Screwdriver torque optimization control method based on neural network

Abstract

The invention discloses a screwdriver torque optimization control method based on a neural network. The method comprises the following steps: constructing a contact state diagram of a whole tightening process; node and edge features are extracted by using an isotropic graph neural network, and the clamping force variation is predicted; calculating an energy consistency condition based on a screw tightening physical model and updating edge features; inputting the updated state into the dimension raising dynamic model, and predicting a future state sequence; constructing a model prediction control optimization problem in combination with existing multi-dimensional constraint conditions, and generating an optimal torque instruction sequence; and judging a slip risk based on online gating, and executing a torque reduction and maintenance strategy. According to the method, the physical constraint and the deep learning model are combined, and high precision and high stability of torque control are achieved.

Description

Technical Field

[0001] This invention relates to the field of industrial assembly automation, and in particular to a method for optimizing the torque control of screwdrivers based on neural networks. Background Technology

[0002] During screw tightening, torque control directly affects connection quality and product reliability. Existing torque control methods mostly rely on a single sensor signal for judgment, typically using preset torque thresholds or strategies based on simple torque-angle relationships. However, because the tightening process involves various dynamic factors, including changes in clamping force, material deformation, and thread friction, a single signal feedback cannot accurately reflect the actual clamping state, easily leading to problems such as insufficient connection or over-tightening damage.

[0003] While some more advanced solutions incorporate multimodal signal acquisition and model-based control algorithms, they generally suffer from three main shortcomings: First, they lack effective modeling of the inherent physical relationships between multimodal features, resulting in redundancy or distortion even after signal fusion. Second, the prediction and control processes fail to effectively incorporate dynamic constraints on key indicators such as clamping force and slippage risk, making them prone to control command fluctuations under complex operating conditions. Third, they do not adequately consider physical consistency conditions such as energy conservation and state reversibility during the tightening process, leading to risks of accuracy drift and stability degradation in the control strategy over long-term operation.

[0004] Furthermore, existing tightening methods based on model predictive control mostly rely on fixed-parameter models, which have weak adaptability to state evolution and sudden anomalies during the process. In the execution phase, they have limited support for real-time response and safety degradation mechanisms for risks such as slippage and overload, making it difficult to maintain control accuracy while ensuring safety. These technical shortcomings are particularly prominent in high-precision assembly, automated production, and variable load, multi-batch operation scenarios, limiting the reliability and consistency of tightening operations.

[0005] Therefore, how to provide a neural network-based screwdriver torque optimization control method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a screwdriver torque optimization control method based on neural networks. This method integrates multimodal feature extraction, isotropic graph neural networks, and predictive modeling techniques based on the Koopman operator to achieve high-precision estimation and multi-time-slice prediction of clamping force and slippage probability. Under the model predictive control framework, it generates optimal torque control commands that satisfy multidimensional constraints such as torque change rate, clamping force range, and slippage risk threshold, and drives the screwdriver actuator to perform real-time torque adjustment. This method has the advantages of high control accuracy, good execution stability, and strong safety.

[0007] A screwdriver torque optimization control method based on a neural network according to an embodiment of the present invention includes the following steps:

[0008] Using sensors in a screwdriver, including a motor current sensor, angular velocity sensor, acoustic sensor, and vibration sensor, real-time multimodal data of the screwdriver during the tightening process is collected. The real-time multimodal data is divided into multiple time slices according to a fixed time window. Multimodal feature extraction is performed on each time slice to generate multimodal feature extraction results. The extracted features include the statistics of motor current, the instantaneous change of angular velocity, the frequency domain features of acoustic signals, and the time domain envelope features of vibration signals. The multimodal feature extraction results are used as node features to construct a contact state diagram. The nodes of the contact state diagram represent the node features within a single time slice, and the edges represent the temporal correlation and physical contact relationship between nodes.

[0009] For each edge in the contact state diagram, edge features are generated. The edge features include the time interval between nodes, the angle change of the screwdriver, and the change in the friction coefficient. The change in the friction coefficient is calculated based on the physical friction model. Physical consistency constraints are introduced during the edge feature update process to ensure that the edge features satisfy the residual minimization condition of the friction model and the relationship equation between torque and clamping force in adjacent time slices.

[0010] The contact state diagram is input into the equivariant graph neural network model. The graph structure is updated using node features and edge features containing physical consistency constraints. The clamping force estimate and slippage probability are output. The clamping force estimate and slippage probability are combined into a contact feature vector. The contact feature vector is concatenated with the multimodal feature extraction results to form an extended feature sequence.

[0011] The extended feature sequence is input into a neural network model based on the Koopman operator. An upgraded state vector is generated through upgraded mapping. The feature channels corresponding to the clamping force estimate and the slip probability are explicitly retained in the upgraded state vector. The upgraded state sequence of multiple future time slices is predicted using the upgraded dynamics model, which includes a trainable parameter matrix and input torque control commands.

[0012] The higher-dimensional state sequence of multiple future time slices is input into the model prediction and control module. Based on the predicted clamping force, slippage probability, and target clamping force setpoint, the optimal torque control command that satisfies the torque change rate limit, clamping force range limit, and slippage risk threshold limit is calculated.

[0013] The optimal torque control command is sent to the screwdriver actuator, which drives the actuator to adjust the torque output in real time during the tightening process, thus completing the screwdriver torque optimization control process.

[0014] Optionally, the construction of the contact state diagram includes:

[0015] The motor current sensor, angular velocity sensor, acoustic sensor, and vibration sensor of the screwdriver are used to continuously collect real-time multimodal data of the screwdriver throughout the tightening process.

[0016] The collected real-time multimodal data is segmented into fixed time windows, and the entire tightening process is divided into multiple consecutive time slices, each containing a fixed number of sampling points.

[0017] Multimodal feature extraction is performed within each time slice: statistical calculations are performed on the motor current signal to extract four types of statistical features: mean, variance, skewness, and kurtosis; instantaneous changes in angular velocity signals are calculated to obtain the average, maximum, and minimum values ​​of the angular velocity differences between adjacent sampling points; frequency domain analysis is performed on the acoustic signal to extract four frequency domain features: dominant frequency, spectral centroid, bandwidth, and energy distribution of characteristic frequency bands; and time-domain envelope features are extracted from the vibration signal, including envelope mean, envelope variance, and envelope energy.

[0018] Multimodal features are concatenated at the time slice level to generate multimodal feature extraction results, and these results are used as node features.

[0019] The temporal relationships are generated based on the chronological order between time slices, and the physical contact relationships are generated based on the physical contact events of the screwdriver during the tightening process. The temporal relationships and physical contact relationships are used as edge information in the contact state graph. The contact state graph is constructed by combining node features, where nodes represent node features within a single time slice, and edges represent the temporal relationships and physical contact relationships between nodes.

