Adaptive optimization method and system for parameters of digital twin-driven joint robot

Abstract

The invention relates to the technical field of robot adaptive control, in particular to a digital twinning driven joint robot parameter adaptive optimization method and system, and the method comprises the following steps: constructing a multi-model digital twinning body of a joint robot, and carrying out the forward prediction of a joint task; obtaining a consensus prediction result according to the prediction result and the confidence coefficient of each model, generating a multi-dimensional risk spectrum vector, obtaining an optimal control parameter, generating an active environment intervention instruction, and issuing the active environment intervention instruction to a joint robot and an environment controller for execution; and when an execution result of the joint robot deviates from a consensus prediction result, a failure source is calibrated based on causal inference and meta-learning calibration locking, and finally structured knowledge formed by the whole process is stored in a dynamic knowledge graph. The control parameters of the joint robot are automatically adjusted, the self-adaptive capacity of the joint robot is improved, the joint stability and quality are further improved, and reliable technical support is provided for intelligent production of textile enterprises.

Description

Technical Field

[0001] This invention relates to the field of robot adaptive control technology, and in particular to a method and system for adaptive optimization of parameters of a joint robot driven by a digital twin. Background Technology

[0002] The main related content of existing technologies for digital twin-driven splicing robots is the automatic splicing system for ring spinning. These primarily employ mechanical devices, track-based robots, and teach-and-playback systems. Mechanical devices use cam mechanisms, programmable controllers, and fixed actuators to achieve standardized splicing actions by executing preset rigid motion trajectories. Track-based robots, based on a track-moving platform and a multi-axis robotic arm, use machine vision to identify the breakage location and execute a pre-programmed fixed path. The teach-and-playback system records and reproduces the manual operation trajectory to achieve the splicing process. Problems and shortcomings of existing technologies include poor environmental adaptability, insufficient control precision, and a lack of autonomous learning capabilities. These problems stem from the rigid and fixed mechanical structure and the simplistic preset control program, preventing the system from adjusting parameters according to dynamic factors such as yarn type changes and environmental temperature and humidity fluctuations. Furthermore, the lack of a real-time force feedback mechanism in the actuators makes it difficult to cope with the complex contact dynamics caused by the flexible nature of yarn. In addition, the closed system architecture cannot autonomously learn and optimize strategies using historical data, and the data isolation between the digital twin model and the physical equipment further limits the system's continuous evolution capabilities. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for adaptive optimization of parameters of a digital twin-driven joint robot.

[0004] To achieve the above objectives, in a first aspect, the present invention provides a digital twin-driven adaptive optimization method for the parameters of a jointing robot. The method includes the following steps: constructing a multi-model digital twin of the jointing robot and performing forward prediction on the jointing task to obtain the prediction results of each model and calculate the confidence level of the corresponding prediction; based on the confidence level, fusing the prediction results using evidence theory to obtain a consensus prediction result, and simultaneously generating a multi-dimensional risk spectrum vector; dynamically adjusting the optimization objective function using the multi-dimensional risk spectrum vector to obtain the optimal control parameters of the jointing robot under constraints, and determining whether to generate an active environmental intervention command through benefit analysis; the jointing robot receiving and executing the optimal control parameters, and the environmental controller receiving and executing the active environmental intervention command, while simultaneously monitoring the execution result of the jointing robot; when the execution result deviates from the consensus prediction result, locking down the failure source based on causal inference, and calibrating the failure source through meta-learning calibration; and storing the context information from the entire process from forward prediction to calibration of the failure source into transferable structured knowledge in a dynamic knowledge graph. This invention enables autonomous adjustment of the control parameters of the jointing robot, improves the robot's self-adaptability, and thus enhances the stability and quality of the joints, providing reliable technical support for textile enterprises to achieve intelligent production.

[0005] Optionally, the multi-model digital twin includes a mechanistic model, a data-driven model, and an environmental disturbance model. The mechanistic model includes the robot's rigid-flexible coupling dynamic equations and a yarn nonlinear dynamic model based on Cosserat rod theory. The data-driven model is constructed based on a spatiotemporal graph convolutional network, with nodes including robot joints, robot end effectors, discrete yarn points, and environmental sensors. Node features include joint position, joint angle, joint torque, joint velocity, end effector pose, six-dimensional end effector force, three-dimensional yarn position, yarn tension, wind speed, ambient temperature, and ambient humidity. Edges include fixed physical connections and preset functional coupling connections. The data-driven model takes historical node state sequences as input and future node states and joint results as outputs. The environmental disturbance model is constructed based on a deep neural network, with inputs being the temperature, humidity, and wind speed at different monitoring points in the workshop, and outputting the wind speed vector within the jointing robot's working domain. The environmental disturbance model provides environmental disturbance inputs for the mechanistic model and the data-driven model.

[0006] Optionally, the step of constructing a multi-model digital twin of the jointing robot and performing forward prediction on the jointing task to obtain the prediction results of each model and calculate the confidence level of the corresponding prediction includes the following steps: constructing a multi-model digital twin of the jointing robot, and then using the multi-model digital twin to perform forward prediction on the jointing task to obtain the prediction results of each model; establishing confidence evaluation functions for the mechanism model, the data-driven model, and the environmental disturbance model respectively, and then calculating the confidence level of the corresponding prediction results respectively.

[0007] Optionally, the step of fusing the prediction results using evidence theory based on the confidence level to obtain a consensus prediction result and simultaneously generating a multidimensional risk spectrum vector includes the following steps: converting the prediction results and their confidence levels of the mechanistic model and the data-driven model into basic probability assignments in DS evidence theory and combining them, while calculating the total conflict quality of all conflict combinations; calculating the combined confidence level of each conflict combination and the total confidence level of all conflict combinations based on the confidence level of the prediction results corresponding to the propositions in the conflict combinations, wherein the propositions include joint failure and joint success; and allocating the total conflict quality according to the proportion of the combined confidence level to the total confidence level. For each of the conflict combinations, the redistributed conflict quality of each conflict combination is obtained; according to the confidence ratio of each proposition in the conflict combination, the redistributed conflict quality is assigned to the propositions in the conflict combination to obtain the redistributed quality of each proposition in the conflict combination; for each proposition, the sum of the redistributed quality is calculated, and the calculation result is added to the non-conflict quality of the proposition to obtain its basic probability allocation value; according to the basic probability allocation value, the pignistic probability of the corresponding proposition is calculated, and then the consensus prediction result is obtained; the physical model mismatch risk, data model extrapolation risk, and environmental perturbation model uncertainty risk are calculated to form the multidimensional risk spectrum vector.

[0008] Optionally, the multidimensional risk spectrum vector satisfies the following relationship: , , , , in, Let the multidimensional risk spectrum vector be... To account for the risk of mismatch in the physical model, The risk of extrapolating the data model, This refers to the uncertainty risk of the environmental disturbance model. Assign a value to the basic probability of the joint failure. The confidence level of the prediction result of the aforementioned mechanism model. The confidence level of the prediction results of the data-driven model. The confidence level of the prediction results of the environmental disturbance model is given. The wind speed vector is denoted as .

[0009] Optionally, the step of dynamically adjusting and optimizing the objective function using the multidimensional risk spectrum vector to obtain the optimal control parameters of the robot under constraints, and determining whether to generate an active environmental intervention command through benefit analysis, includes the following steps: dynamically adjusting and optimizing the objective function using the multidimensional risk spectrum vector, and obtaining the optimal control parameters using a parameter optimization algorithm with the aim of minimizing the optimization objective function under constraints; generating an environmental intervention scheme, predicting the expected loss when executing the optimal control parameters without interfering with the environment, and simultaneously predicting the expected benefit when executing the environmental intervention scheme; and generating the active environmental intervention command based on the environmental intervention scheme when the expected benefit is greater than a benefit threshold and greater than the sum of the expected loss and the intervention cost.

[0010] Optionally, the objective function and the constraints satisfy the following relationship: , , , , in, To optimize the objective function value, Weighted by success rate, As the benchmark weight for joint quality, As time weight, As environmental benchmark weights, Regarding control parameters The success rate objective function For about The quality objective function, For about The time objective function, For about The environmental objective function, As the benchmark weight for success rate, The risk of extrapolating the data model, As a time-based weight, and For risk sensitivity coefficient, This refers to the uncertainty risk of the environmental disturbance model. For the j-th constrained control parameter, This represents the default design margin for the j-th constrained control parameter. Let j be the constraint relaxation factor for the j-th constrained control parameter. This refers to the risk of mismatch in the physical model.

[0011] Optionally, when the execution result deviates from the consensus prediction result, the failure source is locked based on causal inference, and the failure source is calibrated through meta-learning calibration, including the following steps: constructing a lightweight causal inference graph of the connector task and performing causal identification to determine the cause variables, confounding variables, and outcome variables of the connector task; when the execution result deviates from the consensus prediction result, calculating the causal effect value of each failure path in the lightweight causal inference graph, thereby locking the failure source; for the failure source, obtaining multiple reference historical calibration schemes from the dynamic knowledge graph of historical connector tasks through similarity analysis, thereby generating a weighted task distribution; initializing the inner loop optimization process of the model-independent meta-learning algorithm using the weighted task distribution; the model-independent meta-learning algorithm performs gradient iteration on the reference historical calibration schemes to enable the failure source to quickly adapt to a state sensitive to new tasks, and uses data from multiple new connector tasks to enable the model-independent meta-learning algorithm to perform multi-step gradient updates on the reference historical calibration schemes to achieve calibration of the failure source.

