Joint module assembly precision adaptive adjustment control method

Abstract

This invention relates to an adaptive adjustment and control method for the assembly precision of joint modules, specifically in the fields of precision assembly and intelligent manufacturing. By constructing a fully closed-loop control system encompassing multi-physics sensing, intelligent deformation field inversion, digital twin decision-making, physical execution, and online learning, it achieves dynamic maintenance and autonomous optimization of the assembly precision of joint modules. It can sense and compensate for micro-deformations at the mechanical connection interface caused by thermal coupling under complex working conditions in real time, effectively suppressing online precision decay. Deformation field reconstruction and strategy optimization based on swarm intelligence, combined with virtual simulation using digital twins, ensure the accuracy and foresight of compensation decisions. Furthermore, through a residual-driven collaborative online correction mechanism, the sensing model and digital twin can continuously evolve with operation, thereby continuously improving the system's adjustment precision, robustness, and adaptability to equipment performance degradation, ultimately ensuring the long-term operational stability and reliability of the joint modules.

Description

Technical Field

[0001] This invention relates to the field of precision assembly and intelligent manufacturing, and more specifically, to a method for adaptive adjustment and control of the assembly precision of a joint module. Background Technology

[0002] With the rapid development of industrial automation, intelligent manufacturing, and precision equipment, extremely stringent requirements have been placed on the motion accuracy, dynamic performance, and long-term operational reliability of industrial robots. As the core power and execution unit of industrial robots, the assembly accuracy of the joint module directly determines the robot's end-effector positioning accuracy and motion stability. Frameless torque motors, due to their high power density, low torque ripple, and compact structure, are widely used in high-performance joint modules. These motors typically employ a stator-rotor separation and embedded structure, directly connecting to the reducer via a mechanical interface, eliminating the traditional motor housing and bearings, thus placing higher demands on the assembly process. In practical applications, especially under high-speed, high-load, and long-term continuous operation conditions, such as welding and handling in automotive manufacturing or precision insertion operations in the semiconductor industry, the joint module undergoes a complex thermo-coupling process. The stator and rotor of the frameless torque motor... When the module is powered on, it generates uneven Joule heating and iron loss heating. At the same time, harmonic reducers or planetary reducers also generate significant heat due to internal friction when transmitting large torques. In addition, external interference factors such as fluctuations in ambient temperature, sudden changes in processing load, and mechanical vibration further exacerbate the complexity of the thermal and stress field distribution inside the module. This multi-physics coupling leads to a serious challenge: due to the inherent differences in the thermal expansion coefficients of key components made of different materials, such as the motor stator and rotor assembly and the reducer housing, unpredictable microscopic deformation and stress redistribution will occur at their precision mating interfaces under dynamic temperature rise and alternating stress. This micron or even submicron level geometric variation will cause the key assembly precision, such as coaxiality, end face runout, and bearing preload, which are carefully adjusted and solidified in the cold assembly environment, to gradually deviate and decay during actual operation, i.e., the phenomenon of "precision online drift".

[0003] Currently, the industry mainly relies on high-precision offline assembly processes, optimized interference fits and bolt preload design, and passive thermal management measures such as adding heat sinks or forced air cooling to ensure the assembly accuracy of joint modules. Some advanced solutions attempt to integrate temperature sensors and, based on a simplified linear thermal expansion model, provide limited compensation for motor control parameters. However, these existing technologies have significant limitations. First, their core design principles are still based on static or quasi-static assumptions, failing to fully model and compensate for the nonlinear, coupled deformation induced by transient thermal shock, periodic mechanical stress, and complex boundary conditions under real dynamic and time-varying operating conditions. Second, existing methods lack the ability to perceive the microscopic stress and strain state of the critical area—the interface between the motor and the reducer—in real time, in place, and in a distributed manner. Relying solely on single-point or limited temperature monitoring is insufficient to reconstruct the overall deformation. Consequently, traditional methods are incapable of identifying and responding to complex deformation patterns caused by uneven thermal loads and stress concentrations. This leads to problems in high-speed, high-load robot applications, where joint modules often experience a systematic decline in repeatability, increased abnormal vibration and noise, and reduced transmission efficiency after continuous operation for a period of time. The root cause of these problems lies in the inability of static assembly accuracy to adapt to dynamic working conditions, triggering a chain reaction of adverse effects, accelerating bearing wear and gear meshing failure, and ultimately threatening the operational reliability, service life, and work quality of the robot itself. Therefore, there is an urgent need to develop a novel control method capable of online sensing, intelligent identification, and proactive compensation for accuracy degradation caused by thermal deformation and stress accumulation. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an adaptive adjustment and control method for the assembly accuracy of joint modules, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps: Step S1: During the operation of the joint module, the multi-physics field signal is acquired by the sensor network in the area where the mechanical connection interface between the frameless torque motor and the reducer is located inside the joint module. The acquired multi-physics field signal is then processed for time synchronization, spatial coordinate unification and feature extraction to generate a first fusion feature vector that represents the current mechanical state of the mechanical connection interface. Step S2: Based on the pre-constructed parameterized basis function combination, establish a deformation field reconstruction model; use the first swarm intelligent optimization algorithm to generate multiple sets of candidate basis function weight coefficients during the algorithm iteration process; for each set of candidate basis function weight coefficients, calculate a corresponding reconstruction model prediction feature vector from the deformation field reconstruction model; with minimizing the error between the first fused feature vector and each reconstruction model prediction feature vector as the optimization objective, drive the first swarm intelligent optimization algorithm to solve online, and obtain a set of optimal basis function weight coefficients; calculate the full-field deformation distribution of the mechanical connection interface at the current moment based on the set of optimal basis function weight coefficients and the parameterized basis function combination. Step S3: Input the full-field deformation distribution as a boundary condition into the pre-constructed digital twin; in the digital twin, simulate multiple predefined compensation strategies in parallel, and use the second swarm intelligent optimization algorithm to evaluate and optimize the performance index of the digital twin after the execution of each compensation strategy to obtain the optimal compensation control command for the current deformation state; based on the result of the digital twin simulation after executing the optimal compensation control command, calculate a twin prediction feature vector; Step S4: Send the optimal compensation control command to the micro-displacement actuator set at the mechanical connection interface to control the micro-displacement actuator to generate a predetermined displacement to implement physical compensation; after the physical compensation action is completed, collect the multi-physics field signal of the sensor network again and generate a second fusion feature vector; calculate the residual vector between the second fusion feature vector and the twin prediction feature vector; based on the residual vector, perform collaborative online correction on the first swarm intelligent optimization algorithm and the digital twin.

[0006] In a preferred embodiment, the specific process of performing time synchronization, spatial coordinate unification, and feature extraction on the acquired multi-physics field signals in step S1 is as follows: First, spatiotemporal registration processing is performed on the strain signal of the micro strain sensor array, the temperature signal of the temperature sensor, the three-phase current signal of the motor from the current sensor, and the vibration signal from the vibration acceleration sensor in the sensor network. The spatiotemporal registration process is as follows: based on a high-precision hardware clock, the sampling times of strain signals, temperature signals, motor three-phase current signals, and vibration signals are aligned to a unified time reference for time synchronization processing; and the measured values ​​of the micro strain sensor array, temperature sensor, current sensor, and vibration acceleration sensor are mapped to the same preset reference coordinate system of the mechanical connection interface between the frameless torque motor and the reducer inside the joint module for unified spatial coordinate processing. Next, physical features were extracted from the spatiotemporally registered strain signal, temperature signal, three-phase motor current signal, and vibration signal. The physical feature extraction process is as follows: From the strain signal, the average strain value of each measurement point within the current sampling period is extracted to form the spatial strain distribution characteristics; From the temperature signal, the temperature difference between key points of the mechanical connection interface in the radial and axial directions is calculated to form a temperature gradient feature. Spectral analysis of the three-phase current signal of the motor is performed to extract the total harmonic distortion rate and the amplitude of the fifth and seventh harmonic components, thus forming the current harmonic distortion characteristics. The vibration signal is subjected to short-time Fourier transform or wavelet packet decomposition to calculate the signal energy in the main resonant frequency band of the mechanical connection interface, thus forming the vibration mode energy characteristics.

