Multivariable time sequence prediction and closed-loop control method based on digital twinning and deep learning fusion
By constructing a five-dimensional digital twin model and integrating deep learning, the problems of multi-source heterogeneous data fusion and real-time performance in complex industrial processes were solved, achieving high-precision fault diagnosis and control, and improving the system's autonomous optimization capability and robustness.
Patent Information
- Application Number
- CN202511606785.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-01-13
AI Technical Summary
Traditional digital twin modeling struggles to establish accurate analytical models in complex, nonlinear, and strongly coupled industrial processes. It also faces challenges in fusing multi-source heterogeneous data, lacks real-time performance and physical constraints, and suffers from poor generalization ability of deep learning models outside of data scarcity or distribution. Furthermore, it lacks support from big data science and comprehensive interaction between information and physical data.
A multi-level data architecture based on a five-dimensional digital twin model is constructed. Data preprocessing and SMOTE resampling are employed, along with temporal feature extraction based on 1D-CNN, virtual twin data generation using Bi-LSTM and attention mechanisms, multi-modal fault diagnosis based on an RNN-LSTM hybrid network, and adaptive feedback control with Lyapunov stability is introduced. A multi-task learning mechanism is used to optimize virtual twin data generation and fault diagnosis.
It significantly improves the fault diagnosis effect and system autonomous optimization capability of complex industrial processes, enhances prediction accuracy and control performance, solves the problems of data scarcity and real-time performance, and strengthens the robustness and interpretability of the system.
Smart Images
Figure CN121325610A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of industrial intelligent control and predictive maintenance technology, and in particular to a multivariable time-series prediction and closed-loop control method based on the fusion of digital twins and deep learning. Background Technology
[0002] Traditional digital twin modeling relies heavily on precise physical mechanisms and prior knowledge, making it difficult to establish accurate analytical models for complex, nonlinear, and strongly coupled industrial processes. Furthermore, it has inherent bottlenecks in issues such as multi-source heterogeneous data fusion, real-time performance, and "simulation without realism."
[0003] Deep learning technology can automatically extract high-order features from high-dimensional heterogeneous data and establish complex nonlinear mapping relationships, making it possible to replace traditional physical modeling. However, purely data-driven deep learning models lack physical interpretability, have poor generalization ability in scenarios where training data is scarce or out of distribution, and lack an understanding of the basic constraints of physical systems, which may lead to output results that violate physical common sense.
[0004] Digital twins and deep learning are significantly complementary. Digital twins provide physical constraints, mechanistic frameworks, and real-time interactive interfaces, while deep learning offers powerful data representation and pattern recognition capabilities. Their integration can significantly enhance the capabilities of digital twins in perception, modeling, prediction, and control, thereby forming a more intelligent and reliable closed-loop system.
[0005] In existing technologies, the integration of digital twins and deep learning still suffers from the following technical shortcomings: At the data level, there are difficulties in integrating multi-source heterogeneous data and data sparsity issues in key scenarios.
[0006] At the decision-making level, there is a lack of support from big data science; at the execution level, there is a lack of comprehensive interaction between information and physical data.
[0007] At the modeling level, it is difficult to meet real-time requirements and adapt to the needs of complex systems.
[0008] To address the aforementioned technical deficiencies, this application provides a multivariable time-series prediction and closed-loop control method based on the fusion of digital twins and deep learning. Summary of the Invention
[0009] To overcome the problems existing in related technologies, this application provides a multivariable time series prediction and closed-loop control method based on the fusion of digital twins and deep learning, including: Construct a multi-layered data architecture based on a five-dimensional digital twin model; Data preprocessing and SMOTE resampling; Temporal feature extraction based on 1D-CNN; Virtual twin data generation based on Bi-LSTM and attention mechanism; Multimodal fault diagnosis based on RNN-LSTM hybrid network; Adaptive feedback control based on Lyapunov stability; A multi-task learning mechanism is adopted to jointly optimize virtual twin data generation and fault diagnosis.