[0012] Optionally, for the failure source, multiple reference historical calibration schemes are obtained from the dynamic knowledge graph of historical rendezvous tasks through similarity analysis, and then a weighted task distribution is generated. This includes the following steps: constructing the feature vector of the current rendezvous task, and simultaneously extracting the feature vector of the historical rendezvous task from the dynamic knowledge graph of historical rendezvous tasks. The feature vector includes a working condition feature vector, a causal diagnosis feature vector, and a multidimensional risk spectrum vector. Based on the feature vector, the similarity of the working condition features, the similarity of the causal diagnosis features, and the risk similarity between the current rendezvous task and the historical rendezvous task are calculated respectively. The historical calibration scheme of the historical rendezvous task is simulated in the current digital twin environment, and the corresponding optimization objective function value is obtained. The migration value is used as the transferability score. The difference between the optimization objective function value of the current connector task and the migration value is calculated, and the reciprocal of the difference is used as the transferability score of the corresponding historical calibration scheme. The weighted sum of the similarity of the working condition features, the similarity of the causal diagnostic features, the similarity of the risk, and the transferability score is used as the comprehensive similarity between the current connector task and the historical connector task. A similarity threshold is set, and the historical calibration schemes of historical connector tasks whose comprehensive similarity with the current connector task is greater than the similarity threshold are used as reference historical calibration schemes. For the reference historical calibration schemes, the comprehensive similarity is normalized to obtain the scheme weight of the reference historical calibration scheme, and then a weighted task distribution is generated.

[0013] Secondly, the present invention provides a digital twin-driven adaptive optimization system for the parameters of a joint robot. The digital twin-driven adaptive optimization system for the parameters of a joint robot includes: a data acquisition device, a data output device, a processor, and a storage device. The storage device includes a computer-readable storage medium storing a computer program. The computer program includes program instructions, which, when executed by the processor, cause the processor to implement the digital twin-driven adaptive optimization method for the parameters of a joint robot provided by the present invention.

[0014] In summary, the present invention has at least the following beneficial effects: 1. This method achieves real-time interaction between the physical entity and the virtual model by constructing a multi-model digital twin, and performs forward prediction for the jointing task. The prediction results of the multi-model digital twin are not directly used to optimize the control parameters of the jointing robot. Instead, they are fused based on the dynamic confidence levels of the individual model predictions and evidence theory to obtain a consensus prediction result and a multi-dimensional risk spectrum vector. The obtained consensus prediction result remains stable and reasonable even when the prediction results of different models differ significantly, achieving a leap from simple voting to credibility-based debate and consensus formation. The obtained multi-dimensional risk spectrum vector quantifies the uncertainty of the prediction results of different models, providing fine-grained guidance for downstream optimization of the jointing robot's control parameters. This is beneficial for improving the adaptability of the jointing robot, thereby improving the stability and quality of the jointing process.

[0015] 2. This method generates a set of optimal control parameters based on the objective function of the control parameter optimization algorithm dynamically reconstructed by the multidimensional risk spectrum vector. Based on the benefit analysis, it generates active environmental intervention instructions when necessary, realizing the "local instantaneous shaping" of the production environment. This enables the jointing robot to learn and optimize autonomously, further improving the adaptability of the jointing robot and thus improving the stability and quality of the joint.

[0016] 3. After performing the joint operation, this method locks the failure source based on causal inference and calibrates the failure source through meta-learning calibration. Finally, the entire process forms transferable structured knowledge and stores it in a dynamic knowledge graph, realizing joint experience reuse and cold start optimization, and realizing autonomous learning optimization based on historical data.

[0017] 4. A system adapted to the method is provided, which not only improves the practicality of the method but also facilitates its promotion. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the adaptive optimization method for parameters of a digital twin-driven joint robot according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the framework of the digital twin-driven adaptive optimization system for the joint robot parameters according to an embodiment of the present invention. Detailed Implementation

[0020] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0021] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0022] It should be noted in advance that, in one alternative embodiment, except for independent descriptions, the same symbols or letters appearing in all formulas have the same meaning.

[0023] In one optional embodiment, please refer to Figure 1 This invention provides a method for adaptive optimization of parameters of a digital twin-driven joint robot, the method comprising the following steps: S1. Construct a multi-model digital twin of the jointing robot and perform forward prediction on the jointing task to obtain the prediction results of each model and calculate the confidence level of the corresponding prediction.

[0024] Step S1 specifically includes the following steps: S11. Construct a multi-model digital twin of the jointing robot, and then use the multi-model digital twin to perform forward prediction of the jointing task to obtain the prediction results of each model.

[0025] Specifically, in this embodiment, the multi-model digital twin includes a mechanistic model, a data-driven model, and an environmental perturbation model. The mechanistic model comprises the robot's rigid-flexible coupling dynamic equations and a yarn nonlinear dynamic model based on Cosserat bar theory. The robot's rigid-flexible coupling dynamic equations and the yarn nonlinear dynamic model are as follows:

[0026]

[0027] in, The mass inertia matrix depends on the joint position vector q of the joint robot. Let the joint acceleration vector of the connector robot be... For the Coriolis and centripetal force matrices, Let be the joint velocity vector of the connector robot. The gravity vector depends on q. Let J be the joint driving torque vector, J be the geometric Jacobian matrix of the end effector of the joint robot, and T be the transpose. The external force acting on the end effector of the connector robot. Let the mass of the i-th discrete point on the yarn be denoted as . Let i be the acceleration vector of the i-th discrete point on the yarn. Let be the nonlinear tensile force between the i-th discrete point on the yarn and its adjacent discrete points. The bending force at the i-th discrete point on the yarn is based on the discrete curvature penalty. Let be the viscous damping force at the i-th discrete point on the yarn. It is proportional to the velocity of the i-th discrete point on the yarn. Let be the contact and collision force between the i-th discrete point on the yarn and a rigid body such as the yarn guide. Based on the Hertz-Mindlin contact model Let be the air resistance at the i-th discrete point on the yarn. According to Hooke's law, the nonlinear tensile force... , For tensile stiffness, This represents the change in yarn length between adjacent discrete points. The original yarn length between adjacent discrete points. The viscosity coefficient related to the stretching speed, The rate of change of yarn length between adjacent discrete points; air resistance. , air density, Here, A is the drag coefficient, and A is the projected area of ​​the yarn perpendicular to the wind speed direction. This refers to wind speed.

[0028] As an inherent property of yarn materials, it is calibrated through a standard uniaxial tensile test, and its typical range is: ~ N / m, specifically depends on the yarn count, material, and twist; The known geometric parameters preset during modeling are determined based on the yarn discretization resolution and initial configuration; For simulation or real-time calculation, the real-time distance between adjacent discrete points and The difference; The internal damping of the material is characterized by a typical value range of 0.01~0.1 N·s / m, which can be obtained by fitting dynamic tensile test data or online parameter identification. This is a physical constant under standard atmospheric conditions, typically taken as 1.225 kg / m³. 3 ; Depending on the yarn cross-sectional shape (usually considered as a cylinder) and the airflow state (Reynolds number), the typical value is about 1.0 for laminar flow around a cylinder, and can be increased to about 1.2 for turbulent conditions. The specific value can be determined by calculating the Reynolds number based on the yarn diameter and wind speed and then using empirical correlations in fluid dynamics or by looking up tables. Alternatively, it can be calibrated through computational fluid dynamics (CFD) simulation.

[0029] When running a real-time data-driven multi-model digital twin, the mechanistic model numerically solves its inherent robot rigid-flexible coupling dynamics equations and yarn nonlinear dynamics equations to simulate the entire physical process of the current splicing task. This outputs a predicted sequence of key physical quantities, including joint position, joint angle, joint torque, joint velocity, end-effector pose, six-dimensional end-effector force, three-dimensional yarn position, and yarn tension. Simultaneously, the simulation results can directly predict whether the splicing was successful. This prediction, along with the data-driven state and probability predictions from the data-driven model and the wind speed field predictions from the environmental disturbance model, constitutes a multi-perspective forward prediction output, providing input for subsequent evidence-based fusion and decision-making.

[0030] The data-driven model is built upon a Spatiotemporal Graph Convolutional Network (ST-GCN). Nodes include robot joints, robot end effectors, discrete yarn points, and environmental sensors. Edges include fixed physical connections and pre-defined functional coupling connections. Specifically, robot joint node features include joint position, joint angle, joint torque, and joint velocity; robot end effector node features include end effector pose and six-dimensional force; discrete yarn point node features include yarn three-dimensional position and yarn tension; and environmental sensor node features correspond to sensor type, such as ambient temperature for a temperature sensor, ambient humidity for a humidity sensor, and wind speed for a wind speed sensor. The data-driven model takes historical node state sequences as input and outputs future node states and the joint result, with the joint result being an estimate of the probability of joint success. The training process of the data-driven model consists of two steps: First, a dataset is constructed using historical node states and corresponding real junction results. Then, the dataset is divided into a training set and a validation set in a 7:3 ratio to complete the supervised pre-training of the model. The goal of the pre-training is to minimize the mean square error between the prediction results and the real data. Subsequently, adversarial training is introduced in the digital twin environment to make the prediction distribution of the data-driven model move closer to the mechanistic model to enhance physical credibility.

[0031] The adversarial training process is as follows: First, the mechanistic model is run under the same input conditions as the data-driven model, and the output sequence of predicted key physical quantities is used as a "soft label" or reference distribution with physical consistency. Then, a discriminator network is introduced into the adversarial training framework. The input of the discriminator is the future node state obtained by the data-driven model (excluding node features from environmental sensors). Its training objective is to accurately distinguish whether the input data comes from the data-driven model or the mechanistic model. At the same time, the data-driven model adds an adversarial loss to the original supervised loss (mean square error between the predicted result and the actual result). This loss encourages the output future node state to make it impossible for the discriminator to reliably distinguish its source, thereby driving the internal representation and output distribution of the data-driven model to move closer to the physical constraints implied by the mechanistic model. This process essentially uses the mechanistic model as the authoritative source of physical knowledge, and injects physical consistency into the data-driven model in a differentiable way through adversarial regularization. This effectively avoids the physically unreliable predictions that may occur with pure data-driven models and enhances the extrapolation robustness of the model under unseen conditions.

