint torque timing characteristics. When the model predicts that the operator's fatigue trend will continue to rise and exceed the warning threshold within the next ten minutes, even if the current instantaneous load has not yet reached its limit, the system will automatically trigger an individual-level scheduling event, generate a micro-rest insertion instruction, and write it into the manufacturing execution system through the application programming interface. The manufacturing execution system then pushes a rest reminder to the terminal of the workstation during the interval after the current assembly cycle ends, and automatically extends the arrival interval of the next vehicle by fifteen seconds, providing the operator with a brief window for proactive recovery. Specifically, this proactive intervention based on trend prediction avoids the efficiency decline and quality fluctuations caused by operators forcing themselves to work in a state of extreme fatigue, reflecting the health management concept of "prevention before disease occurs." The micro-rest duration is not fixed but dynamically calculated based on the fatigue accumulation rate to ensure a balance between recovery effect and production cycle time.

[0058] In step S704, if the system detects similar high biceps load and wrist deflection characteristics in three different operators at the same instrument panel installation station within three consecutive shifts, it determines that the station has a systemic ergonomic design flaw, thus triggering a station-level scheduling event. The system automatically generates a station ergonomic assessment report and redesign suggestion parameters, pointing out that the current material rack height is too low, causing operators to frequently bend over to retrieve parts, and that the unreasonable angle of the wiring harness insertion position forces the wrist to be in a non-neutral position for a long time. It recommends raising the material rack height by eight centimeters and adjusting the tooling fixture tilt angle by fifteen degrees. Specifically, this level of scheduling goes beyond temporary protection for individuals, eliminating the root causes of fatigue from the environmental and tool levels. By converting quantified physiological load data into specific engineering improvement parameters, this invention breaks down the barriers between occupational health management and process engineering design, enabling ergonomic optimization to no longer rely on subjective experience or post-event questionnaires, but rather on precise decisions based on objective biomechanical data.

[0059] In step S705, at a more macro-level production line, the autonomous mobile embodied intelligent agent, during its inspection, discovers that the effective operating cycle time of the dashboard installation station is significantly longer than that of adjacent stations, and the proportion of ineffective delay time is too high. Spatiotemporal graph convolutional network analysis confirms it as the bottleneck station of the current production line. The system then uses a swarm intelligence optimization algorithm to search for the optimal reconfiguration strategy in the solution space, aiming to minimize the bottleneck characteristic parameters and maximize the overall throughput of the production line. After feasibility verification, it is confirmed that the operational load of each station after reconfiguration does not exceed the limit and the logistics cycle time is matched. The system ultimately determines to split some pre-assembly sub-tasks to the previous station and schedule mobile auxiliary robots to handle the wire harness delivery work. Simultaneously, the current station's production line cycle time is reduced by 3% to adapt to a reasonable load level for the remaining tasks. These reconfiguration strategies are converted into station configuration instructions and line speed adjustment instructions recognizable by the manufacturing execution system via the OPC UA protocol and then executed. Specifically, this dynamic reconfiguration capability at the production line level gives the production system an adaptive adjustment function similar to that of a biological organism. It no longer rigidly executes preset cycles, but seeks the global optimal solution for efficiency and load in real time while ensuring the health and safety of personnel, truly realizing people-oriented intelligent manufacturing.

[0060] In step S706, after the aforementioned scheduling instructions at each level are executed, the manufacturing execution system continuously collects actual feedback signals such as changes in actual production capacity, quality pass rate, and subsequent improvement in the operator's physiological indicators, and transmits these signals back to the ergonomics management platform. The platform uses these actual feedback signals to incrementally update the operational ergonomics evaluation model, correcting the model parameters to more accurately reflect the operator's true tolerance and recovery characteristics under the current working conditions. Specifically, with seasonal changes, personnel turnover, or process fine-tuning, the operator's physiological response pattern may drift. The incremental update mechanism ensures that the evaluation model does not age and become ineffective over time, but rather learns from the consequences of each scheduling execution, continuously improving prediction accuracy and the effectiveness of the scheduling strategy. Thus, the complete closed loop from data acquisition, risk assessment, real-time correction, hierarchical scheduling to feedback optimization has been verified on the automotive assembly line, fully demonstrating the practicality and advancement of the present invention's technical solution in complex industrial environments.

