A cloud-edge collaborative and industrial adaptive digital twin construction method

Abstract

The application relates to the technical field of industrial automation, in particular to a cloud-edge collaboration and industrial adaptive digital twin construction method, which comprises the following steps: constructing an event-damping mapping model based on an attention mechanism; simulating the timeliness of event influence by adopting an exponential decay model; performing tool variable two-stage least square method parameter calibration; encrypting and securely aggregating local gradients; statistically inferring and confidence evaluating physical gain parameters, and introducing a differential privacy and security verification mechanism; constructing a Markov decision process, adopting a deep deterministic policy gradient algorithm to realize optimal control of a continuous action space; performing quantization, layer fusion and memory layout optimization; introducing hierarchical grouping SHAP and Monte Carlo sampling strategies to guarantee real-time analysis efficiency; inputting disturbed signals and real-time state into a controller, and synchronously executing servo power decision and explainable attribution analysis. The application solves the problems of insufficient safety and generalization ability of traditional parameter estimation methods.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, and in particular to a method for constructing an industrial adaptive digital twin based on cloud-edge collaboration. Background Technology

[0002] In smart manufacturing and industrial IoT systems, traditional sensing and monitoring methods are typically limited to collecting physical environmental parameters such as temperature, pressure, and vibration, making it difficult to perceive and respond to the systemic impacts of "soft environment" factors such as macroeconomic policies, supply chain changes, and market sentiment. Unstructured information such as global trade news, industry policy documents, and industry research reports contain a large number of forward-looking risk signals, and existing technological systems generally lack the ability to convert these signals into cross-modal mappings that can drive and control physical parameters.

[0003] Meanwhile, in industrial process modeling and optimization, accurately calibrating the physical gain parameters from raw material input to finished product output is crucial. Traditional parameter estimation methods face two major practical dilemmas: first, industrial data has strict local privacy protection requirements due to the involvement of trade secrets, making cross-plant data aggregation difficult; second, the data scale of a single production node is limited and easily affected by local noise, resulting in insufficient statistical significance and weak generalization ability of the calibration results.

[0004] Furthermore, traditional deep learning-based industrial controllers often face "black box" challenges due to the lack of transparency in their decision-making processes. Engineers find it difficult to understand their internal logic, making debugging and diagnosis challenging. This limits their application in industries with high compliance requirements, such as pharmaceuticals and aviation, and also affects the establishment of trust in human-machine collaboration and the overall acceptability of the system. Summary of the Invention

[0005] To address the shortcomings of existing methods, this invention integrates cloud-based semantic reasoning and nonlinear damping mapping, edge federated parameter collaborative calibration and secure aggregation, interpretable intelligent control and real-time decision attribution, hardware-in-the-loop adversarial testing, and physical system evolution verification modules. Through semantic quantization and physical damping mapping of unstructured text, combined with privacy-preserving federated parameter calibration, interpretable AI decision-making mechanisms, and a hardware-in-the-loop verification system, it achieves end-to-end adaptive control from semantic perception to physical execution. This effectively addresses the technical challenges of traditional industrial systems, such as lack of semantic perception, data privacy risks, insufficient interpretability of AI decisions, and inadequate system robustness verification. It has broad application value in fields such as intelligent manufacturing, supply chain collaboration, and the industrial internet.

[0006] The technical solution adopted in this invention is: a method for constructing an industrial adaptive digital twin based on cloud-edge collaboration, comprising the following steps: Step 1: Extract key semantic events and construct an event-damping mapping model based on an attention mechanism; set basic environmental damping, use an exponential decay model to simulate the timeliness of event impact, and synthesize the final damping signal; As a preferred embodiment of the present invention, step one specifically includes: Step 11: Real-time acquisition of unstructured text data; Step 12: Use a pre-trained NLP model to extract key events; Step 13: Map the semantic vectors to damping increments using an attention mechanism network: And train the mapping network; Step 14: Set the reference environmental damping , This represents the industry average capacity utilization rate. This represents the theoretical maximum production capacity. This is an industry adjustment coefficient; Step 15: Use the exponential decay model ; The attenuation coefficient; Indicates an indicator function; The timestamp of the event; Step 16: Introduce Gaussian noise into the damping signal; Step 17: Send the synthesized damping signal to the edge computing gateway through an encrypted channel.