[0010] In one implementation, a multi-layered data architecture based on a five-dimensional digital twin model is constructed, specifically including: The multi-source heterogeneous data generated during industrial manufacturing processes are reconstructed into physical layer data, control layer data, operation layer data, analysis layer data, and virtual layer data. Physical layer data consists of the basic parameters of physical entities, control layer data consists of control system commands, operation layer data consists of operation parameters set manually or by the system, analysis layer data consists of decision data generated by simulation, AI, or big data analysis, and virtual layer data consists of twin data generated by deep learning models.
[0011] In one implementation, data preprocessing and SMOTE resampling specifically include: Perform zero-mean standardization on numerical features; The SMOTE algorithm is used to oversample the minority class samples in the training set.
[0012] In one implementation, temporal feature extraction based on 1D-CNN specifically includes: using a convolutional neural network to extract discriminative features from enhanced time-series data.
[0013] In one implementation, the generation of virtual twin data based on Bi-LSTM and an attention mechanism specifically includes: A bidirectional long short-term memory network is used to process the input sequence of the operation layer. ; Forward LSTM computation: ; Backward LSTM computation: ; Concatenate the forward and backward hidden states as follows: ; Introducing an attention mechanism to calculate weights:
[0014] The output of the control layer is obtained by weighted fusion and projection of attention weights:
[0015] The output is a high-fidelity control layer response sequence. As virtual twin data.
[0016] In one implementation, multimodal fault diagnosis based on an RNN-LSTM hybrid network specifically includes: An adaptive feedback controller is designed by incorporating Lyapunov stability theory. Define Lyapunov functions To quantify the degree of deviation between the current process status and safety objectives:
[0017] in This represents the anomaly probability predicted by the model, and 0.5 is the safety threshold. The system's goal is to adjust the operating parameters... make ; The parameter update strategy employs gradient descent combined with a chaos compensation term to enhance global search capabilities.
[0018] in, It is the gradient of the Lyapunov function with respect to the parameters, calculated through automatic differentiation:
[0019] in, It is an adaptive step size. It can be dynamically adjusted according to the gradient magnitude; It is a chaos compensation term. ; Introducing controllable randomness to escape local optima and ensure that the updated parameters satisfy physical constraints:
[0020] in, It is a projection operator.
[0021] In one implementation, a multi-task learning mechanism is employed to jointly optimize virtual twin data generation and fault diagnosis, specifically including: The total loss function is a weighted sum of the losses from the two tasks:
[0022] Training uses the Adam optimizer with an initial learning rate of The learning rate is dynamically adjusted using the ReduceLROnPlateau scheduler.
[0023] The technical solution provided in this application may include the following beneficial effects: This application addresses the extreme imbalance problem in industrial data by synthesizing minority class samples using the SMOTE resampling algorithm. It employs a one-dimensional convolutional neural network (1D-CNN) to enhance the local features and temporal dependencies of time series data. A bidirectional long short-term memory network (LSTM) combined with an attention mechanism is used to achieve high-precision temporal mapping from the operational layer to the control layer, generating virtual twin data. This virtual data is then fused with physical layer environmental parameters, and fault diagnosis and state recognition are performed using a hybrid RNN and LSTM network. Finally, an adaptive feedback control mechanism based on Lyapunov stability theory is introduced, using gradient descent and chaos compensation terms to achieve real-time optimization and stability control of process parameters, ensuring system convergence and robustness. This method demonstrates excellent prediction accuracy and control performance on a semiconductor etching process dataset, significantly improving fault diagnosis and system autonomous optimization capabilities in complex industrial processes.
[0024] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0025] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.
[0026] Figure 1 This is a schematic diagram of the perception-modeling-process shown in the embodiments of this application; Figure 2 This is a schematic diagram of the CRD-LSTM closed-loop frame structure shown in the embodiments of this application; Figure 3 This is a comparison chart of the SMOTE resampling effects shown in the embodiments of this application; Figure 4 This is a training loss curve shown in an embodiment of this application; Figure 5 This is a verification loss curve diagram shown in the embodiments of this application; Figure 6 This is a comparison chart of fault diagnosis performance shown in the embodiments of this application. Detailed Implementation
[0027] Preferred embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.
