Force control joint torque self-correction method based on digital twinborn mapping
By using a three-layer digital twin mapping model and a temporal causal attention network, the problem of correction failure of force-controlled joints under complex working conditions was solved, achieving high-precision, stable, and adaptive torque measurement, and improving the system's self-learning ability and long-term operational reliability.
Patent Information
- Application Number
- CN202511592230.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-01-27
AI Technical Summary
Existing force-controlled joint calibration methods are prone to calibration failure during long-term operation. In particular, under the influence of factors such as thermal drift, assembly deviation and material aging, the torque measurement accuracy decreases and cannot achieve adaptive calibration across time scales and multiple working conditions. Furthermore, existing digital twin solutions lack deep learning mechanisms and cannot effectively capture the complex coupling relationship of multi-source signals.
A three-layer digital twin mapping model is adopted, including a physical twin layer, a simulation twin layer, and a cognitive twin layer. Through a temporal causal attention network and a tensor computation mechanism, a virtual-real mapping relationship is constructed to realize dynamic error identification, adaptive compensation, and real-time closed-loop correction. This is combined with the preprocessing of multi-source data and a distributed bootstrap consistency correction process.
It continuously corrects torque measurement errors under complex working conditions, improves calibration accuracy and stability, has self-learning capabilities, and achieves long-term reliability and high-precision torque measurement of the system, breaking through the limitations of traditional static calibration and single-layer simulation.
Smart Images

Figure CN121411142A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital twin and intelligent control technology, and in particular to a force control joint torque self-correction method based on digital twin mapping. Background Technology
[0002] In existing intelligent robots and precision actuators, force-controlled joints are key components for achieving high-precision operation and safe collaboration. Their core lies in the accurate measurement and real-time compensation of output torque. Currently, mainstream force-controlled joint calibration methods mostly rely on static calibration or single-layer simulation correction based on finite element analysis, obtaining the sensor's compensation coefficients through pre-set load experiments. However, traditional methods are prone to calibration failure after long-term system operation, especially under the influence of thermal drift, assembly deviations, and material aging. Calibration parameters are difficult to maintain stability, leading to a gradual decrease in torque measurement accuracy and impacting the response sensitivity and control stability of the force control system.
[0003] With the development of digital twin technology, some studies have begun to explore the introduction of virtual-real mapping models for error prediction and compensation in robot force control systems. However, existing digital twin solutions typically employ only a two-layer structure of physical and simulation layers, lacking deep learning mechanisms for the dynamic evolution of the virtual and real systems, making it difficult to achieve adaptive correction across time scales and multiple operating conditions. Traditional mapping models are mostly based on linear or static correlation functions, which cannot effectively capture the complex coupling relationships between multiple source signals such as torque, angular velocity, current, and temperature, resulting in insufficient compensation accuracy and lagging system updates.
[0004] The existing simulation layer correction process is usually based on unidirectional compensation, which fails to establish a closed-loop feedback mechanism between real working condition data and virtual calculation, resulting in long-term deviation between simulation output and physical feedback, and failing to guarantee the continuous consistency and evolution accuracy of torque estimation.
[0005] Therefore, how to provide a force-controlled joint torque self-calibration method based on digital twin mapping is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose a self-calibration method for force-controlled joint torque based on digital twin mapping. This invention fully utilizes digital twin technology, temporal causal attention networks, and tensor computation mechanisms to construct a three-layer digital twin mapping model consisting of a physical twin layer, a simulation twin layer, and a cognitive twin layer. By establishing a virtual-real mapping relationship, performing spatiotemporal coupling tensor compensation calculations, and implementing a distributed bootstrapping consistency correction process, dynamic error identification, adaptive compensation, and real-time closed-loop correction of force-controlled joints during operation are achieved. This invention can continuously correct torque measurement errors caused by factors such as thermal drift, assembly deviations, and load coupling under complex working conditions, overcoming the limitations of traditional static calibration and single-layer simulation correction. It possesses advantages such as high correction accuracy, strong stability, superior self-learning ability, and high long-term operational reliability.
[0007] A self-calibration method for force-controlled joint torque based on digital twin mapping according to an embodiment of the present invention includes:
[0008] Collect multi-source data of the force-controlled joint during operation, perform preprocessing on the multi-source data, and generate a standardized input dataset;
[0009] A three-layer digital twin mapping model is established based on a standardized input dataset. The physical twin layer reflects the real operating state of the force-controlled joint, the simulation twin layer constructs a virtual simulation structure based on finite element and dynamics calculations, and the cognitive twin layer uses a temporal causal attention network to learn the virtual-real mapping relationship and generate an adaptive compensation parameter vector.
[0010] A bias load device is set in the physical twin layer. The output angle signal and the original signal of the torque sensor are collected by a high-precision encoder. The physical state vector is constructed by combining the drive current signal and the temperature signal. The correspondence between the physical state vector and the simulation input variables is established by a mapping function.
[0011] In the simulation twin layer, spatiotemporal coupling tensor compensation calculation is performed based on the mapping function. The angle, angular velocity, torque and current data obtained by mapping the physical state vector are constructed into a state tensor. The feature matrices of each mode are extracted and tensor calculation is performed to generate dynamic compensation values.
[0012] In the cognitive twin layer, a distributed symmetric bootstrap consistency correction process is performed. Sample data of the main bootstrap chain and the mirror bootstrap chain are periodically generated, the virtual-real consistency index is calculated, the adaptive compensation parameter vector is dynamically updated according to the virtual-real consistency index, and the updated adaptive compensation parameter vector is fused with the dynamic compensation value to form a comprehensive correction parameter set.
[0013] The integrated correction parameter set is input into the simulation twin layer, the parameters of the tensor calculation structure of the simulation twin layer are adjusted, and the correction results are fed back to the physical twin layer to output the corrected high-precision torque value, thus completing the dynamic error identification, adaptive compensation and real-time self-correction of the force-controlled joint during operation.