[0020] Optionally, the generation of the edge features includes:

[0021] Initial edge features are generated for each edge in the contact state diagram. The initial edge features include the time interval between nodes, the angle change of the screwdriver, the change of angular velocity, and the change of friction coefficient. The change of friction coefficient is calculated based on the physical friction model, which adopts the Stribeck friction model and the Coulomb-viscosity combined model. The calculation process is as follows: in each time slice, the clamping force estimate is used as the normal force input, and the relative sliding velocity derived from the angular velocity and angle change is used as the velocity input. First, the friction coefficient of the transition from the static friction zone to the dynamic friction zone is given according to the Stribeck model. Then, the viscous friction component that increases linearly with the velocity is superimposed to obtain the friction coefficient value of the time slice. The difference of the friction coefficient values ​​of adjacent time slices is used as the change of friction coefficient and written into the initial edge features.

[0022] The initial edge features and the node features of the contact state graph are input into the equivariant graph neural network model to perform the graph structure update process of the edge features. During the update process, the features of time interval, angle change, angular velocity change and friction coefficient change are retained.

[0023] The updated edge feature vectors are used to generate the predicted clamping force change. The specific process is as follows: the updated edge feature vectors are concatenated into an input vector in a fixed order, which is the time interval, angle change, angular velocity change and friction coefficient change. This input vector is then input into the edge-level regression module of the equivariant graph neural network model. The edge-level regression module consists of multi-layer fully connected operations and nonlinear activation and shares the previous layer parameters with the graph structure update. It outputs a single scalar as the predicted clamping force change. During the training phase, the parameters are updated using a weighted loss based on the physical consistency residual. During the inference phase, the predicted clamping force change is directly output for use in calculating the physical consistency residual and forming the contact feature vector.

[0024] The residuals of the equations relating the changes in friction coefficient, angular velocity, angle, and clamping force are used as loss function terms for minimization training. The weighted loss of the physical consistency residuals is defined as:

[0025]

[0026] Among them, L phys The sum of the weighted loss function for physical consistency residuals represents the total value used to constrain the consistency between the model's predictions and the physical model. e represents the edge index in the contact state graph, and Δω... e Δθ represents the change in angular velocity between nodes connected by an edge. e This represents the change in angle between nodes connected by the edge. α and β represent non-negative weight hyperparameters used to adjust the weights of the changes in angular velocity and angle on the loss. ΔF represents the change in clamping force predicted by the isotropic graph neural network model. e This indicates the actual change in clamping force on the corresponding side in the training data. This represents the squared difference between the predicted change in clamping force and the actual change in clamping force, and the magnitude of the prediction error is represented by the squared Euclidean distance.

[0027] By optimizing the parameters of the isovariant graph neural network model based on minimizing the weighted loss of the physical consistency residual, edge feature representations that satisfy the physical property constraints of screwdrivers are obtained.

[0028] Optionally, the processing of the isomorphic graph neural network model includes:

[0029] The node features of the contact state graph and the edge features containing physical consistency constraints are input into the equivariant graph neural network model. The equivariant graph neural network model includes a node state update module and an edge feature update module. The node state update module updates the node state based on the node features and the message vector corrected by the edge features. The edge feature update module updates the edge features based on the states of adjacent nodes and the physical properties of the connecting edges.

[0030] Hamiltonian structural constraints and symplectic geometric constraints are introduced into the edge feature update module. Hamiltonian structural constraints are used to maintain the energy conservation property of the system during the iteration process, while symplectic geometric constraints are used to maintain the reversibility and volume preservation property of the state evolution in the phase space. The edge features are updated by constraining the energy consistency condition of torque work and clamping force accumulation based on the existing screw tightening physical model to avoid energy mutation signals before and after slippage.

[0031] During message passing, the node state, edge features, and their physical consistency constraints are input into the node state update module. Through multiple rounds of iterative passing and fusion, an updated node feature representation is generated.

[0032] After the graph neural network model completes the graph structure update, it calculates the clamping force estimate and slippage probability based on the updated node feature representation and edge feature vector. When calculating the clamping force estimate, all node feature representations are globally pooled to obtain the global feature representation of the contact state graph. The global feature representation is then input into the regression output layer to generate the clamping force numerical prediction. When calculating the slippage probability, the global feature representation is input into the binary classification probability output layer and a slippage probability value ranging from 0 to 1 is generated through the Sigmoid function. The clamping force estimate and slippage probability are supervised learning with real clamping force data and slippage event labels during the training process, respectively. The former uses the mean squared error loss function, and the latter uses the binary cross-entropy loss function. Finally, the clamping force estimate and slippage probability are concatenated to form the contact feature vector.

[0033] The contact feature vector and the multimodal feature extraction results are concatenated at the time slice level to generate an extended feature sequence, which serves as the input for subsequent dimensionality-upgrading dynamics modeling.

[0034] Optionally, the prediction of the upgraded state sequence for multiple future time slices includes:

[0035] The extended feature sequence is input into the dimension-upgrading mapping module in the neural network model based on the Koopman operator. The dimension-upgrading mapping module adopts a trainable structure composed of multi-layer fully connected operations and nonlinear activation units to map the extended feature sequence from the original feature space to the high-dimensional dimension-upgrading state space, generate the dimension-upgrading state vector, and explicitly retain the feature channels corresponding to the clamping force estimate and the slippage probability in the dimension-upgrading state vector.

[0036] The up-dimensional state vector is input into the up-dimensional dynamic model, which includes a parameter matrix, a control input matrix, and a state transition unit. The parameter matrix is ​​used to perform linear state propagation in the high-dimensional up-dimensional state space, and the control input matrix is ​​used to map the torque control command to the high-dimensional up-dimensional state space and jointly drive the up-dimensional state vector to iteratively evolve with time slices together with the parameter matrix.

[0037] A multi-step prediction mechanism is adopted in the upgraded dynamics model. By iteratively calculating the parameter matrix and the control input matrix in the high-dimensional upgraded state space, the upgraded state sequence of multiple future time slices is generated. The continuity and traceability of the feature channels corresponding to the clamping force estimate and the slip probability are ensured in the prediction sequence to support the torque optimization calculation of the subsequent model prediction control module.

[0038] Optionally, the generation of the optimal torque control command includes:

[0039] The upgraded state sequence of multiple future time slices is input into the model prediction and control module, which includes a state prediction unit, a target tracking unit, a constraint generation unit, an optimization solution unit, and a torque command output unit.

[0040] The state prediction unit extracts and temporally arranges the clamping force prediction value and slippage probability feature channel in the up-dimensional state sequence to generate the predicted clamping force sequence and the predicted slippage probability sequence.