[0007] In a preferred embodiment, the specific process of generating the first fused feature vector is as follows: The first fused feature vector is achieved through a nonlinear weighted fusion operation, which takes spatial strain distribution characteristics, temperature gradient characteristics, current harmonic distortion characteristics, and vibration mode energy characteristics as inputs. The nonlinear weighted fusion operation includes the following process: applying a first nonlinear function transformation to the spatial strain distribution characteristics to obtain the characteristics after the first nonlinear transformation; Apply a second nonlinear function transformation to the temperature gradient characteristics to obtain the characteristics after the second nonlinear transformation. Applying a third nonlinear function transformation to the vibration mode energy characteristics yields the characteristics after the third nonlinear transformation. Construct a first adaptive coupling weight function, which receives temperature gradient features and outputs a first weight value; construct a second adaptive coupling weight function, which receives spatial strain distribution features and outputs a second weight value; construct a time-varying sensitivity weight function, which receives the derivative of vibration mode energy features with respect to time and outputs a third weight value; construct a joint modulation factor function, which simultaneously receives spatial strain distribution features and temperature gradient features and outputs a fourth weight value. Multiply the first nonlinear transformation feature by the first weight value to obtain the first weighted feature; multiply the second nonlinear transformation feature by the second weight value to obtain the second weighted feature; multiply the current harmonic distortion feature by the third weight value to obtain the third weighted feature. The third nonlinear transformation feature is multiplied by the fourth weight value to obtain the fourth weighted feature; the first weighted feature, the second weighted feature, the third weighted feature and the fourth weighted feature are combined in sequence to generate the first fused feature vector.

[0008] In a preferred embodiment, the specific process of establishing the deformation field reconstruction model based on the pre-constructed combination of parameterized basis functions in step S2 is as follows: In advance, through finite element analysis, the deformation modes of the mechanical connection interface under various typical stress and heat conditions are obtained. From this, a set of basis functions that can characterize at least two of the deformation modes, including radial expansion, axial bending, local ellipticization and higher-order torsion, are extracted to form a parameterized basis function combination. The deformation field reconstruction model expresses the full-field deformation distribution of the mechanical connection interface at the current moment as a linear superposition of all basis functions in the parameterized basis function combination, each multiplied by a corresponding basis function weight coefficient, and then summed. Here, each basis function weight coefficient is a variable to be solved, and all basis function weight coefficients constitute a set of variables to be solved. By solving for a set of optimal basis function weight coefficients, the corresponding full-field deformation distribution can be determined.

[0009] In a preferred embodiment, the optimization objective of minimizing the error between the first fused feature vector and the predicted feature vectors of each reconstructed model is achieved by defining and minimizing a composite loss function. The reconstructed model predicts the feature vector by calculating a forward prediction model, which takes a set of candidate basis function weight coefficients as input and outputs the predicted feature vector that the sensor network should measure under the deformation field determined by the set of candidate basis function weight coefficients. The composite loss function is composed of the weighted data fitting term and the adaptive sparse regularization term. The calculation process of the weighted data fitting term is as follows: First, calculate the difference vector between the predicted feature vector of the reconstructed model and the first fused feature vector. Then, calculate the weighted sum of squares of the difference vector. The weight matrix used for weighting is a diagonal matrix. The weight values ​​in the diagonal matrix corresponding to each of the spatial strain distribution feature, temperature gradient feature, current harmonic distortion feature, and vibration mode energy feature are positive numbers that are pre-set according to the relative importance of the corresponding feature in the state characterization. The calculation process for the adaptive sparse regularization term is as follows: First, take the absolute value of the weight coefficient of each basis function. Then, divide each absolute value by the sum of a positive smoothing constant and the absolute value itself. Finally, sum the calculation results corresponding to all basis functions and multiply by a positive regularization strength coefficient.

[0010] In a preferred embodiment, the specific process of driving the first swarm intelligent optimization algorithm to perform online solution is as follows: The first swarm intelligence optimization algorithm adopts the particle swarm optimization algorithm. First, a certain number of particles are randomly initialized in the parameter space spanned by the basis function weight coefficients. The position vector of each particle represents a set of candidate basis function weight coefficients. In each iteration, the particle position vector is input into the forward prediction model to calculate the corresponding reconstructed model prediction feature vector, and then the fitness value of the particle is calculated based on the composite loss function. The algorithm records the best historical position of each particle and the best global historical position of the entire particle swarm. The particle updates its velocity and position according to its current velocity, its own historical best position and the global historical best position, according to a preset acceleration coefficient and inertia weight, and introduces a chaotic perturbation strategy during the update process. The iterative update process continues until the preset maximum number of iterations is met or the fitness value change is less than a set threshold. At this point, the set of basis function weight coefficients corresponding to the global historical best position is used as the final set of optimal basis function weight coefficients obtained by solving the problem. Then, the obtained set of optimal basis function weight coefficients are substituted into the linear superposition form defined by the deformation field reconstruction model for calculation. That is, each optimal basis function weight coefficient is multiplied by its corresponding basis function, and then all the product results are summed to calculate the full-field deformation distribution of the mechanical connection interface at the current moment.

[0011] In a preferred embodiment, step S3, which involves simulating multiple predefined compensation strategies in parallel within the digital twin and evaluating the performance indicators of the digital twin, is as follows: First, the calculated full-field deformation distribution is applied as a geometric displacement constraint to the geometric model of the joint module contained in the digital twin, so as to reproduce the actual deformation state of the current mechanical connection interface in the virtual environment. Meanwhile, a compensation strategy space is defined, which is composed of the control parameters of multiple micro-displacement actuators. Each predefined compensation strategy in the compensation strategy space is expressed as a multi-dimensional vector. The number of dimensions of the multi-dimensional vector is equal to the number of micro-displacement actuators. The value of each dimension in the multi-dimensional vector represents the control command value for one of the corresponding micro-displacement actuators. Next, in the digital twin, multiple compensation strategies selected from the compensation strategy space are simulated in parallel. For each compensation strategy being simulated, the digital twin simulates the displacement of the micro-displacement actuator under the action of the compensation strategy and calculates the resulting changes in the model structure state of the joint module in the digital twin. After the simulation, for each compensation strategy, multiple performance indicators are extracted from the digital twin, including at least the reciprocal of the standard deviation of the stress distribution at the joint surface obtained from the simulation, the coaxiality error of the spindle after compensation, the total energy consumption of the actuator, and the maximum actuator load rate. Based on a preset multi-objective aggregation evaluation function, these performance indicators are comprehensively calculated to obtain the comprehensive evaluation value corresponding to each candidate compensation strategy.

[0012] In a preferred embodiment, the specific process of using the second swarm intelligence optimization algorithm to perform optimization and calculate the twin prediction feature vector is as follows: The second swarm intelligence optimization algorithm is an improved bat algorithm; a group of bat individuals is initialized in the compensation policy space, and the position vector of each bat individual represents a candidate compensation policy; In each iteration, the fitness is calculated based on the compensation strategy corresponding to the individual bat position vector and its comprehensive evaluation value in the digital twin; The comprehensive evaluation value is calculated through a multi-objective aggregate evaluation function, which is the sum of four components: the first component is the product of the first weighting coefficient and a numerical value, which is the reciprocal of the sum of the standard deviation of the stress distribution on the bonding surface and a minimal positive smoothing constant; the second component is the product of the second weighting coefficient and the compensated spindle coaxiality error; the third component is the product of the third weighting coefficient and the total energy consumption of the actuator; and the fourth component is the product of the fourth weighting coefficient and the value of a natural exponential function with the product of the maximum actuator load rate and the nonlinear penalty coefficient as the exponent. The first weighting coefficient, the second weighting coefficient, the third weighting coefficient, the fourth weighting coefficient, the smoothing constant, and the penalty coefficient are all preset positive real numbers; In the iterative process of the improved bat algorithm, individual bats update their positions based on their own pulse frequency and speed, and accept the new position with pulse emission rate. If the new position is more suitable, the position is updated and the loudness and emission rate are adjusted. The pulse frequency of an individual bat is negatively correlated with the improvement of its historical best fitness. After iteration to convergence, the compensation strategy corresponding to the global best bat individual position vector is determined as the optimal compensation control command. Based on the state after applying the full-field deformation distribution and executing the optimal command based on the digital twin, simulated strain signals, temperature signals, three-phase current signals of the motor, and vibration signals are obtained. The obtained simulated strain signals, simulated temperature signals, simulated three-phase current signals of the motor, and simulated vibration signals are processed according to the process of time synchronization, spatial coordinate unification, feature extraction processing, and generation of the first fused feature vector for multi-physics field signals, and a twin prediction feature vector is calculated.

