A wind valve online self-adaptive adjusting method and system based on digital twinning

Abstract

The application discloses a kind of wind valve online adaptive adjustment method and system of fusion digital twin, belong to wind valve intelligent adjustment field, the adjustment method specific steps are as follows: I, each edge equipment collects and pre-processes the multi-source data of corresponding wind valve system, based on the multi-source data after pre-processing, the virtual simulation environment corresponding to current wind valve system is constructed;II, in virtual simulation environment, pre-training wind valve control model, simultaneously identify the causal relationship between each sensing parameter and wind valve adjustment result, and dynamically optimize each control variable;The application significantly improves the model prediction accuracy and simulation credibility, reduces the dependence on a large number of new data, realizes cross-system rapid adaptive adjustment, can avoid misjudgment and mixed factor interference when optimizing strategy, enhance the explainability and credibility of model decision, while protecting data privacy, accelerating the convergence of global strategy, improve the overall intelligence level of system.

Description

Technical Field

[0001] This invention relates to the field of intelligent valve regulation, and in particular to an online adaptive regulation method and system for valves that integrates digital twins. Background Technology

[0002] Against the backdrop of rapid development in industrial automation and smart building systems, air valves, as key control components for airflow and pressure regulation, directly determine the overall energy efficiency and operational stability of HVAC systems, tunnel ventilation systems, and energy transmission networks. Traditional air valve control methods often rely on proportional-integral-differential algorithms based on empirical parameters or model predictive control strategies with fixed structures. These methods typically assume a constant system dynamic structure and cannot effectively adapt to dynamic changes such as equipment aging, environmental disturbances, parameter drift, and multi-valve coupling, leading to decreased control accuracy, increased energy consumption, and even system instability. With the rise of intelligent manufacturing and digital twin technologies, achieving bidirectional mapping and collaborative optimization between physical equipment operation and virtual simulation systems has become an important direction for breaking through the bottlenecks of traditional air valve control.

[0003] Existing online adaptive control methods and systems for dampers suffer from low model prediction accuracy and simulation reliability, increased reliance on large amounts of new data, inability to achieve rapid cross-system adaptive control, and the presence of misjudgments and confounding factors during strategy optimization, reducing the interpretability and reliability of model decisions. Furthermore, they fail to accelerate the convergence of global strategies while protecting data privacy, thus lowering the overall intelligence level of the system. To address these issues, we propose an online adaptive control method and system for dampers that integrates digital twins. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an online adaptive adjustment method and system for air valves that integrates digital twins.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] An online adaptive adjustment method for dampers integrating digital twins, the specific steps of which are as follows:

[0007] Ⅰ. Each edge device collects and preprocesses multi-source data of the corresponding air valve system, and constructs a virtual simulation environment corresponding to the current air valve system based on the preprocessed multi-source data;

[0008] II. In a virtual simulation environment, a damper control model is pre-trained, and the causal relationship between each sensor parameter and the damper adjustment result is identified, and each control variable is dynamically optimized.

[0009] Ⅲ. Using the air valve control model, the control strategy is adjusted in real time according to the dynamic characteristics of different air valves, and the strategy parameters are dynamically adjusted using real-time operating data;

[0010] IV. Periodically optimize the virtual simulation environment of each edge device in a distributed manner, and update the valve control strategy and correct uncertainties in real time during the online operation phase;

[0011] V. Collect various data throughout the entire life cycle of each air valve system, and predict the aging trend and performance degradation of the equipment in order to dynamically adjust the control strategy;

[0012] VI. During the actual operation phase, real-time control data is input into the virtual simulation environment, the virtual prediction is compared with the actual feedback, the deviation is calculated, and the control strategy is updated online.

[0013] An online adaptive control system for a damper integrating digital twins includes a data acquisition module, a processing and fusion module, a twin modeling module, a simulation calibration module, an analysis and screening module, a reinforcement strategy module, an adaptive control module, a collaborative optimization module, an optimization solution module, a feedback control module, a status monitoring module, and a strategy update module.

[0014] The data acquisition module is used to collect multimodal data from various sensors, external environmental data, and operation logs;

[0015] The processing and fusion module is used to preprocess the original multimodal data and extract feature values ​​from the multimodal data;

[0016] The twin modeling module constructs a virtual simulation environment containing a physical model and a data-driven model in the virtual environment based on the processed multimodal data.

[0017] The simulation calibration module is used to perform offline and online calibration of the virtual simulation environment and to correct the virtual simulation environment parameters in real time.

[0018] The analysis and screening module is used to identify the causal relationship between each sensor parameter and the performance index of the damper.