[0041] The target tracking unit constructs a clamping force deviation sequence based on the target clamping force setpoint and the predicted clamping force sequence. The clamping force deviation sequence is used to reflect the deviation magnitude and trend of the predicted clamping force relative to the target clamping force setpoint.

[0042] The constraint generation unit generates a multi-dimensional constraint vector based on the existing torque change rate limit, clamping force range limit, and slippage risk threshold limit. The multi-dimensional constraint vector includes: upper and lower limits of torque change rate, minimum and maximum clamping force, and slippage probability threshold.

[0043] The optimization solution unit inputs the clamping force deviation sequence and multidimensional constraint vector into the objective function based on the rolling optimization mechanism. The objective function minimizes the weighted sum of clamping force deviation and control increment in the prediction time domain and ensures that the constraint conditions are satisfied in each prediction time slice. The optimization solution unit obtains the optimal torque control sequence in the prediction time domain through iterative calculation.

[0044] The torque command output unit extracts the optimal torque control command corresponding to the current time slice from the optimal torque control sequence and sends the optimal torque control command to the screwdriver actuator for execution, so as to achieve comprehensive optimization control of clamping force and slippage risk in the next time slice.

[0045] Optionally, the screwdriver torque optimization control process includes:

[0046] Receive the optimal torque control command, append a timestamp to the optimal torque control command and align it with the time slice, and generate the execution target for the current time slice;

[0047] Based on the torque change rate limit, the optimal torque control command is shaped by slope limitation to form a torque command sequence that satisfies the torque change rate limit;

[0048] The torque command sequence is encapsulated into a control frame for the screwdriver actuator. The control frame contains the target torque, execution time and slope parameters. The control frame is sent to the screwdriver actuator to establish a real-time control channel.

[0049] During the execution of the torque command sequence, real-time multimodal data from sensors in the screwdriver are collected, including motor current sensor, angular velocity sensor, acoustic sensor and vibration sensor. Time slices are generated according to fixed time windows. The multimodal feature extraction results corresponding to the time slices and the contact state diagram are updated and input into the isomorphic graph neural network model to obtain the clamping force estimate and slippage probability in real time.

[0050] Based on the clamping force range limit and the slippage risk threshold limit, the torque command sequence is gated online. When the clamping force prediction value exceeds the clamping force range limit, the amplitude is clipped. When the slippage probability reaches the slippage risk threshold limit, the torque reduction and maintenance strategy is implemented and the rising phase of the torque command sequence is paused.

[0051] When the torque command sequence is completed or the target clamping force setting value is reached, a stop torque control command is output. The screwdriver actuator ends the current tightening cycle and maintains the final torque. The clamping force prediction value, slippage probability and torque command sequence of the current tightening cycle are recorded for subsequent rolling updates of the model prediction control module.

[0052] If a communication anomaly, missing sensor data, or torque command sequence fails to meet the torque change rate limit, clamping force range limit, or slippage risk threshold limit is detected at any time slice, a safety degradation process is triggered. The safety degradation process includes setting the torque to the safety lower limit and terminating the current tightening cycle.

[0053] The beneficial effects of this invention are:

[0054] First, this invention achieves accurate estimation of clamping force and slippage probability during screw tightening by introducing multimodal feature extraction and equivariant graph neural network collaborative modeling, significantly improving the accuracy and real-time performance of state perception. Second, combined with a Koopman-based dimensionality prediction mechanism, it can accurately predict key states in multiple future time slices, providing a reliable basis for subsequent torque optimization control. Furthermore, within the model predictive control framework, it generates optimal torque control commands by integrating multidimensional constraints such as torque change rate, clamping force range, and slippage risk threshold, and dynamically adjusts and protects against these constraints during execution. This not only improves control accuracy and stability but also effectively reduces the risks of slippage and overtightening, extending the service life of the screw and tightening tool. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is an overall flowchart of a screwdriver torque optimization control method based on neural networks proposed in this invention;

[0057] Figure 2 This is a schematic diagram of the equivariant graph neural network model structure that incorporates physical consistency constraints in this invention.

[0058] Figure 3 This is a schematic diagram of the collaborative computation process of dimensionality prediction and model predictive control based on the Koopman operator in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figure 1-3 A screwdriver torque optimization control method based on neural networks includes the following steps:

[0061] Using sensors in a screwdriver, including a motor current sensor, angular velocity sensor, acoustic sensor, and vibration sensor, real-time multimodal data of the screwdriver during the tightening process is collected. The real-time multimodal data is divided into multiple time slices according to a fixed time window. Multimodal feature extraction is performed on each time slice to generate multimodal feature extraction results. The extracted features include the statistics of motor current, the instantaneous change of angular velocity, the frequency domain features of acoustic signals, and the time domain envelope features of vibration signals. The multimodal feature extraction results are used as node features to construct a contact state diagram. The nodes of the contact state diagram represent the node features within a single time slice, and the edges represent the temporal correlation and physical contact relationship between nodes.

[0062] For each edge in the contact state diagram, edge features are generated. The edge features include the time interval between nodes, the angle change of the screwdriver, and the change in the friction coefficient. The change in the friction coefficient is calculated based on the physical friction model. Physical consistency constraints are introduced during the edge feature update process to ensure that the edge features satisfy the residual minimization condition of the friction model and the relationship equation between torque and clamping force in adjacent time slices.

[0063] The contact state diagram is input into the equivariant graph neural network model. The graph structure is updated using node features and edge features containing physical consistency constraints. The clamping force estimate and slippage probability are output. The clamping force estimate and slippage probability are combined into a contact feature vector. The contact feature vector is concatenated with the multimodal feature extraction results to form an extended feature sequence.

[0064] The extended feature sequence is input into a neural network model based on the Koopman operator. An upgraded state vector is generated through upgraded mapping. The feature channels corresponding to the clamping force estimate and the slip probability are explicitly retained in the upgraded state vector. The upgraded state sequence of multiple future time slices is predicted using the upgraded dynamics model, which includes a trainable parameter matrix and input torque control commands.

[0065] The higher-dimensional state sequence of multiple future time slices is input into the model prediction and control module. Based on the predicted clamping force, slippage probability, and target clamping force setpoint, the optimal torque control command that satisfies the torque change rate limit, clamping force range limit, and slippage risk threshold limit is calculated.

[0066] The optimal torque control command is sent to the screwdriver actuator, which drives the actuator to adjust the torque output in real time during the tightening process, thus completing the screwdriver torque optimization control process.