[0014] Optionally, the multi-source data includes torque sensor output signals, angle encoder signals, current signals, temperature signals, and additional status signals related to external loads, environmental stresses, or assembly deviations.
[0015] Optionally, the preprocessing of multi-source data includes noise filtering, normalization and time synchronization of multi-source data, smoothing and interpolation of data with outliers, and timing alignment of different signal channels according to the sampling period.
[0016] Optionally, generating the adaptive compensation parameter vector includes:
[0017] The torque sensor raw signal, angle encoder signal, drive current signal and temperature signal are extracted sequentially from the standardized input dataset according to a fixed time window, and then arranged into a physical state vector sequence in chronological order.
[0018] A virtual simulation structure for a simulation twin layer is established based on the physical state vector sequence. The virtual simulation structure consists of a finite element module for joint components and a joint kinematics and dynamics calculation unit. An updatable set of structural parameters and boundary conditions are set to output simulation output quantities corresponding to a fixed time window.
[0019] In the three-layer digital twin mapping model, a mapping function is defined, which includes a forward mapping for converting the physical state vector sequence into a set of virtual simulation structure input variables one by one, and an inverse sensitivity mapping for limiting the allowable range of changes in virtual simulation structure parameters when updating the parameters of the cognitive twin layer. Parameter safety boundaries and consistency constraints are set under both mappings.
[0020] A temporal causal attention network is constructed in the cognitive twin layer. The temporal causal attention network consists of a multi-scale causal attention sub-layer, a prototype memory sub-layer, and a consistency gating sub-layer. It takes the physical state vector sequence and the virtual-real difference sequence formed by the difference between the simulation output and the original signal of the torque sensor as input to complete causal association extraction, working condition prototype retrieval and gating weighting, and generate an adaptive compensation parameter vector.
[0021] The adaptive compensation parameter vector is stored and distributed as the output of the cognitive twin layer in the current time window, and together with the mapping function and the virtual simulation structure, it constitutes the initialization configuration of the three-layer digital twin mapping model.
[0022] Optionally, the step of constructing a physical state vector by combining the driving current signal and the temperature signal, and establishing the correspondence between the physical state vector and the simulation input variables through a mapping function, includes:
[0023] A bias load device is set in the physical twin layer, and a high-precision encoder is started to collect the output angle signal. The original signal of the torque sensor, the drive current signal and the temperature signal are collected simultaneously. They are aligned according to a unified timestamp and divided into time windows of fixed length.
[0024] Within each time window, angular velocity data is generated based on the angle signal. A physical state vector is constructed by combining the original signal from the torque sensor, the drive current signal, and the temperature signal. The drive current signal and the temperature signal are then jointly encoded to form an extended field.
[0025] Define a mapping function, which consists of a channel alignment sub-rule, a joint correction sub-rule, and an assembly deviation coding sub-rule. The channel alignment sub-rule is used to compensate for sampling delay and complete phase correction. The joint correction sub-rule determines the bias correction amount based on the joint coding of the drive current signal and the temperature signal. The assembly deviation coding sub-rule writes the installation position and geometric eccentricity information of the bias load device as constraint fields into the mapping process.
[0026] The mapping function is called to convert the angle, angular velocity, torque, driving current and temperature in the physical state vector into a set of simulation input variables. During the conversion process, the bias correction amount is injected according to the joint correction sub-rule, and the assembly constraint field is injected according to the assembly deviation coding sub-rule.
[0027] The simulation input variable set is output to the simulation twin layer at a fixed update cycle. The parameters and constraint fields of the mapping function are recorded as the configuration of the current time window, and the configuration is kept consistent with the timestamp of the physical state vector.
[0028] Optionally, generating the dynamic compensation value includes:
[0029] Receive the set of simulation input variables output by the mapping function, perform phase correction and delay compensation on the angle, angular velocity, torque and current data obtained by mapping from the physical state vector according to the time window, stack and align them according to a fixed channel order, construct the state tensor and record the time window index and channel index;
[0030] A parallel computation structure of main channel and symmetric channel is established on the state tensor. Cross-channel consistency constraints and boundary continuity constraints are set between the two channels. Synchronous advancement and symmetric reconstruction are performed on the data in each time window.
[0031] In the calculation process of the main channel and the symmetrical channel, the bias correction amount and assembly constraint field provided by the mapping function are injected, and the torque energy conservation constraint and joint kinematic consistency constraint are superimposed. Channel consistency alignment and constraint projection are performed on the angle channel, angular velocity channel, torque channel and current channel.
[0032] Based on the current load classification information and temperature segmentation information contained in the physical state vector, the state tensor is subjected to operating domain gated deformation and adaptive complexity control. The calculation path is selected online according to low-rank and sparse routing weights, and the channel weights and gate thresholds are initialized with the current configuration of the adaptive compensation parameter vector.
[0033] For the results after constraint projection and gate control within each time window, anomaly segment gating and symmetric fusion are performed to generate a dynamic compensation value sequence that corresponds one-to-one with the time window. The dynamic compensation value sequence is then associated with the time window index and the timestamp of the physical state vector.
[0034] Optionally, the step of fusing the updated adaptive compensation parameter vector with the dynamic compensation value to form a comprehensive correction parameter set includes:
[0035] In the cognitive twin layer, a master bootstrap chain and a mirror bootstrap chain are established. Master bootstrap sampling data and mirror bootstrap sampling data are generated according to the time window consistent with the physical state vector. The master bootstrap sampling data consists of the physical state vector and dynamic compensation value, while the mirror bootstrap sampling data consists of the virtual response generated by the simulation twin layer according to the safe amplitude perturbation template within the same time window. The time window index and channel correspondence are recorded for the two bootstrap chains respectively.
[0036] The master bootstrap sampling data and the mirror bootstrap sampling data within each time window are compared to calculate the virtual-real consistency index. The virtual-real consistency index consists of five parts: amplitude proximity, trend proximity, phase alignment, cross-channel correlation consistency, and continuity of adjacent window boundaries. Within the cognitive twin layer, the virtual-real consistency index is normalized to a fixed value range and assigned a reliability level.