[0032] The environmental disturbance model is built upon a deep neural network (DNN). It is a DNN proxy model pre-trained through offline CFD simulation and then subjected to lightweight online inference. Its inputs are the temperature, humidity, and wind speed at different monitoring points within the workshop, and its output is the wind speed vector within the working domain of the jointing robot. The construction and training methods of the environmental disturbance model are as follows: First, a deep neural network based on an encoder-decoder architecture using an attention mechanism is employed to construct a wind field prediction network. Specifically, the encoder first performs multi-layer nonlinear transformations on the input temperature, humidity, and wind speed sensor data, mapping the original data to low-dimensional latent space features through fully connected layers, compressing the data dimensionality and extracting key environmental features. The decoder, guided by the target query coordinates (x, y, z), combines latent space features and dynamically calculates the contribution weights of each sensor to the target point through a multi-head attention mechanism—each attention head independently generates a query, key, and value matrix, calculates the attention score, and then weights and fuses the sensor information to capture the complex interaction between temperature, humidity, and wind speed, obtaining attention fusion features. To further improve spatial detail reconstruction capabilities, the decoder introduces Fourier feature encoding technology, mapping the target coordinates to a high-frequency sine / cosine space. The generated spatial code is concatenated with the attention fusion features and input into the final output layer to directly predict the three-dimensional wind speed vector of the target point. The loss function of the wind field prediction network is:

[0033] Where L is the loss value, Predicted wind speed within the working area of ​​the connector robot Wind speed obtained from CFD simulation The mean square error, These are the weighting coefficients. These are physical constraint terms, calculated... The divergence is calculated and penalized to achieve this.

[0034] Then, CFD software was used to conduct large-scale parametric simulations of the workshop, covering various typical and extreme working conditions (such as door and window opening and closing, air conditioning operation, and equipment movement), generating massive amounts of temperature, humidity, and wind speed data within the workshop and constructing a dataset. The constructed dataset was then divided into training and validation sets in a 7:3 ratio to complete the training and validation of the wind field prediction network. After the environmental disturbance model was trained, the readings from the wind speed sensor could be interpolated in real time to represent the airflow distribution at any point within the working domain of the joint robot, providing key environmental disturbance inputs for the mechanistic model and the data-driven model, thus achieving a closed loop from high-fidelity physical simulation to efficient real-time inference.

[0035] After constructing the multi-model digital twin of the jointing robot, real-time acquired data is used to drive the multi-model digital twin's operation and perform forward prediction for the current jointing task, obtaining the prediction results of each model. Specifically, the data acquired includes joint position and angle obtained through joint encoders, joint torque obtained through joint torque sensors, joint velocity obtained through velocity sensors, end-effector six-dimensional force obtained through six-dimensional force sensors on the end effector, yarn tension obtained through miniature tension detectors, ambient temperature, humidity, and wind speed obtained through a network of temperature, humidity, and wind speed sensors deployed in the workshop, and the three-dimensional position of discrete yarn points obtained through a vision system (such as an industrial camera) combined with coordinate system transformation. This is not an exhaustive list; the methods used to acquire data to drive the multi-model digital twin's operation are all existing technologies, and sensor types can be selected according to actual needs. These will not be described in excessive detail here.

[0036] These real-time collected data serve as inputs for the mechanistic model, the data-driven model, and the environmental disturbance model, respectively. The mechanistic model, based on the collected data, performs high-fidelity physical simulation by solving the inherent robot rigid-flexible coupling dynamics equations and yarn nonlinear dynamics equations. This outputs predictions for joint positions, joint angles, joint torques, joint velocities, end-effector poses, six-dimensional end-effector forces, three-dimensional yarn positions, and yarn tension over a future period, providing a definitive judgment on the success or failure of the joint. The data-driven model constructs a historical sequence of spatiotemporal graph node features from the aforementioned data (such as joint positions, joint angles, joint torques, joint velocities, end-effector poses, six-dimensional end-effector forces, three-dimensional yarn positions, yarn tension, ambient temperature, ambient humidity, and wind speed), outputting predictions for future node states and joint results. The environmental disturbance model maps temperature, humidity, and wind speed data monitored at multiple points in the workshop into a precise wind speed field prediction within the robot's working domain, providing crucial environmental disturbance inputs for the aforementioned two models.

[0037] S12. Establish confidence evaluation functions for the mechanism model, the data-driven model, and the environmental disturbance model respectively, and then calculate the confidence of the corresponding prediction results respectively.

[0038] Specifically, in this embodiment, the mechanistic model, the data-driven model, and the environmental perturbation model each obtain a confidence level for each prediction, and the confidence level of the prediction results of each model is calculated through the corresponding confidence level evaluation function.

[0039] For the mechanistic model, the confidence evaluation function is:

[0040] in, The confidence level of the prediction results of the mechanistic model. For a scaling factor that is strictly greater than zero, The readings are from the joint torque sensor. The theoretical joint driving torque is calculated by back-calculation of the mechanism model. The larger the value, the more severe the mismatch in physical parameters, and the lower the confidence level.

[0041] In the confidence evaluation function of the mechanistic model, the core role of the scaling factor is as a "sensitivity regulator," mapping the squared value of the model's prediction error to an appropriate order of magnitude, thereby controlling the... The sensitivity to this error is determined through a systematic engineering calibration process to ensure that changes in confidence level align with engineering intuition and practical needs. This calibration process includes the following steps: First, with the system in good condition, the robot performs standard movements to collect readings from the joint torque sensors. Simultaneously, the theoretical joint driving torque is calculated using a mechanistic model, and then... The baseline prediction error of the mechanistic model is calculated using the following method. Then, based on the requirements for system robustness, a design objective is defined, for example, when the instantaneous prediction error reaches K times the baseline error (e.g., K=4), the confidence level should be reduced to 0.5; finally, the design objective is... , "Substituting these values ​​into the confidence evaluation function of the mechanistic model, the scaling factor is solved. Therefore, the scaling factor is a key calibration parameter that connects the actual statistical performance of the system with the design expectation. Its value allows the confidence index to adaptively and reasonably reflect the model's credibility at the current moment: a small prediction error results in a high confidence level, while an abnormally large prediction error significantly reduces the confidence level. This allows the model to dynamically adjust its weights in the subsequent process of obtaining consensus prediction results through multi-model fusion decision-making."

[0042] For data-driven models, the confidence evaluation function is:

[0043] in, The confidence level of the prediction results of the data-driven model. Real-time node status. The training set for training data-driven models, for and Mahalanobis distance is used to measure whether the current node state is within the empirical range of the model.

[0044] In mathematics, the "model experience range" refers to the region of multivariate probability distributions covered by the training data of the data-driven model. Its core components are the mean vector and covariance matrix of the training set, which define the "main" distribution shape of the data. (Judgment) Whether something falls within this range essentially means checking whether it belongs to the same distribution.

[0045] Based on Mahalanobis distance The calculation method is a rigorous statistical process, specifically including the following steps: First, in The Mahalanobis distance for all samples is calculated to obtain its empirical distribution. Furthermore, if we assume that the sample data approximately follows a multivariate normal distribution, then the squared value of the Mahalanobis distance... Theoretically, it follows a chi-square distribution with the number of degrees of freedom equal to the number of states; a critical value for the Mahalanobis distance is determined based on the required confidence level. This critical value can be directly obtained from the chi-square distribution table. This refers to the statistical boundary of the "empirical range of the model"; for Calculate its relationship with The Mahalanobis distance, if the calculated result is not greater than Then it is believed If the predictions fall within the model's empirical range, the data-driven model's predictions are relatively reliable; otherwise, it indicates... The data has deviated significantly from the distribution of the training set data, indicating an "extrapolation" state, and the model's prediction uncertainty is extremely high. The calculation formula is a continuous quantification of this judgment, thus automatically and quantitatively reflecting the cognitive uncertainty of the data-driven model regarding the current input.

[0046] For environmental disturbance models, the confidence level is quantified based on a multi-factor weighted evaluation framework. The core idea is that the confidence level is determined by the sensor network's ability to capture current flow field characteristics. Specifically, it is derived through a combination of the following calculable indicators: 1. Spatial coverage adequacy score This is the foundation of confidence. An optimal virtual sensor placement scheme is preset for the working domain of the connector robot. Then, by calculating the actual position of currently online, functional, and effective sensors, and... The matching degree is obtained by the formula:

[0047] in, Let Euclidean distance be the distance from the k-th optimal virtual sensor placement point to the nearest effective sensor. For relevant length scale parameters, The optimal number of virtual sensor points. The closer it is to 1, the more complete the spatial coverage.

[0048] Used to quantify the attenuation characteristics of distance influence in the space where sensors are deployed, its physical meaning represents the characteristic distances with significant spatial correlation to environmental parameters (such as wind speed and temperature) within a workshop. It is typically set based on the workshop's geometric dimensions or the range of the working area, with typical values ​​ranging from 1 / 10 to 1 / 5 of the workshop's main longitudinal dimension. For example, for a workshop with dimensions of 20m × 10m × 5m... It can be initialized to 2~4m. The specific value of this parameter can be calibrated and optimized through spatial correlation analysis between historical sensor data, that is, calculating the curve of correlation coefficient decay with distance between sensor readings at different intervals, and fitting the curve to obtain the correlation coefficient decreasing to a certain value. The distance corresponding to the time is used as The empirical value, thereby ensuring The calculation can reasonably reflect the actual monitoring network's ability to capture the state of the working domain environment.

[0049] 2. Data quality and consistency score This score assesses the reliability of the sensor readings themselves. First, it checks whether each sensor reading is within its calibrated normal physical range; any outlier reduces the corresponding sensor's weight to zero. Second, for a given physical quantity (such as wind speed), the readings of multiple neighboring sensors should exhibit spatial consistency. Taking wind speed as an example, this is achieved by calculating the variance of the readings of multiple neighboring wind speed sensors within a local area. After normalization, a consistency factor for wind speed in a local area is obtained to measure the spatial consistency of wind speed within that area. A larger value indicates poorer spatial consistency and lower wind speed uniformity. It is the product of the percentage of effective wind speed sensors among all wind speed sensors and the average consistency factor. Similarly, the ambient temperature can also be obtained. Corresponding to ambient humidity Then, in the same area , and The weighted sum is used as the final data quality and consistency score. Each has a weight of one-third.

[0050] 3. Model prior uncertainty score This score originates from the model's own perception. During the CFD training phase, the prediction error of the data-driven model for the flow field at coordinates (x, y, z) under different samples can be statistically analyzed, and then the variance of the prediction error at (x, y, z) can be calculated. When the real-time position of the end of the splicing robot and the yarn is (x, y, z), the position can be found. Then, the following relationship is used to calculate the result. :

[0051] in, This is the scaling factor. Used to Mapped to Its value range is usually set to 0.5~5.