[0067] This invention proposes a screwdriver torque optimization control method based on neural networks. By integrating multimodal sensor data and physical consistency constraints, a contact state diagram is constructed and multi-source features are extracted throughout the tightening process. By utilizing an isomorphic graph neural network and a higher-dimensional prediction model based on the Koopman operator, accurate prediction of clamping force estimation and slippage probability is achieved. Under the model predictive control framework, the optimal torque control command that satisfies multiple constraints is generated, thereby improving the accuracy, stability and anti-slippage capability of torque control.

[0068] In this embodiment, the construction of the contact state diagram includes:

[0069] The motor current sensor, angular velocity sensor, acoustic sensor, and vibration sensor of the screwdriver are used to continuously collect real-time multimodal data of the screwdriver throughout the tightening process.

[0070] The collected real-time multimodal data is segmented into fixed time windows, and the entire tightening process is divided into multiple consecutive time slices, each containing a fixed number of sampling points.

[0071] Multimodal feature extraction is performed within each time slice: statistical calculations are performed on the motor current signal to extract four types of statistical features: mean, variance, skewness, and kurtosis; instantaneous changes in angular velocity signals are calculated to obtain the average, maximum, and minimum values ​​of the angular velocity differences between adjacent sampling points; frequency domain analysis is performed on the acoustic signal to extract four frequency domain features: dominant frequency, spectral centroid, bandwidth, and energy distribution of characteristic frequency bands; and time-domain envelope features are extracted from the vibration signal, including envelope mean, envelope variance, and envelope energy.

[0072] Multimodal features are concatenated at the time slice level to generate multimodal feature extraction results, and these results are used as node features.

[0073] The temporal relationships are generated based on the chronological order between time slices, and the physical contact relationships are generated based on the physical contact events of the screwdriver during the tightening process. The temporal relationships and physical contact relationships are used as edge information in the contact state graph. The contact state graph is constructed by combining node features, where nodes represent node features within a single time slice, and edges represent the temporal relationships and physical contact relationships between nodes.

[0074] This step collects real-time multimodal data of the screw tightening process using multiple sensors such as motor current, angular velocity, acoustics, and vibration. After time slice division and feature extraction, node features are generated, and a contact state diagram that takes into account both time correlation and physical contact relationship is constructed, realizing fine-grained structured modeling of the dynamic contact state of the tightening process.

[0075] In this embodiment, the generation of the edge features includes:

[0076] Initial edge features are generated for each edge in the contact state diagram. The initial edge features include the time interval between nodes, the angle change of the screwdriver, the change of angular velocity, and the change of friction coefficient. The change of friction coefficient is calculated based on the physical friction model, which adopts the Stribeck friction model and the Coulomb-viscosity combined model. The calculation process is as follows: in each time slice, the clamping force estimate is used as the normal force input, and the relative sliding velocity derived from the angular velocity and angle change is used as the velocity input. First, the friction coefficient of the transition from the static friction zone to the dynamic friction zone is given according to the Stribeck model. Then, the viscous friction component that increases linearly with the velocity is superimposed to obtain the friction coefficient value of the time slice. The difference of the friction coefficient values ​​of adjacent time slices is used as the change of friction coefficient and written into the initial edge features.

[0077] The initial edge features and the node features of the contact state graph are input into the equivariant graph neural network model to perform the graph structure update process of the edge features. During the update process, the features of time interval, angle change, angular velocity change and friction coefficient change are retained.

[0078] The updated edge feature vectors are used to generate the predicted clamping force change. The specific process is as follows: the updated edge feature vectors are concatenated into an input vector in a fixed order, which is the time interval, angle change, angular velocity change and friction coefficient change. This input vector is then input into the edge-level regression module of the equivariant graph neural network model. The edge-level regression module consists of multi-layer fully connected operations and nonlinear activation and shares the previous layer parameters with the graph structure update. It outputs a single scalar as the predicted clamping force change. During the training phase, the parameters are updated using a weighted loss based on the physical consistency residual. During the inference phase, the predicted clamping force change is directly output for use in calculating the physical consistency residual and forming the contact feature vector.

[0079] The residuals of the equations relating the changes in friction coefficient, angular velocity, angle, and clamping force are used as loss function terms for minimization training. The weighted loss of the physical consistency residuals is defined as:

[0080]

[0081] Among them, L phys The sum of the weighted loss function for physical consistency residuals represents the total value used to constrain the consistency between the model's predictions and the physical model. e represents the edge index in the contact state graph, and Δω... e Δθ represents the change in angular velocity between nodes connected by an edge. e This represents the change in angle between nodes connected by the edge. α and β represent non-negative weight hyperparameters used to adjust the weights of the changes in angular velocity and angle on the loss. ΔF represents the change in clamping force predicted by the isotropic graph neural network model. e This indicates the actual change in clamping force on the corresponding side in the training data. This represents the squared difference between the predicted change in clamping force and the actual change in clamping force, and the magnitude of the prediction error is represented by the squared Euclidean distance.

[0082] Formula source:

[0083] Coulomb friction and viscous friction: frictional force F f With normal force F n The relationship is F f =μF n In the higher velocity range, velocity-related dissipation is often characterized by a viscous term. The normal force corresponds to the clamping force F during tightening. c ;

[0084] The approximate formula widely used in engineering for the relationship between torque and clamping force is:

[0085] T≈KF c d;

[0086] Where T is the torque, K is the torque coefficient, and d is the nominal diameter of the thread, affected by the friction and geometry of the thread / bearing surface.

[0087] Kinematic quantities: Angular increment Δθ and angular velocity increment Δω describe the torsional dynamics of adjacent time segments;

[0088] Formula Improvement:

[0089] Within the minute time interval of tightening, the torque increment ΔT can be decomposed into the part that overcomes friction and the part that stores elastic energy, from T≈KF. c d obtains the result for F c First-order differential approximation:

[0090]

[0091] ΔK is related to the change in friction state. Since the precise decomposition of ΔT and ΔK in the field is difficult to measure directly, this invention processes them using a data-physical coupling method: using an isotropic graph neural network to predict the change in clamping force from node features and edge features. ΔF obtained by actual or calibrated methods e To monitor the signal, construct a physically consistent residual;

[0092] During the high-dynamic phase of tightening (large changes in angular velocity or rapid angle transitions), the friction state is more prone to abrupt changes and the risk of slippage is higher. If the model deviates in these segments, it will directly degrade the control effect. Therefore, a weight related to the dynamic intensity is introduced into the standard quadratic residual to obtain the weighted loss of the physical consistency residual:

[0093]

[0094] Here, edge index 'e' represents the adjacent time slice in the corresponding contact state graph. ΔF represents the change in clamping force predicted by the isotropic graph neural network. e Δω represents the change in clamping force in the training data annotations. e Δθ represents the change in angular velocity. e For the angle change, α and β are hyperparameters. The error is weighted with interpretable dynamic amplitude, which allows the training phase to impose stronger constraints on the segments that have the greatest impact on safety and control quality.

