An AI algorithm-based twin modeling simulation analysis system

Abstract

This invention discloses a twin modeling simulation analysis system based on AI algorithms, comprising: a data tensor construction module for acquiring production line workstation operation data and generating input tensors; a frequency domain transformation module for performing frequency domain transformation on the input tensors to obtain complex spectral features; a phase-amplitude decoupling mapping module for performing amplitude and phase decomposition on the complex spectrum and generating updated complex spectral features; a self-consistent conservation projection module for calculating conservation anchor points and performing zero-frequency projection to obtain conservation constraint outputs; a spectral support sparse low-rank adaptation module for determining the frequency support set and constructing low-rank adaptation weights; a residual spectrum update module for updating low-rank adaptation parameters; and a coupling state prediction module for outputting production line coupling state prediction results. This invention achieves frequency domain conservation constraint adaptive modeling, improving production line prediction consistency and model adaptability.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a twin modeling simulation analysis system based on AI algorithms. Background Technology

[0002] Industrial production lines continue to evolve towards automation and digitalization, exhibiting characteristics of multi-workstation coupling, multi-variable linkage, and multi-timescale changes. Digital twin technology, by constructing a virtual model corresponding to the real production line, enables operational status mapping and simulation analysis, and has become an important technological approach in the manufacturing field.

[0003] Existing twin modeling methods mainly include mechanism-driven, data-driven, and hybrid modeling schemes. Mechanism-driven schemes rely heavily on prior knowledge and are difficult to deploy quickly in scenarios with complex process flows and numerous parallel workstations. While data-driven schemes possess nonlinear fitting capabilities, they lack explicit constraints on physical laws, making them prone to cumulative deviations in material flow scenarios. In recent years, neural network models have been used for production line modeling, but most methods focus on time-domain fitting, underutilizing frequency-domain information and failing to effectively represent dynamic behavior at different frequency scales. Furthermore, these models are susceptible to fluctuations in operating conditions during online operation and lack parameter update mechanisms for key frequency bands.

[0004] Existing methods still have shortcomings in terms of production line consistency constraints, and the prediction results may not be consistent with the material conservation relationship, affecting the usability of twin results in scheduling analysis and simulation decision-making.

[0005] Therefore, there is an urgent need for a twin modeling solution for production line scenarios, which establishes a unified mechanism among frequency domain modeling, conservation constraints and adaptive parameter updates to improve the prediction consistency and operational stability of complex workstation coupled systems. Summary of the Invention

[0006] One objective of this invention is to propose a twin modeling and simulation analysis system based on AI algorithms. This invention fully combines the Fourier neural operator backbone network, self-consistent conservation projection, and spectral support sparse low-rank adaptation, and has the advantages of strong frequency domain modeling capability, good conservation consistency, high online update efficiency, and strong adaptability to multi-condition fluctuations.

[0007] According to an embodiment of the present invention, a twin modeling simulation analysis system based on an AI algorithm includes: The data tensor construction module is used to collect the operation history of production line workstations and generate input tensors; The frequency domain transformation module is used to perform frequency domain transformation on the input tensor input to the Fourier neural operator network to obtain the complex spectrum feature matrix; The phase-amplitude decoupling mapping module is used to perform spectral amplitude component and spectral phase component decomposition on the complex spectral feature matrix, and generate an updated complex spectral feature matrix based on the amplitude mapping matrix and the phase mapping matrix. The self-consistent conservation projection module is used to calculate the conservation anchor point value based on the cumulative material count and the differential integral of the change in the quantity of work-in-process within the input tensor time window, and to perform zero-frequency component replacement projection on the prediction output corresponding to the updated complex spectrum feature matrix to obtain the conservation projection output. The spectral support sparse low-rank adaptation module is used to determine the frequency support set based on the energy distribution of the input tensor spectrum, extract the weight submatrix corresponding to the frequency support set, construct a low-rank adaptation increment matrix at the position of the weight submatrix and superimpose it with the original spectrum weights to obtain the spectral support sparse low-rank adaptation weight structure. The residual spectrum update module is used to construct a residual tensor based on the conserved projected output and the true output and generate a residual spectrum energy distribution, and update the parameters of the low-rank adaptation increment matrix based on the residual spectrum energy distribution; The coupling state prediction module is used to write the input tensor back to the Fourier neural operator backbone network, perform inverse frequency domain mapping and zero-frequency component replacement projection, and output the coupling state prediction results of the production line station.

[0008] Optionally, the data tensor construction module includes: Acquire historical operation data for each workstation on the production line. The historical operation data includes process parameter records, equipment operating status records, and quality inspection records. The historical operation data of each workstation is resampled according to a uniform sampling time interval, and timestamp alignment is performed to form time series data. The time series data is processed according to the production line process flow sequence to form workstation sequence data; Perform variable classification mapping on the workstation sequence data, mapping process parameter data, equipment status data, and quality inspection data to variable channel dimensions respectively; The mapped data is processed using a three-dimensional stacking method based on the workstation sequence dimension, time series dimension, and variable channel dimension to generate an input tensor.