[0028] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0029] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0030] like Figures 1 to 6 As shown in the figure, this application provides a multivariable time series prediction and closed-loop control method based on the fusion of digital twin and deep learning, including the following steps: 101. Construct a multi-layered data architecture based on a five-dimensional digital twin model; 102. Data preprocessing and SMOTE resampling; 103. Temporal feature extraction based on 1D-CNN; 104. Virtual twin data generation based on Bi-LSTM and attention mechanism; 105. Multimodal fault diagnosis based on RNN-LSTM hybrid network; 106. Adaptive feedback control based on Lyapunov stability; 107. Employ a multi-task learning mechanism to jointly optimize virtual twin data generation and fault diagnosis.
[0031] The following describes an embodiment of the method shown in this application, using etching processes in semiconductor manufacturing as an example.
[0032] In step 102, data preparation and preprocessing include: The validation was performed using the PHM Data Challenge 18 public dataset (etching tool fault detection). The original training data consists of 20 files with a total of 82,189,440 rows of data, covering 24 sensor parameters and operating condition variables.
[0033] Data cleaning and reconstruction: First, redundant fields were removed, and categorical variables such as Tool_x were merged, resulting in an optimized dataset with 5,110,542 rows × 23 columns. This dataset contains numerical sensor signals (such as voltage, current, and pressure) and categorical operating condition parameters (such as tool number and recipe).
[0034] Dataset partitioning: The data was divided into a training set (90%), a validation set (1%), and a test set (9%) in chronological order. This partitioning method strictly adheres to the temporal characteristics of industrial time-series data to prevent future information leakage.
[0035] Data standardization: Numerical features are standardized to zero mean. For the first... The standardized formula for the feature is:
[0036] in, and The mean and standard deviation are calculated only from the training set, ensuring the unbiasedness of the model evaluation.
[0037] Furthermore, SMOTE resampling is used to address extreme data imbalance.
[0038] The dataset suffers from severe class imbalance, with normal samples (labeled 0) accounting for as much as 99.98%. To address this issue, the SMOTE algorithm is applied only to the minority class (faulty class) samples on the training set.
[0039] Nearest neighbor search: for the training set Each minority class sample Calculate its k nearest neighbor set in the feature space. This implementation sets... .
[0040] Synthetic sample generation: For each Randomly select a sample from its k nearest neighbors. New samples are synthesized using linear interpolation:
[0041] in It is a random number that is uniformly distributed in the range [0, 1].
[0042] Non-numerical feature processing: For the corresponding non-numerical features (such as tool IDs), a random sampling strategy is used to copy them from the nearest neighbor samples.
[0043] Ultimately, through SMOTE resampling, the number of minority class samples in the training set was effectively increased, such as... Figure 3 As shown, the newly generated data maintains the distribution pattern of the original data, providing a balanced learning foundation for the model.
[0044] In step 103, temporal feature extraction includes: Resampled multi-sensor time-series data is input into a one-dimensional convolutional neural network (1D-CNN) to enhance local temporal dependencies and extract discriminative features.
[0045] Convolution operation: For a length of input sequence and a size of convolution kernel Convolution operations are implemented using a sliding window. Output sequence. The Middle The formula for calculating each element is:
[0046] This operation can be represented in matrix form:
[0047] Feature enhancement: Through learning from multiple convolutional kernels, the model can extract various meaningful local patterns (such as rising edges, falling edges, and peaks) from the original signal, and output feature maps. ,in This represents the number of convolution kernels.
[0048] This step aims to achieve a high-precision mapping from operational layer parameters to control layer responses using a deep learning model, generating virtual twin data. Step 104, virtual twin generation, specifically includes: Network input: The sequence of inputs to the operation layer is represented as follows .
[0049] Bidirectional LSTM processing: Forward LSTM processes sequences from front to back: ; Backward LSTM processes sequences from back to front: ; in, , To hide the state dimension.
[0050] Attention mechanism: Concatenate the forward and backward hidden states at each time step. Through a learnable weight vector Calculate the attention weights, which determine the importance of each time step's contribution to the output:
[0051] Attention weights are normalized using the Softmax function, satisfying the following conditions: .