[0095] Weight term 1+α|Δω e |+β|Δθ e | is dimensionless, α and β are chosen as coefficients after normalizing all quantities to dimensionless; residual term The dimension of the sum is the square of the force, which is consistent with the mean square error measure; summing the sums across the edges does not change the consistency of the dimensions.

[0096] By optimizing the parameters of the isovariant graph neural network model based on minimizing the weighted loss of the physical consistency residual, edge feature representations that satisfy the physical property constraints of screwdrivers are obtained.

[0097] After updating the graph structure of edge features in the contact state graph, the updated edge feature vector is obtained. This updated feature vector is used as input in a fixed order: time interval, angle change, angular velocity change, and friction coefficient change. To avoid the influence of different dimensions on regression stability, common normalization or standardization is used to map each component to a uniform numerical range. The processed and updated edge feature vector is then fed into the edge-level regression module of the equivariant graph neural network model. This module uses a publicly available multi-layer fully connected structure combined with commonly used nonlinear activation functions for forward computation. The module shares some parameters or uses shared features as input with the preceding graph structure update layer to ensure that the edge-level regression and graph structure update maintain consistency in the feature space. The edge-level regression module outputs a single scalar as the predicted clamping force change, corresponding to the predicted clamping force change between the two time slices connected by the edge.

[0098] During the training phase, supervised learning is performed between the predicted clamping force change and the actual clamping force change annotations in the training data. A common regression loss and a weighted loss based on the physical consistency residuals defined in claim 3 are used for joint optimization. The actual clamping force change annotations can be derived from clamping force measurement data in tightening tests or calculated according to industry-standard calibration procedures. To improve generalization ability, publicly available regularization methods and mini-batch training strategies can be used in the edge-level regression module. During the inference phase, the predicted clamping force change is output independently for each edge, forming a prediction sequence in time sequence. This prediction sequence is used for calculating and minimizing the physical consistency residuals in claim 3. Furthermore, in step S3 of claim 1, it, along with the clamping force estimate and slippage probability, constitutes one of the sources of the contact feature vector. In subsequent steps, it is concatenated with the multimodal feature extraction results to form an extended feature sequence, providing input for the Koopman-based neural network model.

[0099] In this embodiment, the processing of the isomorphic graph neural network model includes:

[0100] The node features of the contact state graph and the edge features containing physical consistency constraints are input into the equivariant graph neural network model. The equivariant graph neural network model includes a node state update module and an edge feature update module. The node state update module updates the node state based on the node features and the message vector corrected by the edge features. The edge feature update module updates the edge features based on the states of adjacent nodes and the physical properties of the connecting edges.

[0101] Hamiltonian structural constraints and symplectic geometric constraints are introduced into the edge feature update module. Hamiltonian structural constraints are used to maintain the energy conservation property of the system during the iteration process, while symplectic geometric constraints are used to maintain the reversibility and volume preservation property of the state evolution in the phase space. The edge features are updated by constraining the energy consistency condition of torque work and clamping force accumulation based on the existing screw tightening physical model to avoid energy mutation signals before and after slippage.

[0102] During message passing, the node state, edge features, and their physical consistency constraints are input into the node state update module. Through multiple rounds of iterative passing and fusion, an updated node feature representation is generated.

[0103] After the graph neural network model completes the graph structure update, it calculates the clamping force estimate and slippage probability based on the updated node feature representation and edge feature vector. When calculating the clamping force estimate, all node feature representations are globally pooled to obtain the global feature representation of the contact state graph. The global feature representation is then input into the regression output layer to generate the clamping force numerical prediction. When calculating the slippage probability, the global feature representation is input into the binary classification probability output layer and a slippage probability value ranging from 0 to 1 is generated through the Sigmoid function. The clamping force estimate and slippage probability are supervised learning with real clamping force data and slippage event labels during the training process, respectively. The former uses the mean squared error loss function, and the latter uses the binary cross-entropy loss function. Finally, the clamping force estimate and slippage probability are concatenated to form the contact feature vector.

[0104] The contact feature vector and the multimodal feature extraction results are concatenated at the time slice level to generate an extended feature sequence, which serves as the input for subsequent dimensionality-upgrading dynamics modeling.

[0105] The existing physical model for screw tightening refers to a mathematical model widely used in open mechanical engineering and fastener mechanics theories, which describes the screw tightening process through the physical relationship between torque, clamping force, and rotation angle. This model typically assumes that the torque applied during tightening consists of three parts: work done to overcome friction between the threaded pair, friction between the screw head and the supporting surface of the connected part, and work done to induce axial elongation of the screw to form clamping force. The relationship between the torque work and the cumulative clamping force can be described by the principle of energy conservation, i.e., the integral of the externally applied torque over the rotation angle should equal the sum of the increase in elastic energy caused by the clamping force and the energy dissipated by friction. The clamping force and the axial elongation are related through the equivalent stiffness of the screw and the connected part, while the friction component is determined by the thread friction coefficient, the supporting surface friction coefficient, and the normal force. This physical model can reflect the energy distribution law during tightening, providing a theoretical basis for evaluating the change in clamping force and the update of constraint edge characteristics.

[0106] In this embodiment, the prediction of the upgraded state sequence of multiple future time slices includes:

[0107] The extended feature sequence is input into the dimension-upgrading mapping module in the neural network model based on the Koopman operator. The dimension-upgrading mapping module adopts a trainable structure composed of multi-layer fully connected operations and nonlinear activation units to map the extended feature sequence from the original feature space to the high-dimensional dimension-upgrading state space, generate the dimension-upgrading state vector, and explicitly retain the feature channels corresponding to the clamping force estimate and the slippage probability in the dimension-upgrading state vector.

[0108] The up-dimensional state vector is input into the up-dimensional dynamic model, which includes a parameter matrix, a control input matrix, and a state transition unit. The parameter matrix is ​​used to perform linear state propagation in the high-dimensional up-dimensional state space, and the control input matrix is ​​used to map the torque control command to the high-dimensional up-dimensional state space and jointly drive the up-dimensional state vector to iteratively evolve with time slices together with the parameter matrix.

[0109] A multi-step prediction mechanism is adopted in the upgraded dynamics model. By iteratively calculating the parameter matrix and the control input matrix in the high-dimensional upgraded state space, the upgraded state sequence of multiple future time slices is generated. The continuity and traceability of the feature channels corresponding to the clamping force estimate and the slip probability are ensured in the prediction sequence to support the torque optimization calculation of the subsequent model prediction control module.

