Connector structure residual life prediction method based on digital twinning

Abstract

The application discloses a connector structure residual life prediction method based on digital twinning, relates to the field of intelligent operation and maintenance and health management, and comprises the following steps: S1, collecting and preprocessing to obtain a standardized data set; S2, establishing a three-dimensional geometric model and an initial digital twin; S3, introducing a Fourier neural operator to construct a proxy model to accelerate physical field calculation, and obtaining a real-time damage field distribution; S4, copying a twin copy to load future loads, using an improved Koopman neural network and an optimization framework containing a Hamilton constraint loss function to perform linear acceleration deduction, and obtaining an evolution state sequence; S5, monitoring a performance degradation process, and recording a deduction duration when a failure critical value is reached; and S6, combining working condition probability distribution weighted calculation to output residual life prediction and risk positioning information. The application realizes fast physical field calculation and long-time stable deduction, and improves the accuracy and reliability of life prediction under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and health management, and in particular to a method for predicting the remaining life of connector structures based on digital twins. Background Technology

[0002] With the rapid development of industrial and intelligent manufacturing technologies, connectors, as key basic components widely used in aerospace, rail transportation, and new energy fields, directly affect the safety and stability of the entire system due to their operational reliability. Since electrical connectors typically operate under complex electromechanical-thermal coupling load environments, they are highly susceptible to factors such as vibration, shock, temperature and humidity changes, and electrical stress corrosion, leading to material performance degradation, increased contact resistance, and even fracture failure. Therefore, accurate prediction of the remaining lifespan of connector structures is crucial for predictive maintenance and preventing catastrophic accidents. Currently, existing connector lifespan prediction technologies are mainly based on physical finite element simulation or purely data-driven statistical analysis methods. While physical simulation methods can accurately describe the microscopic damage evolution mechanism within materials, constructing high-fidelity digital twins involves solving a large number of nonlinear partial differential equations, making the calculation process extremely time-consuming and difficult to meet real-time requirements. Purely data-driven methods, although computationally fast, often lack consideration of physical constraints and heavily rely on large amounts of tagged historical failure data. When faced with new operating conditions lacking fault samples, their generalization ability and long-term prediction accuracy are poor.

[0003] Furthermore, existing data-driven prediction techniques, such as those based on recurrent neural networks or long short-term memory networks, often suffer from long-term cumulative error divergence when dealing with complex nonlinear dynamic systems, causing predictions to deviate rapidly from the actual trajectory over time. Particularly in multi-field coupled environments, ensuring that prediction models adhere to physical laws such as energy conservation during long-term extrapolation remains a critical challenge. Simultaneously, traditional methods often neglect the randomness of future environmental conditions when extrapolating accelerated life tests, failing to provide probabilistic confidence levels for remaining lifetime predictions and thus hindering effective support for engineering decisions. In summary, existing technologies struggle to achieve a balance between computational efficiency, physical interpretability, long-term prediction stability, and environmental adaptability, limiting their application in practical engineering.

[0004] Therefore, how to provide a method for predicting the remaining life of connector structures based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a digital twin-based method for predicting the remaining life of connector structures. This invention fully integrates key steps such as real-time damage field inversion, Fourier neural operator surrogate modeling, improved Koopman neural network linear acceleration derivation, and multi-condition probabilistic weighted fusion. It constructs a full-process prediction system including rapid physical field calculation, orthogonality constraints on the weight matrix, an optimization framework with Hamiltonian constraint loss functions, and visualization of failure risks. This achieves accurate mapping and long-term stable derivation of connector performance degradation under different future environmental loads. This invention possesses advantages such as high physical computation efficiency, strong long-term prediction stability, adaptive energy conservation constraints, and high reliability of life prediction results. It can significantly improve the timeliness and accuracy of connector remaining life prediction under complex operating conditions, thereby effectively solving problems such as long finite element calculation time, easy divergence in long-term data-driven model derivation, and lack of physical consistency in existing technologies.

[0006] The remaining lifetime prediction method for connector structures based on digital twins according to embodiments of the present invention includes the following steps: S1. Simultaneously acquire electrical contact signals, mechanical signals, and environmental parameter signals, and preprocess them to obtain a standardized input dataset; S2. Establish a three-dimensional geometric model of the connector and discretize it into a mesh, set the basic physical property parameters of the material, implant damage state variables, and establish a dynamic evolution function to form an initial digital twin. S3. Input the standardized input dataset into the initial digital twin to perform physical field calculations, obtain virtual response data, introduce Fourier neural operators to construct a surrogate model of the physical field, perform global convolution operations and frequency domain feature learning, obtain accelerated virtual response data, and iteratively adjust the damage state variables based on the deviation value to obtain the real-time damage field distribution. S4. Based on the real-time damage field distribution replication, a digital twin copy is generated, and preset future environmental load conditions are loaded respectively. An improved Koopman neural network is introduced, the constraint weight matrix is ​​an orthogonal matrix, and an optimization framework containing Hamiltonian constraint loss function is constructed. Linear accelerated deduction is performed to simulate the evolution path under different future scenarios and obtain the deduced evolution state sequence. S5. Monitor the performance degradation process of the accelerated execution of each digital twin replica in real time. When the preset failure threshold is reached, stop the simulation and record the cumulative simulation time, and output the failure time sample set. S6. Based on historical environmental data, statistically analyze the probability distribution of various future operating conditions, assign a weight coefficient to each sample in the failure time sample set, convert the calculation results back to the real-world time scale, and output the remaining life prediction results and failure risk location information.

[0007] Optionally, S1 specifically includes: S11. Use voltage and current sensors to synchronously collect electrical contact signals flowing through the connector circuit, use accelerometers to synchronously collect mechanical signals received by the connector, and use thermometers and hygrometers to synchronously collect environmental parameter signals of the environment in which the connector is located. S12. Perform missing value imputation and moving average filtering on the collected electrical contact signals, mechanical signals and environmental parameter signals, and after linear normalization, align and combine them according to the timestamp to output a standardized input dataset.

[0008] Optionally, S2 specifically includes: S21. Read the two-dimensional engineering drawing of the connector, extrude it to generate a three-dimensional geometric model of the connector, and import it into the finite element preprocessing software. Select the tetrahedral element type, set the element size to one-tenth of the minimum feature size of the connector, and perform mesh generation operation to discretize the continuous solid structure into a finite element mesh model composed of nodes and element connections. S22. Select all elements in the finite element mesh model and input the preset basic physical property parameters of the connector material in the material properties. S23. Create a new state variable field at each element node position in the finite element mesh model and name it damage state variable. Initialize the damage state variable values ​​of all elements to 0. Define the range of the damage state variable values ​​as 0 to 1, where 0 represents that the material is in a completely intact state and 1 represents that the material is in a completely failed state. S24. Construct a dynamic evolution function between damage state variables and material macroscopic physical property parameters. Subtract the damage state variable value of the current element from the value of 1 to obtain the first residual coefficient, and multiply it with the initial Young's modulus to calculate the effective Young's modulus. S25. Multiply the current damage state variable value of the unit by the preset damage sensitivity coefficient to obtain the first exponential factor. Calculate the exponential function value with the natural constant e as the base and the first exponential factor as the exponential. Multiply the exponential function value by the initial resistivity to calculate the effective resistivity. S26. Assign the effective Young's modulus and effective resistivity to the corresponding elements in the finite element mesh model, update the material property parameters, store the complete model file containing the updated material property parameters, damage state variables and geometric mesh information, and generate the initial digital twin.