[0013] In a preferred embodiment, the specific process of calculating the residual vector between the second fused feature vector and the twin prediction feature vector in step S4 is as follows: First, the optimal compensation control command is sent to the micro-displacement actuator located at the mechanical connection interface; the micro-displacement actuator is controlled to generate displacement according to the optimal compensation control command to perform physical compensation on the mechanical connection interface; After the physical compensation action is completed, wait for a stabilization period of a preset duration; After the physical compensation action is completed and a stabilization period has passed, perform the same multi-physics field signal acquisition, time synchronization, spatial coordinate unification, feature extraction and nonlinear weighted fusion operations as in step S1 to generate a second fused feature vector that characterizes the mechanical state of the mechanical connection interface after compensation. Next, the difference between the second fused feature vector and the twin predicted feature vector is calculated to obtain the residual vector; Next, a comprehensive residual metric is calculated, which consists of two terms added together: the first term is the weighted sum of squares of the residual vectors, and the weight matrix used for weighting is a diagonal matrix. Each weight value on the diagonal of this diagonal matrix is ​​pre-set according to the relative importance of the corresponding physical feature in the state representation on which the first fused feature vector is relied upon. The second term is the gradient penalty term, which is the product of the L2 norm of the gradient vector of the forward prediction model with respect to its internal adjustable parameters and a preset positive gradient penalty coefficient. This comprehensive residual metric is used to assess the overall level of deviation between the twin's predicted feature vector and the second fused feature vector obtained after physical compensation.

[0014] In a preferred embodiment, the specific process of collaborative online correction of the first swarm intelligent optimization algorithm and the digital twin is as follows: The correction process for the first swarm intelligent optimization algorithm is as follows: based on the calculated residual vector, dynamically adjust the weight matrix of the weighted data fitting term of the first swarm intelligent optimization algorithm in the composite loss function; The specific adjustment process is as follows: For each component in the residual vector, multiply the negative of the absolute value of the component by a preset positive attenuation coefficient, and multiply it by the original corresponding weight value to obtain a new weight value. The correction process for the digital twin is as follows: the key uncertain physical parameters in the digital twin are used to form a state vector to be estimated; The twin prediction feature vector is defined as the observation function of this state vector; the calculated residual vector is used as the observation information. The extended Kalman filter algorithm is used to process observational information and update the estimated value of the state vector online, thereby calibrating the digital twin; The corrected weight matrix of the first swarm intelligent optimization algorithm obtained through the above correction process and the estimated values ​​of the updated physical parameters of the digital twin will be used in subsequent control cycles during operation.

[0015] The beneficial effects of this invention are as follows: By constructing a fully closed-loop control system encompassing multi-physics field perception, intelligent deformation field inversion, digital twin decision-making, physical execution, and online learning, the dynamic maintenance and autonomous optimization of the assembly accuracy of the joint module are achieved; it can perceive and compensate for micro-deformation of the mechanical connection interface caused by thermal coupling under complex working conditions in real time, effectively suppressing online accuracy decay; deformation field reconstruction and strategy optimization based on swarm intelligence, combined with the virtual simulation of the digital twin, ensure the accuracy and foresight of compensation decisions; and through a residual-driven collaborative online correction mechanism, the perception model and digital twin can continuously evolve with operation, thereby continuously improving the system's adjustment accuracy, robustness, and adaptability to equipment performance degradation, ultimately ensuring the long-term operational stability and reliability of the joint module. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] Example 1 This embodiment provides, for example Figure 1 The method for adaptive adjustment and control of joint module assembly accuracy, as shown, specifically includes the following steps: Step S1: During the operation of the joint module, the multi-physics field signal is acquired by the sensor network in the area where the mechanical connection interface between the frameless torque motor and the reducer is located inside the joint module. The acquired multi-physics field signal is then processed for time synchronization, spatial coordinate unification and feature extraction to generate a first fusion feature vector that represents the current mechanical state of the mechanical connection interface. Step S2: Based on the pre-constructed parameterized basis function combination, establish a deformation field reconstruction model; use the first swarm intelligent optimization algorithm to generate multiple sets of candidate basis function weight coefficients during the algorithm iteration process; for each set of candidate basis function weight coefficients, calculate a corresponding reconstruction model prediction feature vector from the deformation field reconstruction model; with minimizing the error between the first fused feature vector and each reconstruction model prediction feature vector as the optimization objective, drive the first swarm intelligent optimization algorithm to solve online, and obtain a set of optimal basis function weight coefficients; calculate the full-field deformation distribution of the mechanical connection interface at the current moment based on the set of optimal basis function weight coefficients and the parameterized basis function combination. Step S3: Input the full-field deformation distribution as a boundary condition into the pre-constructed digital twin; in the digital twin, simulate multiple predefined compensation strategies in parallel, and use the second swarm intelligent optimization algorithm to evaluate and optimize the performance index of the digital twin after the execution of each compensation strategy to obtain the optimal compensation control command for the current deformation state; based on the result of the digital twin simulation after executing the optimal compensation control command, calculate a twin prediction feature vector; Step S4: Send the optimal compensation control command to the micro-displacement actuator set at the mechanical connection interface to control the micro-displacement actuator to generate a predetermined displacement to implement physical compensation; after the physical compensation action is completed, collect the multi-physics field signal of the sensor network again and generate a second fusion feature vector; calculate the residual vector between the second fusion feature vector and the twin prediction feature vector; based on the residual vector, perform collaborative online correction on the first swarm intelligent optimization algorithm and the digital twin.