[0110] The up-dimensional dynamics model describes the state propagation from time slice to time slice within the high-dimensional space of the up-dimensional state vector. Combined with torque control commands, it achieves controlled evolution. The model consists of a parameter matrix and a control input matrix. The inputs are the up-dimensional state vector output from the up-dimensional mapping module and the torque control command corresponding to the current time slice. The output is the up-dimensional state vector for the next time slice, and the model iteratively generates an up-dimensional state sequence for multiple future time slices along the time axis. During computation, the model performs linear propagation of the up-dimensional state vector using the parameter matrix, and maps the torque control command to the up-dimensional state space using the control input matrix, superimposing the linear propagation results to form a controlled state update. Throughout the propagation and prediction process, the feature channels corresponding to the clamping force estimate and slippage probability are explicitly preserved and synchronously updated to ensure the numerical continuity and traceability of these two types of channels in the prediction sequence. To adapt to the differences in connecting parts and friction states under various working conditions, the parameter matrix and control input matrix are trained using a data-driven approach. The training signal originates from the registration results of the extended feature sequence and the clamping force estimate and slippage probability output by the isotropic graph neural network over time slices. Parameter calibration is completed by minimizing the error between the predicted sequence and the target sequence. During the inference phase, the model receives new upgraded state vectors and torque control commands in a rolling manner, and outputs upgraded state sequences for multiple future time slices in chronological order. This provides the model's predictive control module with predictive inputs including clamping force estimates and slippage probability channels, thereby supporting the subsequent calculation of optimal torque control commands under torque change rate constraints, clamping force range constraints, and slippage risk threshold constraints.

[0111] In this embodiment, the generation of the optimal torque control command includes:

[0112] The upgraded state sequence of multiple future time slices is input into the model prediction and control module, which includes a state prediction unit, a target tracking unit, a constraint generation unit, an optimization solution unit, and a torque command output unit.

[0113] The state prediction unit extracts and temporally arranges the clamping force prediction value and slippage probability feature channel in the up-dimensional state sequence to generate the predicted clamping force sequence and the predicted slippage probability sequence.

[0114] The target tracking unit constructs a clamping force deviation sequence based on the target clamping force setpoint and the predicted clamping force sequence. The clamping force deviation sequence is used to reflect the deviation magnitude and trend of the predicted clamping force relative to the target clamping force setpoint.

[0115] The constraint generation unit generates a multi-dimensional constraint vector based on the existing torque change rate limit, clamping force range limit, and slippage risk threshold limit. The multi-dimensional constraint vector includes: upper and lower limits of torque change rate, minimum and maximum clamping force, and slippage probability threshold.

[0116] The optimization solution unit inputs the clamping force deviation sequence and multidimensional constraint vector into the objective function based on the rolling optimization mechanism. The objective function minimizes the weighted sum of clamping force deviation and control increment in the prediction time domain and ensures that the constraint conditions are satisfied in each prediction time slice. The optimization solution unit obtains the optimal torque control sequence in the prediction time domain through iterative calculation.

[0117] The torque command output unit extracts the optimal torque control command corresponding to the current time slice from the optimal torque control sequence and sends the optimal torque control command to the screwdriver actuator for execution, so as to achieve comprehensive optimization control of clamping force and slippage risk in the next time slice.

[0118] In existing screw tightening process control, torque change rate limits, clamping force range limits, and slippage risk threshold limits are all standard process safety and quality control conditions, widely used in power tools, automated assembly lines, and quality inspection. Torque change rate limits constrain the rate of change of torque output over time, preventing rapid torque fluctuations that could damage threads or cause stress concentration in the connector. This is typically achieved by setting an upper limit on the torque increase or decrease per unit time, for example, specifying that the torque change per second should not exceed a certain percentage of the rated torque. Clamping force range limits ensure that the tightened connector receives sufficient preload to prevent loosening, while avoiding excessive clamping force that could cause material yielding or permanent deformation. Current technologies generally set allowable ranges for minimum and maximum clamping forces based on the connector material, specifications, and design requirements. Slippage risk threshold limits, based on the frictional characteristics and structural load-bearing capacity of the connector during tightening, predefine a safe upper limit for the probability of slippage. When the actual predicted slippage probability exceeds this threshold, the system must adjust the torque or pause the tightening operation to prevent relative slippage damage to the connector or tool. In the method of this invention, these constraints, as publicly available and mature process control parameters, are directly introduced into the model predictive control calculation process to constrain the calculation results of the torque optimization command, thereby ensuring that the tightening process is precisely controlled while meeting the requirements of connection reliability and assembly safety.

[0119] In this embodiment, the screwdriver torque optimization control process includes:

[0120] Receive the optimal torque control command, append a timestamp to the optimal torque control command and align it with the time slice, and generate the execution target for the current time slice;

[0121] Based on the torque change rate limit, the optimal torque control command is shaped by slope limitation to form a torque command sequence that satisfies the torque change rate limit;

[0122] The torque command sequence is encapsulated into a control frame for the screwdriver actuator. The control frame contains the target torque, execution time and slope parameters. The control frame is sent to the screwdriver actuator to establish a real-time control channel.

[0123] During the execution of the torque command sequence, real-time multimodal data from sensors in the screwdriver are collected, including motor current sensor, angular velocity sensor, acoustic sensor and vibration sensor. Time slices are generated according to fixed time windows. The multimodal feature extraction results corresponding to the time slices and the contact state diagram are updated and input into the isomorphic graph neural network model to obtain the clamping force estimate and slippage probability in real time.

[0124] Based on the clamping force range limit and the slippage risk threshold limit, the torque command sequence is gated online. When the clamping force prediction value exceeds the clamping force range limit, the amplitude is clipped. When the slippage probability reaches the slippage risk threshold limit, the torque reduction and maintenance strategy is implemented and the rising phase of the torque command sequence is paused.

[0125] When the torque command sequence is completed or the target clamping force setting value is reached, a stop torque control command is output. The screwdriver actuator ends the current tightening cycle and maintains the final torque. The clamping force prediction value, slippage probability and torque command sequence of the current tightening cycle are recorded for subsequent rolling updates of the model prediction control module.

[0126] If a communication anomaly, missing sensor data, or torque command sequence fails to meet the torque change rate limit, clamping force range limit, or slippage risk threshold limit is detected at any time slice, a safety degradation process is triggered. The safety degradation process includes setting the torque to the safety lower limit and terminating the current tightening cycle.

[0127] The triggering conditions for the torque reduction and maintenance strategy are based on online gating results. When the slippage probability reaches the slippage risk threshold limit, or the clamping force estimate is about to exceed the clamping force range limit, the torque command sequence no longer increases as originally planned, but enters the controlled processing stage. The goal of the controlled processing stage is to reduce the slippage risk and keep the clamping force within the allowable range without violating the torque change rate limit.