[0009] Optionally, the frequency domain transformation module includes: Receives a 3D input tensor generated by the data tensor construction module and inputs the input tensor into a pre-constructed Fourier neural operator network; By using at least one Fourier layer in the Fourier neural operator network, the input tensor is subjected to a fast Fourier transform in the time series dimension, mapping the time-domain signal to the frequency domain. Extract the complex form of the spectrum data after Fourier transform; Based on the hierarchical structure of the Fourier neural operator network, specified linear transformations and nonlinear activation operations are performed on the spectral data to generate frequency domain mapping feature data. The frequency domain feature data is organized into a complex spectrum feature matrix. The row dimension of the matrix corresponds to the combined features of the work station sequence and the variable channel, and the column dimension corresponds to the discrete frequency components obtained after transformation.

[0010] Optionally, the phase amplitude decoupling mapping module includes: For each complex element in the complex spectral characteristic matrix, calculate the modulus to obtain the spectral amplitude component matrix, and calculate the argument to obtain the spectral phase component matrix; For the spectral amplitude component matrix, a linear transformation is performed through a trainable amplitude mapping matrix, followed by a nonlinear mapping, to obtain the updated spectral amplitude component matrix; The updated spectral phase component matrix is ​​obtained by performing a linear transformation on the spectral phase component matrix through a trainable phase mapping matrix. Based on the updated spectral amplitude component matrix and the updated spectral phase component matrix, an updated complex spectral feature matrix is ​​generated by complex number synthesis calculation.

[0011] Optionally, the self-consistent conservation projection module includes: Extract material flow data within a preset time window from the input tensor. The material flow data includes at least the input material count and output material count for each workstation. Calculate the cumulative material count by summing the total amount of materials input at the starting workstation of the production line within the time window; Calculate the differential integral of the change in the quantity of work-in-process, and calculate the instantaneous sum of the quantity of work-in-process at all stations at each sampling moment within the time window; the net change is obtained by differential integral calculation. According to the law of conservation of material, the cumulative amount of material count is subtracted from the differential integral of the change in the quantity of work-in-process, and the result is used as the theoretical total output material conservation anchor value. Perform an inverse fast Fourier transform on the updated complex spectrum feature matrix to transform it back to the time domain and obtain a preliminary time-domain prediction sequence; From the preliminary time-domain prediction sequence, the prediction channel data corresponding to the total output material is extracted, and the data is accumulated and summed over the time window to obtain the total output material value predicted by the model. The deviation between the theoretical total output material conservation anchor point value and the total output material value predicted by the model is calculated. Extract the zero-frequency components corresponding to all channels in the updated complex spectrum feature matrix. The zero-frequency components correspond to the complex coefficients with a frequency of zero in the spectrum. Calculate the zero-frequency component correction value, replace the zero-frequency component of each channel in the updated complex spectrum feature matrix with the correction value, keep all non-zero frequency components unchanged, and generate the projection-adjusted complex spectrum matrix. After zero-frequency replacement, an inverse Fourier transform is performed to generate a conserved projected time-domain output, and the complex spectrum matrix after projection adjustment is used as the conserved projected output.

[0012] Optionally, the spectral support sparse low-rank adaptation module includes: Perform spectral energy distribution calculation on the conserved projection output, calculate the sum of squared amplitudes of all channels corresponding to each discrete frequency component in the complex spectrum matrix, and obtain the total energy value of the frequency component; Select the set of frequencies whose cumulative percentage of total energy value exceeds a preset threshold; Obtain the original weight matrix of the frequency domain linear transformation layer of the corresponding frequency domain transformation module in the Fourier neural operator network; Based on the frequency component indices contained in the frequency support set, column vectors corresponding to these frequency components are extracted from the original weight matrix to form a weight submatrix; Construct a low-rank fitting increment matrix for the weighted submatrix; The low-rank adaptation increment matrix is ​​superimposed with the weight submatrix to generate the adapted weight submatrix. The adapted weight submatrix is ​​written back to the corresponding frequency component column position in the original weight matrix, replacing the original weight values. For the weight columns corresponding to other frequency components not included in the frequency support set, the original values ​​in the original weight matrix are kept unchanged, resulting in a spectral support sparse low-rank adapted weight structure.