[0019] The enhancement strategy module is used to perform strategy pre-training in a virtual simulation environment and learn the dynamic response law of the air valve under various virtual operating conditions.

[0020] The adaptive adjustment module updates the strategy gradient and optimizes the reinforcement strategy module based on real-time data and dynamic feedback.

[0021] The collaborative optimization module is used to deploy local twin models on each edge node, and periodically upload model parameters and global optimization strategies, and then distribute them to each node.

[0022] The optimization solution module is used to search and optimize for uncertainties in the environment;

[0023] The feedback control module is used to send the generated control commands to the damper actuator and monitor the execution results in real time.

[0024] The status monitoring module is used to collect data on the entire life cycle of the air valve during the design, manufacturing, and operation and maintenance stages, and to identify the trend of changes in the health status of the air valve.

[0025] The policy update module is used to feed back the lifecycle monitoring results to the twin modeling module and automatically trigger policy retraining and model recalibration.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] This invention performs time alignment, interpolation, denoising, and anomaly correction on raw data from multiple sensors and logs, and extracts time-domain, frequency-domain, and event features. These features are then standardized and attention-weighted to form a comprehensive feature set, which is used to construct a continuous-time neural differential equation twin combining a physical model and a data-driven model. The ODE parameters are trained offline using historical operating data and iteratively optimized. In a virtual environment, meta-reinforcement learning is used for offline pre-training of the control strategy. Support and query sets are constructed using multi-task samples to achieve rapid policy adaptation and generalization. Subsequently, the causal relationship between key sensing quantities and regulation results is identified, providing an interpretable basis for policy design. Each edge device independently updates its policy model based on its local twin, and parameter aggregation and global optimization are achieved through a federated learning mechanism. During the online operation phase, a variable quantum algorithm is used to perform a global search in a high-dimensional control space to obtain the optimal regulation scheme. It significantly improves model prediction accuracy and simulation credibility, reduces reliance on large amounts of new data, enables rapid cross-system adaptive adjustment, avoids misjudgment and interference from confounding factors during policy optimization, enhances the interpretability and credibility of model decisions, protects data privacy, accelerates the convergence of global policies, and improves the overall intelligence level of the system. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0029] Figure 1 This is a flowchart of an online adaptive adjustment method for a damper that integrates digital twins, as proposed in this invention.

[0030] Figure 2 This is a system block diagram of an online adaptive adjustment system for a damper that integrates digital twins, as proposed in this invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0032] Reference Figure 1 This embodiment discloses an online adaptive adjustment method for a damper that integrates digital twins. The specific steps of this adjustment method are as follows:

[0033] Ⅰ. Each edge device collects and preprocesses multi-source data of the corresponding air valve system, and constructs a virtual simulation environment corresponding to the current air valve system based on the preprocessed multi-source data;

[0034] II. In a virtual simulation environment, a damper control model is pre-trained, and the causal relationship between each sensor parameter and the damper adjustment result is identified, and each control variable is dynamically optimized.

[0035] Ⅲ. Using the air valve control model, the control strategy is adjusted in real time according to the dynamic characteristics of different air valves, and the strategy parameters are dynamically adjusted using real-time operating data;

[0036] IV. Periodically optimize the virtual simulation environment of each edge device in a distributed manner, and update the valve control strategy and correct uncertainties in real time during the online operation phase;

[0037] V. Collect various data throughout the entire life cycle of each air valve system, and predict the aging trend and performance degradation of the equipment in order to dynamically adjust the control strategy;

[0038] VI. During the actual operation phase, real-time control data is input into the virtual simulation environment, the virtual prediction is compared with the actual feedback, the deviation is calculated, and the control strategy is updated online.

[0039] As a further aspect of the present invention, the specific steps for each edge device to collect and preprocess the multi-source data of the corresponding air valve system in step I are as follows:

[0040] S1.1: On-site, raw data from different sources are collected into temporary storage according to the original timestamps. Then, a unified reference time axis is selected, and each piece of raw data is time-standardized. Nearest neighbor time pairing is used to align the data from different sources to a fixed set of time points on the reference time axis to establish different time series.

[0041] S1.2: Check the length and position of the missing segments in each time series after alignment. If the length of the missing segment is less than the preset threshold, it is marked as a short missing interval; otherwise, it is marked as a long missing interval. Linear interpolation is used to fill the short missing interval. The long missing interval is marked as unrecoverable and filled by the data of the nearest neighboring similar device. At the same time, the interpolation flag is retained after interpolation.

[0042] S1.3: Remove low-frequency drift and high-frequency noise from each processed time series to smooth them. Construct robust scaling estimates for each smoothed time series using the MAD method, identify outliers in each time series, and replace the detected outliers with neighbor interpolation. Then, divide each time series into multiple segments using a sliding window with length W and step size S.