[0052] Finally, the confidence evaluation function for the environmental disturbance model is:

[0053] in, The confidence level of the prediction results from the environmental disturbance model. , and As weight, and . , and To allow for regression adjustments based on historical data of ambient temperature, humidity, and wind speed, specifically, it can... and When a safety threshold (e.g., 0.5) is set, when or When it falls below the corresponding safety threshold, Forced to a very low value (e.g., 0.2) to significantly warn of the unreliability of environmental disturbance information, and to solve for... , and This quantification method makes It should become a dynamic, objective, and interpretable reliability indicator.

[0054] S2. Based on the confidence level, the prediction results are fused using evidence theory to obtain a consensus prediction result, and a multidimensional risk spectrum vector is generated.

[0055] Step S2 specifically includes the following steps: S21. The prediction results and their confidence levels of the mechanistic model and the data-driven model are transformed into basic probability assignments in the DS evidence theory and combined, while the total conflict quality of all conflict combinations is calculated.

[0056] Specifically, in this embodiment, the predictions and confidence levels of key indicators by the mechanistic model and the data-driven model are transformed into Basic Probability Assignment (BPA) in Dempster-Shafer evidence theory. The key indicator refers to the binary outcome of the rendezvous task, i.e., the proposition space is explicitly limited to two mutually exclusive basic propositions: "rendezvous success" and "rendezvous failure." This setting aims to simplify the fusion process of Dempster-Shafer evidence theory, avoid the expansion of the joint proposition space and computational complexity caused by introducing multiple performance indicators, ensure the clarity and operability of evidence combination and conflict quality assignment, while still effectively supporting subsequent consensus-based prediction-based decision-making and risk analysis. The mechanistic model provides a deterministic judgment of rendezvous success or failure through physical simulation; a successful rendezvous is considered a success rate of 1, otherwise it is 0, and so on. The degree of support for a deterministic judgment is used to generate a BPA (Best Practice Assessment). For example, if the connection is successful, then... , The data-driven model directly outputs an estimate of the joint success probability. Combined with its confidence level Generate BPA, i.e. , , The environmental disturbance model does not directly predict the joint outcome; its output is a wind speed vector, used to provide environmental input for the aforementioned two models. Therefore, it does not directly participate in BPA allocation, but its confidence level... It will be used for the subsequent calculation of the multidimensional risk spectrum vector, thus playing an indirect role in the decision-making framework. This indicates the degree to which the evidence from the mechanistic model supports the proposition of "successful jointing". For the complete collection, The extent to which the evidence for the mechanistic model supports all possible scenarios other than "successful jointing" (i.e., the uncertainty component) This indicates the degree to which the evidence from the data-driven model supports the proposition of "successful joint". This indicates the degree to which the evidence from the data-driven model supports the proposition of "joint failure". The degree to which the evidence for the data-driven model supports all possible scenarios other than "successful connection" (i.e. the uncertainty part).

[0057] Based on Dempster-Shafer evidence theory, the basic probability assignments of each model are combined. During this combination process, situations arise where the propositional intersection is empty; these combinations are then considered conflicting combinations, and their conflict quality can be calculated. The total conflict quality is obtained by summing the conflict qualities of all conflicting combinations, representing the total irreconcilable discrepancies between models. Unlike traditional Dempster-Shafer evidence theory, which discards this total conflict quality, this embodiment treats it as valuable diagnostic information and redistributes it.

[0058] Furthermore, having calculated the conflict quality, the quality of evidence in the non-conflicting parts can also be calculated, and the quality of each proposition in the non-conflicting parts is recorded as the non-conflicting quality. The methods for calculating the conflict quality and the quality of evidence in the non-conflicting parts are existing technical means and will not be described in detail here.

[0059] S22. Based on the confidence level of the prediction result corresponding to the proposition in the conflict combination, calculate the combination confidence level of each conflict combination and the total confidence level of all conflict combinations, wherein the proposition includes joint failure and joint success.

[0060] Specifically, in this embodiment, in each conflict combination, the proposition is determined based on the prediction results of each model, so the confidence level of each proposition is the confidence level of the prediction result of the corresponding model. Furthermore, the combined confidence level is the sum of the confidence levels of each proposition in the conflict combination, while the total confidence level is the sum of the combined confidence levels of all conflict combinations.

[0061] S23. The total conflict quality is allocated to each of the conflict combinations according to the proportion of the combined confidence level to the total confidence level, thereby obtaining the redistributed conflict quality of each of the conflict combinations.

[0062] Specifically, in this embodiment, the conflict quality allocation weights for each conflict combination are calculated, and the conflict quality allocation weights satisfy the following relationship:

[0063] in, Assign weights to conflict quality. The combined confidence level of conflicting combinations. Let be the total confidence level for all conflict combinations. It can be seen that the conflict quality allocation weight is the proportion of the combination confidence level to the total confidence level. The conflict quality allocation weight for each conflict combination is calculated according to this formula. Then, the product of the conflict quality allocation weight for each conflict combination and the total conflict quality is used as the redistributed conflict quality for the corresponding conflict combination.

[0064] S24. Based on the confidence ratio of each proposition in the conflict combination, the redistribution conflict quality is allocated to the propositions in the conflict combination to obtain the redistribution quality of each proposition in the conflict combination.

[0065] Specifically, in this embodiment, after obtaining the redistributed conflict quality of the conflict combination, the redistributed conflict quality is allocated to the propositions in the conflict combination according to the confidence ratio of each proposition in the conflict combination. Specifically, for a given conflict combination, the ratio of the confidence of each proposition in it to the combined confidence of the conflict combination is calculated, and this ratio is recorded as the proposition weight. The redistributed conflict quality of each proposition in the conflict combination is the product of the proposition weight and the redistributed conflict quality.

[0066] S25. For each of the propositions, calculate the sum of the redistributed qualities, and add the calculation result to the non-conflict quality of the proposition to obtain its basic probability allocation value.

[0067] Specifically, in this embodiment, since there may be multiple conflict combinations, the same proposition may have multiple redistributed conflict qualities. For any proposition, the sum of its redistributed qualities is calculated, and the result is added to the proposition's non-conflicting qualities to obtain the basic probability assignment value of the proposition.

[0068] S26. Based on the basic probability allocation value, calculate the Pengistic probability of the corresponding proposition, and then obtain the consensus prediction result.

[0069] Specifically, in this embodiment, for any proposition, its pignistic probability is calculated, and the result is the consensus prediction result for that proposition. Taking "successful rendezvous" as an example, its pignistic probability is:

[0070] in, Let be the Pignistic probability of the proposition; Assign a value to the basic probability of "successful connection". The basic probability assignment value for the "uncertainty term" represents the total uncertainty mass remaining unassigned to any definite proposition (i.e., "successful rendezvous" or "failed rendezvous") after evidence combination and conflict quality redistribution. This value is naturally derived through the calculation process in steps S21 to S25: after completing the combination of basic probability assignments for each model, the calculation and redistribution of conflict quality, and adding the redistributed mass to the non-conflict quality of each proposition, the sum of the basic probability assignment values ​​for all definite propositions is less than 1. 1 minus the sum of the basic probability assignment values ​​for definite propositions is the result. It quantifies the system uncertainty that cannot be eliminated after multi-model prediction fusion, and distributes it to each proposition according to the principle of uniform distribution when calculating the Pignistic probability (divided by 2 here because there are only two explicit propositions), thereby transforming uncertainty into a probability estimate that can be used for decision-making.

[0071] Thus, this embodiment realizes the process of intelligently redistributing conflict quality using dynamic confidence. The resulting consensus prediction results remain stable and reasonable even when the prediction results of each model are seriously divergent. It achieves a leap from simple voting to confidence-based debate and consensus formation, which is conducive to improving the adaptive ability of the joint robot, thereby improving the stability and quality of the joint.

[0072] S27. Calculate the risk of mismatch in the physical model, the risk of extrapolation in the data model, and the uncertainty risk of the environmental disturbance model to form the multidimensional risk spectrum vector.

[0073] Specifically, in this embodiment, the multidimensional risk spectrum vector satisfies the following relationship:

[0074]

[0075]

[0076]

[0077] in, For a multidimensional risk spectrum vector, To mitigate the risk of physical model mismatch, To mitigate the risks associated with extrapolating data models, This addresses the uncertainty risk associated with environmental disturbance models. Assign values ​​to the basic probability of joint failure. The confidence level of the prediction results of the mechanistic model. The confidence level of the prediction results of the data-driven model. The confidence level of the prediction results of the environmental disturbance model. This is the wind speed vector.

[0078] S3. Use the multidimensional risk spectrum vector to dynamically adjust and optimize the objective function, thereby obtaining the optimal control parameters of the joint robot under constraints, and determine whether to generate an active environmental intervention command through benefit analysis.

[0079] Specifically, step S3 includes the following steps: S31. The objective function is dynamically adjusted and optimized using the multidimensional risk spectrum vector, and the optimal control parameters are obtained using a parameter optimization algorithm with the aim of minimizing the objective function under constraints.

[0080] Specifically, in this embodiment, the objective function and constraints satisfy the following relationship:

[0081]

[0082]

[0083]

[0084] in, To optimize the objective function value, Weighted by success rate, As the benchmark weight for joint quality, As time weight, As environmental benchmark weights, Regarding control parameters The success rate objective function For about The quality objective function, For about The time objective function, For about The environmental objective function, As the benchmark weight for success rate, To mitigate the risks associated with extrapolating data models, As a time-based weight, and For risk sensitivity coefficient, This addresses the uncertainty risk associated with environmental disturbance models. For the j-th constrained control parameter, This represents the default design margin for the j-th constrained control parameter. Let j be the constraint relaxation factor for the j-th constrained control parameter. This refers to the risk of physical model mismatch. The success rate mentioned refers to the joint success rate, and the time refers to the time taken to complete the joint (also known as the joint time).