[0013] Optionally, the residual spectrum update module includes: Obtain the time-domain predicted sequence and the actual production line output sequence corresponding to the conserved projection output; Element-wise difference operations are performed between the time-domain predicted sequence and the actual production line output sequence to obtain the time-domain residual sequence; Perform a discrete Fourier transform on the time-domain residual sequence along the time series dimension to obtain the residual complex spectrum matrix; Perform spectral energy calculation on the residual complex spectrum matrix, and perform amplitude square summation calculation on all work station sequences and variable channels corresponding to each discrete frequency component to obtain the residual frequency energy distribution vector; The matching degree is calculated based on the residual frequency energy distribution vector and the frequency support set. When the residual energy of the frequency component corresponding to the frequency support set accounts for a proportion of the total residual energy greater than the preset update threshold, the low-rank adaptation incremental matrix parameter update is performed. When the residual energy of the frequency component corresponding to the frequency support set accounts for less than the proportion of the total residual energy, the parameters of the low-rank adaptation increment matrix remain unchanged. Write the updated low-rank adaptation increment matrix parameters back to the spectrum to support the sparse low-rank adaptation weight structure.

[0014] Optionally, the coupling state prediction module includes: Receive input tensors and Fourier neural operator network parameters updated with a spectral-supported sparse low-rank adaptive weight structure; The input tensor is input into the Fourier neural operator network, and frequency domain transformation calculation, frequency domain linear mapping calculation, nonlinear mapping calculation and inverse frequency domain transformation calculation are performed in sequence to obtain the time domain coupling state prediction sequence of the production line station. Perform channel mapping processing on the temporal coupled state prediction sequence to extract the prediction results of the production line workstation coupled state variables; Perform time window reconstruction processing on the prediction results of the coupled state variables of the production line station to generate a time-domain prediction output tensor that is consistent with the time window of the input tensor. The time-domain prediction output tensor is input into the self-consistent conservation projection module to perform zero-frequency component conservation projection processing, thereby obtaining the conservation constraint prediction output. The output of the conservation constraint prediction is used as the output of the production line workstation coupling state prediction result.

[0015] The beneficial effects of this invention are: By constructing a frequency domain modeling structure based on Fourier neural operators, a high-precision representation of the coupled dynamic behavior of multiple workstations on a production line is achieved. Compared with traditional modeling methods that rely solely on time domain features, this approach can simultaneously characterize the dynamic changes of the system at different frequency scales. It can maintain stable prediction consistency under scenarios involving production line cycle fluctuations, multi-workstation linkage disturbances, and periodic load changes, thereby enhancing the adaptability of complex production line system modeling.

[0016] By introducing a self-consistent conservation projection mechanism, the material flow conservation relationship of the production line is embedded into the model prediction process, ensuring that the prediction results remain consistent at the global material balance level. By injecting conservation anchor point constraints at the zero-frequency component position in the frequency domain, the model output maintains consistency with physical laws while satisfying data fitting capabilities, avoiding the cumulative deviation problem that occurs in traditional data-driven models during long-term operation, and enhancing the credibility of the digital twin model in simulation analysis and operational decision-making.

[0017] By constructing a spectral-supported sparse low-rank adaptation mechanism and a residual spectrum-driven update mechanism, adaptive adjustment of model parameters for key frequency structures is achieved. While ensuring the stability of the main model structure, low-rank parameter updates are performed only in the frequency support region, reducing disturbances caused by ineffective parameter adjustments. Combined with a residual spectrum energy distribution-triggered update strategy, the model's continuous adaptability to changing operating conditions and disturbances is improved, while reducing the computational overhead of online model updates. Attached Figure Description

[0018] 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: Fig. 1 This is a flowchart of a twin modeling simulation analysis system based on AI algorithms proposed in this invention. Fig. 2 This is a frequency domain feature decoupling structure diagram of a twin modeling simulation analysis system based on AI algorithms proposed in this invention.

[0019] Fig. 3 This is a schematic diagram of the spectrum adaptive update and prediction closed loop of a twin modeling simulation analysis system based on AI algorithms proposed in this invention. Detailed Implementation

[0020] 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.

[0021] refer to Figs. 1-3 A twin modeling simulation analysis system based on AI algorithms includes: The data tensor construction module is used to collect the operation history of production line workstations and generate input tensors; The frequency domain transformation module is used to perform frequency domain transformation on the input tensor input to the Fourier neural operator network to obtain the complex spectrum feature matrix; The phase-amplitude decoupling mapping module is used to perform spectral amplitude component and spectral phase component decomposition on the complex spectral feature matrix, and generate an updated complex spectral feature matrix based on the amplitude mapping matrix and the phase mapping matrix. The self-consistent conservation projection module is used to calculate the conservation anchor point value based on the cumulative material count and the differential integral of the change in the quantity of work-in-process within the input tensor time window, and to perform zero-frequency component replacement projection on the prediction output corresponding to the updated complex spectrum feature matrix to obtain the conservation projection output. The spectral support sparse low-rank adaptation module is used to determine the frequency support set based on the energy distribution of the input tensor spectrum, extract the weight submatrix corresponding to the frequency support set, construct a low-rank adaptation increment matrix at the position of the weight submatrix and superimpose it with the original spectrum weights to obtain the spectral support sparse low-rank adaptation weight structure. The residual spectrum update module is used to construct a residual tensor based on the conserved projected output and the true output and generate a residual spectrum energy distribution, and update the parameters of the low-rank adaptation increment matrix based on the residual spectrum energy distribution; The coupling state prediction module is used to write the input tensor back to the Fourier neural operator backbone network, perform inverse frequency domain mapping and zero-frequency component replacement projection, and output the coupling state prediction results of the production line station.