[0009] Optionally, S3 specifically includes: S31. Extract the voltage, current, acceleration, temperature and humidity values ​​from the standardized input dataset. Apply the voltage and current values ​​to the electrical contact ports of the initial digital twin as electrical boundary conditions, apply the acceleration values ​​to the mounting points as mechanical boundary conditions, and apply the temperature and humidity values ​​to all external surfaces as environmental boundary conditions. Perform physical field calculations to obtain virtual response data. S32. Extract the measured contact resistance, measured stress, and measured temperature at the corresponding time from the standardized input dataset. Calculate the absolute value of the first difference between the virtual contact resistance and the measured contact resistance, the absolute value of the second difference between the virtual stress and the measured stress, and the absolute value of the third difference between the virtual temperature field and the measured temperature. Add the first, second, and third absolute values ​​of the difference to obtain the total deviation value. S33. Construct a Fourier neural operator network structure, take the geometric grid coordinates of the initial digital twin as the network input to perform a fast Fourier transform operation, transform the data to Fourier space, perform global convolution operation and frequency domain feature learning on the solution of the partial differential equation in Fourier space, perform inverse fast Fourier transform operation on the learned frequency domain features, output the solution of the physical field, construct a surrogate model of the physical field, use the surrogate model to replace the time-consuming finite element calculation, input boundary conditions to obtain accelerated virtual response data; S34. Calculate the gradient information of the total deviation value relative to the damage state variable, update the value of the damage state variable using the gradient information, substitute the updated damage state variable into the initial digital twin, recalculate the new total deviation value, determine whether the new total deviation value is less than the preset deviation threshold, if it is not less than, continue to iterate in reverse until the deviation value converges to the preset range. S35. Extract the damage state variable values ​​at the final convergence, map the damage state variable values ​​to the grid nodes of the initial digital twin, and generate the real-time damage field distribution.

[0010] Optionally, S4 specifically includes: S41. Read the grid node number and corresponding damage state variable value stored in the real-time damage field distribution, define the damage state variable value as the node damage value, execute the file copy command to copy all grid information and physical attribute file of the initial digital twin, and generate a digital twin copy. S42. Locate the damage state variable field of each unit in the digital twin replica, assign the node damage value to the corresponding damage state variable, substitute the damage state variable value into the calculation formula of effective Young's modulus and the calculation formula of effective resistivity, and recalculate and update the effective Young's modulus and effective resistivity of all units in the digital twin replica. S43. Set future environmental load conditions, sample the temperature, humidity, vibration acceleration and current values ​​that change with time under the future environmental load conditions according to a preset time interval, multiply the minimum timing resolution value by a coefficient of 10 to obtain the sampling step size, plot the sampled data points into a load data curve, extract the value of each time point in the load data curve, apply the temperature and humidity values ​​to all outer surface nodes of the digital twin replica, apply the vibration acceleration value to the mounting and fixing points, and apply the current to the electrical contact port; S44. Construct an improved Koopman neural network, use the orthogonality constraint loss function and Hamilton constraint loss function to obtain the total optimization objective function by weighted summation, and use the backpropagation algorithm to update the parameters of the improved Koopman neural network to complete the construction of the optimization framework. S45. Read the node displacement vector, node velocity vector and damage state variable values ​​of the digital twin copy at the current moment and combine them into a state feature vector. Perform nonlinear mapping on the state feature vector to elevate it to an infinite-dimensional Koopman feature function space to obtain a high-dimensional feature vector. S46. Perform linear matrix multiplication in the high-dimensional feature space, multiply the high-dimensional feature vector with the pre-learned Koopman operator matrix to obtain the high-dimensional prediction features for the next time step, and perform nonlinear inverse mapping operation to project back to the original physical state space, and calculate and output the predicted node displacement vector, predicted node velocity vector and predicted damage state variable values ​​for the next time step. S47. Write the predicted node displacement vector, predicted node velocity vector, and predicted damage state variable values ​​for the next time step into the digital twin copy, overwriting the values ​​at the current time step. Use the updated digital twin copy as the new input, perform the linear mapping operation cyclically, continuously extrapolate a preset number of time steps, extract the predicted node displacement vector, predicted node velocity vector, and predicted damage state variable values ​​output at each time step, and arrange and combine them in chronological order to generate an extrapolated evolution state sequence.

[0011] Optionally, S44 specifically includes: S441. Construct an improved Koopman neural network, initialize the fully connected weight matrix, calculate the transpose of the fully connected weight matrix, perform matrix multiplication between the transpose and the fully connected weight matrix to obtain the product matrix, calculate the difference matrix between the product matrix and the identity matrix, calculate the sum of squares of all elements in the difference matrix as the orthogonality constraint loss function, and update the value of the fully connected weight matrix using the gradient descent algorithm until the orthogonality constraint loss function converges to the preset convergence threshold. S442. Calculate the total mechanical energy of all nodes in the digital twin replica, calculate the product of the displacement vector of all nodes and the global stiffness matrix to obtain the elastic potential energy, add the total mechanical energy and the elastic potential energy to obtain the total system energy, calculate the first derivative of the total system energy with respect to time to obtain the Hamiltonian, and construct a Hamiltonian constraint loss function containing the constraint condition that the value of the Hamiltonian approaches the preset zero-point tolerance of energy. S443. The Hamiltonian constraint loss function and the orthogonality constraint loss function are weighted and summed to obtain the overall optimization objective function. The gradient of the overall optimization objective function with respect to the parameters of the improved Koopman neural network is calculated. The parameters of the improved Koopman neural network are updated using the backpropagation algorithm. The calculation and update steps are iteratively executed until the value of the overall optimization objective function no longer decreases, thus completing the construction of the optimization framework.

[0012] Optionally, S5 specifically includes: S51. Read the preset key point number in the digital twin copy, and obtain the node displacement vector, node velocity vector and damage state variable value of the grid node corresponding to the preset key point number at the current simulation time. S52. Extract the values ​​of the damage state variables and determine whether the values ​​of the damage state variables are greater than or equal to the preset failure threshold. If the values ​​of the damage state variables are less than the preset failure threshold, continue to the next step of the deduction; otherwise, stop the deduction. S53. At the moment when the simulation stops, read the system clock of the digital twin copy, extract the current cumulative running time value recorded by the system clock as the simulation duration, multiply the simulation duration by the time acceleration coefficient, and convert it into the real-world time scale. S54. Associate and store the converted real-world time value with the type label of the currently loaded future environmental load condition to form a failure time sample data. Summarize the failure time sample data generated by all digital twin copies under different future environmental load conditions and output the failure time sample set.