[0043] S1.4: Calculate the time domain features of the time series in each sliding window, and record the start and end times of the sliding window and the source sensor ID corresponding to each time domain feature. Then, perform discrete Fourier transform on the time series in each sliding window to obtain the corresponding spectral components. Calculate the overall power spectral density, the energy proportion in each frequency band, the main frequency component and the frequency centroid of each frequency domain feature. Then, save the frequency domain features and time domain features in parallel to obtain the corresponding sensor features.

[0044] S1.5: Extract discrete events from historical operation logs and event records, perform time alignment on the text entries of each discrete event, perform one-hot encoding on text-type or category-type events in discrete events, extract frequency statistics on periodic or count-type events in discrete events to obtain event features, then aggregate each event feature by window and concatenate it with sensor window features.

[0045] S1.6: Based on the preset division criteria, the concatenated features are divided into numerical features and binary features. Then, the Z-score standardization method is used to unify the numerical features to the [0, 1] interval, and the binary features are mapped to 0 or 1 to unify the dimensions of various features. Subsequently, the same mean or variance is used to normalize various features, aligning various features to the same index at the sample level, and using learnable attention weights to weightedly combine various features to generate the corresponding comprehensive features.

[0046] As a further aspect of the present invention, the specific calculation formula for nearest neighbor time pairing in S1.1 is as follows:

[0047] ;

[0048] In the formula, Representative and the The sensor records the index of the closest reference time point; Represents Class 1, No. The original timestamps recorded by the sensors; Represents the reference timeline. A timestamp at a specific point in time;

[0049] The specific calculation formula for the linear interpolation described in S1.2 is as follows:

[0050] ;

[0051] In the formula, The interpolated value is at the reference time point. The estimated value; as well as These represent the most recent observed values ​​at both ends of the missing interval, and correspond to the timestamps respectively. as well as ; Representative reference time point Timestamp.

[0052] As a further aspect of the present invention, the specific steps for constructing the virtual simulation environment corresponding to the current air valve system in step I are as follows:

[0053] S2.1: Read the observations used for physical modeling from each comprehensive feature, and determine the set of physical quantities that need to be explicitly simulated in the physical model according to the preset requirements. Based on the determined set of physical quantities, establish corresponding physical sub-models respectively, and process each physical relationship in the physical sub-model into a continuous time form, and distinguish between available physical parameters and unknown adjustable parameters.

[0054] S2.3: Set the data-driven continuous-time state vector and determine the physical quantities to be corrected by the data. Then, represent the continuous state derivative in the form of the physical model derivative neural network correction term to establish the ODE model. Extract the corresponding valve real state and control input sequence from historical operation data. Then, input the control input sequence into the ODE model and output the valve prediction state under the initial model parameters. Calculate the loss value between the valve prediction state and the corresponding valve real state through the preset offline training loss function.

[0055] S2.4: Based on the calculated loss value, adjust the initial parameters of the ODE model using a differentiable numerical integrator, and save the normalization parameters and time step strategy during training. Repeat the training and adjustment of the ODE model, and compare the current loss value of the ODE model with the loss value of the previous round. If the change in the model loss value converges to the preset range, stop training, and select the model parameters with the smallest loss value as the optimal parameters of the ODE model.

[0056] S2.5: Couple the physical sub-model and the ODE model into a simulateable virtual running model, load the offline normalized parameters and the trained model parameters into the virtual running model, run the simulation for each set of historical working conditions, and output the simulation results using the same sampling time point as the historical observation, record the simulation-observation residual sequence and classify and statistically analyze the performance indicators according to the working conditions.

[0057] S2.6: Decompose the observation residuals by frequency band and classify them by operating conditions to determine whether they are structural errors of the model, parameter biases, or historical data problems. If they are structural errors, mark them as subsequent model topology improvement items; if they are parameter biases, prepare for parameter re-estimation; if they are data quality problems, mark them and return to the data preprocessing stage.

[0058] S2.7: When the virtual running model reaches a new time window, an instant loss is constructed using the data of that time window, and the model parameters are updated online with a preset small step size. After the model parameters are updated, the virtual running model is continuously predicted for its state. The instant loss of the current time window is calculated in each subsequent time window. When the instant loss exceeds the preset tolerance, the model parameters are rolled back to the model parameters of the previous time window, the time window size is shortened, and the model parameters are updated again.