[0021] In this embodiment, the specific process of time synchronization, spatial coordinate unification, and feature extraction of the acquired multi-physics field signals in step S1 is as follows: First, spatiotemporal registration processing is performed on the strain signal of the micro strain sensor array, the temperature signal of the temperature sensor, the three-phase current signal of the motor from the current sensor, and the vibration signal from the vibration acceleration sensor in the sensor network. The spatiotemporal registration process is as follows: Based on a high-precision hardware clock or interpolation algorithm, the sampling times of strain signals, temperature signals, three-phase motor current signals, and vibration signals are aligned to a unified time reference for time synchronization. The high-precision hardware clock can employ a synchronous clock source with an accuracy better than 1 microsecond to ensure the timing consistency of multi-channel data acquisition, thus providing an accurate time correlation basis for subsequent analysis. Linear interpolation or spline interpolation algorithms can be used to unify signals with different sampling rates onto the highest sampling rate or a common, faster virtual time axis, eliminating phase errors caused by minor deviations in sampling times. Additionally, the micro-strain transmitter... The measured values ​​of the sensor array, temperature sensor, current sensor, and vibration acceleration sensor are mapped to the same preset reference coordinate system of the mechanical connection interface between the frameless torque motor and the reducer inside the joint module for unified spatial coordinate processing. The preset reference coordinate system is usually based on the geometric center of the mechanical connection interface as the origin, the motor axis as the Z-axis, and two orthogonal directions in the plane perpendicular to the axis as the X-axis and Y-axis. The spatial position of each sensor is determined by calibration measurement, and its measured value is transformed to this unified coordinate system through a coordinate transformation matrix, so that the distribution of field quantities such as strain and temperature can be fused and compared for analysis within the same spatial framework. Next, physical features were extracted from the spatiotemporally registered strain signal, temperature signal, three-phase motor current signal, and vibration signal. The physical feature extraction process is as follows: The average strain value of each measurement point within the current sampling period is extracted from the strain signal to form the spatial strain distribution characteristics. The current sampling period is set according to the working frequency of the joint module and the signal change rate, usually between 1 millisecond and 100 milliseconds, in order to balance real-time performance and data stability. From the temperature signal, calculate the temperature difference between key points in the radial and axial directions of the mechanical connection interface to form a temperature gradient feature. The key points are selected near the heat source and at the structural boundary, such as the position of the motor winding corresponding to the housing and the position of the outer flange of the reducer. The radial temperature difference reflects the uneven radial heat dissipation caused by the internal heat source, and the axial temperature difference reflects the heat conduction state along the force transmission path. Spectral analysis of the three-phase current signal of the motor is performed to extract the total harmonic distortion rate and the amplitudes of the fifth and seventh harmonic components, thus forming the current harmonic distortion characteristics. The total harmonic distortion rate reflects the overall non-sinusoidal degree of the current waveform and is related to the motor load, efficiency, and uniformity of the stator and rotor air gap. The fifth and seventh harmonics are characteristic harmonics caused by common spatial harmonics in permanent magnet synchronous motors, and their amplitude changes can indirectly reflect the change in the symmetry of the magnetic field of the motor under non-uniform force or deformation. The vibration signal is subjected to short-time Fourier transform or wavelet packet decomposition to calculate the signal energy within the main resonant frequency band of the mechanical connection interface, thus forming the vibration modal energy characteristics. The main resonant frequency band is obtained in advance through experimental modal analysis or finite element simulation, and usually includes a certain bandwidth near the first or second natural frequencies of the mechanical connection interface in the axial, radial, and torsional directions. The window function length of the short-time Fourier transform and the basis functions and decomposition levels of the wavelet packet decomposition are selected according to the frequency band range and resolution requirements. The specific process of generating the first fused feature vector is as follows: The first fused feature vector is achieved through a nonlinear weighted fusion operation, which takes spatial strain distribution characteristics, temperature gradient characteristics, current harmonic distortion characteristics, and vibration mode energy characteristics as inputs. The nonlinear weighted fusion operation includes the following processing steps: applying a first nonlinear function transformation to the spatial strain distribution characteristics to obtain the first nonlinear transformed characteristics; the first nonlinear function transformation can be the Sigmoid function, the hyperbolic tangent function or its variants, which maps the strain values ​​to a bounded, more sensitive interval, thereby enhancing the model's ability to learn nonlinear deformation responses; Apply a second nonlinear function transformation to the temperature gradient characteristics to obtain the second nonlinear transformed characteristics; the second nonlinear function transformation can use the same or different function as the first nonlinear function transformation to highlight the nonlinear effect of temperature gradient change, such as the nonlinear segment of material properties changing with temperature; A third nonlinear function transformation is applied to the vibration modal energy characteristics to obtain the characteristics after the third nonlinear transformation. The third nonlinear function transformation can be either the ReLU function or the Sigmoid function with a threshold, which is used to suppress background noise energy and amplify significant vibration modal components that exceed a specific energy threshold. The threshold can be set according to the background vibration energy level during no-load steady-state operation. A first adaptive coupling weight function is constructed, which receives temperature gradient features and outputs a first weight value. This first adaptive coupling weight function can be designed as a positive correlation function, such as a linear function or an exponential function, with the norm or maximum component of the temperature gradient feature vector as input. Its physical meaning is that when a large temperature gradient is detected, it indicates a significant thermal effect, and the contribution weight of strain features in the fusion should be increased to reflect the dominant role of thermal stress coupling. A second adaptive coupling weight function is constructed, which receives spatial strain distribution features and outputs a second weight value. This second adaptive coupling weight function can be designed as a function with the norm or uniformity index of the spatial strain distribution feature vector as input. When the strain distribution is non-uniform or there is high stress concentration, the weight of the temperature gradient features is increased because high-stress areas are more sensitive to temperature changes and the thermoelastic effect is more pronounced. A time-varying sensitivity weight function is constructed. The time-varying sensitivity weight function receives the time derivative of the vibration modal energy characteristics and outputs a third weight value. This time-varying sensitivity weight function can be designed to nonlinearly amplify the absolute value of the derivative, for example, by using a square function or an exponential function, so that when the vibration energy changes rapidly, the weight of the current harmonic distortion characteristics increases sharply, thereby quickly capturing the electromechanical coupling response caused by transient events such as impact loads. A joint modulation factor function is constructed, which simultaneously receives the spatial strain distribution characteristics and the temperature gradient characteristics and outputs a fourth weight value. This joint modulation factor function can be designed as a function that measures the two characteristics in some way, for example, by calculating the dot product of the two and mapping it to the [0,1] interval through a Sigmoid function, which is used to quantify the degree of coordination between the current thermal field and the stress field, and modulate the contribution of the vibration energy characteristics accordingly. The vibration characteristics under high thermo-mechanical coordination often contain richer information on the degradation of the connection state. Multiply the first nonlinear transformation feature by the first weight value to obtain the first weighted feature; multiply the second nonlinear transformation feature by the second weight value to obtain the second weighted feature; multiply the current harmonic distortion feature by the third weight value to obtain the third weighted feature. The third nonlinear transformation feature is multiplied by the fourth weight value to obtain the fourth weighted feature. The first, second, third, and fourth weighted features are combined in sequence to generate the first fused feature vector. This sequential combination is usually a vector concatenation operation. The final first fused feature vector not only contains the independent information of each physical field, but also embeds the dynamic coupling relationship between fields through an adaptive weighting mechanism, making it a more discriminative and robust input descriptor for subsequent deformation field reconstruction.