[0052] Output Generation: Finally, the control layer prediction output at this time step is obtained by linear projection of the attention-weighted fusion of the hidden states.
[0053] in To output the projection matrix, This represents element-wise multiplication. The output of this stage is a sequence of control layer responses. As high-quality virtual twin data.
[0054] This application embodiment fuses virtual data with physical environment data for fault diagnosis. Step 105 specifically includes: Multimodal Fault Diagnosis: RNN-LSTM Hybrid Network By fusing virtual data with physical environment data, fault diagnosis can be performed.
[0055] Data fusion: integrating virtual twin sequences With physical layer environmental parameters The sequences are concatenated to form a multimodal input sequence. .
[0056] RNN feature extraction: Input is fed into an RNN network for initial temporal feature extraction:
[0057] LSTM deep modeling: converting the output of an RNN into a deep model. Input to an LSTM network to capture long-term dependencies:
[0058] Fault probability prediction: The fault probability is finally output through a fully connected layer and a sigmoid activation function.
[0059] The loss function for this stage is binary cross-entropy loss:
[0060] To achieve closed-loop control, an adaptive controller based on Lyapunov stability is introduced. Step 106 specifically includes: Define Lyapunov functions To quantify the deviation between the current state of the system and the security objective:
[0061] in The fault diagnosis module is at the time step The predicted anomaly probability is 0.5, which is a preset safety threshold. The control objective is to ensure... .
[0062] Gradient calculation: via automatic differentiation For operating parameters gradient:
[0063] Parameter update: Iterative optimization is performed using gradient descent with chaos compensation.
[0064] Wherein: Adaptive step size: Chaos compensation term: ; It is a projection operator that ensures the updated parameters satisfy physical constraints (such as the range of values); this update rule guarantees the convergence of the system. Experiments show that the system typically converges within 3-5 iterations.
[0065] In this embodiment, a multi-task learning mechanism is employed in step 107 to jointly optimize virtual twin data generation and fault diagnosis. The model training and hyperparameter setting methods are as follows: Multi-task loss: The total loss function is a weighted sum of the regression loss (MSE) and the classification loss (BCE).
[0066] in .
[0067] Optimizer: Using the Adam optimizer, the initial learning rate is... ; Learning rate scheduling: The ReduceLROnPlateau scheduler is used, and the learning rate decreases by a factor when the validation loss stops decreasing; Regularization and early stopping: An early stopping mechanism (Patience=5) is used to prevent overfitting, that is, training is terminated when the verification loss no longer decreases for 5 consecutive epochs.
[0068] The experimental results and performance analysis of the embodiments of this application are as follows: After 1400 training epochs, the model demonstrated excellent performance: Convergence: The training loss from Down to (A decrease of 13.16%), the validation loss decreased from 1.3507 to 1.3153 (a decrease of 2.62%). The validation loss exhibited a strictly monotonically decreasing trend, eventually stabilizing at... The coefficient of variation is only 0.0008%, proving that the model has superstable convergence.
[0069] Generalization ability: The training and validation losses decrease simultaneously, and both plateau in the later stages, with a generalization gap ( The magnitude is far below the overfitting alarm threshold. This indicates that the model has extremely strong generalization ability and the risk of overfitting is negligible.
[0070] Diagnostic performance: On the test set, the fault diagnosis MSE of the two-stage model of this invention is 0.4496, which is significantly better than the single-stage LSTM benchmark model (MSE=0.9826), representing a performance improvement of 54.2%. Its error compatibility index ECI=0.0012 proves that the results are highly reliable.
[0071] Control effect: The Lyapunov controller successfully optimized and adjusted key process parameters (as shown in Table 4), such as adjusting ETCHBEAMCURRENT by -1.1461 and FLOWCOOLFLOWRATE by +1.1746, bringing the system state closer to the safe and stable region.
[0072] The solution of this application has been described in detail above with reference to the accompanying drawings. In the above embodiments, the descriptions of each embodiment have different emphases; parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. Those skilled in the art should also understand that the actions and modules involved in the specification are not necessarily essential to this application. Furthermore, it is understood that the steps in the method of this application embodiment can be adjusted, combined, and deleted according to actual needs, and the modules in the device of this application embodiment can be combined, divided, and deleted according to actual needs.