[0059] As a further aspect of the present invention, the specific steps of the pre-trained damper control model in step II are as follows:

[0060] S3.1: In the virtual twin environment, based on historical working conditions and design possibilities, multiple representative tasks are divided into groups, and the generated multiple representative tasks are integrated into a task set. Then, sub-tasks are sampled from the task set, the sampling probability distribution of each sampled sub-task is recorded, and the parameter range, sampling frequency and importance weight of each task are saved.

[0061] S3.2: Use the meta-parameter initialization strategy of the current sampling subtask in the virtual twin, and perform virtual simulation within a preset short time interval. At the same time, collect the valve control trajectory within the interval to form a support dataset, calculate the instantaneous reward of the support data, and use it as the task objective. Then, use negative cumulative reward to calculate the inner loop loss, and update the meta-parameters in multiple time steps based on the preset learning rate to generate task-specific parameters. At the same time, save the intermediate parameters after each step of adaptation and the corresponding support set performance.

[0062] S3.3: Under the updated or adapted task specialization parameters of the support set, virtual simulation is performed under the same sampling subtask within a preset long time interval. At the same time, the valve control trajectory within the interval is collected, a query set is established, and the advantage of each time step in each valve control trajectory in the query set is calculated by the generalized advantage estimation method. The advantage is then packaged with the corresponding policy-action pair as the outer loop sample weight.

[0063] S3.4: Based on the task specialization parameters, the outer loop loss of each sampled subtask is obtained through negative weighted reward. Then, based on the outer loop sample weights, the outer loop losses of each sampled subtask are weighted according to task importance and summed to obtain the overall meta-objective. Then, based on the overall meta-objective, the loss gradient of the current task specialization parameters is calculated, and the task specialization parameters are updated once or multiple times at the meta-level through the Adam optimizer.

[0064] S3.5: After the inner-outer loop parameter update is completed in each round, each unused representative task is used as a validation task set. The adaptation speed and final query performance under the updated task specialization parameters are evaluated on an independent validation task set. If the overall meta-objective converges to the preset range, the query performance on the validation task set reaches the preset expected value, or the preset number of training rounds is reached, the parameter update is stopped, and the original meta-parameters are replaced with the current task specialization parameters.

[0065] As a further aspect of the present invention, the representative task described in S3.1 is specifically a combination of valve dynamics or environmental conditions that includes boundary condition distribution, typical disturbance sequence, initial state set, and corresponding controllable target weights.

[0066] By using a damper control model, the control strategy is adjusted in real time based on the dynamic characteristics of different dampers, and the strategy parameters are dynamically adjusted using real-time operating data.

[0067] The virtual simulation environment of each edge device is periodically optimized in a distributed manner, and the valve control strategy is updated and uncertainties are corrected in real time during the online operation phase.

[0068] Collect various data throughout the entire lifecycle of each air valve system, and predict equipment aging trends and performance degradation status in order to dynamically adjust control strategies.

[0069] During actual operation, real-time control data is input into the virtual simulation environment, the virtual predictions are compared with the actual feedback, the deviation is calculated, and the control strategy is updated online.

[0070] Reference Figure 2 This embodiment discloses an online adaptive adjustment system for a damper that integrates digital twins, including a data acquisition module, a processing and fusion module, a twin modeling module, a simulation calibration module, an analysis and screening module, a reinforcement strategy module, an adaptive adjustment module, a collaborative optimization module, an optimization solution module, a feedback control module, a status monitoring module, and a strategy update module.

[0071] The data acquisition module is used to collect multimodal data from various sensors, external environmental data, and operation logs;

[0072] The processing and fusion module is used to preprocess the original multimodal data and extract feature values ​​from the multimodal data;

[0073] The twin modeling module constructs a virtual simulation environment containing a physical model and a data-driven model in the virtual environment based on the processed multimodal data.

[0074] The simulation calibration module is used to perform offline and online calibration of the virtual simulation environment and to correct the virtual simulation environment parameters in real time.

[0075] The analysis and screening module is used to identify the causal relationship between each sensor parameter and the performance index of the damper.

[0076] The enhancement strategy module is used to perform strategy pre-training in a virtual simulation environment and learn the dynamic response law of the air valve under various virtual operating conditions.

[0077] The adaptive adjustment module updates the strategy gradient and optimizes the reinforcement strategy module based on real-time data and dynamic feedback.

[0078] The collaborative optimization module is used to deploy local twin models on each edge node, and periodically upload model parameters and global optimization strategies, and then distribute them to each node.

[0079] The optimization solution module is used to search and optimize for uncertainties in the environment;

[0080] The feedback control module is used to send the generated control commands to the damper actuator and monitor the execution results in real time.

[0081] The status monitoring module is used to collect data on the entire life cycle of the air valve during the design, manufacturing, and operation and maintenance stages, and to identify the trend of changes in the health status of the air valve.