[0022] In this embodiment, the specific process of establishing the deformation field reconstruction model based on the pre-constructed parameterized basis function combination in step S2 is as follows: Beforehand, deformation modes of the mechanical connection interface under various typical stress and heat conditions are obtained through finite element analysis or experimental modal analysis. A set of basis functions that can characterize at least two deformation modes among radial expansion, axial bending, local ellipticization, and higher-order torsion are extracted to form a parameterized basis function combination. Finite element analysis can establish a detailed model in the software, including the stator of the frameless torque motor, the reducer housing, and its connection structure, and apply various load conditions, including axial preload, radial torque, and non-uniform temperature field, to perform static and thermodynamic coupled simulation. The displacement field of the key interface of the model under each condition is extracted as a candidate deformation mode. Experimental modal analysis can perform modal testing on the actual object or prototype through excitation hammer or vibrator, measure the frequency response function through sensors, and identify the main vibration modes of the interface. The deformation field reconstruction model expresses the full-field deformation distribution of the mechanical connection interface at the current moment as a linear superposition of all basis functions in the parameterized basis function combination, each multiplied by a corresponding basis function weight coefficient, and then summed. Here, each basis function weight coefficient is a variable to be solved, and all basis function weight coefficients constitute a set of variables to be solved. By solving for a set of optimal basis function weight coefficients, the corresponding full-field deformation distribution can be determined. The physical significance of this linear superposition form is that it decomposes the complex, continuous spatial deformation field into a weighted combination of several known, basic deformation modes (basis functions), thereby transforming the infinite-dimensional field estimation problem into a finite-dimensional parameter (weight coefficient) estimation problem, which greatly reduces the complexity of subsequent optimization and enhances the physical interpretability of the model. The optimization objective is to minimize the error between the first fused feature vector and the predicted feature vectors of each reconstructed model, which is achieved by defining and minimizing a composite loss function. The reconstructed model predicts the feature vector through a forward prediction model. This forward prediction model takes a set of candidate basis function weight coefficients as input and outputs the predicted feature vector that the sensor network should measure under the deformation field determined by the set of candidate basis function weight coefficients. This forward prediction model is essentially a mapping from the deformation field to the sensor reading. Its establishment depends on the physical location, sensitivity, and feature extraction logic in step S1 of the sensor network. It can be constructed by embedding a "virtual sensor" corresponding to the physical sensor in the digital twin, or by training a surrogate model with a large amount of calibration data. The composite loss function is composed of the weighted data fitting term and the adaptive sparse regularization term. The calculation process for the weighted data fitting term is as follows: First, calculate the difference vector between the predicted feature vector of the reconstructed model and the first fused feature vector. Then, calculate the weighted sum of squares of this difference vector. The weight matrix used for weighting is a diagonal matrix. The weight values ​​in this diagonal matrix corresponding to each of the spatial strain distribution feature, temperature gradient feature, current harmonic distortion feature, and vibration mode energy feature are positive numbers pre-set based on the relative importance or reliability of the corresponding feature in the state characterization. For example, since the strain feature directly reflects mechanical deformation, its reliability is usually the highest, so it can be given a larger weight value, such as set between 0.8 and 1.0. The temperature gradient feature is second, and the weight range can be between 0.5 and 0.8. The current and vibration features are indirectly reflected and are subject to more noise interference, so the weight range can be set between 0.2 and 0.5. The specific values ​​can be determined during the system calibration stage by evaluating the correlation between each feature and the reference deformation measurement. The calculation process for the adaptive sparse regularization term is as follows: First, take the absolute value of the weight coefficient of each basis function. Then, divide each absolute value by the sum of a positive smoothing constant and the absolute value itself. Finally, sum the calculation results corresponding to all basis functions and multiply by a positive regularization strength coefficient. The smoothing constant is usually set to a decimal much smaller than 1, such as 0.01 or 0.001, to ensure that the denominator is not zero and that the function is differentiable when the weight coefficients are close to zero. The regularization strength coefficient controls the strength of the sparse constraint, and its value can be determined by cross-validation, typically ranging from 0.01 to 0.1. The role of this term is to guide the optimization solution to tend towards a "sparse solution", that is, most weight coefficients tend to be close to zero, and only a few basis functions most related to the current actual deformation are significantly activated. This is consistent with physical reality (deformation is usually superimposed by a few dominant modes) and can effectively prevent overfitting to measurement noise and improve the generalization ability of the model. The specific process of driving the first swarm intelligent optimization algorithm to solve online is as follows: The first swarm intelligence optimization algorithm adopts the particle swarm optimization algorithm. First, a certain number of particles are randomly initialized in the parameter space spanned by the basis function weight coefficients. The position vector of each particle represents a set of candidate basis function weight coefficients. The number of particles is usually set to 5 to 20 times the number of weight coefficients to be solved (i.e., the number of basis functions N) to ensure the sufficiency of the search. In each iteration, the particle position vector is input into the forward prediction model to calculate the corresponding reconstructed model prediction feature vector, and then the fitness value of the particle is calculated based on the composite loss function. The algorithm records the best historical position of each particle and the best global historical position of the entire particle swarm. Particles update their velocity and position based on their current velocity, their own historical best position, and the global historical best position, according to a preset acceleration coefficient and inertia weight. A chaotic perturbation strategy is introduced during the update process. The acceleration coefficient typically includes a cognitive acceleration coefficient and a social acceleration coefficient, both of which can be set between 1.5 and 2.0. The inertia weight can adopt a linear decreasing strategy, for example, decreasing from 0.9 to 0.4, to strengthen global exploration in the early stages of the search and local development in the later stages. The chaotic perturbation strategy can use chaotic sequences such as Logistic mapping to slightly perturb the positions of some particles, to help the population escape local optima and enhance global search capabilities. The iterative update process continues until the preset maximum number of iterations or the change in fitness value is less than a set threshold is met. The maximum number of iterations can be set according to the complexity of the problem and the real-time requirements, for example, 50 to 200 times. The fitness value change threshold can be set to a small positive number, such as 1e-6. When the improvement of the global optimal fitness is less than this threshold for several consecutive generations, the algorithm is considered to have converged. At this time, the set of basis function weight coefficients corresponding to the global historical optimal position is used as the final set of optimal basis function weight coefficients obtained by solving the problem. Next, the obtained set of optimal basis function weight coefficients are substituted into the linear superposition form defined by the deformation field reconstruction model for calculation. That is, each optimal basis function weight coefficient is multiplied by its corresponding basis function, and then all the product results are summed to calculate the full-field deformation distribution of the mechanical connection interface at the current moment. This calculation can be completed instantaneously after obtaining the weight coefficients. The final output full-field deformation distribution is a two-dimensional scalar field defined on the coordinates of the mechanical connection interface, expressed in the form of a data matrix or interpolation function, providing accurate and spatially continuous state input for the compensation decision in step S3.

[0023] In this embodiment, the specific process of simulating multiple predefined compensation strategies in parallel within the digital twin and evaluating the performance indicators of the digital twin in step S3 is as follows: First, the calculated full-field deformation distribution is applied as a geometric displacement constraint to the geometric model of the joint module contained in the digital twin, so as to reproduce the actual deformation state of the current mechanical connection interface in the virtual environment. The application process is realized through the boundary condition setting function of the finite element analysis software to ensure that the deformation field is accurately mapped to the displacement of the key nodes or surfaces of the model. Simultaneously, a compensation strategy space is defined, consisting of a combination of control parameters from multiple micro-displacement actuators. Each predefined compensation strategy in the compensation strategy space is expressed as a multi-dimensional vector. The number of dimensions of the multi-dimensional vector is equal to the number of micro-displacement actuators. The value of each dimension in the multi-dimensional vector represents the control command value for one of the corresponding micro-displacement actuators. This control command value is usually a voltage or current signal, and its range is determined according to the rated operating parameters of the actuator. For example, for piezoelectric ceramic actuators, the voltage range may be between 0V and 150V. Next, in the digital twin, multiple compensation strategies selected from the compensation strategy space are simulated in parallel. For each compensation strategy being simulated, the digital twin simulates the displacement of the micro-displacement actuator under the action of the compensation strategy and calculates the resulting changes in the model structure state of the joint module in the digital twin. The simulation calculation is performed using a multiphysics coupled solver, taking into account mechanical deformation, contact nonlinearity, and piezoelectric inverse effects, in order to obtain a high-fidelity structural response. After the simulation, for each compensation strategy, several performance indicators are extracted from the digital twin, including the reciprocal of the standard deviation of the stress distribution at the joint surface obtained from the simulation, the coaxiality error of the spindle after compensation, the total energy consumption of the actuator, and the maximum actuator load rate. The reciprocal of the standard deviation of the stress distribution at the joint surface is obtained by calculating the equivalent stress value of all contact units or nodes, first calculating its standard deviation, and then taking the reciprocal. The larger this value, the more uniform the stress distribution. The coaxiality error of the spindle is calculated by comparing the deviation of the center position of the motor shaft and the reducer input shaft at the key section after compensation. The total energy consumption of the actuator is estimated by integrating the instantaneous power consumed by all actuators during operation. The maximum actuator load rate is the highest value of the ratio of the output force (or torque) of each actuator to its rated value. Based on a preset multi-objective aggregation evaluation function, these performance indicators are comprehensively calculated to obtain the comprehensive evaluation value corresponding to each candidate compensation strategy. The specific process of using the second swarm intelligence optimization algorithm to find the optimal solution and calculate the twin prediction feature vector is as follows: The second swarm intelligence optimization algorithm adopts an improved bat algorithm; within the compensation policy space, a group of bat individuals is initialized, and the position vector of each bat individual represents a candidate compensation policy; the number of bat individuals is usually set to 10 to 30 times the dimension of the compensation policy space (i.e. the number of executors) to ensure population diversity; In each iteration, the fitness of each bat individual is calculated based on the compensation strategy corresponding to its current position vector and the comprehensive evaluation value obtained from simulation in the digital twin. The comprehensive evaluation value is calculated through a multi-objective aggregate evaluation function, which is the sum of four components. The first component is the product of the first weighting coefficient and a numerical value, which is the reciprocal of the sum of the standard deviation of the stress distribution on the bonding surface and a minimum positive smoothing constant. The minimum positive smoothing constant is usually set to a value much less than 1, such as 0.001, to prevent the denominator from being zero. The second item is the product of the second weighting coefficient and the compensated spindle coaxiality error; The third item is the product of the third weighting coefficient and the total energy consumption of the actuator; The fourth sub-item is the product of the fourth weighting coefficient and the power function of the natural constant e, where the exponent of the power function is the product of a positive nonlinear penalty coefficient and the maximum actuator load rate. Among them, the first weighting coefficient, the second weighting coefficient, the third weighting coefficient, the fourth weighting coefficient, the minimum positive smoothing constant, and the nonlinear penalty coefficient are all preset positive real numbers; each weighting coefficient can be adjusted according to the focus of the application scenario. For example, in a scenario where accuracy is prioritized, the second weighting coefficient can be set to a larger value, such as 1.0, while the third weighting coefficient can be set to a smaller value, such as 0.1; the nonlinear penalty coefficient can be set between 1 and 5 to achieve a significant penalty for high load rates; During the iterative process of the improved bat algorithm, each individual bat updates its position vector in the compensation strategy space based on its own pulse frequency and current flight speed, thereby generating a new candidate compensation strategy, i.e., a new position vector. For the newly generated position vector, the improved bat algorithm accepts the new position vector with a pulse emission rate, even if its fitness may be worse than the original position vector; the initial pulse emission rate can be set to around 0.5 and dynamically adjusted with iteration; If the fitness of the new position vector is better than the bat's own historical best fitness, then not only is the bat's current position vector updated with the new position vector, but the loudness of the bat is also reduced and the impulse emission rate of the bat is increased, so that the bat tends to perform a more refined local search near the better position vector; the initial value of the loudness can be set to 1.0, and it is attenuated by a certain proportion when a better solution is found, for example, multiplied by 0.9. Among them, the pulse frequency of an individual bat is negatively correlated with the improvement of its own historical best fitness in multiple iterations. When the improvement decreases, the range of pulse frequency variation increases to enhance global exploration capabilities. The pulse frequency range can be set between 0 and 2, and its upper and lower limits can be dynamically adjusted according to the improvement. The iterative process continues until the preset convergence condition is met. The convergence condition can be reaching the maximum number of iterations (e.g., 100 times) or the improvement of the global optimal fitness over multiple generations (e.g., 10 generations) being less than a threshold (e.g., 1e-4). Finally, the position vector corresponding to the individual with the best fitness in the entire bat population, i.e. the compensation strategy, is determined as the optimal compensation control command. Subsequently, based on the final state of the digital twin after applying full-field deformation distribution and executing optimal compensation control commands, simulated strain signals, temperature signals, motor three-phase current signals, and vibration signals are acquired through the simulation function of the digital twin. The simulated signals are obtained by querying the simulation results (such as displacement, temperature, current density, and acceleration) at the virtual sensor installation location at the end of the simulation. The acquired simulated strain signals, simulated temperature signals, simulated motor three-phase current signals, and simulated vibration signals are processed according to the process of time synchronization, spatial coordinate unification, feature extraction, and generation of the first fused feature vector for multi-physics field signals, and a twin predicted feature vector is calculated. This twin predicted feature vector represents the ideal sensing characteristic state that the system should exhibit under the ideal model and optimal compensation, providing an accurate digital benchmark for verifying the actual compensation effect in step S4.