[0022] In this embodiment, the data tensor construction module includes: Acquire historical operation data for each workstation on the production line. The historical operation data includes process parameter records, equipment operating status records, and quality inspection records. The historical operation data of each workstation is resampled according to a uniform sampling time interval, and timestamp alignment is performed to form time series data. In this embodiment, time resampling adopts a fixed time interval resampling method, with the resampling time interval ranging from 1 second to 60 seconds. When the original sampling time interval is less than the resampling time interval, downsampling is performed using a sliding window averaging method. When the original sampling time interval is greater than the resampling time interval, upsampling is performed using a linear interpolation method. When missing data occurs, it is filled with the data from the most recent valid time point. The time series data is processed according to the production line process flow sequence to form workstation sequence data; the workstation sequence is determined according to the production line process flow path; when there are parallel workstations, the workstation sequence is arranged using a fixed sequence numbering method, and the parallel workstation data is mapped to the same workstation sequence level; Perform variable classification mapping on the workstation sequence data, mapping process parameter data, equipment status data, and quality inspection data to variable channel dimensions respectively; The mapped data is processed using a three-dimensional stacking method based on the workstation sequence dimension, time series dimension, and variable channel dimension to generate an input tensor. The input tensor has a time window length of 300 sampling time points. When the number of workstations is less than the preset number of workstations, zero-value padding is used; when the number of workstations is greater than the preset number of workstations, truncation is used to retain the workstation data of the main process path.

[0023] In this embodiment, the frequency domain transformation module includes: Receives a 3D input tensor generated by the data tensor construction module and inputs the input tensor into a pre-constructed Fourier neural operator network; The input tensor is subjected to a Fast Fourier Transform (FFT) in the time-series dimension through at least one Fourier layer in the Fourier neural operator network, mapping the time-domain signal to the frequency domain. In this embodiment, there are 3 Fourier layers. Each Fourier layer performs a Discrete Fourier Transform (DFT) calculation in the time-series dimension, and the frequency domain transform length is 256 sampling points. When the time window length is less than 256 sampling points, it is extended to 256 sampling points using zero-padding. Extract the complex spectral data after Fourier transform; the complex spectral data includes a real part spectral matrix and an imaginary part spectral matrix; the real and imaginary parts are stored in a dual-channel tensor format; Based on the hierarchical structure of the Fourier neural operator network, specified linear transformations and nonlinear activation operations are performed on the spectral data to generate frequency domain mapping feature data. The linear transformation is calculated using frequency domain weight matrix multiplication; the nonlinear mapping is performed element-wise using the ReLU function; the frequency domain feature data retains the first 128 discrete frequency components; frequency data above the 128th discrete frequency component is truncated. The frequency domain feature data is organized into a complex spectral feature matrix. The row dimension of the matrix corresponds to the combined features of the workstation sequence and variable channels, and the column dimension corresponds to the discrete frequency components obtained after transformation. The frequency domain transformation calculation is performed in batch mode; the batch size is 32.

[0024] In this embodiment, the phase amplitude decoupling mapping module includes: For each complex element in the complex spectral characteristic matrix, calculate the modulus to obtain the spectral amplitude component matrix, and calculate the argument to obtain the spectral phase component matrix; For the spectral amplitude component matrix, a linear transformation is performed through a trainable amplitude mapping matrix, followed by a nonlinear mapping using the ReLU activation function to obtain the updated spectral amplitude component matrix; the dimension of the amplitude mapping matrix matches that of the spectral amplitude component matrix. For the spectral phase component matrix, a linear transformation is performed through a trainable phase mapping matrix to obtain the updated spectral phase component matrix; the dimension of the phase mapping matrix matches that of the spectral phase component matrix; in this embodiment, both the amplitude mapping matrix and the phase mapping matrix adopt a low-rank decomposition structure, that is, it is represented as the product of two low-dimensional matrices, reducing the number of parameters. Based on the updated spectral amplitude component matrix and the updated spectral phase component matrix, an updated complex spectral feature matrix is ​​recombined through complex number synthesis calculation. The complex number synthesis calculation method is as follows: using the updated amplitude matrix as the modulus and the updated phase matrix as the argument, the corresponding real and imaginary parts of the complex number are calculated.