[0013] Optionally, S6 specifically includes: S61. Read the stored historical environmental data file, divide the temperature data, humidity data and vibration acceleration data in the historical environmental data file into numerical ranges, count the frequency of data occurrence in each numerical range, divide the frequency by the total number of historical environmental data to calculate the probability of occurrence of each numerical range, and combine the numerical ranges with the corresponding probability of occurrence to generate a probability distribution table of future working conditions. S62. Extract the type label of the future environmental load condition carried by each failure time sample data in the failure time sample set, find the occurrence probability that matches the type label in the probability distribution table of future working conditions, and directly assign the found occurrence probability value to the corresponding failure time sample data as a weighting coefficient. S63. Read the converted real-world time value of each failure time sample data, multiply the real-world time value with the corresponding weight coefficient to obtain the weighted failure time, sum the weighted failure times of all failure time sample data arithmetically, divide by the sum of all weight coefficients, and calculate and output the remaining lifetime prediction result. S64. Extract the spatial coordinates of the preset key site number corresponding to each failure time sample data in the real-time damage field distribution, draw a spherical region with the failure critical value as the radius at the spatial coordinates, project all spherical regions onto the surface of the connector's three-dimensional geometric model, generate a red highlighted display area on the surface of the three-dimensional geometric model, and output the failure risk location information.

[0014] The beneficial effects of this invention are: This invention introduces a Fourier neural operator to construct a physical field proxy model. Addressing the issues of long computation times and difficulty in meeting real-time requirements in traditional finite element physical field calculations, it transforms the solution of partial differential equations into frequency domain global convolution and feature learning in Fourier space, generating accelerated virtual response data. By constructing an improved Koopman neural network and combining orthogonality constraints on the weight matrix with an optimization framework including a Hamiltonian-constrained loss function, it addresses the error accumulation and lack of physical consistency problems in long-term simulations of purely data-driven models. It elevates the nonlinear state mapping to a high-dimensional feature space to perform linear matrix multiplication operations, and enforces energy conservation and bounded operator norm constraints. Furthermore, it combines historical environmental probability distributions to weightedly fuse failure time samples under different operating conditions, converting the simulation results back to the real-world timescale and locating failure risk regions. Ultimately, it achieves real-time inversion of the internal damage state of connector structures, stable simulation of long-term performance degradation, and accurate prediction of remaining lifetime, effectively improving the computational efficiency, long-term numerical stability, and reliability of lifetime prediction results in complex multi-field coupled environments. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the connector structure remaining lifetime prediction method based on digital twin proposed in this invention; Figure 2This is a flowchart of the physical field proxy model construction and real-time damage field inversion based on Fourier neural operators proposed in this invention; Figure 3 This is a flowchart illustrating the construction and linear acceleration derivation of the improved Koopman neural network optimization framework based on the Hamiltonian constraint loss function proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 A method for predicting the remaining life of connector structures based on digital twins includes the following steps: S1. Simultaneously acquire electrical contact signals, mechanical signals, and environmental parameter signals, and preprocess them to obtain a standardized input dataset; S2. Establish a three-dimensional geometric model of the connector and perform mesh discretization. Set the basic physical property parameters of the material, implant damage state variables that can characterize the degree of aging inside the mesh, and establish a dynamic evolution function between the damage state variables and the macroscopic physical property parameters of the material to form an initial digital twin that can automatically adjust the physical property parameters according to the degree of damage. S3. Input the standardized input dataset into the initial digital twin as boundary conditions for physical field calculation, obtain the virtual response data output by the initial digital twin, and calculate the deviation value between the virtual response data and the corresponding measured data in the standardized input dataset. Introduce Fourier neural operators to construct a surrogate model of the physical field, perform global convolution operation and frequency domain feature learning on the solution of the partial differential equation in Fourier space, use the surrogate model to replace the time-consuming finite element calculation to obtain accelerated virtual response data, and iteratively adjust the damage state variables according to the deviation value until the deviation value converges to the preset range to obtain the real-time damage field distribution reflecting the current real internal damage of the connector. S4. Based on the real-time damage field distribution replication, a digital twin copy is generated, and preset future environmental load conditions are loaded respectively. An improved Koopman neural network is introduced, the constraint weight matrix is ​​an orthogonal matrix, and an optimization framework containing Hamiltonian constraint loss function is constructed. Linear accelerated deduction is performed to simulate the evolution path under different future scenarios and obtain the deduced evolution state sequence. S5. During the performance degradation process of the accelerated execution of each digital twin replica, monitor the physical state values ​​of key points in real time. When the preset failure threshold is reached, stop the simulation and record the cumulative simulation time, and output the failure time sample set corresponding to different environmental load conditions. S6. Based on historical environmental data, statistically analyze the probability distribution of various future operating conditions, assign a weight coefficient to each sample in the failure time sample set, convert the calculation results back to the real-world time scale, and output the remaining life prediction results and failure risk location information.

[0018] This implementation significantly improves the accuracy and efficiency of connector remaining life prediction. By embedding damage state variables into the mesh and establishing a dynamic evolution function, an initial digital twin with adaptively adjustable physical properties is constructed, enabling a fine characterization of the connector's internal aging process. A Fourier neural operator is introduced to construct a physical field proxy model, performing global convolution and feature learning in the frequency domain, replacing time-consuming finite element calculations. Real-time damage field distribution is rapidly obtained through reverse iteration based on deviation values, significantly improving computational speed. Furthermore, an improved Koopman neural network is utilized, combined with orthogonal matrix constraints and Hamiltonian constraint loss functions for linear acceleration, ensuring that the long-term evolution process conforms to the law of physical energy conservation and effectively avoiding accumulated errors. By simulating evolution paths in multiple scenarios and weighting the results based on operating condition probabilities, the results are converted back to a real-world time scale, outputting high-confidence remaining life predictions and failure location information. This invention not only solves the problems of slow computation and large prediction bias in traditional methods but also achieves accurate capture of performance degradation trends under complex operating conditions, significantly enhancing the robustness and engineering applicability of the prediction model.

[0019] In this embodiment, S1 specifically includes: S11. Use voltage and current sensors to synchronously collect electrical contact signals flowing through the connector circuit, use accelerometers to synchronously collect mechanical signals received by the connector, and use thermometers and hygrometers to synchronously collect environmental parameter signals of the environment in which the connector is located. S12. Perform missing value imputation and moving average filtering on the collected electrical contact signals, mechanical signals and environmental parameter signals, and after linear normalization, align and combine them according to the timestamp to output a standardized input dataset.

[0020] In this embodiment, S2 specifically includes: S21. Read the two-dimensional engineering drawing of the connector, extrude it to generate a three-dimensional geometric model of the connector, and import it into the finite element preprocessing software. Select the tetrahedral element type, set the element size to one-tenth of the minimum feature size of the connector, and perform mesh generation operation to discretize the continuous solid structure into a finite element mesh model composed of nodes and element connections. S22. Select all elements in the finite element mesh model, and input the preset basic physical property parameters of the connector material in the material properties. The preset basic physical property parameters include Young's modulus, Poisson's ratio, thermal conductivity and resistivity. Obtain the specific values ​​of the preset basic physical property parameters by consulting the metal material handbook. The value range of Young's modulus is 110GPa to 130GPa, and the value range of Poisson's ratio is 0.3 to 0.35. The preset basic physical property parameters of the material are the initial parameters when the connector has not aged. S23. Create a new state variable field at each element node position in the finite element mesh model and name it damage state variable. Initialize the damage state variable values ​​of all elements to 0. Define the range of the damage state variable values ​​as 0 to 1, where 0 represents that the material is in a completely intact state and 1 represents that the material is in a completely failed state. S24. Construct a dynamic evolution function between damage state variables and material macroscopic physical property parameters. Subtract the damage state variable value of the current element from the value of 1 to obtain the first residual coefficient, and multiply it with the initial Young's modulus to calculate the effective Young's modulus that automatically adjusts with the degree of damage. S25. Multiply the current unit's damage state variable value by a preset damage sensitivity coefficient to obtain a first exponential factor. Calculate the exponential function value with the natural constant e as the base and the first exponential factor as the exponential factor. Multiply the exponential function value by the initial resistivity to calculate the effective resistivity that automatically adjusts with the degree of damage. The preset damage sensitivity coefficient ranges from 5 to 10. The preset damage sensitivity coefficient is a parameter characterizing the rate at which the material's resistivity deteriorates with damage. S26. Assign the effective Young's modulus and effective resistivity to the corresponding elements in the finite element mesh model, update the material property parameters, store the complete model file containing the updated material property parameters, damage state variables and geometric mesh information, and generate the initial digital twin.