[0024] In this embodiment, the specific process of calculating the residual vector between the second fused feature vector and the twin prediction feature vector in step S4 is as follows: First, the optimal compensation control command is sent to the micro-displacement actuator located at the mechanical connection interface; the micro-displacement actuator is controlled to generate displacement according to the optimal compensation control command to perform physical compensation on the mechanical connection interface; After the physical compensation action is completed, a stabilization period of a preset duration is waited for. The stabilization period is set to allow the internal stress of the mechanical structure to redistribute and reach a new dynamic equilibrium after being actively intervened. This duration is usually determined according to the size, material and actuator response speed of the joint module, and is generally between 10 milliseconds and 100 milliseconds, to ensure that the sensor signals collected subsequently can reflect the stable state after compensation. After the physical compensation action is completed and a stabilization period has passed, perform the same multi-physics field signal acquisition, time synchronization, spatial coordinate unification, feature extraction and nonlinear weighted fusion operations as in step S1 to generate a second fused feature vector that characterizes the mechanical state of the mechanical connection interface after compensation. Next, the difference between the second fused feature vector and the twin predicted feature vector is calculated to obtain the residual vector; Next, a comprehensive residual metric is calculated, which consists of two terms: the first term is the weighted sum of squares of the residual vectors, and the weight matrix used for weighting is a diagonal matrix. Each weight value on the diagonal of this diagonal matrix is ​​pre-set based on the relative importance or reliability of the corresponding physical feature in the state representation on which the first fused feature vector was relied upon. For example, strain features and temperature gradient features that directly reflect deformation are given higher weights, while current and vibration features that are affected by the coupling of multiple factors are given relatively lower weights. The second term is the gradient penalty term, which is the product of the L2 norm of the gradient vector of the forward prediction model with respect to its internal adjustable parameters and a preset positive gradient penalty coefficient. The gradient penalty coefficient is usually between 0.001 and 0.1, and its role is to balance the residual minimization objective with the stability of model parameter updates. This term is added to prevent the model parameters from being adjusted to the sensitive region of maximum gradient in order to overfit the current residual during the correction process, thereby avoiding the model becoming fragile and the generalization ability decreasing, and enhancing the robustness of the online learning process. This comprehensive residual metric is used to assess the overall level of deviation between the twin's predicted feature vector and the second fused feature vector obtained after physical compensation. The specific process of collaborative online correction of the first group intelligent optimization algorithm and the digital twin is as follows: The correction process for the first swarm intelligent optimization algorithm is as follows: based on the calculated residual vector, dynamically adjust the weight matrix of the weighted data fitting term of the first swarm intelligent optimization algorithm in the composite loss function; The specific adjustment process is as follows: For each component in the residual vector, the negative of the absolute value of the component is multiplied by a preset positive attenuation coefficient. Then, the exponential function value with the natural constant e as the base and the product as the exponent is calculated. Finally, the old weight value corresponding to the index position of the residual component on the diagonal of the weight matrix is ​​multiplied by this exponential function value to obtain the new weight value. The attenuation coefficient controls the attenuation rate of the weight as the residual increases, with a typical value between 0.1 and 2.0. This adjustment mechanism realizes online re-evaluation of sensor feature confidence based on feedback: When there is a large deviation between the prediction and the actual measurement on a certain feature dimension (large residual), the system automatically reduces the weight of that feature in the subsequent deformation inversion optimization, making the optimization process more dependent on other more accurate prediction features, so that robust deformation estimation ability can still be maintained when some sensor performance degrades or is disturbed. The correction process for the digital twin is as follows: the key uncertain physical parameters in the digital twin, including the material elastic modulus, contact surface stiffness, and damping coefficient, are used to form a state vector to be estimated. The twin prediction feature vector is defined as a function of this state vector, called the observation function; the calculated residual vector is used as the observation information. The Extended Kalman Filter (EKF) algorithm is used to process observational information and update the estimated values ​​of the state vector, which consists of parameters such as material elastic modulus, contact surface stiffness, and damping coefficient, online. This calibrates the digital twin, making the output of the observation function closer to the second fused feature vector obtained from actual observation. The EKF algorithm works in two steps: prediction and updating. The prediction step estimates the state vector based on the system dynamics model. The updating step uses the residual vector (observational information) and the observation noise covariance to calculate the Kalman gain, and uses this to correct the estimated value of the state vector and the estimation error covariance. This process can effectively extract estimates of physical parameter drift (such as stiffness reduction due to material fatigue and stiffness changes due to contact surface wear) from the noisy residual signal, enabling the digital twin model to evolve synchronously with the slow degradation of the physical entity and maintain its prediction fidelity. The corrected weight matrix of the first group intelligent optimization algorithm obtained through the above correction process, along with the updated physical parameter estimates of the digital twin, will be used in subsequent sensing, estimation, decision-making, and compensation control cycles during the operation of the joint module. This collaborative online correction mechanism transforms the entire adaptive adjustment system from a static, open-loop controller into an intelligent system with continuous learning and self-optimization capabilities. It can compensate for model errors online, adapt to slow changes in component characteristics, and automatically deweight sensor information with decreased confidence, thereby maintaining a high-precision assembly state continuously and autonomously during long-term operation, significantly improving the reliability and service life of the joint module.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for adaptive adjustment and control of joint module assembly accuracy, characterized in that, Specifically, the steps include the following: Step S1: During the operation of the joint module, the multi-physics field signal is acquired by the sensor network in the area where the mechanical connection interface between the frameless torque motor and the reducer is located inside the joint module. The acquired multi-physics field signal is then processed for time synchronization, spatial coordinate unification and feature extraction to generate a first fusion feature vector that represents the current mechanical state of the mechanical connection interface. Step S2: Based on the pre-constructed parameterized basis function combination, establish a deformation field reconstruction model; use the first swarm intelligent optimization algorithm to generate multiple sets of candidate basis function weight coefficients during the algorithm iteration process; for each set of candidate basis function weight coefficients, calculate a corresponding reconstruction model prediction feature vector from the deformation field reconstruction model; with minimizing the error between the first fused feature vector and each reconstruction model prediction feature vector as the optimization objective, drive the first swarm intelligent optimization algorithm to solve online, and obtain a set of optimal basis function weight coefficients; calculate the full-field deformation distribution of the mechanical connection interface at the current moment based on the set of optimal basis function weight coefficients and the parameterized basis function combination. Step S3: Input the full-field deformation distribution as a boundary condition into the pre-constructed digital twin; In the digital twin, simulate multiple predefined compensation strategies in parallel, and use the second swarm intelligent optimization algorithm to evaluate and optimize the performance index of the digital twin after each compensation strategy is executed, so as to obtain the optimal compensation control command for the current deformation state. Based on the result of executing the optimal compensation control command through digital twin simulation, a twin prediction feature vector is calculated. Step S4: Send the optimal compensation control command to the micro-displacement actuator set at the mechanical connection interface to control the micro-displacement actuator to generate a predetermined displacement to implement physical compensation; after the physical compensation action is completed, collect the multi-physics field signal of the sensor network again and generate a second fusion feature vector; calculate the residual vector between the second fusion feature vector and the twin prediction feature vector; based on the residual vector, perform collaborative online correction on the first swarm intelligent optimization algorithm and the digital twin.