[0037] Based on the reliability level and grading threshold of the virtual-real consistency index, the temporal causal attention network of the cognitive twin layer is triggered to update the parameters, and an updated adaptive compensation parameter vector is generated within the parameter safety boundary and consistency constraints.
[0038] Within a window consistent with the time window index, the updated adaptive compensation parameter vector and the dynamic compensation value are fused together. The fusion weight is determined according to the reliability level of the virtual-real consistency index to form a comprehensive correction parameter set. Field integrity verification, timestamp alignment and version registration are performed on the comprehensive correction parameter set.
[0039] The integrated correction parameter set is used as the output of the cognitive twin layer. When the reliability level of the virtual-real consistency index is lower than the set lower limit, the adaptive compensation parameter vector of the previous version remains unchanged, and the timestamp of the integrated correction parameter set is kept consistent with that of the physical state vector.
[0040] Optionally, the step of inputting the comprehensive correction parameter set into the simulation twin layer, performing parameter adjustments on the tensor computation structure of the simulation twin layer, and feeding back the correction results to the physical twin layer includes:
[0041] Receive the integrated correction parameter set, register and cache it according to the time window index, align the integrated correction parameter set with the state tensor timestamp of the corresponding time window of the simulation twin layer, and establish a shadow parameter area;
[0042] In the tensor computation structure of the simulated twin layer, hierarchical parameter injection is performed. The channel weights, gating thresholds, constraint fields, symmetric fusion coefficients and cross-scale consistency boundaries in the comprehensive correction parameter set are sequentially loaded into the shadow parameter area to complete the legality and consistency verification.
[0043] Initiate adaptive routing adjustment, select the calculation path based on current load classification information and temperature segmentation information within the tensor calculation structure, set steady-state protection threshold to limit parameter jump amplitude, perform trial run recalculation on the state tensor corresponding to the time window, and generate candidate output of correction results;
[0044] A dual-phase confirmation mechanism is adopted to process the parameters in the shadow parameter area. First, the candidate output of the correction result is checked against the historical output for boundary continuity and cross-channel correlation. Then, it is written into the main parameter area for immediate effect. The correction result is fed back to the physical twin layer, aligned with the timestamp of the physical state vector, and merged with the dynamic compensation value and the original signal of the torque sensor to output the corrected high-precision torque value.
[0045] Establish a closed-loop monitoring and rollback process, record the calibrated high-precision torque value, parameter version identifier and trigger conditions. When the consistency index is lower than the preset lower limit or exceeds the safety boundary, roll back to the comprehensive calibration parameter set of the previous version. When the stability conditions are met, solidify the main parameter area.
[0046] The beneficial effects of this invention are:
[0047] This invention introduces a three-layer digital twin mapping structure into a traditional force-controlled joint system, achieving deep fusion of virtual and real information and a closed-loop self-calibration process. The physical twin layer is used to acquire multi-source operational data in real time; the simulation twin layer performs multi-dimensional coupled calculations on signals such as torque, angle, current, and temperature; and the cognitive twin layer uses a temporal causal attention network to achieve adaptive learning of virtual-real differences and dynamic parameter optimization. The three layers work together to maintain consistency and sustainability of torque measurement accuracy under different operating conditions and long-term operation, overcoming the problems of strong calibration dependence and insufficient real-time performance in existing technologies.
[0048] This invention effectively solves the nonlinear mismatch problem of multi-source signals in time and space dimensions by introducing a spatiotemporal coupling tensor compensation mechanism into the simulation twin layer, making the coupling calculation of torque, angular velocity, and driving current more accurate. The dynamic compensation value generation mechanism enables the system to automatically adjust the correction amount according to real-time state changes, ensuring that the force control system maintains high precision and robustness under complex working conditions such as temperature drift, structural deformation, and sudden load changes. The system can achieve autonomous correction without relying on external high-precision calibration equipment, greatly improving the self-maintenance and self-adaptation capabilities of the force control system.
[0049] This invention constructs a distributed symmetric bootstrap consistency correction mechanism in the cognitive twin layer. By establishing a main bootstrap chain and a mirror bootstrap chain, it achieves dynamic calculation and bidirectional verification of the virtual-real consistency index, ensuring a stable correspondence between the virtual simulation output and the physical measured data. This allows the system to continuously optimize the mapping function and parameter configuration during operation, thereby achieving closed-loop self-evolutionary updates. In summary, this invention not only significantly improves the torque measurement accuracy and stability of force-controlled joints in complex environments but also achieves long-term self-learning and adaptive evolution of the system, possessing the beneficial effects of high precision, strong robustness, and sustainable operation. Attached Figure Description
[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0051] Figure 1 This is a flowchart of a force control joint torque self-correction method based on digital twin mapping proposed in this invention;
[0052] Figure 2 This is a schematic diagram of the structure of a three-layer digital twin mapping model for a force control joint torque self-correction method based on digital twin mapping proposed in this invention. Detailed Implementation
[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0054] refer to Figure 1 and Figure 2 A self-calibration method for force-controlled joint torque based on digital twin mapping includes:
[0055] Collect multi-source data of the force-controlled joint during operation, perform preprocessing on the multi-source data, and generate a standardized input dataset;
[0056] A three-layer digital twin mapping model is established based on a standardized input dataset. The physical twin layer reflects the real operating state of the force-controlled joint, the simulation twin layer constructs a virtual simulation structure based on finite element and dynamics calculations, and the cognitive twin layer uses a temporal causal attention network to learn the virtual-real mapping relationship and generate an adaptive compensation parameter vector.
[0057] A bias load device is set in the physical twin layer. The output angle signal and the original signal of the torque sensor are collected by a high-precision encoder. The physical state vector is constructed by combining the drive current signal and the temperature signal. The correspondence between the physical state vector and the simulation input variables is established by a mapping function.