[0061] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A digital management method for operational efficiency based on embodied intelligent data acquisition, characterized in that, include: The operator's muscle activation status data, joint motion status data, and anthropometric parameters are collected by wearable sensing devices, and the operator's individualized electromyography-mechanical mapping relationship is obtained by calibrating the action. Using the anthropometric parameters and the individualized electromyography-mechanical mapping relationship, the musculoskeletal model is adaptively calibrated online to obtain the calibrated musculoskeletal model. The kinematic parameters of each joint of the operator are calculated based on the joint motion state data, and the confidence weight of each signal source is dynamically calculated according to the signal quality index. Based on the calibrated musculoskeletal model, the kinematic parameters, and the confidence weights, a fusion calculation is performed using an inverse dynamics model. During the fusion calculation, the contribution ratio of each signal source is allocated according to the confidence weights to obtain the net joint torque of each joint of the operator and generate joint torque temporal features. The joint torque time sequence characteristics are input into a pre-built operation efficiency evaluation model, and the operator's operation load score and fatigue trend prediction value are output. In response to the operation load score or fatigue trend prediction value meeting preset scheduling conditions, corresponding task scheduling optimization instructions are generated according to the scheduling level, which includes individual-level scheduling, workstation-level scheduling, and production line-level scheduling. Specifically: individual-level scheduling includes adjusting the task intensity of the corresponding operator or inserting a short break; workstation-level scheduling includes triggering workstation ergonomic redesign suggestions when multiple operators experience high loads consecutively; and production line-level scheduling includes triggering an overall reduction in the production line cycle time when multiple workstations experience high loads simultaneously. The task scheduling optimization instruction is sent to the manufacturing execution system for execution, and the actual feedback signal after execution is obtained; The operational efficiency evaluation model is incrementally updated based on the actual feedback signal. Also includes: Augmented reality smart glasses and multi-finger haptic feedback gloves are used to simultaneously collect the operator's eye movement gaze point distribution data and fine hand operation data; The eye-tracking fixation point distribution data includes fixation point coordinates, fixation duration, and saccade path; the fine hand operation data includes the angles of each finger joint, fingertip tactile pressure values, and pinching force. Using the muscle activation state data as a baseline physiological signal, the operator's global physical load characteristics are extracted; The cognitive load characteristics of the operator are extracted from the eye movement fixation point distribution data, including fixation concentration, scan rate, and pupil change rate. The fine operational load characteristics of the operator are extracted from the fine hand operation data, including the range of motion of the finger joints, the rate of change of pinching force, and the stability of the operation rhythm. The global physical load characteristics, the cognitive load characteristics, and the fine operation load characteristics are weighted and fused to generate a multi-dimensional comprehensive work efficiency score; The operational load score output by the operational efficiency evaluation model is corrected and calibrated using the multi-dimensional efficiency comprehensive score. The correction and calibration is a real-time output compensation of the operational load score during the inference phase, without changing the model parameters of the operational efficiency evaluation model.

2. The method for digital management of operational efficiency based on embodied intelligent data acquisition according to claim 1, characterized in that, The generated joint torque timing features include: The net joint torque is smoothed and outliers are removed to obtain the temporal characteristics of the joint torque.

3. The method for digital management of operational efficiency based on embodied intelligent data acquisition according to claim 1, characterized in that, The calibrated musculoskeletal model includes muscle origin and insertion points, muscle physiological cross-sectional area, and muscle force-length relationship parameters. The process of fusing and calculating based on the calibrated musculoskeletal model, the kinematic parameters, and the confidence weights through an inverse dynamics model includes: The lever arm vector of each muscle group is determined based on the origin and insertion points of the muscles. The lever arm vector represents the geometric distance and direction of the muscle action line relative to the joint rotation center. Based on the muscle physiological cross-sectional area and the muscle force-length relationship parameters, combined with the surface electromyography signal amplitude in the muscle activation state data, the instantaneous muscle force of each muscle group is calculated through the muscle activation degree-force mapping function. The muscle activation degree-force mapping function characterizes the nonlinear relationship between muscle activation level and active force output. Based on the lever arm vector and the instantaneous muscle force, the muscle torque components generated by each muscle group at the corresponding joint are calculated, and the muscle torque components of all related muscle groups are vector-superimposed to obtain the net joint torque. The vector superposition process takes into account the co-contraction effect of antagonistic muscle groups and agonistic muscle groups.

4. The method for digital management of operational efficiency based on embodied intelligent data acquisition according to claim 1, characterized in that, The pre-built operational efficiency evaluation model includes a multi-head self-attention mechanism and a temporal position encoding encoder network; The multi-head self-attention mechanism is used to capture the long-range dependencies in the timing features of the joint torques, and the timing position encoding is used to preserve the temporal order information of the signal. The step of inputting the joint torque time-series characteristics into a pre-constructed operational efficiency evaluation model and outputting the operator's operational load score and fatigue trend prediction value includes: The joint torque temporal features are embedded into the temporal features by means of the temporal position encoding to obtain position encoded features; The location encoding features are input into an encoder network containing the multi-head self-attention mechanism for feature interaction, and the operation load score and the fatigue trend prediction value are output.

5. The method for digital management of operational efficiency based on embodied intelligent data acquisition according to claim 1, characterized in that, The step of generating corresponding task scheduling optimization instructions according to the scheduling level in response to the operation load score or the fatigue trend prediction value meeting the preset scheduling conditions includes: When the operational load score falls into the high load range or the fatigue trend prediction value exceeds the warning threshold, a hierarchical scheduling event is triggered. In response to the hierarchical scheduling event, the following actions are performed according to the scheduling level: For the individual-level scheduling, generate task allocation adjustment instructions and micro-rest insertion instructions for the operator; for the workstation-level scheduling, output a workstation ergonomics evaluation report and redesign suggestion parameters; for the production line-level scheduling, with the objective function of minimizing overall fatigue differences and maximizing production line efficiency, use a multi-objective optimization algorithm to calculate the task allocation weight adjustment amount for each operator, the correction value of the current workstation production line takt time, and the optimization scheme of the shift rotation sequence. The task allocation adjustment instruction, the micro-rest insertion instruction, the redesign suggestion parameters, the task allocation weight adjustment amount, the correction value, and the optimization scheme are encapsulated into the task scheduling optimization instruction, and written into the manufacturing execution system for execution through the application programming interface.