[0082] The policy update module is used to feed back the lifecycle monitoring results to the twin modeling module and automatically trigger policy retraining and model recalibration.

[0083] As a further aspect of the present invention, the specific steps of the analysis and screening module in identifying the causal relationship between each sensing parameter and the performance index of the damper are as follows:

[0084] S4.1: Extract the sensor measurements and target variables that meet the requirements from the preprocessed multi-source data, establish a set of observed variables, and record the candidates for confounding variables as a potential adjustment set. Based on domain knowledge, specify the directional relationships that meet the domain knowledge, and mark the edges with uncertain directions as "to be checked". Then, use each variable in the set of observed variables as nodes and the prior relationships as directed or undirected edges to construct a preliminary prior causal graph, and record the data type, sampling frequency and missing flag of each variable.

[0085] S4.2: Fix the directed edges in the prior cause-effect graph, and determine the direction of each undirected edge in the prior cause-effect graph using the NOTEARS method. Based on the prior cause-effect graph, perform conditional independence tests on the data, delete redundant undirected or directed edges according to the test results, and record the confidence level of each edge.

[0086] S4.3: Using the parent node in the prior cause-effect graph as the explanatory variable, and based on the directed edges retained in the prior cause-effect graph and each explained variable, establish a structural equation model, evaluate the fitting quality of each equation and check the independence of the residuals. If the independence of the residuals is lower than the preset threshold, then readjust the prior cause-effect graph.

[0087] S4.4: Check whether there is an adjustment set that satisfies the back-door condition in the prior causal graph. If it does, calculate the estimated value of each candidate causal edge in the adjustment set through regression adjustment, inverse probability weighting, or double robust estimation. If it does not satisfy the condition, mark it as a candidate causal edge that needs to be verified by experiment / human intervention.

[0088] S4.5: Generate counterfactual trajectories in the structural equation model and compare them with observed trajectories to check whether the structural equation model gives consistent and interpretable evolution under different conditions. At the same time, simulate potential confounding and observe changes in causal effect estimates. If the change in causal effect estimates is higher than a preset threshold, the corresponding causal edge is marked as low confidence and intervention verification is recommended. Otherwise, the causal edge is confirmed as a causal relationship that can be used for control strategy design, and the causal effect value and uncertainty measure are output.

[0089] As a further aspect of the present invention, the specific steps of the global optimization strategy of the collaborative optimization module are as follows:

[0090] P1.1: Each edge device maintains a policy model parameter vector in the local virtual simulation environment. At the same time, within a local training cycle, the device uses its own local dataset to perform several mini-batch gradient descent iterations on the local virtual simulation environment through optimizers such as SGD and Adam, minimizes the local loss function, defines the difference vector of parameters before and after local training, and uses it as the object to be uploaded.

[0091] P1.2: After each local training cycle ends, the device decides whether to immediately prepare to upload or cache and wait for the next federated round based on its own resources and privacy policy. After that, the edge device performs optional sparsification / quantization processing on the difference vector.

[0092] P1.3: During the upload phase, the edge device fragments / masks the local update vector to be uploaded and exchanges mask keys with other devices. It then sends the mask keys to the central node through a secure channel. After receiving the mask updates from a batch of devices, the central node first verifies the signature / integrity. If some devices lose packets or time out, a fallback strategy is adopted and the communication status is recorded for subsequent retry or robust aggregation.

[0093] P1.4: The received local update set is stored in the central temporary queue for aggregation calculation. At the same time, the weight information corresponding to each update is recorded. After the central node receives a batch of unmasked weighted updates, it first checks for abnormal updates. If an abnormality is detected, the update is downweighted or removed.

[0094] P1.5: Use weighted average or aggregation rules with regularization / momentum to generate global update items, then apply the global update items to the central global model synchronously, and generate the model snapshot for the next round of distribution. At the same time, record the aggregation statistics and use them for subsequent adaptive weight adjustment strategies. After completing the global update and generating a new global model snapshot, send the model parameters or increments to the edge device set in a compressed / incremental differential manner.

[0095] P1.6: Each edge device that receives the global model will use the global model as a new starting point and perform short-term local personalized fine-tuning. The central node will adjust the aggregation weight strategy for the next round based on the local verification performance feedback of the device.

[0096] As a further aspect of the present invention, the specific steps of the optimization solution module in searching for uncertainties in the optimization environment are as follows:

[0097] P2.1: Obtain the current control variables from the online status of the virtual simulation environment, and discretize the continuous control variables. Use binary encoding to map the discrete control variables into quantum representations in quantum space. Construct the corresponding cost Hamiltonian according to the preset objective function. Select the corresponding parameterized circuit structure according to the valve control requirements and available quantum resources. Preset the number of measurement runs, measurement variance threshold and allowable quantum execution time budget.