[0025] In this embodiment, the self-consistent conservation projection module includes: Extract material flow data within a preset time window from the input tensor. The material flow data includes at least the input material count and output material count for each workstation. The preset time window length is 256 sampling points. Calculate the cumulative material count by summing the total amount of materials input at the starting workstation of the production line within the time window; Calculate the differential integral of the change in the quantity of work-in-process, and calculate the instantaneous sum of the quantity of work-in-process at all stations at each sampling moment within the time window; calculate the net change by differential integration; calculate the difference between the start and end values ​​of the instantaneous sum sequence; and perform numerical integration on the difference over the time window to obtain the net change. According to the law of conservation of material, the cumulative amount of material count is subtracted from the differential integral of the change in the quantity of work-in-process, and the result is used as the theoretical total output material conservation anchor value. Perform an inverse fast Fourier transform on the updated complex spectrum feature matrix to transform it back to the time domain and obtain a preliminary time-domain prediction sequence; From the preliminary time-domain prediction sequence, the prediction channel data corresponding to the total output material is extracted, and the data is accumulated and summed over the time window to obtain the total output material value predicted by the model. The deviation between the theoretical total output material conservation anchor point value and the model-predicted total output material value is calculated. In this embodiment, the model-predicted total output material value is obtained by summing the total output channels in the preliminary time-domain prediction sequence over a time window. The deviation value is obtained by subtracting the model-predicted total output material value from the theoretical conservation anchor point value. Extract the zero-frequency components corresponding to all channels in the updated complex spectrum feature matrix. The zero-frequency components correspond to the complex coefficients with a frequency of zero in the spectrum. The zero-frequency component correction value is calculated, which is proportional to the deviation and inversely proportional to the time window length and the number of workstations. In this embodiment, the zero-frequency component correction value is obtained by dividing the deviation value by the time window length. The zero-frequency components of each channel in the updated complex spectrum feature matrix are replaced with the correction value, keeping all non-zero frequency components unchanged, and a projection-adjusted complex spectrum matrix is ​​generated. In this embodiment, the zero-frequency component replacement operation is performed once after each time window prediction is completed. The zero-frequency component correction only applies to the spectral coefficients corresponding to the conserved channels. After zero-frequency replacement, an inverse Fourier transform is performed to generate a conserved projected time-domain output, and the complex spectrum matrix after projection adjustment is used as the conserved projected output.

[0026] In this embodiment, the spectral support sparse low-rank adaptation module includes: Perform spectral energy distribution calculation on the conserved projection output, calculate the sum of squared amplitudes of all channels corresponding to each discrete frequency component in the complex spectrum matrix, and obtain the total energy value of the frequency component; Select the lowest frequency set whose cumulative percentage of total energy value exceeds a preset threshold of 0.9; Obtain the original weight matrix of the frequency domain linear transformation layer of the corresponding frequency domain transformation module in the Fourier neural operator network; Based on the frequency component indices contained in the frequency support set, column vectors corresponding to these frequency components are extracted from the original weight matrix to form a weight submatrix; Construct a low-rank fitting increment matrix for the weight submatrix; initialize two low-dimensional matrices A and B, where the number of rows in matrix A is equal to the number of rows in the weight submatrix, and the number of columns is rank r; the number of rows in matrix B is rank r, and the number of columns is equal to the number of columns in the weight submatrix, and r is much smaller than the number of rows and columns of the weight submatrix; the low-rank fitting increment matrix is ​​calculated by multiplying matrix A and matrix B. The low-rank adaptation increment matrix and the weight submatrix are superimposed to generate the adapted weight submatrix. The superposition method is: adapted weight submatrix = original weight submatrix + α × (low-rank adaptation increment matrix), where α is the adaptation coefficient; in this embodiment, α is 0.1. The adapted weight submatrix is ​​written back to the corresponding frequency component column position in the original weight matrix, replacing the original weight values. For the weight columns corresponding to other frequency components not included in the frequency support set, the original values ​​in the original weight matrix are kept unchanged, resulting in a spectral support sparse low-rank adapted weight structure.

[0027] In this embodiment, the residual spectrum update module includes: Obtain the time-domain predicted sequence and the actual production line output sequence corresponding to the conserved projection output; Element-wise difference operations are performed between the time-domain predicted sequence and the actual production line output sequence to obtain the time-domain residual sequence; A discrete Fourier transform is performed on the time-domain residual sequence along the time series dimension to obtain the residual complex spectrum matrix. In this embodiment, the discrete Fourier transform is implemented using the fast Fourier transform. The residual complex spectrum matrix is ​​stored using a dual-channel tensor structure with a real part matrix and an imaginary part matrix. Spectral energy calculation is performed on the residual complex spectrum matrix. For each discrete frequency component, the amplitude square summation is performed on all workstation sequences and variable channels to obtain the residual frequency energy distribution vector. The spectral energy calculation adopts the complex modulus square calculation method, that is, the summation of the real part square and the imaginary part square. The matching degree is calculated based on the residual frequency energy distribution vector and the frequency support set. When the proportion of the residual energy of the frequency component corresponding to the frequency support set to the total residual energy is greater than the preset update threshold of 0.35, the low-rank adaptation incremental matrix parameter update is performed. In this embodiment, the total residual energy is the sum of the energies of all discrete frequency components. The proportion of the residual energy corresponding to the frequency support set is calculated by dividing the sum of the energies of each frequency in the frequency support set by the total residual energy. Only the low-rank adaptation incremental matrix parameter corresponding to the frequency support set is updated; the corresponding parameters outside the frequency support set remain unchanged. When the residual energy of the frequency component corresponding to the frequency support set is less than the proportion of the total residual energy, the parameters of the low-rank adaptation incremental matrix remain unchanged. In this embodiment, when the total residual energy is less than the preset residual stability threshold, the update of the parameters of the low-rank adaptation incremental matrix is ​​paused, and the residual stability threshold is 0.05. The updated low-rank adaptation increment matrix parameters are written back to the spectral support sparse low-rank adaptation weight structure. In this implementation, the parameter write-back operation is performed once after the prediction is completed in each time window.