[0128] The torque reduction strategy involves a combined process of amplitude clipping and slope limiting on the current torque command sequence. Amplitude clipping immediately suppresses the target torque to the upper limit of the safe range or proportionally reverts it to the torque level corresponding to the previous stable time slice. Slope limiting ensures that the torque change rate limit is met during the command descent, preventing mechanical shock or secondary oscillations caused by sudden and significant reductions. When the torque reduction strategy is executed, the model predictive control module retains existing multidimensional constraints and performs rolling verification of the target clamping force setpoint based on the latest upgraded state sequence and contact feature vector. If the clamping force estimate is still higher than the allowable range, it continues to gradually reduce to the safe lower limit in the form of slope limiting until the slippage probability is lower than the slippage risk threshold limit and the clamping force returns to the clamping force range limit.

[0129] The maintenance strategy aims to keep the current torque at a stable level and continuously monitor the contact state until the risk is eliminated. The maintenance level is preferentially selected based on the minimum torque value that ensures the clamping force estimate falls within the clamping force range limit. If a stable point exists in the previous time slice, the torque corresponding to that point is used; otherwise, the torque at the end of the torque reduction strategy is used as the maintenance value. During maintenance, sensors continue to collect real-time multimodal data, which is processed through time slices to update the contact state map. The isomorphic graph neural network model recalculates the clamping force estimate and slippage probability. A neural network model based on the Koopman operator updates the upgraded state sequence for future time slices, and the model predictive control module generates new optimal torque control commands accordingly. When the slippage probability remains below the slippage risk threshold and the clamping force stabilizes within the clamping force range limit, the maintenance strategy exits, and the torque command sequence smoothly recovers and converges towards the target clamping force setpoint according to the torque change rate limit.

[0130] If, during the execution of the torque reduction and maintenance strategy, the slippage probability increases again or the clamping force estimate deviates from the allowable range, and the torque command sequence can no longer simultaneously meet the torque change rate limit, clamping force range limit, and slippage risk threshold limit, then the safety degradation process defined in claim 7 is entered, the torque is set to the safety lower limit, and the current tightening cycle is terminated.

[0131] Example 1:

[0132] To verify the feasibility of this invention in practice, it was applied to the torque control system of a screwdriver used in an industrial production line. The goal was to improve clamping force accuracy, reduce slippage probability, and shorten tightening time during batch assembly. This production line requires high-frequency, high-precision screw tightening operations, with approximately 5000 tightening operations per shift.

[0133] During implementation, the screwdriver's sensor array continuously collects multimodal data on motor current, angular velocity, acoustic signals, and vibration signals, dividing the data into multiple time slices within a fixed time window. Within each time slice, the system extracts the mean, variance, skewness, and kurtosis statistics of the motor current; the average and extreme values ​​of the angular velocity variation; the dominant frequency, spectral centroid, bandwidth, and characteristic energy distribution of the acoustic signal; and the envelope mean, variance, and energy characteristics of the vibration signal. These features are combined to construct a contact state diagram using node features. The edge information incorporates time intervals, angular changes, and friction coefficient changes calculated based on a physical friction model, and is optimized through physical consistency constraints.

[0134] The equivariant graph neural network updates the graph structure using node and edge features, outputting clamping force estimates and slippage probabilities, which are then concatenated with the original multimodal features to generate an extended feature sequence. This extended feature sequence is input into a neural network model based on the Koopman operator, and through dimensionality-upgrading mapping, generates an upgraded state vector. It explicitly preserves the feature channels of the clamping force estimates and slippage probabilities, and predicts the upgraded state sequence for multiple future time slices.

[0135] The model predictive control module uses the predicted clamping force, slippage probability, and target clamping force setpoint, combined with torque change rate limits, clamping force range limits, and slippage risk threshold limits, to solve for the optimal torque control command. This command is sent to the actuator in real time, enabling the screwdriver to dynamically adjust the torque output during tightening, ensuring that the target clamping force is met throughout the tightening curve, while avoiding slippage and over-tightening.

[0136] Experimental results show that this invention significantly improves control accuracy and assembly efficiency under the same production conditions. The clamping force control error was reduced from ±8% of the traditional method to ±2%, the slippage rate decreased from 4.2% to 0.6%, the average tightening time was shortened by 0.35 seconds, and the production cycle time was increased by approximately 12%. Meanwhile, long-term operational data indicates that this invention can reduce the number of reworks caused by tightening defects, reducing the number of reworks per shift from 42 to 7, significantly reducing labor and material costs.

[0137] Table 1 Comparison between traditional control methods and the method of this invention

[0138]

[0139]

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A screwdriver torque optimization control method based on neural networks, characterized in that, Includes the following steps: By using sensors in the screwdriver, real-time multimodal data of the screwdriver during the tightening process is collected, multimodal feature extraction is performed, multimodal feature extraction results are generated, and the multimodal feature extraction results are used as node features to construct a contact state diagram; For each edge in the contact state graph, generate edge features and introduce physical consistency constraints during the edge feature update process; The contact state diagram is input into the equivariant graph neural network model. The graph structure is updated using node features and edge features. The clamping force estimate and slippage probability are output. The clamping force estimate and slippage probability are combined into a contact feature vector. The contact feature vector is concatenated with the multimodal feature extraction results to form an extended feature sequence. The extended feature sequence is input into a neural network model based on the Koopman operator. An upgraded state vector is generated through upgraded mapping. The feature channels corresponding to the clamping force estimate and the slip probability are explicitly retained in the upgraded state vector. The upgraded dynamics model is used to predict the upgraded state sequence of multiple future time slices. The higher-dimensional state sequence of multiple future time slices is input into the model prediction and control module. Based on the predicted clamping force, slippage probability, and target clamping force setpoint, the optimal torque control command that satisfies the torque change rate limit, clamping force range limit, and slippage risk threshold limit is calculated. The optimal torque control command is sent to the screwdriver actuator, which drives the actuator to adjust the torque output in real time during the tightening process.

2. The screwdriver torque optimization control method based on neural networks according to claim 1, characterized in that, The construction of the contact state diagram includes: The motor current sensor, angular velocity sensor, acoustic sensor and vibration sensor of the screwdriver are used to continuously collect real-time multimodal data of the screwdriver. The collected real-time multimodal data is segmented into fixed time windows, and the entire tightening process is divided into multiple consecutive time slices, each containing a fixed number of sampling points. Perform multimodal feature extraction operations within each time slice; Multimodal features are concatenated at the time slice level to generate multimodal feature extraction results, and these results are used as node features. The temporal relationships are generated based on the chronological order of time slices, and the physical contact relationships are generated based on the physical contact events of the screwdriver during the tightening process. The temporal relationships and physical contact relationships are used as edge information in the contact state graph, and the contact state graph is constructed by combining node features.