[0098] P2.2: Initialize the parameters in the control strategy based on prior knowledge and small random noise. Based on the initialized parameter information, use the quantum processing unit to run the parameterized circuit in the preset measurement rounds and measure the corresponding observable operator terms. If the cost Hamiltonian can be decomposed into the sum of multiple sets of intrinsic measurable terms, then measure and count the expectation of each term according to the grouping scheme.

[0099] P2.3: Replace the precise expected value with the sample mean and variance estimate, calculate the cost estimate obtained from the measurement, and then use measurement retry, group optimization or classical post-processing to reduce measurement noise. Record the measurement uncertainty and sample variance. If the variance exceeds the tolerance, automatically increase the preset measurement rounds or trigger measurement retry.

[0100] P2.4: For each parameter component, run the circuit forward and backward with the offset step size and estimate the partial derivatives using the difference between the measurement mean values. Then, the classical optimizer updates the parameter vector based on the obtained gradient. After reaching convergence or the minimum cost allowed by the budget, the optimal parameters are used to measure the variational circuit multiple times to obtain the final measurement distribution. Then, each measurement bit string is mapped back to the control candidate solution set, and the candidate solutions are sorted from high to low according to their observation cost.

[0101] S2.5: Select multiple sets of low-cost control candidate solutions, evaluate the performance of each control candidate solution in the virtual simulation stage through classical fast simulation, and select a single or small batch of adjustment schemes for application in actual systems.

[0102] The feedback control module sends the generated control commands to the damper actuator and monitors the execution results in real time; the status monitoring module collects the full life cycle data of the damper during the design, manufacturing, and operation and maintenance stages, and identifies the trend of the damper's health status changes; the strategy update module feeds back the life cycle monitoring results to the twin modeling module and automatically triggers strategy retraining and model recalibration.