[0021] In this embodiment, S3 specifically includes: S31. Extract voltage, current, acceleration, temperature, and humidity values ​​from the standardized input dataset. Apply voltage and current values ​​to the electrical contact ports of the initial digital twin as electrical boundary conditions, apply acceleration values ​​to the mounting points of the initial digital twin as mechanical boundary conditions, and apply temperature and humidity values ​​to all outer surfaces of the initial digital twin as environmental boundary conditions. Perform physical field calculations to obtain virtual response data. The physical field calculation involves reading the grid node coordinates and element connection relationships, performing numerical integration on each element using the element volume formula to obtain the element volume, constructing an elastic matrix D based on the effective Young's modulus and Poisson's ratio, multiplying the elastic matrix D by the transpose of the strain matrix B, then multiplying it by the strain matrix B again, and finally multiplying it by the element volume to calculate the element stiffness matrix. Assemble all element stiffness matrices according to the node degrees of freedom numbering to form a global stiffness matrix K. External loads are mapped to nodes to form load vector F. The linear equation system Ku=F is solved iteratively using Gaussian elimination to obtain the node displacement vector u. The first-order partial derivatives of the shape function of the mesh element with respect to the coordinate axes x, y, and z are calculated, and the first-order partial derivatives are combined and arranged to construct the strain matrix B. The node displacement vector u is multiplied by the strain matrix B to obtain the element strain vector ε. The element strain vector ε is multiplied by the elasticity matrix D to obtain the element stress vector σ. The displacement components of the contact nodes in the node displacement vector u are calculated to obtain the normal deformation. The normal deformation is multiplied by the contact stiffness coefficient to obtain the contact pressure. The contact pressure and contact area are input into the Holm contact resistance formula to calculate the virtual contact resistance. The node temperature is calculated using thermal conductivity according to Fourier's law of heat conduction, thereby obtaining the virtual response data output by the initial digital twin. The virtual response data includes virtual contact resistance, virtual stress distribution, and virtual temperature field. S32. Extract the measured contact resistance, measured stress, and measured temperature at the corresponding time from the standardized input dataset. Calculate the absolute value of the first difference between the virtual contact resistance and the measured contact resistance, the absolute value of the second difference between the virtual stress and the measured stress, and the absolute value of the third difference between the virtual temperature field and the measured temperature. Add the first, second, and third absolute values ​​of the difference to obtain the total deviation value. S33. Construct a Fourier neural operator network structure. Using the geometric grid coordinates of the initial digital twin as network input, perform a Fast Fourier Transform (FFT) operation to transform the data into Fourier space. In Fourier space, perform global convolution operations and frequency domain feature learning on the solutions to partial differential equations. Perform an inverse Fast Fourier Transform (IFFT) operation on the learned frequency domain features to output the solution to the physical field. Construct a surrogate model of the physical field and use this surrogate model to replace time-consuming finite element calculations. Input boundary conditions to obtain accelerated virtual response data. The process described in Fourier space... The method involves performing a global convolution operation and learning frequency domain features on the solution of a partial differential equation. This involves using the frequency domain data transformed to Fourier space as the frequency domain solution of the partial differential equation to be solved. The frequency domain solution of the partial differential equation is then multiplied element-wise with a preset learnable frequency domain weight matrix to simulate the integral transformation process of the partial differential equation in the frequency domain. The convolution theorem of Fourier transform is used to transform the global convolution operation into a frequency domain multiplication operation. The multiplication result is then subjected to linear transformation and nonlinear activation processing, thereby learning the spectral features of the solution of the partial differential equation in the frequency domain. S34. Calculate the gradient information of the total deviation value relative to the damage state variable, update the value of the damage state variable using the gradient information, substitute the updated damage state variable into the initial digital twin, recalculate the new total deviation value, determine whether the new total deviation value is less than the preset deviation threshold, if it is not less than, continue the reverse iterative adjustment until the deviation value converges to the preset range, the preset deviation threshold value range is 0.001 to 0.01; S35. Extract the damage state variable values ​​at the final convergence, map the damage state variable values ​​to the grid nodes of the initial digital twin, and generate a real-time damage field distribution that reflects the current real internal damage of the connector.

[0022] This implementation introduces Fourier neural operators as an innovative technology, offering significant differences and advantages compared to traditional finite element numerical simulations and purely data-driven deep neural networks. Traditional finite element methods require repeated assembly of the stiffness matrix and iterative solutions when inverting the damage field, resulting in extremely long computation times and failing to meet real-time requirements. While standard CNNs or RNNs offer fast inference speeds, they struggle to strictly adhere to physical laws under complex boundary conditions, exhibiting poor generalization ability. This invention utilizes the frequency-domain global convolution mechanism of Fourier neural operators, employing the Fast Fourier Transform to transform the solution of complex partial differential equations into element-wise multiplication operations in the frequency domain. This ensures high accuracy in physical field calculations while achieving extremely rapid approximation of the solution space. In particular, the physical field proxy model constructed in S33 can acquire virtual response data in milliseconds, replacing the time-consuming numerical integration process. This mechanism effectively supports gradient-based backward iterative updates of damage state variables, enabling the acquisition of real-time damage field distributions to possess both physical interpretability and extremely high computational efficiency, significantly improving the accuracy and speed of damage inversion.