[0058] In the simulation twin layer, spatiotemporal coupling tensor compensation calculation is performed based on the mapping function. The angle, angular velocity, torque and current data obtained by mapping the physical state vector are constructed into a state tensor. The feature matrices of each mode are extracted and tensor calculation is performed to generate dynamic compensation values.
[0059] In the cognitive twin layer, a distributed symmetric bootstrap consistency correction process is performed. Sample data of the main bootstrap chain and the mirror bootstrap chain are periodically generated, the virtual-real consistency index is calculated, the adaptive compensation parameter vector is dynamically updated according to the virtual-real consistency index, and the updated adaptive compensation parameter vector is fused with the dynamic compensation value to form a comprehensive correction parameter set.
[0060] The integrated correction parameter set is input into the simulation twin layer, the parameters of the tensor calculation structure of the simulation twin layer are adjusted, and the correction results are fed back to the physical twin layer to output the corrected high-precision torque value, thus completing the dynamic error identification, adaptive compensation and real-time self-correction of the force-controlled joint during operation.
[0061] In this embodiment, the multi-source data includes torque sensor output signal, angle encoder signal, current signal, temperature signal, and additional state signals related to external load, environmental stress, or assembly deviation.
[0062] In this embodiment, the preprocessing of multi-source data includes noise filtering, normalization and time synchronization of multi-source data, smoothing and interpolation of data with outliers, and timing alignment of different signal channels according to the sampling period.
[0063] In this embodiment, generating the adaptive compensation parameter vector includes:
[0064] The torque sensor raw signal, angle encoder signal, drive current signal and temperature signal are extracted sequentially from the standardized input dataset according to a fixed time window, and then arranged into a physical state vector sequence in chronological order.
[0065] A virtual simulation structure for a simulation twin layer is established based on the physical state vector sequence. The virtual simulation structure consists of a finite element module for joint components and a joint kinematics and dynamics calculation unit. An updatable set of structural parameters and boundary conditions are set to output simulation output quantities corresponding to a fixed time window.
[0066] In the three-layer digital twin mapping model, a mapping function is defined, which includes a forward mapping for converting the physical state vector sequence into a set of virtual simulation structure input variables one by one, and an inverse sensitivity mapping for limiting the allowable range of changes in virtual simulation structure parameters when updating the parameters of the cognitive twin layer. Parameter safety boundaries and consistency constraints are set under both mappings.
[0067] A temporal causal attention network is constructed in the cognitive twin layer. This network consists of a multi-scale causal attention sublayer, a prototype memory sublayer, and a consistency gating sublayer. It takes a sequence of physical state vectors and a sequence of virtual-real differences formed by the differences between the simulation output and the original signals from the torque sensor as input. This process completes causal correlation extraction, prototype retrieval, and gating weighting, generating an adaptive compensation parameter vector, where:
[0068] The multi-scale causal attention sublayer receives the virtual-real difference sequence and physical state vector sequence from the simulation twin layer. The input sequence is processed in a multi-scale segmentation according to the time window. Each time scale corresponds to a different sampling interval and delay weight. The causal dependency measure between adjacent time points is calculated in turn. The results of each time scale are aggregated by time weighting to generate a multi-scale time series feature matrix.
[0069] The prototype memory sublayer receives a multi-scale temporal feature matrix, performs similarity clustering on the multi-scale temporal feature matrix, maintains a prototype feature set containing typical operating states, and compares the current feature vector with the historical prototype set one by one through a combination of Euclidean distance and cosine similarity. The prototype vectors with the highest similarity are linearly combined to form the prototype alignment feature matrix at the current time.
[0070] The consistency gating sublayer receives the prototype aligned feature matrix, normalizes each feature channel, sets an independent gating coefficient for each channel, adjusts the gating opening according to the time change rate of the virtual-real difference sequence, performs weighted filtering on the feature channels, and the filtered feature matrix is aggregated by the channels to form the final compensation weight matrix. The matrix is then multiplied with the input variables of the simulation twin layer to generate an adaptive compensation parameter vector.
[0071] The adaptive compensation parameter vector is stored and distributed as the output of the cognitive twin layer in the current time window, and together with the mapping function and the virtual simulation structure, it constitutes the initialization configuration of the three-layer digital twin mapping model.
[0072] In this embodiment, the step of constructing a physical state vector by combining the driving current signal and the temperature signal, and establishing the correspondence between the physical state vector and the simulation input variables through a mapping function, includes:
[0073] A bias load device is set in the physical twin layer, and a high-precision encoder is started to collect the output angle signal. The original signal of the torque sensor, the drive current signal and the temperature signal are collected simultaneously. They are aligned according to a unified timestamp and divided into time windows of fixed length.
[0074] Within each time window, angular velocity data is generated based on the angle signal. A physical state vector is constructed by combining the original signal from the torque sensor, the drive current signal, and the temperature signal. The drive current signal and the temperature signal are then jointly encoded to form an extended field.
[0075] Define a mapping function, which consists of a channel alignment sub-rule, a joint correction sub-rule, and an assembly deviation coding sub-rule. The channel alignment sub-rule is used to compensate for sampling delay and complete phase correction. The joint correction sub-rule determines the bias correction amount based on the joint coding of the drive current signal and the temperature signal. The assembly deviation coding sub-rule writes the installation position and geometric eccentricity information of the bias load device as constraint fields into the mapping process.
[0076] The mapping function is called to convert the angle, angular velocity, torque, driving current and temperature in the physical state vector into a set of simulation input variables. During the conversion process, the bias correction amount is injected according to the joint correction sub-rule, and the assembly constraint field is injected according to the assembly deviation coding sub-rule.
[0077] The simulation input variable set is output to the simulation twin layer at a fixed update cycle. The parameters and constraint fields of the mapping function are recorded as the configuration of the current time window, and the configuration is kept consistent with the timestamp of the physical state vector.