[0007] In a preferred embodiment of the present invention, the upper and lower bound constraints of the synthesized damping signal are defined by the following formula: , This is the synthesized damping signal.

[0008] Step 2: Construct a physical model of the industrial process and perform two-stage least squares parameter calibration of instrumental variables; encrypt and securely aggregate local gradients; perform statistical inference and confidence assessment of physical gain parameters, and introduce differential privacy and security verification mechanisms. In a preferred embodiment of the present invention, step two specifically includes: Step 21: Construct a physical model of the industrial process ; For output variables; For input variables; For environmental damping variables; This represents the physical gain parameter to be calibrated; This is the damping effect coefficient; Represents the random error term; Step 22, Build For instrumental variables and linear regression model ; For the intercept term; Instrumental variable right The influence coefficient; For environmental damping right The influence coefficient; The parameters are estimated using the least squares method based on the first-stage regression residuals. , to obtain the predicted value ; Step 23, Construction right and regression model ; For the second-stage regression residuals; Step 24, let the first... Local datasets of edge nodes; Step 25: Perform parameter estimation in both stages. , ; Step 26: The edge nodes encrypt the local parameter gradients using the Paillier public key encryption algorithm.

[0009] In a preferred embodiment of the present invention, step 26 specifically includes: Randomly select two prime numbers and ,satisfy , The greatest common divisor; calculate pq modulus and Carmichael function value ; Randomly select an integer ,satisfy ; Calculate the modular inverse ; Set the public key to The private key is ; For plaintext messages Random selection : ; Decryption algorithm, for ciphertext ,have: ; No. The parameter gradient of each edge node is calculated based on the local loss function. ; gradient calculation ; The gradient is encrypted using Paillier's public key. ; The cloud-based federated server receives encryption gradients from all edge nodes. ; Decrypt the aggregation result using the private key. ; Global parameters are updated using the FedAvg algorithm. ; Indicates the learning rate. This is the global physical gain parameter.

[0010] Step 3: Construct a Markov decision process and use a deep deterministic policy gradient algorithm to achieve optimal control of the continuous action space; perform quantization, layer fusion and memory layout optimization; introduce hierarchical grouping SHAP and Monte Carlo sampling strategies to ensure real-time analysis efficiency; As a preferred embodiment of the present invention, constructing a Markov decision process and employing a deep deterministic policy gradient algorithm to achieve optimal control in the continuous action space specifically includes: Step 31: Construct a Markov decision process ; Step 32: Define the system's state vector , , This refers to the level of production capacity and inventory. To maintain the level of technological reserves, Environmental risk level, This represents the environmental damping value. For the first A historical state characteristic; Step 33: Set the controller's output motion space to the servo motor power coefficient; Step 34: Construct a composite reward function ;, The weighting coefficient is non-negative; the generation efficiency reward is... , Physical gain calibrated for federated learning The instantaneous throughput of raw materials; energy consumption penalty is , Energy consumption cost coefficient; safety constraint reward is , and These represent the lower limit of safe production capacity and inventory, and the upper limit of safe risk level; the smoothing reward is... ; Step 35: Use the DDPG algorithm to control the motion space.

[0011] As a preferred embodiment of the present invention, quantization, layer fusion, and memory layout optimization include: INT8 quantization is used to merge the BatchNorm layer into the convolutional layer and convert the NCHW memory layout to the NHWC layout.

[0012] As a preferred embodiment of the present invention, the hierarchical grouping SHAP includes: Build includes Model prediction of features ,feature SHAP value ; Constructing a linear model ; A simplified input indicating whether a feature exists; Construct weights ; By minimizing the weighted squared loss Solve for the SHAP value.