[0023] In this embodiment, S4 specifically includes: S41. Read the grid node number and corresponding damage state variable value stored in the real-time damage field distribution, define the damage state variable value as the node damage value, execute the file copy command to copy all grid information and physical attribute file of the initial digital twin, and generate a digital twin copy. S42. Locate the damage state variable field of each unit in the digital twin replica, assign the node damage value to the corresponding damage state variable, substitute the damage state variable value into the calculation formula of effective Young's modulus and the calculation formula of effective resistivity, and recalculate and update the effective Young's modulus and effective resistivity of all units in the digital twin replica. S43. Set future environmental load conditions, sample the temperature, humidity, vibration acceleration, and current values ​​that change over time under the future environmental load conditions according to a preset time interval, multiply the minimum timing resolution value by a coefficient of 10 to obtain the sampling step size, plot the sampled data points into a load data curve, extract the value of each time point in the load data curve, apply the temperature and humidity values ​​to all external surface nodes of the digital twin replica, apply the vibration acceleration value to the mounting fixing point, and apply the current value to the electrical contact port. The preset time interval is the minimum timing resolution for reading the system clock of the digital twin replica, and the sampling step size ranges from 0.1 seconds to 1 second. S44. Construct an improved Koopman neural network, use the orthogonality constraint loss function and Hamilton constraint loss function to obtain the total optimization objective function by weighted summation, and use the backpropagation algorithm to update the parameters of the improved Koopman neural network to complete the construction of the optimization framework. S45. Read the node displacement vector, node velocity vector and damage state variable values ​​of the digital twin copy at the current moment and combine them into a state feature vector. Perform nonlinear mapping on the state feature vector to elevate it to an infinite-dimensional Koopman feature function space to obtain a high-dimensional feature vector. S46. Perform linear matrix multiplication in the high-dimensional feature space, multiplying the high-dimensional feature vector with the pre-learned Koopman operator matrix to obtain the high-dimensional predicted features for the next time step. Then, perform a nonlinear inverse mapping operation to project back to the original physical state space, and calculate and output the predicted node displacement vector, predicted node velocity vector, and predicted damage state variable values ​​for the next time step. The pre-learned Koopman operator matrix is ​​composed of the node displacement vector, node velocity vector, and damage state variable values ​​extracted from historical time steps to form a first sample vector. Each element of the first sample vector is input into the first neural network layer containing the hyperbolic tangent activation function, and weighted summation and nonlinear transformation operations are performed to output the first high-dimensional mapping vector. The node displacement vector, node velocity vector, and damage state variable values ​​extracted for the next historical time step are then used to form the second high-dimensional mapping vector. Two sample vectors are used. Each element of the second sample vector is input into the first neural network layer. Weighted summation and nonlinear transformation operations are performed to output a second high-dimensional mapping vector. A matrix is ​​initialized as the matrix to be updated. Matrix multiplication is performed between the matrix to be updated and the first high-dimensional mapping vector to obtain the predicted mapping vector. The sum of squares of the differences between corresponding elements in the predicted mapping vector and the second high-dimensional mapping vector is calculated as the error value. The partial derivative of the error value with respect to each element in the matrix to be updated is calculated. The product of the preset learning rate and the partial derivative is subtracted from the partial derivative value to update the value of each element in the matrix to be updated. The matrix multiplication, error calculation and element update steps are repeated until the error value is less than a preset threshold. The finally updated matrix is ​​used as the Koopman operator matrix. The value range of the Koopman operator matrix is ​​-2 to 2. S47. Write the predicted node displacement vector, predicted node velocity vector, and predicted damage state variable values ​​for the next time step into the digital twin copy, overwriting the values ​​at the current time step. Use the updated digital twin copy as the new input, and repeatedly perform the linear mapping operation to continuously extrapolate a preset number of time steps. Extract the predicted node displacement vector, predicted node velocity vector, and predicted damage state variable values ​​output at each time step, and arrange and combine them in chronological order to generate an extrapolated evolution state sequence. The preset number of time steps ranges from 1000 to 10000 steps and is calculated by dividing the required duration of the prediction task by the preset time interval.

[0024] In this embodiment, S44 specifically includes: S441. Construct an improved Koopman neural network. Initialize the fully connected weight matrix, calculate the transpose of the fully connected weight matrix, and perform matrix multiplication between the transpose and the fully connected weight matrix to obtain a product matrix. Calculate the difference matrix between the product matrix and the identity matrix. Calculate the sum of squares of all elements in the difference matrix as the orthogonality constraint loss function. Update the values ​​of the fully connected weight matrix using the gradient descent algorithm until the orthogonality constraint loss function converges to a preset convergence threshold. The preset convergence threshold is one-thousandth of the initial orthogonality constraint loss function value, and its range is 1.0 × 10⁻⁶. -6 Up to 1.0×10 -4 This forces the weight matrix to be an orthogonal matrix. This operation aims to ensure that the norm of the Koopman operator is bounded during the deduction process and prevent numerical errors from accumulating infinitely in long-term linear deduction. This is a unique operation of this improvement targeting the long-term prediction instability of the traditional data-driven Koopman method. S442. Calculate the total mechanical energy of all nodes in the digital twin replica. Calculate the elastic potential energy by multiplying the displacement vector of all nodes by the global stiffness matrix. Add the total mechanical energy to the elastic potential energy to obtain the total system energy. Calculate the first derivative of the total system energy with respect to time to obtain the Hamiltonian. Construct a Hamiltonian constraint loss function that includes constraints where the Hamiltonian value approaches a preset zero-point tolerance of energy. The calculation of the total mechanical energy of all nodes in the digital twin replica involves extracting the node mass value and the velocity components in three directions for each node in the digital twin replica, and then multiplying the node mass value by X... The X, Y, and Z momentum components are obtained by multiplying the velocity components in the X, Y, and Z directions. The X, Y, and Z momentum components are arranged and combined according to the node numbers to form an intermediate vector. The square of each element in the intermediate vector is calculated and summed. The sum is multiplied by a coefficient of half to obtain the instantaneous kinetic energy value of the node. The total mechanical energy value of the system is calculated by summing the instantaneous kinetic energy values ​​of all nodes. The instantaneous kinetic energy value of the node ranges from 0 joules to 500 joules. The preset zero-point tolerance is a small amount selected according to the magnitude of the total energy of the system. The preset zero-point tolerance ranges from 0.01 joules to 0.1 joules. S443. The Hamiltonian constraint loss function and the orthogonality constraint loss function are weighted and summed to obtain the overall optimization objective function. The gradient of the overall optimization objective function with respect to the parameters of the improved Koopman neural network is calculated. The parameters of the improved Koopman neural network are updated using the backpropagation algorithm. The calculation and update steps are iteratively executed until the value of the overall optimization objective function no longer decreases, thus completing the construction of the optimization framework. The weight coefficients used in the weighted summation are selected from the range of 0.1 to 1.0 using the grid search method to minimize the error of the validation set.

[0025] This implementation method achieves accelerated long-term linear derivation of connector performance degradation by introducing an improved Koopman neural network combined with orthogonality and Hamiltonian constraint mechanisms. The orthogonality constraint forces the weight matrices to be orthogonal, effectively limiting the operator norm and preventing the infinite accumulation of numerical errors during long-term derivation, thus solving the problem of long-term prediction instability in traditional methods. Simultaneously, the introduction of a Hamiltonian constraint loss function forces the total system energy to remain conserved during evolution, ensuring that the derivation path strictly follows physical laws. By performing linear matrix multiplication in an infinite-dimensional feature space, the complex nonlinear damage evolution is transformed into an efficient derivation, significantly improving computational speed while ensuring the physical rationality and high accuracy of the prediction results.

[0026] The improved Koopman neural network of this invention is similar to the original Koopman neural network in that both retain the core architecture of Koopman analysis, namely, using a nonlinear mapping function to elevate the finite-dimensional physical state space to an infinite-dimensional Koopman feature function space, and using a linear operator matrix in the high-dimensional space to describe the evolution of the system state over time. Finally, the prediction results are projected back to the original physical space through a nonlinear inverse mapping. Both are based on a data-driven approach to approximate the system's feature functions and linear dynamics through a deep neural network.