[0078] In this embodiment, generating dynamic compensation values includes:
[0079] Receive the set of simulation input variables output by the mapping function, perform phase correction and delay compensation on the angle, angular velocity, torque and current data obtained by mapping from the physical state vector according to the time window, stack and align them according to a fixed channel order, construct the state tensor and record the time window index and channel index;
[0080] A parallel computation structure of main channel and symmetric channel is established on the state tensor. Cross-channel consistency constraints and boundary continuity constraints are set between the two channels. Synchronous advancement and symmetric reconstruction are performed on the data within each time window, wherein:
[0081] Cross-channel consistency constraint refers to the mutual constraint processing of the feature mapping results of corresponding tensor elements within the same time window during the calculation process of the main channel and the symmetric channel, so that the two channels maintain a symmetric relationship in feature amplitude, phase response and dynamic gradient direction. In each iteration, the output difference between the main channel and the symmetric channel is normalized and measured. If the difference exceeds the set threshold, the weight distribution and feature propagation rate between the channels are automatically adjusted.
[0082] Boundary continuity constraint refers to the smooth transition processing of boundary data of adjacent time periods or adjacent feature blocks during the switching of time windows or the splicing of state tensors. This prevents sudden torque changes or computational breaks caused by window switching. A transition coefficient is introduced at the junction of time windows, and a weighted average is performed on the tensor at the end of the previous window and the tensor at the beginning of the next window, so that the channel output maintains the continuity of the first derivative in the time series.
[0083] In the calculation process of the main channel and the symmetrical channel, the bias correction amount and assembly constraint field provided by the mapping function are injected, and the torque energy conservation constraint and joint kinematic consistency constraint are superimposed. Channel consistency alignment and constraint projection are performed on the angle channel, angular velocity channel, torque channel and current channel, where:
[0084] Torque energy conservation constraint refers to introducing a conservation balance condition between torque and input energy during the tensor calculation of angle, torque and current channels, and dynamically constraining the energy transfer between channels. Within each time window, based on the coupling relationship between motor drive current and output torque, the power change rate is monitored synchronously, and the instantaneous change of output torque is limited to not exceeding the allowable range of input energy change.
[0085] Joint kinematic consistency constraints refer to maintaining temporal consistency and spatial alignment between the kinematic state calculated in the simulation and the physical measurement results during the calculation of angle and angular velocity channels. This is achieved by verifying the mapping relationship between joint angular displacement, angular velocity and torque, maintaining continuity and differentiability in the state space, and adjusting the phase delay and gain factor of the corresponding channel if the deviation exceeds the set threshold.
[0086] Based on the current load classification information and temperature segmentation information contained in the physical state vector, the state tensor is subjected to operating condition domain gating deformation and adaptive complexity control. The computation path is selected online according to low-rank and sparse routing weights, and the channel weights and gating thresholds are initialized with the current configuration of the adaptive compensation parameter vector. Specifically, the operating condition domain gating deformation and adaptive complexity control of the state tensor involves:
[0087] The operating conditions are divided into different domains based on current load classification information and temperature segmentation information. Independent gating weights are set for each operating condition domain. The channel weight distribution of the state tensor is adjusted under different operating conditions so that the tensor features maintain the corresponding deformation relationship when the load changes.
[0088] The computational complexity of the state tensor is automatically adjusted according to the degree of signal fluctuation. When the load is stable, a low-complexity path is executed, and when the load fluctuation increases, a high-complexity path is activated, thereby achieving adaptive allocation of computing resources.
[0089] The channel weights and gating thresholds are initialized based on the current configuration of the adaptive compensation parameter vector, so that the state tensor can smoothly transition between different working conditions and maintain computational stability and response consistency.
[0090] For the results after constraint projection and gate control within each time window, anomaly segment gating and symmetric fusion are performed to generate a dynamic compensation value sequence that corresponds one-to-one with the time window. The dynamic compensation value sequence is then associated with the time window index and the timestamp of the physical state vector.
[0091] In this embodiment, fusing the updated adaptive compensation parameter vector with the dynamic compensation value to form a comprehensive correction parameter set includes:
[0092] In the cognitive twin layer, a master bootstrap chain and a mirror bootstrap chain are established. Master bootstrap sampling data and mirror bootstrap sampling data are generated according to the time window consistent with the physical state vector. The master bootstrap sampling data consists of the physical state vector and dynamic compensation value, while the mirror bootstrap sampling data consists of the virtual response generated by the simulation twin layer according to the safe amplitude perturbation template within the same time window. The time window index and channel correspondence are recorded for the two bootstrap chains respectively.
[0093] The master bootstrap sampling data and the mirror bootstrap sampling data within each time window are compared to calculate the virtual-real consistency index. The virtual-real consistency index consists of five parts: amplitude proximity, trend proximity, phase alignment, cross-channel correlation consistency, and continuity of adjacent window boundaries. Within the cognitive twin layer, the virtual-real consistency index is normalized to a fixed value range and assigned a reliability level.
[0094] Based on the reliability level and grading threshold of the virtual-real consistency index, the temporal causal attention network of the cognitive twin layer is triggered to update the parameters, and an updated adaptive compensation parameter vector is generated within the parameter safety boundary and consistency constraints.
[0095] Within a window consistent with the time window index, the updated adaptive compensation parameter vector and the dynamic compensation value are fused together. The fusion weight is determined according to the reliability level of the virtual-real consistency index to form a comprehensive correction parameter set. Field integrity verification, timestamp alignment and version registration are performed on the comprehensive correction parameter set.
[0096] The integrated correction parameter set is used as the output of the cognitive twin layer. When the reliability level of the virtual-real consistency index is lower than the set lower limit, the adaptive compensation parameter vector of the previous version remains unchanged, and the timestamp of the integrated correction parameter set is kept consistent with that of the physical state vector.