[0013] Step 4: Generate cloud-based environmental damping signal stream; input the disturbed signal and real-time status into the controller, and simultaneously perform servo power decision-making and interpretable attribution analysis; In a preferred embodiment of the present invention, step four specifically includes: Step 41: Obtain environmental damping signal ; As the reference damping, For semantic event increments, It is Gaussian white noise; Step 42, in Superimposed composite interference inspired by GAN networks : ; Step 43: The controller executes servo power decisions based on a deep reinforcement learning strategy. , For technological reserves; Step 44: Calculate the contribution of each state variable to the decision in real time based on the SHAP framework. ; ; ; Step 45: Calculate raw material throughput, update production capacity inventory and technology reserves.

[0014] As a preferred embodiment of the present invention, a cloud-edge collaborative and industrial adaptive digital twin construction system includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement the cloud-edge collaborative and industrial adaptive digital twin construction method.

[0015] The beneficial effects of this invention are: 1. This invention achieves high-precision collaborative calibration under privacy protection, effectively resolving the contradiction between data silos and modeling accuracy. Traditional methods require centralized uploading of original production data from each factory (such as wafer count and yield rate), posing a risk of leakage of trade secrets. In contrast, this invention, by deploying edge gateways and federated aggregation servers, combined with homomorphic encrypted gradient transmission and a two-stage estimation algorithm for instrumental variables, effectively removes intrinsic noise interference such as equipment aging and hidden losses without the need to collect original data, significantly improving the accuracy and robustness of physical parameter calibration in cross-regional and cross-factory scenarios. 2. This invention establishes cross-modal semantic perception capabilities, enabling forward mapping from unstructured information to physical control signals. Traditional industrial control systems can only process structured numerical signals and cannot perceive textual information such as macroeconomic policies and market sentiment. This invention introduces a cloud-based semantic reasoning module, which uses natural language processing technology to quantify trade news, policy texts, etc., into physical environment damping signals in real time. This gives the control system "semantic understanding" capabilities, allowing it to adjust production strategies in advance based on semantic warnings before physical risks such as raw material supply disruptions occur, significantly improving the system's resilience in the face of uncertain events. 3. This invention provides interpretable AI decision support, overcoming the application barriers of "black box" algorithms in industrial scenarios. Addressing the issue of insufficient interpretability of deep reinforcement learning strategies, this invention integrates a SHAP attribution analysis engine into the industrial controller. This engine not only outputs control commands but also generates a real-time decision contribution graph, clearly demonstrating the marginal impact of features such as risk, capacity, and technology on the current action. This enables on-site personnel to intuitively distinguish between "normal risk avoidance" and "algorithm anomalies," significantly enhancing the credibility and safety of intelligent control in industrial implementation. 4. This invention constructs a flexible architecture that is collaborative between the cloud, edge, and terminal, enabling optimized configuration of computing power and data. This invention deploys semantic reasoning in the cloud, places parameter calibration at the edge, and pushes real-time control decisions down to the PLC, forming a hierarchical and collaborative computing power layout. This architecture supports both global perception of the macro environment and meets the requirements of millisecond-level real-time control, effectively overcoming the problems of large network latency and high risk of single point of failure in traditional centralized architectures. Attached Figure Description