[0027] The difference lies in that this invention breaks through the limitations of the original Koopman neural network, which relies on pure data fitting, lacks physical constraints, and suffers from long-term numerical instability. It introduces a dual constraint mechanism of orthogonality and Hamiltonian, constructing a physically conserved optimization framework. Building upon the original model's direct use of backpropagation to minimize prediction error, this invention adds an orthogonality constraint in step S441, forcing the calculation of the difference between the product of the fully connected weight matrix and its transpose and the identity matrix, using this difference as the orthogonality constraint loss function. Subsequently, in step S442, a Hamiltonian calculation based on physical energy is introduced. The total system energy is constructed by extracting node kinetic energy and system elastic potential energy, and its time derivative is used as the Hamiltonian constraint loss function. Finally, in step S443, the weighted sum of the orthogonality constraint loss function and the Hamiltonian constraint loss function is incorporated into the overall optimization objective function, jointly updating the network parameters.

[0028] Based on the aforementioned improvements, the beneficial effects of this invention lie in the fact that, through orthogonal matrix constraints and the Hamiltonian energy conservation mechanism, the improved Koopman neural network can effectively suppress the exponential accumulation of numerical errors during long-term step-linear extrapolation, solving the problem of long-term prediction divergence in traditional methods. Simultaneously, the Hamiltonian constraint forces the extrapolation path to follow the law of conservation of total system energy, ensuring that the prediction results conform to the fundamental principles of thermodynamics and kinetics, and avoiding the generation of non-physical phenomena. This design significantly enhances the model's generalization ability under complex dynamic loads and the reliability of long-term extrapolation, achieving high-precision and highly robust lifetime evolution prediction while maintaining computational efficiency.

[0029] In this embodiment, S5 specifically includes: S51. Read the preset key point number in the digital twin copy, and obtain the node displacement vector, node velocity vector and damage state variable value of the grid node corresponding to the preset key point number at the current simulation time. The preset key point number is the edge node number of the connector socket area. S52. Extract the damage state variable values ​​and determine whether the damage state variable values ​​are greater than or equal to the preset failure threshold. If the damage state variable values ​​are less than the preset failure threshold, continue to the next step of the deduction; otherwise, stop the deduction. The preset failure threshold is 0.9. S53. At the moment when the simulation stops, read the system clock of the digital twin copy, extract the current cumulative running time value recorded by the system clock as the simulation duration, multiply the simulation duration by the time acceleration factor, and convert it into the real-world time scale. The time acceleration factor is the ratio of a preset time interval to the simulation step size. The preset time interval is 0.5 seconds. S54. Associate and store the converted real-world time value with the type label of the currently loaded future environmental load condition to form a failure time sample data. Summarize the failure time sample data generated by all digital twin copies under different future environmental load conditions and output the failure time sample set.

[0030] In this embodiment, S6 specifically includes: S61. Read the stored historical environmental data file, divide the temperature data, humidity data and vibration acceleration data in the historical environmental data file into numerical ranges, count the frequency of data occurrence in each numerical range, divide the frequency by the total number of historical environmental data to calculate the probability of occurrence of each numerical range, and combine the numerical ranges with the corresponding probability of occurrence to generate a probability distribution table of future working conditions. S62. Extract the type label of the future environmental load condition carried by each failure time sample data in the failure time sample set, find the occurrence probability that matches the type label in the probability distribution table of future working conditions, and directly assign the found occurrence probability value to the corresponding failure time sample data as a weighting coefficient. S63. Read the converted real-world time value of each failure time sample data, multiply the real-world time value with the corresponding weight coefficient to obtain the weighted failure time, sum the weighted failure times of all failure time sample data arithmetically, divide by the sum of all weight coefficients, and calculate and output the remaining lifetime prediction result. S64. Extract the spatial coordinates of the preset key site number corresponding to each failure time sample data in the real-time damage field distribution, draw a spherical region with the failure critical value as the radius at the spatial coordinates, project all spherical regions onto the surface of the connector's three-dimensional geometric model, generate a red highlighted display area on the surface of the three-dimensional geometric model, and output the failure risk location information.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to the intelligent assembly workshop of a heavy equipment manufacturing company. The main purpose was to solve the problem of predicting the remaining lifespan of critical electrical connectors on a highly flexible automated production line in the workshop under complex electromechanical-thermal coupling conditions. This workshop is responsible for the assembly of large industrial robot arms and servo drive systems. The production line contains 240 highly dynamic electrical connection units, covering various types such as power transmission, signal control, and data communication. These connectors operate under harsh conditions of high-frequency vibration, large current fluctuations, and large day-night temperature differences. Previously, the factory mainly relied on traditional periodic preventative maintenance strategies, i.e., shutdowns for maintenance at fixed intervals (e.g., every three or six months). This approach not only led to serious problems of over-maintenance or under-maintenance, consuming a large amount of manpower and resources, but also resulted in unplanned production line shutdowns multiple times due to undetected internal contact failures of connectors, with direct economic losses of hundreds of thousands of yuan per shutdown. Traditional finite element simulation methods based on a single physical model have acceptable accuracy, but they are computationally very time-consuming and cannot meet the needs of real-time online monitoring. On the other hand, deep learning methods based on pure data-driven approaches have advantages in prediction speed, but they often lack physical interpretability. When faced with new working conditions that have never appeared before, the prediction results are biased and it is difficult to directly guide engineering decisions.

[0032] In practical applications, this invention fully utilizes a multi-source sensor network deployed at key nodes of the production line. The system receives over 500GB of real-time data daily, including voltage, current, vibration acceleration, temperature, and humidity data, as well as the connector's three-dimensional geometric model and material properties. First, a data preprocessing module cleans and aligns this heterogeneous data before inputting it into the constructed initial digital twin. To overcome the bottleneck of low efficiency in traditional finite element method calculations, this invention innovatively introduces a Fourier neural operator (FNO) to construct a physical field proxy model. This compresses the partial differential equation solution process, which originally took hours, to the millisecond level, enabling real-time inversion of the connector's internal contact resistance, stress distribution, and temperature field. This successfully generates a "real-time damage field distribution" reflecting the equipment's current true health status. Based on this, the system automatically copies the digital twin and loads predicted environmental load conditions for the next week. To address the issue of numerical accumulation and divergence in long-term prediction, this invention employs an improved Koopman neural network. By introducing orthogonality constraints on the weight matrix and an optimization framework incorporating a Hamiltonian-constrained loss function, the complex nonlinear damage evolution process is transformed into a linear extrapolation in a high-dimensional space. This allows for the rapid simulation of the connector's performance degradation path under different future operating conditions while ensuring the conservation of physical energy. The system monitors the damage state variables at key locations in real time. Once the values ​​reach a preset failure threshold, the extrapolation duration is immediately recorded. Combined with the probability distribution of historical operating conditions, the system outputs a weighted remaining lifetime prediction result and failure risk location information.