[0097] In this embodiment, the step of inputting the comprehensive correction parameter set into the simulation twin layer, adjusting the parameters of the tensor calculation structure of the simulation twin layer, and feeding back the correction result to the physical twin layer includes:
[0098] Receive the integrated correction parameter set, register and cache it according to the time window index, align the integrated correction parameter set with the state tensor timestamp of the corresponding time window of the simulation twin layer, and establish a shadow parameter area;
[0099] In the tensor computation structure of the simulated twin layer, hierarchical parameter injection is performed. The channel weights, gating thresholds, constraint fields, symmetric fusion coefficients and cross-scale consistency boundaries in the comprehensive correction parameter set are sequentially loaded into the shadow parameter area to complete the legality and consistency verification.
[0100] Initiate adaptive routing adjustment, select the calculation path based on current load classification information and temperature segmentation information within the tensor calculation structure, set steady-state protection threshold to limit parameter jump amplitude, perform trial run recalculation on the state tensor corresponding to the time window, and generate candidate output of correction results;
[0101] A dual-phase confirmation mechanism is adopted to process the parameters in the shadow parameter area. First, the candidate output of the correction result is checked against the historical output for boundary continuity and cross-channel correlation. Then, it is written into the main parameter area for immediate effect. The correction result is fed back to the physical twin layer, aligned with the timestamp of the physical state vector, and merged with the dynamic compensation value and the original signal of the torque sensor to output the corrected high-precision torque value.
[0102] Establish a closed-loop monitoring and rollback process, record the calibrated high-precision torque value, parameter version identifier and trigger conditions. When the consistency index is lower than the preset lower limit or exceeds the safety boundary, roll back to the comprehensive calibration parameter set of the previous version. When the stability conditions are met, solidify the main parameter area.
[0103] Example 1:
[0104] To verify the feasibility of this invention in practice, it was applied to a composite assembly robot experimental platform. This platform is mainly used for the dynamic calibration and assembly error compensation testing of force-controlled joint modules with torque sensors. The joint under test is mounted on a fixed frame, with an offset load structure connected to its end. The offset load has a mass of 3.8 kg, a lever arm length of 0.22 m from the center of rotation, and corresponds to a rated load torque of approximately 8.2 N·m. The experimental environment temperature fluctuates within the range of 22℃ to 45℃, and the test cycle is 72 consecutive hours. The system adopts the three-layer digital twin mapping structure proposed in this invention, including a physical twin layer, a simulation twin layer, and a cognitive twin layer, and combines an offset load calibration mechanism and a dynamic error compensation algorithm to achieve the torque self-correction process.
[0105] During the experiment, the physical twin layer acquired angle signals and raw signals from the torque sensor in real time using a high-precision encoder (25 arcsecond resolution), and simultaneously acquired current and temperature data to construct a standardized input dataset. The simulation twin layer performed spatiotemporal coupling tensor compensation calculations based on the input data, fusing multi-scale features from channels such as angle, torque, current, and temperature to generate dynamic compensation values for real-time correction of the sensor's raw torque output. The cognitive twin layer calculated the virtual-real difference sequence through a temporal causal attention network and performed distributed symmetric bootstrapping consistency correction, adaptively learning the deviation between the virtual simulation results and the measured data, and periodically updating the adaptive compensation parameter vector. Finally, the comprehensive correction parameter set was fed back to the simulation twin layer, performing hierarchical parameter adjustments on its tensor calculation structure, and feeding the correction results back to the physical twin layer, outputting a corrected high-precision torque value. The entire process realized a complete correction workflow from error identification and dynamic compensation to closed-loop self-learning.
[0106] To verify the performance of this invention, a comparative experiment was conducted with the traditional static calibration method and the single-layer simulation correction method. Under the same test conditions (bias load torque of 8.2 N·m and joint speed of 2.5° / s), the average torque error of the traditional static calibration method was ±0.16 N·m, the error of the single-layer simulation correction method was ±0.09 N·m, while the average error was reduced to ±0.035 N·m after using the method of this invention. When the system temperature rises to 45°C, the torque drift of the traditional method reaches 2.1%, while the method of this invention controls it to within 0.6%. After 72 hours of continuous testing, the virtual-real consistency index of the system of this invention is stable above 0.95, which is about 14% higher than the 0.83 of the traditional two-layer simulation method. The system torque response delay is shortened from 37 milliseconds to 21 milliseconds, and the long-term operational stability is significantly improved.
[0107] Table 1. Comparison data of force-controlled joint torque self-calibration experiment
[0108]
[0109] As can be seen from the comparison of data in Table 1, the digital twin mapping method of this invention has a significant advantage over traditional methods in terms of force control joint torque self-calibration performance. Regarding torque measurement accuracy, the method of this invention, under continuous operation for 72 hours, has an average measurement error of only ±0.035 N·m, which is 78% and 61% lower than the ±0.16 N·m of the traditional static calibration method and the ±0.09 N·m of the single-layer simulation calibration method, respectively. This indicates that the three-layer digital twin mapping model can effectively compensate for nonlinear errors caused by temperature, assembly deviations, and dynamic load coupling, enabling the system to maintain high torque consistency during long-term operation.
[0110] In terms of stability and dynamic response performance, the method of this invention shows significant improvements. As the ambient temperature rises to 45℃, the torque drift amplitude of the traditional static calibration system reaches 2.1%, while that of the digital twin mapping system is only 0.6%, representing a nearly 3.5-fold improvement in stability. The torque response delay is reduced from 37 milliseconds to 21 milliseconds, indicating that the spatiotemporal coupling tensor compensation mechanism of the simulation twin layer and the adaptive learning strategy of the cognitive twin layer can achieve faster parameter convergence and real-time compensation, improving the dynamic response efficiency of the system in complex environments. The system's virtual-real consistency index reaches 0.95, far exceeding the 0.78 and 0.83 of the other two methods, verifying the effectiveness of the distributed symmetric bootstrap consistency correction mechanism in maintaining virtual-real synchronization accuracy.