[0085] Control parameters refer to the low-level or high-level command variables that can be directly adjusted by the joint robot control system. These include, but are not limited to: setpoints for robot joint positions, joint velocity curve parameters, gain or target values ​​of the joint torque control loop, coordinates of the motion trajectory interpolation points of the end effector, attitude adjustment parameters, and the timing and force thresholds of the gripping or manipulating mechanisms. These parameters directly affect the success rate, quality, efficiency, and response to environmental disturbances of the joint by altering the robot's kinematic and dynamic behavior. Therefore, they can be used as optimization variables, iteratively adjusted by optimization algorithms in a digital twin simulation environment to evaluate their impact on the objective function. , , and The influence of simulation can be used to achieve parameter optimization based on simulation.

[0086] objective function , , , Not about control parameters Instead of an explicit analytical expression, it uses a performance evaluation function based on digital twin simulation. Specifically, during the optimization process, for each set of control parameters, multiple fast forward simulations are performed in a multi-model digital twin, and simulation results are collected to calculate the values ​​of each objective function. This represents the proportion of successful joint operations in the simulation out of the total number of simulations (which may include random disturbances) under this set of parameters. , The average tensile stiffness is obtained by averaging the tensile stiffness at the joint points through virtual tensile testing. To score the uniformity of the joint appearance, the reciprocal of the standard deviation of the joint segment diameter is calculated, and then the reciprocal is normalized using the maximum-minimum normalization method. The average time to complete the joint; The integration interval is the simulation period. Therefore, these function values ​​are indirectly but deterministically related to the control parameters through simulation experiments. The optimization algorithm constructs the objective function response surface and performs optimization by querying the simulation results corresponding to different control parameters.

[0087] Constraints are designed to ensure that the optimized control parameters are within the physically feasible and safe range of the equipment. Typical control parameters that need to be constrained include: joint torque, i.e., the torque output by each joint drive motor must not exceed its peak or rated torque; joint speed, i.e., the movement speed of each joint must not exceed the mechanical limit; and workspace constraints, i.e., the position of the end effector must not exceed the robot's reachable space or interfere with surrounding equipment.

[0088] In the optimization objective function, a set of benchmark weights (success rate benchmark weight, joint quality benchmark weight, time benchmark weight, and environmental benchmark weight) defines the fixed priority of each optimization objective (success rate, quality, time, and environment) under risk-free or benchmark risk levels. This priority is determined based on offline multi-objective Pareto front analysis, specifically as follows: A massive number of parameter combinations are generated for all adjustable parameters in the optimization objective function, and joint simulation is performed in a digital twin environment to generate a Pareto front in a four-dimensional objective space containing success rate, joint quality, time, and environment. Subsequently, domain experts select an optimal trade-off point on this front, representing the best balance between the objectives under the current constraints. Finally, based on the optimal trade-off point, a particle swarm optimization algorithm is used to find a set of optimal benchmark weights. Specifically, when using the particle swarm optimization algorithm to find a set of optimal benchmark weights, the number of particles is set to 30, with each particle representing a set of candidate benchmark weights. The constraints are and The initial positions of the particles are randomly generated under constraints, the initial velocity of the particles is set to 0, the fitness function is set to the Euclidean distance between the particle and the optimal tradeoff point, the maximum number of iterations is set to 100, and the convergence condition is that the fitness value changes by less than 0.001 for 10 consecutive iterations. When the particle swarm optimization algorithm meets the convergence condition or reaches the maximum number of iterations, it outputs a set of optimal baseline weights.

[0089] Risk sensitivity coefficient and This controls the adjustment range of the corresponding weights when a specific risk increases. The risk sensitivity coefficient can be obtained by constructing a risk gradient scenario in the simulation and performing regression analysis. For example, when simulating a joint task in a digital twin environment, different risk levels can be set... Observation in maintaining Under the condition of no change The value of is used to draw the image. The change curve is fitted to Similarly, by setting different The proportion that should be strengthened in the time objective function is analyzed, thereby fitting the desired result. These coefficients enable the optimization objective to be rebalanced in real time and proportionally with the risk spectrum.

[0090] Constraint relaxation factor Decided to be When the value increases, the corresponding constraint conditions are... The extent to which the upper limit of joint torque (e.g., the maximum joint torque) can be relaxed is directly derived from conservative design principles in safety engineering. The specific steps are as follows: First, determine the absolute physical limit of the j-th constraint through hardware testing. (e.g., 85% of the rated torque); subsequently, defined in Default design margin for the j-th constraint (usually far away) The constraint relaxation factor is based on... Calculations show that even if the risk reaches the preset maximum value, the relaxed constraints will never touch the absolute physical limit. This allows the optimizer to explore more aggressive strategies while ensuring an absolute safety baseline when optimizing the control parameters of the joint robot.

[0091] This embodiment uses a risk-driven strategy to reduce the time that connector tasks are exposed to harsh environments. For example, when At higher levels, The increased weighting of [the component] prompts the development of faster connector strategies; simultaneously, [it is related to] [other factors]. Relevant constraints (For example, the upper limit of joint torque) is dynamically relaxed. Risk-driven strategies include dynamic adjustments. The dynamic relaxation of physical constraint limits is implemented as follows: 1. Dynamic adjustment: The weights are not fixed. Instead, it is based on environmental risks. Real-time adjustment, i.e. Therefore, when When increasing, The weight will be clearly increased. This drives the optimizer to generate faster connector strategies to mitigate risks.

[0092] 2. Dynamic relaxation of physical constraint limits: The upper limit is dynamically relaxed, and the new upper limit is This mechanism is in The increased speed provides the optimizer with the necessary search space, while the hardware limits ensure an absolute safety baseline.

[0093] Finally, the control parameters of the joint robot are optimized using the Deep Deterministic Policy Gradient (DDPG) algorithm, yielding a set of optimal control parameters. Here, the action of the DDPG algorithm is defined as the control parameter to be optimized, and the reward is... .

[0094] S32. Generate an environmental intervention plan, predict the expected loss when implementing the optimal control parameters without interfering with the environment, and predict the expected benefit of implementing the environmental intervention plan.

[0095] Specifically, in this embodiment, a profit analysis is used to determine whether to generate an active environmental intervention instruction. This active intervention aims to reduce the time the rendezvous task is exposed to harsh environments. The specific implementation process is as follows: generating an environmental intervention plan (such as activating a local wind curtain); predicting the expected losses when executing the current optimal control parameters without environmental intervention. , , First, a comprehensive cost estimate for a single joint failure is performed. Second, an environmental disturbance model is used to quickly extrapolate and predict environmental intervention schemes to obtain wind speed data, and the costs after implementing the environmental intervention scheme are calculated. Before and after the implementation of the environmental intervention program Calculate the expected reduction in environmental risk value that the implementation of the environmental intervention plan can achieve. Then, calculate the expected benefits of proactive environmental intervention. , ; Calculate the intervention cost of implementing the environmental intervention plan. The intervention cost is the sum of energy cost, equipment depreciation cost, and opportunity cost. Energy consumption is precisely calculated based on the command power and duration, and then the energy cost is calculated by combining energy prices. Equipment depreciation cost uses the experience amortization value of the equipment's historical use. Opportunity cost is estimated based on experience by assessing the expected losses incurred by other potential beneficiary areas due to the inability to use shared resources (such as air curtains) during the period when they are occupied in the environmental intervention plan.

[0096] The specific implementation of the generated environmental intervention plan is based on a predefined and extensible environmental intervention strategy library, which stores various feasible environmental control measures and their corresponding digital impact models. Taking "activating a local wind curtain" as an example, it is a pre-defined strategy in the environmental intervention strategy library, associated with a lightweight surrogate model. This model describes the quantitative impact of activating a wind curtain at a specific location on environmental parameters such as wind speed, temperature, and humidity within the robot's work area. When the system determines that environmental intervention is necessary, it determines the dominant risk type (by...) based on the currently identified dominant risk type... Based on the high indicated wind speed uncertainty risk, the risk spatial location, and the status of available environmental actuators (such as air curtains, air conditioning vents, and ventilation fans), one or more candidate intervention schemes are matched from the environmental intervention strategy library.

[0097] The lightweight proxy model is a data-driven, computable mathematical model used to quickly predict the distribution changes of key environmental parameters (such as wind speed, ambient temperature, and ambient humidity) within the working domain after the execution of specific environmental intervention commands. Its specific construction and operation are as follows: In the offline phase, for each preset strategy (such as "activate a local wind curtain") in the environmental intervention strategy library, a large number of simulations or data collections are conducted under different initial environmental states and different intervention command parameters (such as wind curtain speed and angle) using high-fidelity CFD simulations or historical measured data. This establishes a mapping dataset from "intervention command parameters" and "current environmental parameters" to "post-intervention environmental parameters." A multilayer perceptron is trained using this dataset as the preset... A dedicated surrogate model for the strategy learns the complex nonlinear relationship between the preset strategy and the flow field response, but only requires simple tensor operations during online inference, achieving millisecond-level prediction speed. During online decision-making, the environmental parameters read by the current sensors and the candidate intervention command parameters are input into this surrogate model, and the predicted results of wind speed, ambient temperature and humidity in the robot's working domain after executing the command can be obtained instantly. Then, combined with the environmental disturbance model, the uncertainty risk of the new environmental disturbance model can be quickly calculated for benefit-cost analysis. This lightweight surrogate model can ensure the real-time performance and engineering practicality of the environmental intervention function. Compared with using high-fidelity simulation directly, it can achieve similar decision-making results while meeting the control timing requirements.

[0098] S33. When the expected benefit is greater than the benefit threshold and greater than the sum of the expected loss and the intervention cost, the active environmental intervention instruction is generated based on the environmental intervention scheme.

[0099] Specifically, in this embodiment, the comparison and Only when And expected returns Only when the profit threshold is exceeded will the environmental intervention plan generate an active environmental intervention command to drive the corresponding execution device, and send it to the corresponding execution device along with the optimal control parameters, so as to realize the "local instantaneous shaping" of the production environment and ensure that the operation has substantial significance, avoiding increasing the complexity of operation for the sake of small profits.

[0100] Furthermore, during the proactive intervention in the environment, wind speed sensors continuously monitor and report actual airflow changes to verify the effectiveness of "local instantaneous shaping" and thereby update the environmental disturbance model, costs, and benefits, forming a closed-loop learning and optimization process for environmental intervention.