[0033] After two months of continuous field testing and data comparison, this invention demonstrated superior predictive performance and engineering practical value. During the testing period, the system successfully detected and warned of four potential connector contact failure risks, preventing three accidents that could have caused production line downtime. To objectively evaluate the technical advantages of this invention, we conducted a detailed comparison of the prediction results of the method used in this invention with traditional calculation methods based on physical fatigue life formulas and pure data-driven methods based on Long Short-Term Memory (LSTM) networks currently used in factories. This comparison covered multiple dimensions, including contact resistance prediction error, remaining life prediction deviation, and single-inference time. Table 1 below shows a detailed performance comparison of this invention and traditional methods in the connector performance degradation prediction task: Table 1. Performance Comparison Data Between the Invention and Traditional Methods in Connector Performance Degradation Prediction Task

[0034] Based on the comparative data shown in Table 1, it can be seen that the prediction method proposed in this invention, based on digital twins and an improved Koopman operator, exhibits significant performance advantages over traditional methods in predicting connector performance degradation, especially in key indicators such as prediction accuracy, computational efficiency, long-term stability, and adaptability to multiple operating conditions. Regarding prediction accuracy, this invention demonstrates extremely high accuracy in both contact resistance prediction and remaining lifetime estimation. For example, in terms of contact resistance prediction error, the average absolute error of this invention is controlled at 0.96 mΩ, far lower than the 4.25 mΩ of the traditional finite element method; while in terms of remaining lifetime prediction accuracy, this invention achieves 95.4%, far exceeding the 72.3% of the traditional method and the 81.6% of the ordinary LSTM method. This is mainly due to the physical field proxy model constructed by the Fourier neural operator, whose high-precision damage field inversion capability effectively solves the problem of ambiguous perception of internal state in traditional methods.

[0035] Regarding response timeliness, although the physical mechanisms of traditional finite element simulation are clear, a single simulation takes as long as 3620 seconds, which cannot meet the needs of real-time monitoring. This invention, by introducing a deep learning architecture, reduces the simulation time to 0.82 seconds, improving computational efficiency by thousands of times, while maintaining the interpretability of the physical model. More notably, the model demonstrates superior long-term stability. Traditional LSTM methods experience an error accumulation rate as high as 12.5% ​​in 24-hour continuous simulations, leading to long-term prediction failures. This invention, by introducing an optimization framework including a Hamiltonian constraint loss function, enforces physical energy conservation, controlling the error accumulation rate to 0.08%, ensuring the physical consistency of the prediction results. Furthermore, in the new working condition adaptability test, the error increment of this invention is only 2.1%, compared to 18.4% for the LSTM method, demonstrating extremely strong generalization ability. Overall, this invention successfully integrates the interpretability of the physical model with the efficiency of data-driven approaches, effectively solving the technical shortcomings of existing technologies such as low prediction accuracy, slow speed, and long-term instability in complex industrial environments.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining lifetime of connector structures based on digital twins, characterized in that, Includes the following steps: S1. Simultaneously acquire electrical contact signals, mechanical signals, and environmental parameter signals, and preprocess them to obtain a standardized input dataset; S2. Establish a three-dimensional geometric model of the connector and discretize it into a mesh, set the basic physical property parameters of the material, implant damage state variables, and establish a dynamic evolution function to form an initial digital twin. S3. Input the standardized input dataset into the initial digital twin to perform physical field calculations, obtain virtual response data, introduce Fourier neural operators to construct a surrogate model of the physical field, perform global convolution operations and frequency domain feature learning, obtain accelerated virtual response data, and iteratively adjust the damage state variables based on the deviation value to obtain the real-time damage field distribution. S4. Based on the real-time damage field distribution replication, a digital twin copy is generated, and preset future environmental load conditions are loaded respectively. An improved Koopman neural network is introduced, the constraint weight matrix is ​​an orthogonal matrix, and an optimization framework containing Hamiltonian constraint loss function is constructed. Linear accelerated deduction is performed to simulate the evolution path under different future scenarios and obtain the deduced evolution state sequence. S5. Monitor the performance degradation process of the accelerated execution of each digital twin replica in real time. When the preset failure threshold is reached, stop the simulation and record the cumulative simulation time, and output the failure time sample set. S6. Based on historical environmental data, statistically analyze the probability distribution of various future operating conditions, assign a weight coefficient to each sample in the failure time sample set, convert the calculation results back to the real-world time scale, and output the remaining life prediction results and failure risk location information.

2. The method for predicting the remaining lifetime of connector structures based on digital twins according to claim 1, characterized in that, S1 specifically includes: S11. Use voltage and current sensors to synchronously collect electrical contact signals flowing through the connector circuit, use accelerometers to synchronously collect mechanical signals received by the connector, and use thermometers and hygrometers to synchronously collect environmental parameter signals of the environment in which the connector is located. S12. Perform missing value imputation and moving average filtering on the collected electrical contact signals, mechanical signals and environmental parameter signals, and after linear normalization, align and combine them according to the timestamp to output a standardized input dataset.

3. The method for predicting the remaining lifetime of connector structures based on digital twins according to claim 1, characterized in that, S2 specifically includes: S21. Read the two-dimensional engineering drawing of the connector, extrude it to generate a three-dimensional geometric model of the connector, and import it into the finite element preprocessing software. Select the tetrahedral element type, set the element size to one-tenth of the minimum feature size of the connector, and perform mesh generation operation to discretize the continuous solid structure into a finite element mesh model composed of nodes and element connections. S22. Select all elements in the finite element mesh model and input the preset basic physical property parameters of the connector material in the material properties. S23. Create a new state variable field at each element node position in the finite element mesh model and name it damage state variable. Initialize the damage state variable values ​​of all elements to 0. Define the range of the damage state variable values ​​as 0 to 1, where 0 represents that the material is in a completely intact state and 1 represents that the material is in a completely failed state. S24. Construct a dynamic evolution function between damage state variables and material macroscopic physical property parameters. Subtract the damage state variable value of the current element from the value of 1 to obtain the first residual coefficient, and multiply it with the initial Young's modulus to calculate the effective Young's modulus. S25. Multiply the current damage state variable value of the unit by the preset damage sensitivity coefficient to obtain the first exponential factor. Calculate the exponential function value with the natural constant e as the base and the first exponential factor as the exponential. Multiply the exponential function value by the initial resistivity to calculate the effective resistivity. S26. Assign the effective Young's modulus and effective resistivity to the corresponding elements in the finite element mesh model, update the material property parameters, store the complete model file containing the updated material property parameters, damage state variables and geometric mesh information, and generate the initial digital twin.

4. The method for predicting the remaining lifetime of connector structures based on digital twins according to claim 1, characterized in that, S3 specifically includes: S31. Extract the voltage, current, acceleration, temperature and humidity values ​​from the standardized input dataset. Apply the voltage and current values ​​to the electrical contact ports of the initial digital twin as electrical boundary conditions, apply the acceleration values ​​to the mounting points as mechanical boundary conditions, and apply the temperature and humidity values ​​to all external surfaces as environmental boundary conditions. Perform physical field calculations to obtain virtual response data. S32. Extract the measured contact resistance, measured stress, and measured temperature at the corresponding time from the standardized input dataset. Calculate the absolute value of the first difference between the virtual contact resistance and the measured contact resistance, the absolute value of the second difference between the virtual stress and the measured stress, and the absolute value of the third difference between the virtual temperature field and the measured temperature. Add the first, second, and third absolute values ​​of the difference to obtain the total deviation value. S33. Construct a Fourier neural operator network structure, take the geometric grid coordinates of the initial digital twin as the network input to perform a fast Fourier transform operation, transform the data to Fourier space, perform global convolution operation and frequency domain feature learning on the solution of the partial differential equation in Fourier space, perform inverse fast Fourier transform operation on the learned frequency domain features, output the solution of the physical field, construct a surrogate model of the physical field, use the surrogate model to replace the time-consuming finite element calculation, input boundary conditions to obtain accelerated virtual response data; S34. Calculate the gradient information of the total deviation value relative to the damage state variable, update the value of the damage state variable using the gradient information, substitute the updated damage state variable into the initial digital twin, recalculate the new total deviation value, determine whether the new total deviation value is less than the preset deviation threshold, if it is not less than, continue to iterate in reverse until the deviation value converges to the preset range. S35. Extract the damage state variable values ​​at the final convergence, map the damage state variable values ​​to the grid nodes of the initial digital twin, and generate the real-time damage field distribution.