[0111] Finally, from an engineering application perspective, the method of this invention demonstrates higher assembly quality and long-term operational reliability in actual press-fitting tasks. After 500 assembly tests, the standard deviation of the assembly torque under this method was only 0.04 N·m, with an assembly success rate of 99.2%, an improvement of nearly 3 percentage points compared to traditional calibration systems. The system's stable operating time was extended to 400 hours, during which no manual intervention or recalibration was required, demonstrating the method's high adaptability and maintenance economy in automated production. This invention, through a three-layer digital twin mapping and self-correcting closed-loop control mechanism, achieves full-process optimization from error identification and dynamic compensation to long-term self-learning, improving the accuracy, stability, and intelligence level of the force control system under complex working conditions.
[0112] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for self-calibrating force-controlled joint torque based on digital twin mapping, characterized in that, include: Collect multi-source data of the force-controlled joint during operation, perform preprocessing on the multi-source data, and generate a standardized input dataset; A three-layer digital twin mapping model is established based on a standardized input dataset. The physical twin layer reflects the real operating state of the force-controlled joint, the simulation twin layer constructs a virtual simulation structure based on finite element and dynamics calculations, and the cognitive twin layer uses a temporal causal attention network to learn the virtual-real mapping relationship and generate an adaptive compensation parameter vector. A bias load device is set in the physical twin layer. The output angle signal and the original signal of the torque sensor are collected by a high-precision encoder. The physical state vector is constructed by combining the drive current signal and the temperature signal. The correspondence between the physical state vector and the simulation input variables is established by a mapping function. In the simulation twin layer, spatiotemporal coupling tensor compensation calculation is performed based on the mapping function. The angle, angular velocity, torque and current data obtained by mapping the physical state vector are constructed into a state tensor. The feature matrices of each mode are extracted and tensor calculation is performed to generate dynamic compensation values. In the cognitive twin layer, a distributed symmetric bootstrap consistency correction process is performed. Sample data of the main bootstrap chain and the mirror bootstrap chain are periodically generated, the virtual-real consistency index is calculated, the adaptive compensation parameter vector is dynamically updated according to the virtual-real consistency index, and the updated adaptive compensation parameter vector is fused with the dynamic compensation value to form a comprehensive correction parameter set. The integrated correction parameter set is input into the simulation twin layer, the parameters of the tensor calculation structure of the simulation twin layer are adjusted, and the correction results are fed back to the physical twin layer to output the corrected high-precision torque value, thus completing the dynamic error identification, adaptive compensation and real-time self-correction of the force-controlled joint during operation.
2. The self-calibration method for force-controlled joint torque based on digital twin mapping according to claim 1, characterized in that, The multi-source data includes torque sensor output signals, angle encoder signals, current signals, temperature signals, and additional status signals related to external loads, environmental stresses, or assembly deviations.
3. The self-calibration method for force-controlled joint torque based on digital twin mapping according to claim 1, characterized in that, The preprocessing of multi-source data includes noise filtering, normalization and time synchronization of the multi-source data, smoothing and interpolation of data with outliers, and timing alignment of different signal channels according to the sampling period.
4. The self-calibration method for force-controlled joint torque based on digital twin mapping according to claim 1, characterized in that, The generation of the adaptive compensation parameter vector includes: The torque sensor raw signal, angle encoder signal, drive current signal and temperature signal are extracted sequentially from the standardized input dataset according to a fixed time window, and then arranged into a physical state vector sequence in chronological order. A virtual simulation structure for a simulation twin layer is established based on the physical state vector sequence. The virtual simulation structure consists of a finite element module for joint components and a joint kinematics and dynamics calculation unit. An updatable set of structural parameters and boundary conditions are set to output simulation output quantities corresponding to a fixed time window. In the three-layer digital twin mapping model, a mapping function is defined, which includes a forward mapping for converting the physical state vector sequence into a set of virtual simulation structure input variables one by one, and an inverse sensitivity mapping for limiting the allowable range of changes in virtual simulation structure parameters when updating the parameters of the cognitive twin layer. Parameter safety boundaries and consistency constraints are set under both mappings. A temporal causal attention network is constructed in the cognitive twin layer. The temporal causal attention network consists of a multi-scale causal attention sub-layer, a prototype memory sub-layer, and a consistency gating sub-layer. It takes the physical state vector sequence and the virtual-real difference sequence formed by the difference between the simulation output and the original signal of the torque sensor as input to complete causal association extraction, working condition prototype retrieval and gating weighting, and generate an adaptive compensation parameter vector. The adaptive compensation parameter vector is stored and distributed as the output of the cognitive twin layer in the current time window, and together with the mapping function and the virtual simulation structure, it constitutes the initialization configuration of the three-layer digital twin mapping model.
5. The self-calibration method for force-controlled joint torque based on digital twin mapping according to claim 1, characterized in that, The process of constructing a physical state vector by combining the driving current signal and the temperature signal, and establishing the correspondence between the physical state vector and the simulation input variables through a mapping function, includes: A bias load device is set in the physical twin layer, and a high-precision encoder is started to collect the output angle signal. The original signal of the torque sensor, the drive current signal and the temperature signal are collected simultaneously. They are aligned according to a unified timestamp and divided into time windows of fixed length. Within each time window, angular velocity data is generated based on the angle signal. A physical state vector is constructed by combining the original signal from the torque sensor, the drive current signal, and the temperature signal. The drive current signal and the temperature signal are then jointly encoded to form an extended field. Define a mapping function, which consists of a channel alignment sub-rule, a joint correction sub-rule, and an assembly deviation coding sub-rule. The channel alignment sub-rule is used to compensate for sampling delay and complete phase correction. The joint correction sub-rule determines the bias correction amount based on the joint coding of the drive current signal and the temperature signal. The assembly deviation coding sub-rule writes the installation position and geometric eccentricity information of the bias load device as constraint fields into the mapping process. The mapping function is called to convert the angle, angular velocity, torque, driving current and temperature in the physical state vector into a set of simulation input variables. During the conversion process, the bias correction amount is injected according to the joint correction sub-rule, and the assembly constraint field is injected according to the assembly deviation coding sub-rule. The simulation input variable set is output to the simulation twin layer at a fixed update cycle. The parameters and constraint fields of the mapping function are recorded as the configuration of the current time window, and the configuration is kept consistent with the timestamp of the physical state vector.