[0016] Figure 1 This invention relates to a flowchart of a cloud-edge collaborative and industrial adaptive digital twin construction method. Figure 2 This is a diagram of the environmental damping signal output by the cloud-based semantic reasoning server involved in the embodiment; Figure 3 This is a joint calibration diagram of the federated parameters of the edge computing gateway involved in the embodiment; Figure 4This is a diagram illustrating the interpretability of AI decision attribution analysis based on SHAP values ​​in the embodiments. Figure 5 The physical state evolution diagram of the hardware in-the-loop (HIL) simulation involved in the embodiment; Figure 6 The system state space phase plane trajectory diagram involved in the embodiment. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0018] like Figure 1 As shown, a method for constructing a cloud-edge collaborative and industrial adaptive digital twin includes the following steps: Step 1: Use a pre-trained NLP model to extract key semantic events and construct an event-damping mapping model based on an attention mechanism; set basic environmental damping based on industry benchmarks and historical data, use an exponential decay model to simulate the timeliness of event impact, synthesize the final damping signal and send it to the edge computing gateway through an encrypted channel. Cloud-based semantic reasoning server: The cloud-based semantic reasoning server is deployed in a private cloud cluster. It uses natural language processing technology to analyze global unstructured text data in real time, identify key semantic events, and quantify them into continuously adjustable physical damping signals. These signals are then distributed to the edge execution layer by the cloud-edge collaborative architecture to achieve a closed-loop response from macro-environmental changes to micro-physical control. First, semantic event extraction and encoding; the system collects unstructured text data in real time through channels such as global trade news APIs (e.g., Reuters, Bloomberg), government policy announcement websites, industry analysis reports, and supply chain early warning information; Secondly, pre-trained NLP models (BERT / GPT series) were used to extract key events, including changes in tariff policies, supply chain disruptions, technological breakthroughs, and the release of industry subsidies. Event vectorization encoding was then performed, with the pre-trained model converting textual events into vectorized data. Encode it as a high-dimensional semantic vector, the formula is: (1) in, For BERT-based models; This indicates the output dimension, which is 768.

[0019] Secondly, an event-damping mapping model is constructed; semantic vectors are mapped to damping increments through an attention mechanism network, defined as: (2) in, This is the weight matrix; For bias terms; The knowledge base matrix contains historical event patterns; The function is defined as ; The loss function is defined as follows: The mapping network is trained using supervised learning. (3) in, Labels representing the actual damping impact of historical events; The expected damping increment for historical events; This is the regularization coefficient.

[0020] Secondly, the damping signal is set. The basic damping is set based on industry benchmarks and historical data, using the following formula: (4) in, This represents the reference environmental damping (usually set as a constant); This represents the industry average capacity utilization rate. This represents the theoretical maximum production capacity. This is an industry adjustment coefficient.

[0021] An exponential decay model is used to simulate the timeliness of information, and the formula is defined as: (5) in, This is the attenuation coefficient, with a default value of 0.05. Indicates an indicator function; This is the timestamp of the event.

[0022] Furthermore, the damping signal is the superposition of the effects of all events, with Gaussian noise added to simulate uncertainty, and is defined by the formula: (6) in, In time The total number of events previously identified; Represents Gaussian white noise ; Finally, the synthesized damping signal is sent to the edge computing gateway through an encrypted channel for real-time adjustment of the control strategy. The upper and lower bounds of the damping signal are defined as follows: (7) Among them, the default settings , ; Edge Federated Learning Nodes: Step 2: Construct a physical model of the industrial process, perform two-stage least squares parameter calibration of instrumental variables locally on the edge computing gateway; encrypt the local gradient using the Paillier homomorphic encryption algorithm, and complete secure aggregation in the encrypted state on the cloud federated server; perform statistical inference and confidence assessment on the physical gain parameters, and introduce differential privacy and security verification mechanisms. By using an edge computing gateway, instrumental variable two-stage least squares (IV-2SLS) parameter calibration is performed locally, and only the encrypted parameter gradients are uploaded to the cloud federated server. The cloud server aggregates the gradients of multiple edge nodes and updates the global physical parameters in an encrypted state, thereby achieving collaborative learning and high-precision parameter calibration of multi-source heterogeneous data while ensuring zero leakage of industrial data.

[0023] First, we construct a physical model of the industrial process, assuming the physical process of the industrial system is as follows: (8) in, For output variables (yield rate), dimensions Indicates the number of samples; For input variables (raw material input); For environmental damping variables (from the cloud semantic reasoning server); This represents the physical gain parameter to be calibrated; This is the damping effect coefficient; Represents the random error term. ; In industrial processes, input variables (Often related to unobserved factors, such as equipment condition, operator experience, etc.), these factors are included in the error term. This leads to endogeneity problems, namely: .

[0024] Endogeneity will lead to a decline in the traditional least squares estimator There is a discrepancy: .