Claims

1. A method for online adaptive adjustment of a damper integrating digital twins, characterized in that, The specific steps of this adjustment method are as follows: Ⅰ. Each edge device collects and preprocesses multi-source data of the corresponding air valve system, and constructs a virtual simulation environment corresponding to the current air valve system based on the preprocessed multi-source data; The specific steps for each edge device to collect and preprocess the multi-source data of the corresponding air valve system are as follows: S1.1: On-site, raw data from different sources are collected into temporary storage according to the original timestamps. Then, a unified reference time axis is selected, and each piece of raw data is time-standardized. Nearest neighbor time pairing is used to align the data from different sources to a fixed set of time points on the reference time axis to establish different time series. S1.2: Check the length and position of the missing segments in each time series after alignment. If the length of the missing segment is less than the preset threshold, it is marked as a short missing interval; otherwise, it is marked as a long missing interval. Linear interpolation is used to fill the short missing interval. The long missing interval is marked as unrecoverable and filled with data from the nearest neighbor of the same type of device. At the same time, the interpolation flag is retained after interpolation. S1.3: Remove low-frequency drift and high-frequency noise from each processed time series to smooth each time series. Construct robust scaling estimates for each smoothed time series using the MAD method, identify outliers in each time series, and replace the detected outliers with neighbor interpolation. Then, divide each time series into multiple segments using a sliding window with length W and step size S. S1.4: Calculate the time domain features of the time series in each sliding window, and record the start and end times of the sliding window and the source sensor ID corresponding to each time domain feature. Then, perform discrete Fourier transform on the time series in each sliding window to obtain the corresponding spectral components. Calculate the overall power spectral density, the energy proportion in each frequency band, the main frequency component and the frequency centroid of each frequency domain feature. Then, save the frequency domain features and time domain features in parallel to obtain the corresponding sensor features. S1.5: Extract discrete events from historical operation logs and event records, perform time alignment on the text entries of each discrete event, perform one-hot encoding on text-type or category-type events in discrete events, extract frequency statistics on periodic or count-type events in discrete events to obtain event features, then aggregate each event feature by window and concatenate it with sensor window features. S1.6: Based on the preset division criteria, the concatenated features are divided into numerical features and binary features. Then, the Z-score standardization method is used to unify the numerical features to the [0, 1] interval, and the binary features are mapped to 0 or 1 to unify the dimensions of various features. Subsequently, the same mean or variance is used to normalize various features, aligning various features to the same index at the sample level, and using learnable attention weights to weightedly combine various features to generate the corresponding comprehensive features. The specific steps for constructing the virtual simulation environment corresponding to the current air valve system are as follows: S2.1: Read the observations used for physical modeling from each comprehensive feature, and determine the set of physical quantities that need to be explicitly simulated in the physical model according to the preset requirements. Based on the determined set of physical quantities, establish corresponding physical sub-models respectively, and process each physical relationship in the physical sub-model into a continuous time form, and distinguish between available physical parameters and unknown adjustable parameters. S2.2: Set up a data-driven continuous-time state vector and determine the physical quantities to be corrected by data. Then, express the continuous state derivative in the form of physical model derivative and neural network correction term to establish an ODE model. Extract the corresponding valve real state and control input sequence from historical operation data. Then, input the control input sequence into the ODE model and output the valve prediction state under the initial model parameters. Calculate the loss value between the valve prediction state and the corresponding valve real state through a preset offline training loss function. S2.3: Based on the calculated loss value, adjust the initial parameters of the ODE model using a differentiable numerical integrator, and save the normalization parameters and time step strategy during training. Repeat the training and adjustment of the ODE model, and compare the current loss value of the ODE model with the loss value of the previous round. If the change in the model loss value converges to the preset range, stop training, and select the model parameters with the smallest loss value as the optimal parameters of the ODE model. S2.4: Couple the physical sub-model and the ODE model into a simulateable virtual running model, load the offline normalized parameters and the trained model parameters into the virtual running model, run the simulation for each set of historical working conditions, and output the simulation results using the same sampling time points as the historical observations. Record the simulation-observation residual sequence and classify and statistically analyze the performance indicators according to the working conditions. S2.5: Decompose the observation residuals by frequency band and classify them by operating conditions to determine whether they are structural errors of the model, parameter deviations or historical data problems. If they are structural errors, mark them as subsequent model topology improvement items. If the problem is parameter deviation, then prepare for parameter re-estimation; if the problem is data quality, then label it and return to the data preprocessing stage. S2.6: When the virtual running model reaches a new time window, an instant loss is constructed using the data of that time window, and the model parameters are updated online with a preset small step size. After the model parameters are updated, the virtual running model is continuously predicted for its state. In each subsequent time window, the instant loss of the current time window is calculated. When the instant loss exceeds the preset tolerance, the model parameters are rolled back to the model parameters of the previous time window, the time window size is shortened, and the model parameters are updated again. II. In a virtual simulation environment, a damper control model is pre-trained, and the causal relationship between each sensor parameter and the damper adjustment result is identified, and each control variable is dynamically optimized. Ⅲ. Using the air valve control model, the control strategy is adjusted in real time according to the dynamic characteristics of different air valves, and the strategy parameters are dynamically adjusted using real-time operating data; IV. Periodically optimize the virtual simulation environment of each edge device in a distributed manner, and update the valve control strategy and correct uncertainties in real time during the online operation phase; V. Collect various data throughout the entire life cycle of each air valve system, and predict the aging trend and performance degradation of the equipment in order to dynamically adjust the control strategy; VI. During the actual operation phase, real-time control data is input into the virtual simulation environment, the virtual prediction is compared with the actual feedback, the deviation is calculated, and the control strategy is updated online.

2. The method for online adaptive adjustment of a damper integrating digital twins according to claim 1, characterized in that, The specific steps of the pre-trained damper control model described in step II are as follows: S3.1: In the virtual twin environment, based on historical working conditions and design possibilities, multiple representative tasks are divided into groups, and the generated multiple representative tasks are integrated into a task set. Then, sub-tasks are sampled from the task set, the sampling probability distribution of each sampled sub-task is recorded, and the parameter range, sampling frequency and importance weight of each task are saved. S3.2: Use the meta-parameter initialization strategy of the current sampling subtask in the virtual twin, and perform virtual simulation within a preset short time interval. At the same time, collect the valve control trajectory within the interval to form a support dataset, calculate the instantaneous reward of the support data, and use it as the task objective. Then, use negative cumulative reward to calculate the inner loop loss, and update the meta-parameters in multiple time steps based on the preset learning rate to generate task-specific parameters. At the same time, save the intermediate parameters after each step of adaptation and the corresponding support set performance. S3.3: Under the updated or adapted task specialization parameters of the support set, virtual simulation is performed under the same sampling subtask within a preset long time interval. At the same time, the valve control trajectory within the interval is collected, a query set is established, and the advantage of each time step in each valve control trajectory in the query set is calculated by the generalized advantage estimation method. The advantage is then packaged with the corresponding policy-action pair as the outer loop sample weight. S3.4: Based on the task specialization parameters, the outer loop loss of each sampled subtask is obtained through negative weighted reward. Then, based on the outer loop sample weights, the outer loop losses of each sampled subtask are weighted according to task importance and summed to obtain the overall meta-objective. Then, based on the overall meta-objective, the loss gradient of the current task specialization parameters is calculated, and the task specialization parameters are updated once or multiple times at the meta-level through the Adam optimizer. S3.5: After the inner-outer loop parameter update is completed in each round, each unused representative task is used as a validation task set. The adaptation speed and final query performance under the updated task specialization parameters are evaluated on an independent validation task set. If the overall meta-objective converges to the preset range, the query performance on the validation task set reaches the preset expected value, or the preset number of training rounds is reached, the parameter update is stopped, and the original meta-parameters are replaced with the current task specialization parameters.