[0101] S4. The connector robot receives and executes the optimal control parameters, the environmental controller receives and executes the active environmental intervention command, and monitors the execution result of the connector robot.

[0102] S5. When the execution result deviates from the consensus prediction result, the failure source is locked based on causal inference, and the failure source is calibrated through meta-learning calibration.

[0103] Step S5 includes the following steps: S51. Construct a lightweight causal inference graph for the connector task and perform causal identification to determine the causal variables, confounding variables, and outcome variables of the connector task.

[0104] Specifically, in this embodiment, the complex joint system in the industrial joint scenario is abstracted into a lightweight structural causal model (SCM) based on domain knowledge. The nodes are divided into model prediction nodes (prediction error of mechanism model, prediction error of data-driven model and prediction error of environmental disturbance model), environmental state nodes (ambient temperature, ambient humidity and wind speed, etc.), equipment nodes (joint position, end pose and wear state, etc.), joint state (joint position and wear state), end state of the jointing robot (end pose and wear state), material nodes (yarn position, yarn type and yarn tension, etc.), and result nodes (joint failure and joint success). Directed edges represent the direct causal relationship between variables.

[0105] Furthermore, in SCM, backdoor paths refer to paths from the causal variable to the outcome variable where there is an arrow pointing back to the causal variable. These paths are potential confounding variables that may cause the correlation between the causal and outcome variables to be inconsistent with causation. These backdoor paths can be easily identified by constructing a causal graph. For example, "mechanistic model prediction error" is the causal variable, and "joint failure" is the outcome variable. Simultaneously, all key confounding variables, such as ambient temperature, ambient humidity, yarn batch, and wear condition, can be identified and measured. These factors may simultaneously affect both the predictive accuracy of the mechanistic model and the joint outcome. The backdoor criterion, as a mature method in modern causal inference science, tells us that if this set of confounding variables is controlled, all backdoor paths from the causal variable to the outcome variable can be blocked, thus allowing for an unbiased estimation of the causal effect of the causal variable on the outcome variable using monitoring data.

[0106] S52. When the execution result deviates from the consensus prediction result, calculate the causal effect value of each failure path in the lightweight causal inference graph, and then lock the failure source.

[0107] Specifically, in this embodiment, for the current connector task, when the execution result deviates from the consensus prediction result, it indicates that the connector system may have an abnormal path. At this point, after determining the causal variable, confounding variable, and outcome variable, the effect estimation stage can begin. In any lightweight causal model, the path from the causal variable to the outcome variable could be a failure path. The final failure path can be determined by calculating the causal effect value of each failure path. The most commonly used method for obtaining the causal effect value is estimation based on the backdoor adjustment formula, i.e.:

[0108] Where ACE is the causal effect value of the failure path. This indicates that the causal variable X is artificially intervened to force it to be... , This indicates that the causal variable X is artificially intervened to force it to be... , Indicates forced intervention The expected value of the outcome variable Y. Indicates forced intervention The expected value of the outcome variable Y.

[0109] However, confounding variables may distort the relationship between X and Y, leading to biased results from the aforementioned ACE calculation. Therefore, in practice, a dual machine learning estimation method is often used to address this issue. The specific steps are as follows: First, using a large amount of historical connector data, two backpropagation (BP) neural networks are trained with the mean squared error between predicted and actual values ​​as the loss function. The first BP neural network predicts the expected value of the outcome variable given the confounding variables; the second BP neural network predicts the probability (propensity score) of the cause variable being in different states / values ​​given the confounding variables. Then, based on the predictions from the two BP neural networks, an unbiased estimator under the mean squared error (such as an enhanced inverse probability weighted estimator) is constructed as the unbiased ACE for the failure path, and the confidence interval of the unbiased ACE is obtained through bootstrap. The unbiased ACE specifically satisfies the following relationship:

[0110] in, An unbiased ACE; For the first The observed values ​​of the causal variables for each sample (connection task) are typically 1 or 0, representing a certain state or above a threshold. Let be the observed value of the outcome variable for the i-th sample; Let i be the vector of all measured confounding variables for the i-th sample; ( ) represents the predictions obtained by the second BP neural network, given the confounding variables. Cause variables The probability estimate of =1; ( ) and ( ) represent the predictions obtained by the first BP neural network under a given confounding variable. And assuming the causal variable is intervened at a value of 1 or 0, the outcome variable... The expected value estimate is given by N, where N is the total number of samples used for estimation. This estimator is doubly robust, providing an unbiased or consistent estimate of causal effects as long as either the propensity score model or the outcome prediction model is correctly specified.

[0111] Furthermore, the unbiased ACE of each failure path is calculated. If the lower bound of the 95% confidence interval of the unbiased ACE of a failure path is greater than 0, and the unbiased ACE is greater than a preset significance threshold, then this failure path is considered a candidate failure path. The causal variables among the three candidate failure paths with the largest unbiased ACE are taken as the failure sources. This embodiment achieves interpretable and quantitative root cause analysis of joint system failures, surpassing traditional joint failure diagnosis methods based on correlation or human experience.

[0112] The significance threshold is typically set based on standard statistical testing criteria and domain experience, with a typical value of 0.05. This threshold means that when the calculated unbiased ACE is greater than 0.05, the causal variable is considered to have a significant, non-negligible causal influence on the outcome variable (joint failure). This value draws on the significance level (α=0.05) commonly used in statistical hypothesis testing, while also considering the effect size requirements in engineering practice. It aims to exclude weak causal paths that, although statistically significant (lower bound of the confidence interval greater than 0), have excessively small effect sizes and lack engineering intervention value, thereby ensuring that the identified failure sources have clear engineering diagnostic and calibration significance.

[0113] S53. For the failure source, multiple reference historical calibration schemes are obtained from the dynamic knowledge graph of historical joint tasks through similarity analysis, and then a weighted task distribution is generated.

[0114] Specifically, step S53 includes the following steps: S531. Construct the feature vector of the current joint task, and extract the feature vector of the historical joint task from the dynamic knowledge graph of the historical joint task. The feature vector includes the working condition feature vector, the causal diagnosis feature vector, and the multidimensional risk spectrum vector.

[0115] Specifically, in this embodiment, a feature vector for the current splicing task is constructed, and feature vectors for historical splicing tasks are extracted from the dynamic knowledge graph of historical splicing tasks. The operating condition feature vector includes yarn parameters, environmental indicators, equipment status, and task objectives; the causal diagnosis feature vector includes all causal variables, with only the failure source being non-zero; the multidimensional risk spectrum vector has been explained in step S27.

[0116] More specifically, the nodes of the dynamic knowledge graph comprise four main categories of entities, each with a structured attribute vector. The first category is the task node, the core of the dynamic knowledge graph. Its attribute, the "working condition feature vector," is a multi-dimensional structured vector, specifically including: yarn parameters (such as yarn count, twist, material type, measured moisture regain, etc.), environmental indicators (such as ambient temperature, humidity, and wind speed, etc.), equipment status (such as joint clearance compensation value, servo motor temperature, cumulative operating time of key components, etc.), and task objectives (preset success rate, joint quality, time, etc.). The second category is the model node, recording the performance of each specific model in the multi-model digital twin in the jointing task. Its attributes include its input data, prediction results, confidence level, and prediction error. The third category is the risk and decision node, whose attributes include the multi-dimensional risk spectrum vector and optimal control parameters of the jointing task. The fourth category is the causal and calibration node, used to record the conclusions of root cause analysis (final failure path, failure source) and the implemented calibration scheme.

[0117] The edge relationships in a dynamic knowledge graph define the specific, quantifiable interactions and logical connections between nodes. There are four main types of edges. The first type is causal edges, used to connect risk and decision nodes with task result nodes. The edge weight stores the calculated unbiased ACE (Accuracy, Aspect, and Principle), with positive values ​​indicating facilitation and negative values ​​indicating inhibition, and the magnitude representing the strength of facilitation or inhibition. The second type is temporal edges, connecting consecutively occurring task nodes to form a timeline of system operation, used to analyze trends and pattern evolution. The third type is generative edges, used to connect task nodes with causal and calibration nodes, indicating what optimization strategies were generated under specific conditions. Edge attributes can include the applicability score of the strategy. The fourth type is similarity edges, used to connect task nodes with similar condition feature vectors or multidimensional risk spectrum vectors. The edge weight is obtained by calculating the Mahalanobis distance or cosine similarity between the vectors of the two nodes and then normalizing it. This is one of the direct bases for obtaining reference historical calibration schemes through experience retrieval.

[0118] The task result node, as described, is not an independent node category in the complete node classification system of the dynamic knowledge graph. Instead, it actually refers to the set of core attributes (such as Boolean results for success / failure, time, etc.) within the task node that characterizes the final output state of the rendezvous task. Therefore, the accurate description of the first type of causal edge connection is: from the risk and decision nodes representing risk states and decision information, pointing to the task node carrying the final output record of the task (specifically, its embedded result attributes, such as the rendezvous result), the unbiased ACE stored in the edge weight quantifies the impact of specific risks or decision factors on the task result. This statement clarifies the logical correspondence between nodes and edges, ensuring the accuracy and consistency of information association in the dynamic knowledge graph.

[0119] The dynamic update and evolution mechanism of the dynamic knowledge graph relies on an automated knowledge extraction and association engine. After each connection task, the multi-source data is formatted according to the above node and edge relationship template, and then associated with existing nodes in the graph or create new nodes through an entity alignment algorithm.

[0120] S532. Calculate the similarity of working condition features, causal diagnosis features, and risk based on the feature vectors of the current joint task and the historical joint task.

[0121] Specifically, in this embodiment, condition feature vectors are extracted from the dynamic knowledge graph of all historical splicing tasks, and the global covariance matrix of each condition feature dimension is calculated to standardize different dimensions. Simultaneously, weights are manually assigned to different types of data based on domain knowledge (for example, the material type of yarn may have a more critical impact on the splicing result than ambient temperature, thus receiving a higher weight), thereby obtaining the diagonal weight matrix of each type of data in the condition feature vector. Then, the similarity of condition features between the current splicing task and historical splicing tasks is calculated using the following formula:

[0122] in, For similarity of working conditions, This represents the feature vector of the current joint task's operating conditions. This represents the feature vector of historical rendezvous tasks. It is a diagonal weight matrix. This is the global covariance matrix. The larger the value, the closer the current jointing task is to the working conditions of historical jointing tasks.