6. The self-calibration method for force-controlled joint torque based on digital twin mapping according to claim 1, characterized in that, The generation of dynamic compensation values includes: Receive the set of simulation input variables output by the mapping function, perform phase correction and delay compensation on the angle, angular velocity, torque and current data obtained by mapping from the physical state vector according to the time window, stack and align them according to a fixed channel order, construct the state tensor and record the time window index and channel index; A parallel computation structure of main channel and symmetric channel is established on the state tensor. Cross-channel consistency constraints and boundary continuity constraints are set between the two channels. Synchronous advancement and symmetric reconstruction are performed on the data in each time window. In the calculation process of the main channel and the symmetrical channel, the bias correction amount and assembly constraint field provided by the mapping function are injected, and the torque energy conservation constraint and joint kinematic consistency constraint are superimposed. Channel consistency alignment and constraint projection are performed on the angle channel, angular velocity channel, torque channel and current channel. Based on the current load classification information and temperature segmentation information contained in the physical state vector, the state tensor is subjected to operating domain gated deformation and adaptive complexity control. The calculation path is selected online according to low-rank and sparse routing weights, and the channel weights and gate thresholds are initialized with the current configuration of the adaptive compensation parameter vector. For the results after constraint projection and gate control within each time window, anomaly segment gating and symmetric fusion are performed to generate a dynamic compensation value sequence that corresponds one-to-one with the time window. The dynamic compensation value sequence is then associated with the time window index and the timestamp of the physical state vector.
7. The self-calibration method for force-controlled joint torque based on digital twin mapping according to claim 1, characterized in that, The process of fusing the updated adaptive compensation parameter vector with the dynamic compensation value to form a comprehensive correction parameter set includes: In the cognitive twin layer, a master bootstrap chain and a mirror bootstrap chain are established. Master bootstrap sampling data and mirror bootstrap sampling data are generated according to the time window consistent with the physical state vector. The master bootstrap sampling data consists of the physical state vector and dynamic compensation value, while the mirror bootstrap sampling data consists of the virtual response generated by the simulation twin layer according to the safe amplitude perturbation template within the same time window. The time window index and channel correspondence are recorded for the two bootstrap chains respectively. The master bootstrap sampling data and the mirror bootstrap sampling data within each time window are compared to calculate the virtual-real consistency index. The virtual-real consistency index consists of five parts: amplitude proximity, trend proximity, phase alignment, cross-channel correlation consistency, and continuity of adjacent window boundaries. Within the cognitive twin layer, the virtual-real consistency index is normalized to a fixed value range and assigned a reliability level. Based on the reliability level and grading threshold of the virtual-real consistency index, the temporal causal attention network of the cognitive twin layer is triggered to update the parameters, and an updated adaptive compensation parameter vector is generated within the parameter safety boundary and consistency constraints. Within a window consistent with the time window index, the updated adaptive compensation parameter vector and the dynamic compensation value are fused together. The fusion weight is determined according to the reliability level of the virtual-real consistency index to form a comprehensive correction parameter set. Field integrity verification, timestamp alignment and version registration are performed on the comprehensive correction parameter set. The integrated correction parameter set is used as the output of the cognitive twin layer. When the reliability level of the virtual-real consistency index is lower than the set lower limit, the adaptive compensation parameter vector of the previous version remains unchanged, and the timestamp of the integrated correction parameter set is kept consistent with that of the physical state vector.
8. The self-calibration method for force-controlled joint torque based on digital twin mapping according to claim 1, characterized in that, The process of inputting the comprehensive correction parameter set into the simulation twin layer, adjusting the parameters of the tensor computation structure of the simulation twin layer, and feeding back the correction results to the physical twin layer includes: Receive the integrated correction parameter set, register and cache it according to the time window index, align the integrated correction parameter set with the state tensor timestamp of the corresponding time window of the simulation twin layer, and establish a shadow parameter area; In the tensor computation structure of the simulated twin layer, hierarchical parameter injection is performed. The channel weights, gating thresholds, constraint fields, symmetric fusion coefficients and cross-scale consistency boundaries in the comprehensive correction parameter set are sequentially loaded into the shadow parameter area to complete the legality and consistency verification. Initiate adaptive routing adjustment, select the calculation path based on current load classification information and temperature segmentation information within the tensor calculation structure, set steady-state protection threshold to limit parameter jump amplitude, perform trial run recalculation on the state tensor corresponding to the time window, and generate candidate output of correction results; A dual-phase confirmation mechanism is adopted to process the parameters in the shadow parameter area. First, the candidate output of the correction result is checked against the historical output for boundary continuity and cross-channel correlation. Then, it is written into the main parameter area for immediate effect. The correction result is fed back to the physical twin layer, aligned with the timestamp of the physical state vector, and merged with the dynamic compensation value and the original signal of the torque sensor to output the corrected high-precision torque value. Establish a closed-loop monitoring and rollback process, record the calibrated high-precision torque value, parameter version identifier and trigger conditions. When the consistency index is lower than the preset lower limit or exceeds the safety boundary, roll back to the comprehensive calibration parameter set of the previous version. When the stability conditions are met, solidify the main parameter area.
Citation Information
Cited By
Warehouse virtual simulation scheduling method and system based on digital twinning
CN121599591A
Signal calibration method and system for semiconductor test equipment
CN122017711A
A complex equipment system visualization simulation method and system based on digital twinning
CN122113658B
Multi-user self-learning self-adaptive exoskeleton parameter self-tuning control method
CN122143068A
High-pressure glass fiber reinforced plastic pipe full-automatic layer winding simulation system based on digital twinning
CN122263553A