2. The adaptive adjustment and control method for joint module assembly accuracy according to claim 1, characterized in that: In step S1, the specific process of time synchronization, spatial coordinate unification, and feature extraction of the acquired multi-physics field signals is as follows: First, spatiotemporal registration processing is performed on the strain signal of the micro strain sensor array, the temperature signal of the temperature sensor, the three-phase current signal of the motor from the current sensor, and the vibration signal from the vibration acceleration sensor in the sensor network. The spatiotemporal registration process is as follows: based on a high-precision hardware clock, the sampling times of strain signals, temperature signals, motor three-phase current signals, and vibration signals are aligned to a unified time reference for time synchronization processing; and the measured values ​​of the micro strain sensor array, temperature sensor, current sensor, and vibration acceleration sensor are mapped to the same preset reference coordinate system of the mechanical connection interface between the frameless torque motor and the reducer inside the joint module for unified spatial coordinate processing. Next, physical features were extracted from the spatiotemporally registered strain signal, temperature signal, three-phase motor current signal, and vibration signal. The physical feature extraction process is as follows: From the strain signal, the average strain value of each measurement point within the current sampling period is extracted to form the spatial strain distribution characteristics; From the temperature signal, the temperature difference between key points of the mechanical connection interface in the radial and axial directions is calculated to form a temperature gradient feature. Spectral analysis of the three-phase current signal of the motor is performed to extract the total harmonic distortion rate and the amplitude of the fifth and seventh harmonic components, thus forming the current harmonic distortion characteristics. The vibration signal is subjected to short-time Fourier transform or wavelet packet decomposition to calculate the signal energy in the main resonant frequency band of the mechanical connection interface, thus forming the vibration mode energy characteristics.

3. The adaptive adjustment and control method for joint module assembly accuracy according to claim 2, characterized in that: The specific process for generating the first fused feature vector is as follows: The first fused feature vector is achieved through a nonlinear weighted fusion operation, which takes spatial strain distribution characteristics, temperature gradient characteristics, current harmonic distortion characteristics, and vibration mode energy characteristics as inputs. The nonlinear weighted fusion operation includes the following process: applying a first nonlinear function transformation to the spatial strain distribution characteristics to obtain the characteristics after the first nonlinear transformation; Apply a second nonlinear function transformation to the temperature gradient characteristics to obtain the characteristics after the second nonlinear transformation. Applying a third nonlinear function transformation to the vibration mode energy characteristics yields the characteristics after the third nonlinear transformation. Construct a first adaptive coupling weight function, which receives temperature gradient features and outputs a first weight value; Construct a second adaptive coupling weight function, which receives the spatial strain distribution characteristics and outputs a second weight value; construct a time-varying sensitivity weight function, which receives the derivative of the vibration mode energy characteristics with respect to time and outputs a third weight value; construct a joint modulation factor function, which simultaneously receives the spatial strain distribution characteristics and temperature gradient characteristics and outputs a fourth weight value. Multiply the first nonlinear transformation feature by the first weight value to obtain the first weighted feature; multiply the second nonlinear transformation feature by the second weight value to obtain the second weighted feature; multiply the current harmonic distortion feature by the third weight value to obtain the third weighted feature. The third nonlinear transformation feature is multiplied by the fourth weight value to obtain the fourth weighted feature; the first weighted feature, the second weighted feature, the third weighted feature and the fourth weighted feature are combined in sequence to generate the first fused feature vector.

4. The adaptive adjustment and control method for joint module assembly accuracy according to claim 3, characterized in that: In step S2, the specific process of establishing the deformation field reconstruction model based on the pre-constructed combination of parameterized basis functions is as follows: In advance, through finite element analysis, the deformation modes of the mechanical connection interface under various typical stress and heat conditions are obtained. From this, a set of basis functions that can characterize at least two of the deformation modes, including radial expansion, axial bending, local ellipticization and higher-order torsion, are extracted to form a parameterized basis function combination. The deformation field reconstruction model expresses the full-field deformation distribution of the mechanical connection interface at the current moment as a linear superposition of all basis functions in the parameterized basis function combination, each multiplied by a corresponding basis function weight coefficient, and then summed. Here, each basis function weight coefficient is a variable to be solved, and all basis function weight coefficients constitute a set of variables to be solved. By solving for a set of optimal basis function weight coefficients, the corresponding full-field deformation distribution can be determined.

5. The adaptive adjustment and control method for joint module assembly accuracy according to claim 4, characterized in that: The optimization objective is to minimize the error between the first fused feature vector and the predicted feature vectors of each reconstructed model, which is achieved by defining and minimizing a composite loss function. The reconstructed model predicts the feature vector by calculating a forward prediction model, which takes a set of candidate basis function weight coefficients as input and outputs the predicted feature vector that the sensor network should measure under the deformation field determined by the set of candidate basis function weight coefficients. The composite loss function is composed of the weighted data fitting term and the adaptive sparse regularization term. The calculation process of the weighted data fitting term is as follows: First, calculate the difference vector between the predicted feature vector of the reconstructed model and the first fused feature vector. Then, calculate the weighted sum of squares of the difference vector. The weight matrix used for weighting is a diagonal matrix. The weight values ​​in the diagonal matrix corresponding to each of the spatial strain distribution feature, temperature gradient feature, current harmonic distortion feature, and vibration mode energy feature are positive numbers that are pre-set according to the relative importance of the corresponding feature in the state characterization. The calculation process for the adaptive sparse regularization term is as follows: First, take the absolute value of the weight coefficient of each basis function. Then, divide each absolute value by the sum of a positive smoothing constant and the absolute value itself. Finally, sum the calculation results corresponding to all basis functions and multiply by a positive regularization strength coefficient.