3. The screwdriver torque optimization control method based on neural networks according to claim 1, characterized in that, The generation of the edge features includes: Initial edge features are generated for each edge in the contact state diagram. The initial edge features include the time interval between nodes, the angle change of the screwdriver, the change of angular velocity, and the change of friction coefficient. The change of friction coefficient is calculated based on the physical friction model. The initial edge features and the node features of the contact state graph are input into the equivariant graph neural network model to perform the graph structure update process of the edge features. During the update process, the features of time interval, angle change, angular velocity change and friction coefficient change are retained. The updated edge feature vector is used to generate the predicted clamping force change. The predicted clamping force change and the initial edge features together constitute the input for the calculation of physical consistency constraints. The residuals of the equations relating the changes in friction coefficient, angular velocity, angle, and clamping force are used as loss function terms for minimization training. The weighted loss of the physical consistency residuals is defined as: Among them, L phys The sum of the weighted loss function for physical consistency residuals represents the total value, e represents the edge index in the contact state graph, and Δω represents the total value. e Δθ represents the change in angular velocity between nodes connected by an edge. e The angular change between nodes connected by the edge is represented by α and β, which are non-negative weight hyperparameters. ΔF represents the change in clamping force predicted by the isotropic graph neural network model. e This indicates the actual change in clamping force on the corresponding side in the training data. This represents the squared difference between the predicted change in clamping force and the actual change in clamping force. By optimizing the parameters of the isovariant graph neural network model based on minimizing the weighted loss of the physical consistency residual, edge feature representations that satisfy the physical property constraints of screwdrivers are obtained.

4. The screwdriver torque optimization control method based on neural networks according to claim 1, characterized in that, The processing of the isomorphic graph neural network model includes: The contact state graph is input into the equivariant graph neural network model, which includes a node state update module and an edge feature update module. The node state update module updates the node state based on the node features and the message vector corrected by the edge features. The edge feature update module updates the edge features based on the states of adjacent nodes and the physical properties of the connecting edges. Hamiltonian structural constraints and symplectic geometric constraints are introduced into the edge feature update module. Hamiltonian structural constraints are used to maintain the energy conservation property of the system during the iteration process, and symplectic geometric constraints are used to maintain the reversibility and volume preservation property of the state evolution in the phase space. The edge features are updated by constraint based on the energy consistency condition of torque work and clamping force accumulation in the existing screw tightening physical model. During message passing, the node state, edge features, and their physical consistency constraints are input into the node state update module. Through multiple rounds of iterative passing and fusion, an updated node feature representation is generated. After the graph neural network model completes the graph structure update, it calculates the clamping force estimate and slippage probability based on the updated node feature representation and edge feature vector, and finally concatenates the clamping force estimate and slippage probability into a contact feature vector. The contact feature vector and the multimodal feature extraction results are concatenated at the time slice level to generate an extended feature sequence.

5. The screwdriver torque optimization control method based on neural networks according to claim 1, characterized in that, The prediction of the upgraded state sequence for multiple future time slices includes: The extended feature sequence is input into the up-dimensional mapping module in the neural network model based on the Koopman operator, and the extended feature sequence is mapped from the original feature space to the high-dimensional up-dimensional state space to generate the up-dimensional state vector. The feature channels corresponding to the clamping force estimate and the slip probability are explicitly retained in the up-dimensional state vector. The upgraded state vector is input into the upgraded dynamic model, which includes a parameter matrix and a control input matrix. The parameter matrix is ​​used to perform linear state propagation in the high-dimensional upgraded state space, and the control input matrix is ​​used to map the torque control command to the high-dimensional upgraded state space, and together with the parameter matrix, drive the upgraded state vector to iteratively evolve with time slices. In the up-dimensional dynamics model, a multi-step prediction mechanism is adopted. By iteratively calculating the parameter matrix and the control input matrix in the high-dimensional up-dimensional state space, an up-dimensional state sequence for multiple future time slices is generated, and the continuity and traceability of the feature channels corresponding to the clamping force estimate and the slip probability are ensured in the prediction sequence.

6. The screwdriver torque optimization control method based on neural networks according to claim 1, characterized in that, The generation of the optimal torque control command includes: The upgraded state sequence of multiple future time slices is input into the model prediction and control module, which includes a state prediction unit, a target tracking unit, a constraint generation unit, an optimization solution unit, and a torque command output unit. The state prediction unit extracts and temporally arranges the clamping force prediction value and slippage probability feature channel in the up-dimensional state sequence to generate the predicted clamping force sequence and the predicted slippage probability sequence. The target tracking unit constructs a clamping force deviation sequence based on the target clamping force setpoint and the predicted clamping force sequence; The constraint generation unit generates a multi-dimensional constraint vector based on existing torque change rate limits, clamping force range limits, and slippage risk threshold limits; The optimization solution unit inputs the clamping force deviation sequence and multidimensional constraint vector into the objective function based on the rolling optimization mechanism. The objective function minimizes the weighted sum of clamping force deviation and control increment in the prediction time domain and ensures that the constraint conditions are satisfied in each prediction time slice. The optimization solution unit obtains the optimal torque control sequence in the prediction time domain through iterative calculation. The torque command output unit extracts the optimal torque control command corresponding to the current time slice from the optimal torque control sequence and sends the optimal torque control command to the screwdriver actuator for execution.

7. The screwdriver torque optimization control method based on neural networks according to claim 1, characterized in that, The screwdriver torque optimization control process includes: Receive the optimal torque control command, append a timestamp to the optimal torque control command and align it with the time slice, and generate the execution target for the current time slice; Based on the torque change rate limit, the optimal torque control command is shaped by slope limitation to form a torque command sequence that satisfies the torque change rate limit; The torque command sequence is encapsulated into a control frame for the screwdriver actuator, and the control frame is sent to the screwdriver actuator to establish a real-time control channel. During the execution of the torque command sequence, real-time multimodal data is collected, time slices are generated according to fixed time windows, and the multimodal feature extraction results corresponding to the time slices and the contact state diagram are updated and input into the isotropic graph neural network model to obtain the clamping force estimate and slippage probability in real time. The torque command sequence is controlled online based on the clamping force range limit and the slippage risk threshold limit; When the torque command sequence is completed or the target clamping force setting value is reached, a stop torque control command is output, and the screwdriver actuator ends the current tightening cycle and maintains the final torque. If a communication anomaly, missing sensor data, or torque command sequence fails to meet the torque change rate limit, clamping force range limit, or slippage risk threshold limit is detected at any time slice, a safety degradation process is triggered, the torque is set to the safety lower limit, and the current tightening cycle is terminated.

Images