[0028] In this embodiment, the coupling state prediction module includes: Receive input tensors and Fourier neural operator network parameters updated with a spectral-supported sparse low-rank adaptive weight structure; The input tensor is input into the Fourier neural operator network, and frequency domain transformation calculation, frequency domain linear mapping calculation, nonlinear mapping calculation and inverse frequency domain transformation calculation are performed in sequence to obtain the time domain coupling state prediction sequence of the production line station. Perform channel mapping processing on the temporal coupled state prediction sequence to extract the prediction results of the production line workstation coupled state variables; Perform time window reconstruction processing on the prediction results of the coupled state variables of the production line station to generate a time-domain prediction output tensor that is consistent with the time window of the input tensor. The time-domain prediction output tensor is input into the self-consistent conservation projection module to perform zero-frequency component conservation projection processing, thereby obtaining the conservation constraint prediction output. The output of the conservation constraint prediction is used as the output of the production line workstation coupling state prediction result.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a scenario of predicting and simulating the operating status of a multi-station continuous manufacturing line. In this scenario, the production line contains multiple continuous processing stations, and there are material transfer relationships and cycle time coupling relationships between the stations. In traditional modeling methods, time series prediction models are usually built based solely on historical statistical data. When the production line load changes, equipment status fluctuates, or process parameters are adjusted, the prediction error increases. Furthermore, after long-term operation, the predicted cumulative output results are prone to deviating from the actual output results, failing to maintain the overall material balance, leading to significant differences between simulation results and actual operating results. This invention focuses on solving the problems of insufficient utilization of frequency domain information, difficulty in maintaining material conservation relationships, and insufficient online adaptability of models in traditional methods.

[0030] In practical applications, historical and real-time data from each workstation are first collected, and a multi-dimensional input tensor is constructed to fully describe the changes in the workstation's operating status. Then, a Fourier neural operator is used to perform frequency domain mapping on the input data, decomposing the production line's operating status into dynamic features at different frequency scales. A phase amplitude decoupling structure is then used to map the spectral amplitude and phase separately, enabling the model to independently learn the characteristics of changes in operating intensity and cycle time. After obtaining the frequency domain prediction results, a self-consistent conservation projection structure is used to adjust the prediction results with conservation constraints, ensuring that the predicted cumulative output is consistent with the material statistics, thus avoiding long-term cumulative deviations.

[0031] During continuous model operation, residual spectrum analysis identifies the frequency bands where errors are mainly concentrated, and low-rank parameter updates are performed only on key frequency ranges to avoid instability issues caused by frequent adjustments to the entire model parameters. Simultaneously, a frequency-supported sparsity adaptation mechanism enables the model to perform local adaptive optimization while maintaining overall structural stability, improving its responsiveness to changes in operating conditions. Through these methods, the production line operating status prediction results maintain high stability in both short-term fluctuations and long-term trends.

[0032] In continuous operation tests, traditional models showed a gradual increase in cumulative output prediction bias, reaching a high level in the later stages of operation. However, the method of this invention consistently maintained a low deviation in cumulative output prediction, while also demonstrating better adaptability to sudden fluctuations in operating conditions. In multiple rounds of operation tests, the method of this invention outperformed traditional modeling methods in terms of output prediction accuracy, state prediction stability, and long-term operational consistency.

[0033] Table 1: Comparison of Production Line Twin Modeling Effects

[0034] As shown in Table 1, under continuous multi-set testing conditions, the method of this invention demonstrates advantages in both prediction accuracy and long-term operational stability. The prediction error of the traditional model is generally distributed between 4.5% and 5.5%, with an average prediction error of approximately 5.06%, while the prediction error of the method of this invention is stably controlled between 1.6% and 2.3%, with an average prediction error of approximately 1.95%, representing a reduction in prediction error of over 60%. Regarding the cumulative deviation index, the cumulative deviation of the traditional model is concentrated in the range of 6.4% to 7.5%, with an average cumulative deviation of approximately 6.94%, while the cumulative deviation of the method of this invention remains within the range of 1.9% to 2.6%, with an average cumulative deviation of approximately 2.24%, indicating that this invention can effectively suppress error accumulation during long-term operation.