5. The method for predicting the remaining life of connector structures based on digital twins according to claim 1, characterized in that, S4 specifically includes: S41. Read the grid node number and corresponding damage state variable value stored in the real-time damage field distribution, define the damage state variable value as the node damage value, execute the file copy command to copy all grid information and physical attribute file of the initial digital twin, and generate a digital twin copy. S42. Locate the damage state variable field of each unit in the digital twin replica, assign the node damage value to the corresponding damage state variable, substitute the damage state variable value into the calculation formula of effective Young's modulus and the calculation formula of effective resistivity, and recalculate and update the effective Young's modulus and effective resistivity of all units in the digital twin replica. S43. Set future environmental load conditions, sample the temperature, humidity, vibration acceleration and current values ​​that change with time under the future environmental load conditions according to a preset time interval, multiply the minimum timing resolution value by a coefficient of 10 to obtain the sampling step size, plot the sampled data points into a load data curve, extract the value of each time point in the load data curve, apply the temperature and humidity values ​​to all outer surface nodes of the digital twin replica, apply the vibration acceleration value to the mounting and fixing points, and apply the current to the electrical contact port; S44. Construct an improved Koopman neural network, use the orthogonality constraint loss function and Hamilton constraint loss function to obtain the total optimization objective function by weighted summation, and use the backpropagation algorithm to update the parameters of the improved Koopman neural network to complete the construction of the optimization framework. S45. Read the node displacement vector, node velocity vector and damage state variable values ​​of the digital twin copy at the current moment and combine them into a state feature vector. Perform nonlinear mapping on the state feature vector to elevate it to an infinite-dimensional Koopman feature function space to obtain a high-dimensional feature vector. S46. Perform linear matrix multiplication in the high-dimensional feature space, multiply the high-dimensional feature vector with the pre-learned Koopman operator matrix to obtain the high-dimensional prediction features for the next time step, and perform nonlinear inverse mapping operation to project back to the original physical state space, and calculate and output the predicted node displacement vector, predicted node velocity vector and predicted damage state variable values ​​for the next time step. S47. Write the predicted node displacement vector, predicted node velocity vector, and predicted damage state variable values ​​for the next time step into the digital twin copy, overwriting the values ​​at the current time step. Use the updated digital twin copy as the new input, perform the linear mapping operation cyclically, continuously extrapolate a preset number of time steps, extract the predicted node displacement vector, predicted node velocity vector, and predicted damage state variable values ​​output at each time step, and arrange and combine them in chronological order to generate an extrapolated evolution state sequence.

6. The method for predicting the remaining lifetime of connector structures based on digital twins according to claim 5, characterized in that, S44 specifically includes: S441. Construct an improved Koopman neural network, initialize the fully connected weight matrix, calculate the transpose of the fully connected weight matrix, perform matrix multiplication between the transpose and the fully connected weight matrix to obtain the product matrix, calculate the difference matrix between the product matrix and the identity matrix, calculate the sum of squares of all elements in the difference matrix as the orthogonality constraint loss function, and update the value of the fully connected weight matrix using the gradient descent algorithm until the orthogonality constraint loss function converges to the preset convergence threshold. S442. Calculate the total mechanical energy of all nodes in the digital twin replica, calculate the product of the displacement vector of all nodes and the global stiffness matrix to obtain the elastic potential energy, add the total mechanical energy and the elastic potential energy to obtain the total system energy, calculate the first derivative of the total system energy with respect to time to obtain the Hamiltonian, and construct a Hamiltonian constraint loss function containing the constraint condition that the value of the Hamiltonian approaches the preset zero-point tolerance of energy. S443. The Hamiltonian constraint loss function and the orthogonality constraint loss function are weighted and summed to obtain the overall optimization objective function. The gradient of the overall optimization objective function with respect to the parameters of the improved Koopman neural network is calculated. The parameters of the improved Koopman neural network are updated using the backpropagation algorithm. The calculation and update steps are iteratively executed until the value of the overall optimization objective function no longer decreases, thus completing the construction of the optimization framework.

7. The method for predicting the remaining lifetime of connector structures based on digital twins according to claim 1, characterized in that, S5 specifically includes: S51. Read the preset key point number in the digital twin copy, and obtain the node displacement vector, node velocity vector and damage state variable value of the grid node corresponding to the preset key point number at the current simulation time. S52. Extract the values ​​of the damage state variables and determine whether the values ​​of the damage state variables are greater than or equal to the preset failure threshold. If the values ​​of the damage state variables are less than the preset failure threshold, continue to the next step of the deduction; otherwise, stop the deduction. S53. At the moment when the simulation stops, read the system clock of the digital twin copy, extract the current cumulative running time value recorded by the system clock as the simulation duration, multiply the simulation duration by the time acceleration coefficient, and convert it into the real-world time scale. S54. Associate and store the converted real-world time value with the type label of the currently loaded future environmental load condition to form a failure time sample data. Summarize the failure time sample data generated by all digital twin copies under different future environmental load conditions and output the failure time sample set.

8. The method for predicting the remaining lifetime of connector structures based on digital twins according to claim 1, characterized in that, S6 specifically includes: S61. Read the stored historical environmental data file, divide the temperature data, humidity data and vibration acceleration data in the historical environmental data file into numerical ranges, count the frequency of data occurrence in each numerical range, divide the frequency by the total number of historical environmental data to calculate the probability of occurrence of each numerical range, and combine the numerical ranges with the corresponding probability of occurrence to generate a probability distribution table of future working conditions. S62. Extract the type label of the future environmental load condition carried by each failure time sample data in the failure time sample set, find the occurrence probability that matches the type label in the probability distribution table of future working conditions, and directly assign the found occurrence probability value to the corresponding failure time sample data as a weighting coefficient. S63. Read the converted real-world time value of each failure time sample data, multiply the real-world time value with the corresponding weight coefficient to obtain the weighted failure time, sum the weighted failure times of all failure time sample data arithmetically, divide by the sum of all weight coefficients, and calculate and output the remaining lifetime prediction result. S64. Extract the spatial coordinates of the preset key site number corresponding to each failure time sample data in the real-time damage field distribution, draw a spherical region with the failure critical value as the radius at the spatial coordinates, project all spherical regions onto the surface of the connector's three-dimensional geometric model, generate a red highlighted display area on the surface of the three-dimensional geometric model, and output the failure risk location information.

Images