6. The method for digital management of operational efficiency based on embodied intelligent data acquisition according to claim 1, characterized in that, Also includes: The eye-tracking fixation point distribution data and the fine hand operation data are synchronized and aligned on the time axis to form a multimodal temporal data pair for eye-hand coordination; The multimodal temporal data is input into a lightweight Transformer-based edge embody intelligent model, which includes an eye-tracking attention encoding branch and a hand posture encoding branch. The two branches interact with each other through a cross-modal attention mechanism and output the operation intention category and ergonomic risk index. When the efficiency risk index exceeds the preset risk threshold, the current operation is determined to be in a high-risk state; In response to the determination of the high-risk state, visual guidance prompts are superimposed on the operator's field of vision through the augmented reality smart glasses. The visual guidance prompts include highlighting the target operation area and virtual guide lines to correct the movement trajectory. At the same time, vibration warning signals are emitted from the corresponding finger parts of the multi-finger haptic feedback glove. The vibration frequency of the vibration warning signal is positively correlated with the work efficiency risk index, and the vibration duration covers the duration of the high-risk state, so as to realize real-time correction and guidance of the operation process.

7. The method for digital management of operational efficiency based on embodied intelligent data acquisition according to claim 1, characterized in that, Also includes: Multiple heterogeneous embodied acquisition intelligent agents are deployed on the production line for collaborative acquisition. The multiple heterogeneous embodied acquisition intelligent agents include a depth camera array fixedly installed around the workstation and an autonomous mobile embodied intelligent agent carrying a robotic arm. The autonomous mobile embodied intelligent agent navigates autonomously in the assembly line environment and actively seeks target acquisition perspectives, and works in conjunction with the depth camera array to acquire the operator's complete skeletal motion data and hand operation trajectory; The process involves spatiotemporal registration and fusion of multi-view depth images acquired by the depth camera array and dynamic view depth images acquired by the autonomous mobile avatar to reconstruct the operator's three-dimensional skeletal motion sequence. This includes: extracting the coordinates of skeletal key points from the multi-view depth images and the dynamic view depth images; synchronizing the skeletal key point coordinates based on timestamps and aligning the spatial coordinate system using an iterative nearest-point algorithm; and inputting the aligned skeletal key point coordinates into a graph neural network for topological relationship inference to reconstruct the three-dimensional skeletal motion sequence containing complete joint connections.

8. The method for digital management of operational efficiency based on embodied intelligent data acquisition according to claim 7, characterized in that, Also includes: The spatiotemporal graph convolutional network is used to perform real-time motion recognition and cycle segmentation on the three-dimensional skeleton motion sequence, and to extract the effective operation cycle duration, invalid delay duration and operation rhythm features of each workstation. The effective operation cycle duration, the invalid delay duration, and the operation rhythm characteristics of each workstation are mapped onto the production line layout diagram to generate a real-time production line efficiency heat map. Based on the real-time production line efficiency heatmap, workstations with efficiency lower than a first preset threshold or delay exceeding a second preset threshold are identified as bottleneck workstations, and bottleneck workstation identifiers and their corresponding bottleneck characteristic parameters are output. Minimizing the bottleneck characteristic parameters is the first optimization objective, and maximizing the overall throughput of the production line is the second optimization objective. The optimal reconstruction strategy is searched in the solution space using a swarm intelligence optimization algorithm. The reconstruction strategy includes a task splitting and redistribution scheme between workstations, a personnel sharing and scheduling scheme between workstations, and a mobile auxiliary robot intervention scheme. The optimal reconstruction strategy obtained by the search is verified for feasibility. The feasibility verification includes verifying whether the operational load of each workstation after reconstruction exceeds the operator's capacity limit and verifying whether the logistics cycle time after reconstruction is matched. The optimal reconfiguration strategy that has passed the feasibility verification is converted into workstation configuration instructions and line speed adjustment instructions that the manufacturing execution system can recognize via the OPC UA protocol, and written into the manufacturing execution system to perform dynamic adjustment of production line workstation configuration and online adjustment of production line cycle time.

Citation Information

Patent Citations

  • Articular sports fatigue model based on sports biomechanics, and application thereof

    CN105404792A

  • Nurse dynamic man-hour optimization system based on real-time load perception

    CN121393798A

  • Weight lifting capacity prediction method and system based on muscle force-joint load dynamic parameters

    CN122196347A

  • Human intelligence-based work efficiency risk monitoring method, edge computing device, system and medium

    CN122222386A