3. A digital twin-integrated online adaptive control system for a damper, used to implement the digital twin-integrated online adaptive control method for a damper as described in claim 1 or 2, characterized in that, It includes a data acquisition module, a processing and fusion module, a twin modeling module, a simulation calibration module, an analysis and screening module, a reinforcement strategy module, an adaptive adjustment module, a collaborative optimization module, an optimization solution module, a feedback control module, a state monitoring module, and a strategy update module; The data acquisition module is used to collect multimodal data from various sensors, external environmental data, and operation logs; The processing and fusion module is used to preprocess the original multimodal data and extract feature values ​​from the multimodal data; The twin modeling module constructs a virtual simulation environment containing a physical model and a data-driven model in the virtual environment based on the processed multimodal data. The simulation calibration module is used to perform offline and online calibration of the virtual simulation environment and to correct the virtual simulation environment parameters in real time. The analysis and screening module is used to identify the causal relationship between each sensor parameter and the performance index of the damper. The enhancement strategy module is used to perform strategy pre-training in a virtual simulation environment and learn the dynamic response law of the air valve under various virtual operating conditions. The adaptive adjustment module updates the strategy gradient and optimizes the reinforcement strategy module based on real-time data and dynamic feedback. The collaborative optimization module is used to deploy local twin models on each edge node, while periodically uploading model parameters, dynamically adjusting the global optimization strategy, and then distributing it to each node. The optimization solution module is used to search for and optimize uncertainties in the environment. The feedback control module is used to send the generated control commands to the damper actuator and monitor the execution results in real time. The status monitoring module is used to collect data on the entire life cycle of the air valve during the design, manufacturing, and operation and maintenance stages, and to identify the trend of changes in the health status of the air valve. The policy update module is used to feed back the lifecycle monitoring results to the twin modeling module and automatically trigger policy retraining and model recalibration.

4. The online adaptive adjustment system for a damper integrating digital twins according to claim 3, characterized in that, The specific steps of the analysis and screening module in identifying the causal relationship between each sensor parameter and the performance index of the damper are as follows: S4.1: Extract the sensor measurements and target variables that meet the requirements from the preprocessed multi-source data, establish a set of observed variables, and record the candidates for confounding variables as a potential adjustment set. Based on domain knowledge, specify the directional relationships that meet the domain knowledge, and mark the edges with uncertain directions as "to be checked". Then, use each variable in the set of observed variables as nodes and the prior relationships as directed or undirected edges to construct a preliminary prior causal graph, and record the data type, sampling frequency and missing flag of each variable. S4.2: Fix the directed edges in the prior cause-effect graph, and determine the direction of each undirected edge in the prior cause-effect graph using the NOTEARS method. Based on the prior cause-effect graph, perform conditional independence tests on the data, delete redundant undirected or directed edges according to the test results, and record the confidence level of each edge. S4.3: Using the parent node in the prior cause-effect graph as the explanatory variable, and based on the directed edges retained in the prior cause-effect graph and each explained variable, establish a structural equation model, evaluate the fitting quality of each equation and check the independence of the residuals. If the independence of the residuals is lower than the preset threshold, then readjust the prior cause-effect graph. S4.4: Check whether there is an adjustment set that satisfies the back-door condition in the prior causal graph. If it does, calculate the estimated value of each candidate causal edge in the adjustment set through regression adjustment, inverse probability weighting, or double robust estimation. If it does not satisfy the condition, mark it as a candidate causal edge that needs to be verified by experiment / human intervention. S4.5: Generate counterfactual trajectories in the structural equation model and compare them with observed trajectories to check whether the structural equation model gives consistent and interpretable evolution under different conditions. At the same time, simulate potential confounding and observe changes in causal effect estimates. If the change in causal effect estimates is higher than a preset threshold, mark the corresponding causal edge as low confidence and suggest intervention verification. Conversely, if the causal edge is not found, it is recognized as a causal relationship that can be used for control strategy design, and the causal effect value and uncertainty measure are output.

Images