[0123] The "root cause analysis conclusions" stored in the dynamic knowledge graph are encoded as structured causal fragments, such as mechanistic model error ← high humidity → joint failure. Furthermore, graph neural networks are used to obtain the causal diagnostic feature similarity between the current preliminary causal pattern and historically stored causal patterns. A historical node that occurs under the same risk spectrum and is caused by the same root cause (e.g., joint failure due to a mechanistic model in a humid environment) will obtain extremely high [status / value]. .calculate The specific implementation is as follows: The currently preliminarily diagnosed causal pattern (a directed graph consisting of the entire path from the failure source to the joint result) and the historical causal patterns stored in the dynamic knowledge graph are both abstracted into attribute graphs with node type and edge relationship weights; a pre-trained graph encoder (GraphSAGE) is used to map each causal graph into a low-dimensional graph embedding vector, and then the cosine similarity between the current graph embedding vector and the graph embedding vectors of the historical causal patterns is calculated to obtain... .

[0124] Calculate the cosine similarity between the multidimensional risk spectrum vector of the current rendezvous task and the multidimensional risk spectrum vectors of historical rendezvous tasks. This is based on the risk similarity between the current rendezvous mission and historical rendezvous missions. The larger the value, the more similar the risk patterns of the current rendezvous mission are to those of historical rendezvous missions.

[0125] S533. Simulate the historical calibration scheme of the historical connector task in the current digital twin environment and obtain the corresponding optimized objective function value, which is denoted as the migration value.

[0126] Specifically, in this embodiment, even if the operating conditions and risk modes are similar, the historical calibration scheme cannot be migrated if the execution conditions required by the historical calibration scheme (such as the use of a specific type of end effector) do not match the current conditions. Therefore, it is necessary to assess the potential feasibility of migrating the historical calibration scheme to the current connector task.

[0127] This embodiment simulates the historical calibration scheme of the historical node in the digital twin environment of the current joint task, calculates the optimization objective function value when using the historical calibration scheme, and records it as the migration value for subsequent evaluation of the potential feasibility of migrating the historical calibration scheme to the current joint task.

[0128] S534. Calculate the difference between the optimized objective function value of the current connector task and the migration value, and use the reciprocal of the difference as the transferability score of the corresponding historical calibration scheme.

[0129] Specifically, in this embodiment, the difference between the optimized objective function value and the migration value of the current connector task is calculated as the transferability score of the corresponding historical calibration scheme. . The larger the value, the greater the portability of the historical calibration scheme.

[0130] S535. The weighted sum of the working condition feature similarity, the causal diagnosis feature similarity, the risk similarity, and the transferability score is used as the comprehensive similarity between the current junction task and the historical junction task.

[0131] Specifically, in this embodiment, the comprehensive similarity between the current rendezvous task and historical rendezvous tasks is... The following relationship must be satisfied:

[0132] in, , , and These are weights, and their values ​​can be dynamically adjusted based on historical migration success rates.

[0133] S536. Set a similarity threshold, and use the historical calibration scheme of historical joint tasks whose overall similarity with the current joint task is greater than the similarity threshold as a reference historical calibration scheme.

[0134] Specifically, in this embodiment, the similarity threshold is typically set empirically between 0.6 and 0.75, and its specific value can be determined by analyzing the comprehensive similarity in historical tasks. The threshold is determined by the relationship between the distribution of the reference scheme and the success rate of the corresponding calibration scheme migration. For example, it can be initially set to 0.7, and then, based on feedback from actual applications, if too many historical schemes are recalled but their applicability is low, the threshold is increased to enhance targeting; if too few effective historical schemes are found, the threshold is appropriately reduced to expand the search scope. The threshold can also be designed to be adaptive, that is, dynamically adjusted according to the density and quality of available historical tasks in the current knowledge graph, so as to ensure the relevance of the reference scheme while taking into account the system's cold start capability and experience reuse coverage.

[0135] Furthermore, in some alternative embodiments, it can be based on Perform multi-hop similarity retrieval, that is: first find those directly similar to the current header task ( Historical linker tasks (above a similarity threshold) are identified. If no such task is found, neighboring nodes are explored one or two hops away along the edges of the dynamic knowledge graph (especially "similarity edges" and "causal edges"), and the combined similarity between the historical linker tasks of these neighboring nodes and the current linker task is calculated. This approach enables the discovery of historical experiences that are not immediately apparent in terms of direct features but are highly relevant in terms of internal logic and solutions, achieving experience reuse and cold start optimization.

[0136] S537. For the reference historical calibration scheme, the comprehensive similarity is normalized to obtain the scheme weight of the reference historical calibration scheme, and then a weighted task distribution is generated.

[0137] Specifically, in this embodiment, the comprehensive similarity of the reference historical calibration schemes is normalized using the maximum-minimum normalization method to obtain the scheme weights of the reference historical calibration schemes. In other words, all reference historical calibration schemes are treated as a sample set, and the scheme weight is the "sampling probability" or "importance" of each sample (reference historical calibration scheme), representing the relative probability of selecting each reference historical calibration scheme for reference in subsequent optimization processes.

[0138] S54. Initialize the inner loop optimization process of the model-independent meta-learning algorithm using the weighted task distribution.

[0139] Specifically, in this embodiment, during the initialization of the model-independent meta-learning algorithm's inner loop, a task is sampled from the weighted task distribution, and its data is used to fine-tune the model. Through weighted sampling, the model pays more attention to the reference historical calibration scheme with high weights during the meta-learning process, enabling the algorithm's initial parameters to preferentially absorb the experience most relevant to the current new task, laying the foundation for rapid adaptation to new tasks.

[0140] S55. The model-independent meta-learning algorithm performs gradient iteration on the reference historical calibration scheme, enabling the failure source to quickly adapt to a state sensitive to new tasks. Using data from multiple new connector tasks, the model-independent meta-learning algorithm performs multi-step gradient updates on the reference historical calibration scheme to achieve calibration of the failure source.

[0141] Specifically, in this embodiment, the gradient update step size strategy is a key design in the fast calibration framework of the model-independent meta-learning algorithm. It is not a single fixed value, but a multi-stage, conditional, adaptive dynamic process. Its core objective is to boldly explore in the initial stage to adapt to new tasks, and to carefully fine-tune to meet physical constraints when approaching convergence.

[0142] Specifically, the learning rates for the inner and outer loops are designed separately. The inner loop phase employs a dynamic learning rate strategy. In the initial few steps (e.g., the first three steps), a higher learning rate (e.g., 0.01) is used to boldly explore directions, allowing the failure sources to quickly adapt to a state sensitive to new tasks. Subsequently, a mini-validation set is constructed using data from multiple new connector tasks (the data type must be consistent with the data type of the reference historical calibration scheme). The validation algorithm loss is monitored in real time. If the loss continues to decrease, a high learning rate is maintained to accelerate convergence. If fluctuations occur or the physical consistency check fails (e.g., the yarn tension predicted by the model exceeds the reasonable range derived from the energy conservation principle in the mechanistic model), the learning rate is immediately switched to a low learning rate (e.g., 0.001) to enter the fine-tuning stage, avoiding deviation from the reasonable parameter space due to excessively large step sizes.

[0143] Compared to the inner loop, the outer loop uses a more conservative and fixed learning rate. It is set to a small value (e.g., 0.001) to ensure that the "experience" learned from new connector tasks does not overwrite or destroy the general, robust feature representations learned from a vast pool of alternative historical calibration schemes when updating parameters that need calibration (from failure sources). This "small, gradual" outer loop update ensures the stability of the meta-knowledge base.

[0144] Furthermore, the entire gradient update process is subject to real-time supervision by physical constraints. After each inner loop update, a rapid forward physical consistency check is performed. Taking the data-driven model and yarn tension as an example, if the prediction error of the data-driven model is the source of failure, and the calibration scheme involves adjusting the model parameters of the data-driven model, the predicted value of yarn tension during the calibration process is compared with the reasonable range of yarn tension. Once the predicted value continuously exceeds its reasonable physical range, the calibration path is rejected, and a step-size backtracking and reset mechanism is triggered: that is, it reverts to the model parameters of the previous step and significantly reduces the inner loop learning rate to retry in a more conservative manner. This strategy uses physical prior knowledge as a "guardrail" for the calibration process, ensuring the effectiveness of the calibration while strictly remaining within the physically reasonable solution space.

[0145] S6. The context information of the entire process from forward prediction to calibration of the failure source is formed into transferable structured knowledge and stored in a dynamic knowledge graph.

[0146] Specifically, in this embodiment, the context information of the entire process from forward prediction to calibration of the failure source, namely the information collected and generated in steps S1 to S5, is organized into transferable structured knowledge and stored in a dynamic knowledge graph. The dynamic knowledge graph has been described in detail in step S531.

[0147] It should be noted that in some cases, the actions described in the specification can be performed in different orders and still achieve the desired results. In this embodiment, the order of steps is given only to make the embodiment clearer and easier to explain, and not to limit it.

[0148] In one optional embodiment, please refer to Figure 2 To improve the practicality of this method and facilitate its promotion, the present invention also provides a digital twin-driven adaptive optimization system for the parameters of a joint robot. The digital twin-driven adaptive optimization system for the parameters of a joint robot includes: a data acquisition device 1, a data output device 2, a processor 3, and a storage device 4. The storage device 4 includes a computer-readable storage medium storing a computer program. The computer program includes program instructions, which, when executed by the processor 3, cause the processor 3 to implement the contents described in steps S1 to S6.