[0035] Regarding the state stability index, the method of this invention consistently maintains a value between 0.93 and 0.96, which is higher than the corresponding level of traditional models. This indicates that the present invention can maintain higher consistency and resistance to fluctuations in the process of multi-station coupled state prediction. In summary, through frequency domain modeling, conservation constraint injection, and spectrum-driven adaptive update mechanisms, the present invention enables the model to respond quickly to changes in operating state while maintaining long-term stability of operating results. The data results verify the practical application value of this invention in complex production line modeling and prediction scenarios.

[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 twin modeling simulation analysis system based on AI algorithms, characterized in that, include: The data tensor construction module is used to collect the operation history of production line workstations and generate input tensors; The frequency domain transformation module is used to perform frequency domain transformation on the input tensor input to the Fourier neural operator network to obtain the complex spectrum feature matrix; The phase-amplitude decoupling mapping module is used to perform spectral amplitude component and spectral phase component decomposition on the complex spectral feature matrix, and generate an updated complex spectral feature matrix based on the amplitude mapping matrix and the phase mapping matrix. The self-consistent conservation projection module is used to calculate the conservation anchor point value based on the cumulative material count and the differential integral of the change in the quantity of work-in-process within the input tensor time window, and to perform zero-frequency component replacement projection on the prediction output corresponding to the updated complex spectrum feature matrix to obtain the conservation projection output. The spectral support sparse low-rank adaptation module is used to determine the frequency support set based on the energy distribution of the input tensor spectrum, extract the weight submatrix corresponding to the frequency support set, construct a low-rank adaptation increment matrix at the position of the weight submatrix and superimpose it with the original spectrum weights to obtain the spectral support sparse low-rank adaptation weight structure. The residual spectrum update module is used to construct a residual tensor based on the conserved projected output and the true output and generate a residual spectrum energy distribution, and update the parameters of the low-rank adaptation increment matrix based on the residual spectrum energy distribution; The coupling state prediction module is used to write the input tensor back to the Fourier neural operator backbone network, perform inverse frequency domain mapping and zero-frequency component replacement projection, and output the coupling state prediction results of the production line station.

2. The twin modeling simulation analysis system based on AI algorithm according to claim 1, characterized in that, The data tensor construction module includes: Acquire historical operation data for each workstation on the production line. The historical operation data includes process parameter records, equipment operating status records, and quality inspection records. The historical operation data of each workstation is resampled according to a uniform sampling time interval, and timestamp alignment is performed to form time series data. The time series data is processed according to the production line process flow sequence to form workstation sequence data; Perform variable classification mapping on the workstation sequence data, mapping process parameter data, equipment status data, and quality inspection data to variable channel dimensions respectively; The mapped data is processed using a three-dimensional stacking method based on the workstation sequence dimension, time series dimension, and variable channel dimension to generate an input tensor.

3. The twin modeling simulation analysis system based on AI algorithm according to claim 2, characterized in that, The frequency domain transformation module includes: Receives a 3D input tensor generated by the data tensor construction module and inputs the input tensor into a pre-constructed Fourier neural operator network; By using at least one Fourier layer in the Fourier neural operator network, the input tensor is subjected to a fast Fourier transform in the time series dimension, mapping the time-domain signal to the frequency domain. Extract the complex form of the spectrum data after Fourier transform; Based on the hierarchical structure of the Fourier neural operator network, specified linear transformations and nonlinear activation operations are performed on the spectral data to generate frequency domain mapping feature data. The frequency domain feature data is organized into a complex spectrum feature matrix. The row dimension of the matrix corresponds to the combined features of the work station sequence and the variable channel, and the column dimension corresponds to the discrete frequency components obtained after transformation.

4. The twin modeling simulation analysis system based on AI algorithm according to claim 3, characterized in that, The phase amplitude decoupling mapping module includes: For each complex element in the complex spectral characteristic matrix, calculate the modulus to obtain the spectral amplitude component matrix, and calculate the argument to obtain the spectral phase component matrix; For the spectral amplitude component matrix, a linear transformation is performed through a trainable amplitude mapping matrix, followed by a nonlinear mapping, to obtain the updated spectral amplitude component matrix; The updated spectral phase component matrix is ​​obtained by performing a linear transformation on the spectral phase component matrix through a trainable phase mapping matrix. Based on the updated spectral amplitude component matrix and the updated spectral phase component matrix, an updated complex spectral feature matrix is ​​generated by complex number synthesis calculation.