6. The adaptive adjustment and control method for joint module assembly accuracy according to claim 5, characterized in that: The specific process of driving the first swarm intelligent optimization algorithm to solve the problem online is as follows: The first swarm intelligence optimization algorithm adopts the particle swarm optimization algorithm. First, a certain number of particles are randomly initialized in the parameter space spanned by the basis function weight coefficients. The position vector of each particle represents a set of candidate basis function weight coefficients. In each iteration, the particle position vector is input into the forward prediction model to calculate the corresponding reconstructed model prediction feature vector, and then the fitness value of the particle is calculated based on the composite loss function. The algorithm records the best historical position of each particle and the best global historical position of the entire particle swarm. The particle updates its velocity and position according to its current velocity, its own historical best position and the global historical best position, according to a preset acceleration coefficient and inertia weight, and introduces a chaotic perturbation strategy during the update process. The iterative update process continues until the preset maximum number of iterations is met or the fitness value change is less than a set threshold. At this point, the set of basis function weight coefficients corresponding to the global historical best position is used as the final set of optimal basis function weight coefficients obtained by solving the problem. Then, the obtained set of optimal basis function weight coefficients are substituted into the linear superposition form defined by the deformation field reconstruction model for calculation. That is, each optimal basis function weight coefficient is multiplied by its corresponding basis function, and then all the product results are summed to calculate the full-field deformation distribution of the mechanical connection interface at the current moment.

7. The adaptive adjustment and control method for joint module assembly accuracy according to claim 6, characterized in that: In step S3, the specific process of simulating multiple predefined compensation strategies in parallel within the digital twin and evaluating the performance indicators of the digital twin is as follows: First, the calculated full-field deformation distribution is applied as a geometric displacement constraint to the geometric model of the joint module contained in the digital twin, so as to reproduce the actual deformation state of the current mechanical connection interface in the virtual environment. Meanwhile, a compensation strategy space is defined, which is composed of the control parameters of multiple micro-displacement actuators. Each predefined compensation strategy in the compensation strategy space is expressed as a multi-dimensional vector. The number of dimensions of the multi-dimensional vector is equal to the number of micro-displacement actuators. The value of each dimension in the multi-dimensional vector represents the control command value for one of the corresponding micro-displacement actuators. Next, in the digital twin, multiple compensation strategies selected from the compensation strategy space are simulated in parallel. For each compensation strategy being simulated, the digital twin simulates the displacement of the micro-displacement actuator under the action of the compensation strategy and calculates the resulting changes in the model structure state of the joint module in the digital twin. After the simulation, for each compensation strategy, multiple performance indicators are extracted from the digital twin, including at least the reciprocal of the standard deviation of the stress distribution at the joint surface obtained from the simulation, the coaxiality error of the spindle after compensation, the total energy consumption of the actuator, and the maximum actuator load rate. Based on a preset multi-objective aggregation evaluation function, these performance indicators are comprehensively calculated to obtain the comprehensive evaluation value corresponding to each candidate compensation strategy.

8. The adaptive adjustment and control method for joint module assembly accuracy according to claim 7, characterized in that: The specific process of using the second swarm intelligence optimization algorithm to perform optimization and calculate the twin prediction feature vector is as follows: The second swarm intelligence optimization algorithm is an improved bat algorithm; a group of bat individuals is initialized in the compensation policy space, and the position vector of each bat individual represents a candidate compensation policy; In each iteration, the fitness is calculated based on the compensation strategy corresponding to the individual bat position vector and its comprehensive evaluation value in the digital twin; The comprehensive evaluation value is calculated through a multi-objective aggregate evaluation function, which is the sum of four components: the first component is the product of the first weighting coefficient and a numerical value, which is the reciprocal of the sum of the standard deviation of the stress distribution on the bonding surface and a minimal positive smoothing constant; the second component is the product of the second weighting coefficient and the compensated spindle coaxiality error; the third component is the product of the third weighting coefficient and the total energy consumption of the actuator; and the fourth component is the product of the fourth weighting coefficient and the value of a natural exponential function with the product of the maximum actuator load rate and the nonlinear penalty coefficient as the exponent. The first weighting coefficient, the second weighting coefficient, the third weighting coefficient, the fourth weighting coefficient, the smoothing constant, and the penalty coefficient are all preset positive real numbers; In the iterative process of the improved bat algorithm, individual bats update their positions based on their own pulse frequency and speed, and accept the new position with pulse emission rate. If the new position is more suitable, the position is updated and the loudness and emission rate are adjusted. The pulse frequency of an individual bat is negatively correlated with the improvement of its historical best fitness. After iteration to convergence, the compensation strategy corresponding to the global best bat individual position vector is determined as the optimal compensation control command. Based on the state after applying the full-field deformation distribution and executing the optimal command based on the digital twin, simulated strain signals, temperature signals, three-phase current signals of the motor, and vibration signals are obtained. The obtained simulated strain signals, simulated temperature signals, simulated three-phase current signals of the motor, and simulated vibration signals are processed according to the process of time synchronization, spatial coordinate unification, feature extraction processing, and generation of the first fused feature vector for multi-physics field signals, and a twin prediction feature vector is calculated.

9. The adaptive adjustment and control method for joint module assembly accuracy according to claim 8, characterized in that: In step S4, the specific process of calculating the residual vector between the second fused feature vector and the twin prediction feature vector is as follows: First, the optimal compensation control command is sent to the micro-displacement actuator located at the mechanical connection interface; the micro-displacement actuator is controlled to generate displacement according to the optimal compensation control command to perform physical compensation on the mechanical connection interface; After the physical compensation action is completed, wait for a stabilization period of a preset duration; After the physical compensation action is completed and a stabilization period has passed, perform the same multi-physics field signal acquisition, time synchronization, spatial coordinate unification, feature extraction and nonlinear weighted fusion operations as in step S1 to generate a second fused feature vector that characterizes the mechanical state of the mechanical connection interface after compensation. Next, the difference between the second fused feature vector and the twin predicted feature vector is calculated to obtain the residual vector; Next, a comprehensive residual metric is calculated, which consists of two terms added together: the first term is the weighted sum of squares of the residual vectors, and the weight matrix used for weighting is a diagonal matrix. Each weight value on the diagonal of this diagonal matrix is ​​pre-set according to the relative importance of the corresponding physical feature in the state representation on which the first fused feature vector is relied upon. The second term is the gradient penalty term, which is the product of the L2 norm of the gradient vector of the forward prediction model with respect to its internal adjustable parameters and a preset positive gradient penalty coefficient. This comprehensive residual metric is used to assess the overall level of deviation between the twin's predicted feature vector and the second fused feature vector obtained after physical compensation.

10. The adaptive adjustment and control method for joint module assembly accuracy according to claim 9, characterized in that: The specific process of collaborative online correction of the first group intelligent optimization algorithm and the digital twin is as follows: The correction process for the first swarm intelligent optimization algorithm is as follows: based on the calculated residual vector, dynamically adjust the weight matrix of the weighted data fitting term of the first swarm intelligent optimization algorithm in the composite loss function; The specific adjustment process is as follows: For each component in the residual vector, multiply the negative of the absolute value of the component by a preset positive attenuation coefficient, and multiply it by the original corresponding weight value to obtain a new weight value. The correction process for the digital twin is as follows: the key uncertain physical parameters in the digital twin are used to form a state vector to be estimated; The twin prediction feature vector is defined as the observation function of this state vector; the calculated residual vector is used as the observation information. The extended Kalman filter algorithm is used to process observational information and update the estimated value of the state vector online, thereby calibrating the digital twin; The corrected weight matrix of the first swarm intelligent optimization algorithm obtained through the above correction process and the estimated values ​​of the updated physical parameters of the digital twin will be used in subsequent control cycles during operation.

Images