[0073] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.
[0074] Alternatively, this application may be implemented as a non-transitory machine-readable storage medium (or computer-readable storage medium, or machine-readable storage medium) storing executable code (or computer program, or computer instruction code) thereon, which, when executed by a processor of an electronic device (or electronic device, server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.
[0075] Those skilled in the art will also understand that the various exemplary logic blocks, modules, circuits, and algorithm steps described in connection with the present application can be implemented as electronic hardware, computer software, or a combination of both.
[0076] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems and methods according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0077] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A multivariable time-series prediction and closed-loop control method based on the fusion of digital twin and deep learning, characterized in that, include. Construct a multi-layered data architecture based on a five-dimensional digital twin model; Data preprocessing and SMOTE resampling; Temporal feature extraction based on 1D-CNN; Virtual twin data generation based on Bi-LSTM and attention mechanism; Multimodal fault diagnosis based on RNN-LSTM hybrid network; Adaptive feedback control based on Lyapunov stability; A multi-task learning mechanism is adopted to jointly optimize virtual twin data generation and fault diagnosis.
2. The multivariable time-series prediction and closed-loop control method as described in claim 1, characterized in that, Constructing a multi-layered data architecture based on a five-dimensional digital twin model, specifically including: The multi-source heterogeneous data generated during industrial manufacturing processes are reconstructed into physical layer data, control layer data, operation layer data, analysis layer data, and virtual layer data. Physical layer data consists of the basic parameters of physical entities, control layer data consists of control system commands, operation layer data consists of operation parameters set manually or by the system, analysis layer data consists of decision data generated by simulation, AI, or big data analysis, and virtual layer data consists of twin data generated by deep learning models.
3. The multivariable time-series prediction and closed-loop control method as described in claim 1, characterized in that, Data preprocessing and SMOTE resampling specifically include: Perform zero-mean standardization on numerical features; The SMOTE algorithm is used to oversample the minority class samples in the training set.
4. The multivariable time-series prediction and closed-loop control method as described in claim 1, characterized in that, Temporal feature extraction based on 1D-CNN specifically includes: using convolutional neural networks to extract discriminative features from enhanced time series data.
5. The multivariable time-series prediction and closed-loop control method as described in claim 1, characterized in that, Virtual twin data generation based on Bi-LSTM and attention mechanisms specifically includes: A bidirectional long short-term memory network is used to process the input sequence of the operation layer. ; Forward LSTM computation: ; Backward LSTM computation: ; Concatenate the forward and backward hidden states as follows: ; Introducing an attention mechanism to calculate weights: The output of the control layer is obtained by weighted fusion and projection of attention weights: The output is a high-fidelity control layer response sequence. As virtual twin data.
6. The multivariable time-series prediction and closed-loop control method as described in claim 1, characterized in that, Multimodal fault diagnosis based on RNN-LSTM hybrid networks specifically includes: An adaptive feedback controller is designed by incorporating Lyapunov stability theory. Define Lyapunov functions To quantify the degree of deviation between the current process status and safety objectives: in This represents the anomaly probability predicted by the model, and 0.5 is the safety threshold. The system's goal is to adjust the operating parameters... make ; The parameter update strategy employs gradient descent combined with a chaos compensation term to enhance global search capabilities. in, It is the gradient of the Lyapunov function with respect to the parameters, calculated through automatic differentiation: in, It is an adaptive step size. It can be dynamically adjusted according to the gradient magnitude; It is a chaos compensation term. ; Introducing controllable randomness to escape local optima and ensure that the updated parameters satisfy physical constraints: in, It is a projection operator.
7. The multivariable time-series prediction and closed-loop control method as described in claim 1, characterized in that, A multi-task learning mechanism is adopted to jointly optimize virtual twin data generation and fault diagnosis, specifically including: The total loss function is a weighted sum of the losses from the two tasks: Training uses the Adam optimizer with an initial learning rate of The learning rate is dynamically adjusted using the ReduceLROnPlateau scheduler.