[0149] In summary, the present invention has at least the following beneficial effects: 1. This method achieves real-time interaction between the physical entity and the virtual model by constructing a multi-model digital twin, and performs forward prediction for the jointing task. The prediction results of the multi-model digital twin are not directly used to optimize the control parameters of the jointing robot. Instead, they are fused based on the dynamic confidence levels of the individual model predictions and evidence theory to obtain a consensus prediction result and a multi-dimensional risk spectrum vector. The obtained consensus prediction result remains stable and reasonable even when the prediction results of different models differ significantly, achieving a leap from simple voting to credibility-based debate and consensus formation. The obtained multi-dimensional risk spectrum vector quantifies the uncertainty of the prediction results of different models, providing fine-grained guidance for downstream optimization of the jointing robot's control parameters. This is beneficial for improving the adaptability of the jointing robot, thereby improving the stability and quality of the jointing process.

[0150] 2. This method generates a set of optimal control parameters based on the objective function of the control parameter optimization algorithm dynamically reconstructed by the multidimensional risk spectrum vector. Based on the benefit analysis, it generates active environmental intervention instructions when necessary, realizing the "local instantaneous shaping" of the production environment. This enables the jointing robot to learn and optimize autonomously, further improving the adaptability of the jointing robot and thus improving the stability and quality of the joint.

[0151] 3. After performing the joint operation, this method locks the failure source based on causal inference and calibrates the failure source through meta-learning calibration. Finally, the entire process forms transferable structured knowledge and stores it in a dynamic knowledge graph, realizing joint experience reuse and cold start optimization, and realizing autonomous learning optimization based on historical data.

[0152] 4. A system adapted to the method is provided, which not only improves the practicality of the method but also facilitates its promotion.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for adaptive optimization of parameters of a digital twin-driven joint robot, characterized in that, Includes the following steps: Construct a multi-model digital twin of the jointing robot and perform forward prediction on the jointing task to obtain the prediction results of each model and calculate the confidence level of the corresponding prediction. Based on the confidence level, the prediction results are fused using evidence theory to obtain a consensus prediction result, and a multidimensional risk spectrum vector is generated simultaneously. The objective function is dynamically adjusted and optimized using the multidimensional risk spectrum vector, thereby obtaining the optimal control parameters of the joint robot under constraints, and determining whether to generate an active environmental intervention command through benefit analysis. The connector robot receives and executes the optimal control parameters, and the environmental controller receives and executes the active environmental intervention command, while monitoring the execution results of the connector robot. When the execution result deviates from the consensus prediction result, the failure source is identified based on causal inference, and the failure source is calibrated through meta-learning calibration. The contextual information from the entire process from forward prediction to calibration of the failure source will be formed into transferable structured knowledge and stored in a dynamic knowledge graph.

2. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 1, characterized in that: The multi-model digital twin includes a mechanistic model, a data-driven model, and an environmental disturbance model. The mechanistic model comprises the robot's rigid-flexible coupling dynamics equations and a yarn nonlinear dynamics model based on Cosserat bar theory. The data-driven model is constructed based on a spatiotemporal graph convolutional network, with nodes including robot joints, robot end effectors, discrete yarn points, and environmental sensors. Node features include joint position, joint angle, joint torque, joint velocity, end effector pose, six-dimensional end effector force, three-dimensional yarn position, yarn tension, wind speed, ambient temperature, and ambient humidity. Edges include fixed physical connections and preset functional coupling connections. The data-driven model takes historical node state sequences as input and future node states and jointing results as outputs. The environmental disturbance model is constructed based on a deep neural network, with inputs being the temperature, humidity, and wind speed at different monitoring points within the workshop, and outputting the wind speed vector within the jointing robot's working domain. The environmental disturbance model provides environmental disturbance inputs for the mechanistic model and the data-driven model.

3. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 2, characterized in that, The process of constructing a multi-model digital twin of the jointing robot and performing forward prediction on the jointing task to obtain the prediction results of each model and calculate the confidence level of the corresponding prediction includes the following steps: A multi-model digital twin of the jointing robot is constructed, and then the multi-model digital twin is used to perform forward prediction of the jointing task to obtain the prediction results of each model. Confidence evaluation functions for the aforementioned mechanism model, data-driven model, and environmental disturbance model are established respectively, and the confidence levels of the corresponding prediction results are calculated accordingly.

4. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 3, characterized in that, The step of fusing the prediction results based on the confidence level using evidence theory to obtain a consensus prediction result and simultaneously generating a multi-dimensional risk spectrum vector includes the following steps: The prediction results and their confidence levels of the mechanistic model and the data-driven model are transformed into basic probability assignments in the DS evidence theory and combined, while the total conflict quality of all conflict combinations is calculated. Based on the confidence level of the prediction results corresponding to the propositions in the conflict combinations, calculate the combination confidence level of each conflict combination and the total confidence level of all conflict combinations, wherein the propositions include joint failure and joint success; The total conflict quality is allocated to each of the conflict combinations according to the proportion of the combined confidence level to the total confidence level, thereby obtaining the redistributed conflict quality of each of the conflict combinations. Based on the confidence ratio of each proposition in the conflict combination, the redistribution conflict quality is assigned to the propositions in the conflict combination to obtain the redistribution quality of each proposition in the conflict combination. For each of the propositions, the sum of the redistributed qualities is calculated, and the result is added to the non-conflict quality of the proposition to obtain its basic probability allocation value; Based on the basic probability allocation values, the Pignistic probability of the corresponding proposition is calculated, and then the consensus prediction result is obtained. The risk of physical model mismatch, the risk of data model extrapolation, and the risk of uncertainty in the environmental disturbance model are calculated to form the multidimensional risk spectrum vector.

5. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 4, characterized in that, The multidimensional risk spectrum vector satisfies the following relationship: ， ， ， ， in, Let the multidimensional risk spectrum vector be... To account for the risk of mismatch in the physical model, The risk of extrapolating the data model, This refers to the uncertainty risk of the environmental disturbance model. Assign a value to the basic probability of failure of the connector. The confidence level of the prediction result of the aforementioned mechanism model. The confidence level of the prediction results of the data-driven model. The confidence level of the prediction results of the environmental disturbance model is given. The wind speed vector is denoted as .

6. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 4, characterized in that, The process of dynamically adjusting and optimizing the objective function using the multidimensional risk spectrum vector to obtain the optimal control parameters of the joint robot under constraints, and determining whether to generate an active environmental intervention command through benefit analysis, includes the following steps: The multidimensional risk spectrum vector is used to dynamically adjust and optimize the objective function, and under constraints, the optimal control parameters are obtained by using a parameter optimization algorithm with the aim of minimizing the objective function. Generate an environmental intervention plan, predict the expected loss when implementing the optimal control parameters without interfering with the environment, and predict the expected benefit of implementing the environmental intervention plan. When the expected benefit is greater than the benefit threshold and also greater than the sum of the expected loss and the intervention cost, the active environmental intervention instruction is generated based on the environmental intervention scheme.

7. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 6, characterized in that, The objective function and the constraints satisfy the following relationship: ， ， ， ， in, To optimize the objective function value, Weighted by success rate, As the benchmark weight for joint quality, As time weight, As environmental benchmark weights, Regarding control parameters The success rate objective function For about The quality objective function, For about The time objective function, For about The environmental objective function, As the benchmark weight for success rate, The risk of extrapolating the data model, As a time-based weight, and For risk sensitivity coefficient, This refers to the uncertainty risk of the environmental disturbance model. For the j-th constrained control parameter, This represents the default design margin for the j-th constrained control parameter. Let j be the constraint relaxation factor for the j-th constrained control parameter. This refers to the risk of mismatch in the physical model.

8. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 1, characterized in that, When the execution result deviates from the consensus prediction result, the failure source is identified based on causal inference, and the failure source is calibrated through meta-learning calibration, including the following steps: Construct a lightweight causal inference graph for the connector task and perform causal identification to determine the causal variables, confounding variables, and outcome variables of the connector task; When the execution result deviates from the consensus prediction result, the causal effect value of each failure path in the lightweight causal inference graph is calculated to pinpoint the failure source. For the aforementioned failure source, multiple reference historical calibration schemes are obtained from the dynamic knowledge graph of historical joint tasks through similarity analysis, thereby generating a weighted task distribution; The inner loop optimization process of the model-independent meta-learning algorithm is initialized using the weighted task distribution; The model-independent meta-learning algorithm performs gradient iteration on the reference historical calibration scheme, enabling the failure source to quickly adapt to a state sensitive to new tasks. It also uses data from multiple new connector tasks to enable the model-independent meta-learning algorithm to perform multi-step gradient updates on the reference historical calibration scheme, thereby achieving calibration of the failure source.

9. The adaptive optimization method for parameters of a digital twin-driven joint robot according to claim 8, characterized in that, For the failure source, multiple reference historical calibration schemes are obtained from the dynamic knowledge graph of historical joint tasks through similarity analysis, and then a weighted task distribution is generated, including the following steps: Construct the feature vector of the current rendezvous task, and extract the feature vector of the historical rendezvous task from the dynamic knowledge graph of historical rendezvous tasks. The feature vector includes the working condition feature vector, the causal diagnosis feature vector, and the multidimensional risk spectrum vector. Based on the feature vectors, calculate the similarity of working condition features, causal diagnosis features, and risk between the current joint task and the historical joint tasks; The historical calibration scheme of the historical connector task is simulated in the current digital twin environment, and the corresponding optimized objective function value is obtained and recorded as the migration value. Calculate the difference between the optimization objective function value of the current connector task and the migration value, and use the reciprocal of the difference as the transferability score of the corresponding historical calibration scheme; The weighted sum of the operating condition feature similarity, the causal diagnosis feature similarity, the risk similarity, and the transferability score is used as the comprehensive similarity between the current rendezvous task and the historical rendezvous task. Set a similarity threshold, and use the historical calibration schemes of historical joint tasks whose overall similarity with the current joint task is greater than the similarity threshold as reference historical calibration schemes; For the reference historical calibration scheme, the comprehensive similarity is normalized to obtain the scheme weight of the reference historical calibration scheme, and then a weighted task distribution is generated.

10. A digital twin-driven adaptive optimization system for the parameters of a joint robot, characterized in that, The digital twin-driven adaptive optimization system for joint robot parameters includes: a data acquisition device, a data output device, a processor, and a storage device. The storage device includes a computer-readable storage medium storing a computer program. The computer program includes program instructions, which, when executed by the processor, cause the processor to implement the adaptive optimization method for joint robot parameters as described in any one of claims 1-9.

Images