5. The twin modeling simulation analysis system based on AI algorithm according to claim 4, characterized in that, The self-consistent conservation projection module includes: Extract material flow data within a preset time window from the input tensor. The material flow data includes at least the input material count and output material count for each workstation. Calculate the cumulative material count by summing the total amount of materials input at the starting workstation of the production line within the time window; Calculate the differential integral of the change in the quantity of work-in-process, and calculate the instantaneous sum of the quantity of work-in-process at all stations at each sampling moment within the time window; the net change is obtained by differential integral calculation. According to the law of conservation of material, the cumulative amount of material count is subtracted from the differential integral of the change in the quantity of work-in-process, and the result is used as the theoretical total output material conservation anchor value. Perform an inverse fast Fourier transform on the updated complex spectrum feature matrix to transform it back to the time domain and obtain a preliminary time-domain prediction sequence; From the preliminary time-domain prediction sequence, the prediction channel data corresponding to the total output material is extracted, and the data is accumulated and summed over the time window to obtain the total output material value predicted by the model. The deviation between the theoretical total output material conservation anchor point value and the total output material value predicted by the model is calculated. Extract the zero-frequency components corresponding to all channels in the updated complex spectrum feature matrix. The zero-frequency components correspond to the complex coefficients with a frequency of zero in the spectrum. Calculate the zero-frequency component correction value, replace the zero-frequency component of each channel in the updated complex spectrum feature matrix with the correction value, keep all non-zero frequency components unchanged, and generate the projection-adjusted complex spectrum matrix. After zero-frequency replacement, an inverse Fourier transform is performed to generate a conserved projected time-domain output, and the complex spectrum matrix after projection adjustment is used as the conserved projected output.

6. The twin modeling simulation analysis system based on AI algorithm according to claim 5, characterized in that, The spectral support sparse low-rank adaptation module includes: Perform spectral energy distribution calculation on the conserved projection output, calculate the sum of squared amplitudes of all channels corresponding to each discrete frequency component in the complex spectrum matrix, and obtain the total energy value of the frequency component; Select the set of frequencies whose cumulative percentage of total energy value exceeds a preset threshold; Obtain the original weight matrix of the frequency domain linear transformation layer of the corresponding frequency domain transformation module in the Fourier neural operator network; Based on the frequency component indices contained in the frequency support set, column vectors corresponding to these frequency components are extracted from the original weight matrix to form a weight submatrix; Construct a low-rank fitting increment matrix for the weighted submatrix; The low-rank adaptation increment matrix is ​​superimposed with the weight submatrix to generate the adapted weight submatrix. The adapted weight submatrix is ​​written back to the corresponding frequency component column position in the original weight matrix, replacing the original weight values. For the weight columns corresponding to other frequency components not included in the frequency support set, the original values ​​in the original weight matrix are kept unchanged, resulting in a spectral support sparse low-rank adapted weight structure.

7. The twin modeling simulation analysis system based on AI algorithm according to claim 6, characterized in that, The residual spectrum update module includes: Obtain the time-domain predicted sequence and the actual production line output sequence corresponding to the conserved projection output; Element-wise difference operations are performed between the time-domain predicted sequence and the actual production line output sequence to obtain the time-domain residual sequence; Perform a discrete Fourier transform on the time-domain residual sequence along the time series dimension to obtain the residual complex spectrum matrix; Perform spectral energy calculation on the residual complex spectrum matrix, and perform amplitude square summation calculation on all work station sequences and variable channels corresponding to each discrete frequency component to obtain the residual frequency energy distribution vector; The matching degree is calculated based on the residual frequency energy distribution vector and the frequency support set. When the residual energy of the frequency component corresponding to the frequency support set accounts for a proportion of the total residual energy greater than the preset update threshold, the low-rank adaptation incremental matrix parameter update is performed. When the residual energy of the frequency component corresponding to the frequency support set accounts for less than the proportion of the total residual energy, the parameters of the low-rank adaptation increment matrix remain unchanged. Write the updated low-rank adaptation increment matrix parameters back to the spectrum to support the sparse low-rank adaptation weight structure.

8. The twin modeling simulation analysis system based on AI algorithm according to claim 7, characterized in that, The coupling state prediction module includes: Receive input tensors and Fourier neural operator network parameters updated with a spectral-supported sparse low-rank adaptive weight structure; The input tensor is input into the Fourier neural operator network, and frequency domain transformation calculation, frequency domain linear mapping calculation, nonlinear mapping calculation and inverse frequency domain transformation calculation are performed in sequence to obtain the time domain coupling state prediction sequence of the production line station. Perform channel mapping processing on the temporal coupled state prediction sequence to extract the prediction results of the production line workstation coupled state variables; Perform time window reconstruction processing on the prediction results of the coupled state variables of the production line station to generate a time-domain prediction output tensor that is consistent with the time window of the input tensor. The time-domain prediction output tensor is input into the self-consistent conservation projection module to perform zero-frequency component conservation projection processing, thereby obtaining the conservation constraint prediction output. The output of the conservation constraint prediction is used as the output of the production line workstation coupling state prediction result.

Images