[0025] To address the endogeneity problem, instrumental variables are introduced. Satisfying the correlation condition Instrumental variables are correlated with endogenous explanatory variables; exogenous conditions Instrumental variables are uncorrelated with the error term; Exclusivity constraint: instrumental variables are only generated through endogenous variables. Impact on output .

[0026] Secondly, the instrumental variable two-stage least squares (IV-2SLS) parameter calibration is performed locally via the edge computing gateway; the first stage is auxiliary regression, which constructs the input variables. For instrumental variables and environmental damping The linear regression model is defined as follows: (9) in, For the intercept term; Instrumental variable right The influence coefficient; For environmental damping right The influence coefficient; This represents the regression residuals for the first stage.

[0027] Parameters are estimated using the least squares method The predicted value is: (10) Secondly, in the second stage, main regression is used to construct output variables. For predicted input and environmental damping The regression model is defined by the following formula: (11) in, For the second-stage regression residuals; The damping effect coefficient is given; and the physical gain parameter is estimated using the least squares method. .

[0028] To derive the matrix form, let the first... The local dataset for each edge node is: , (12) in, For input variables; For output variables.

[0029] The parameter estimation formula for the first stage is: (13) in, This is the actual input vector.

[0030] The parameter estimation formula for the second stage is:

[0031] in, .

[0032] Secondly, through homomorphic encrypted gradient transmission, edge nodes use the Paillier public-key encryption algorithm to encrypt the local parameter gradients. The key generation process includes the following: The key generation algorithm randomly selects two prime numbers. and ,satisfy: , , (14) in, P It is the set of large prime numbers.

[0033] Calculate the modulus separately and Carmichael function value The calculation formula is: , (15) in, It is the least common multiple.

[0034] Randomly select an integer ,satisfy ,in , It is the greatest common divisor.

[0035] Calculate the modular inverse The calculation formula is: .

[0036] The public key is The private key is .

[0037] Encryption algorithms for plaintext messages Random selection : ; Decryption algorithm, for ciphertext ,have: ; Homomorphic properties include additive homomorphism and scalar multiplication homomorphism, with the following formulas: ; ; Encryption gradient calculation, the first The gradient of parameters for each edge node is calculated based on the local loss function, as shown in the formula: (16) The gradient calculation formula is: (17) Quantize the floating-point gradient into an integer to avoid the precision loss of homomorphic encryption. The formula is as follows: (18) in, This is the quantization accuracy parameter.

[0038] The gradient is encrypted using the Paillier public key, and the calculation formula is as follows: (19) in, This is the random blinding factor.

[0039] Secondly, and furthermore, a federated aggregation algorithm is adopted; Encryption state aggregation: The cloud federated server receives the encryption gradients from all edge nodes, using the following formula: (twenty one) in, This represents the node weight, which is proportional to the node's data.

[0040] After aggregation, decrypt the result using the private key. The formula is: (twenty two) Global parameters are updated using the FedAvg algorithm, with the following formula: (twenty three) in, Indicates the learning rate. This is the global physical gain parameter.

[0041] Then, statistical inference and confidence assessment are performed: Parameter variance estimation: Calculate the variance estimate of the physical gain parameter using the following formula: (twenty four) in, .

[0042] Confidence space construction: Based on the central limit theorem, construct the parameter confidence intervals: (25) in, The quantiles of the standard normal distribution Physical gain parameter The global estimated average value.

[0043] Finally, a security and privacy protection mechanism is adopted, employing differential privacy enhancement by adding Gaussian noise before gradient uploading, defined as: (26) The noise variance satisfies the following: (27) in, Budget for privacy; Gradient Sensitivity.

[0044] To prevent malicious nodes from uploading erroneous gradients, a verification mechanism based on secure multi-party computation is introduced, defined as follows: (28) in, To encrypt gradient ciphertext, Zero-knowledge concise non-interactive knowledge argumentation.

[0045] Step 3: Construct a Markov decision process for the industrial control system, and use a deep deterministic policy gradient algorithm to achieve optimal control in the continuous action space; implement three inference acceleration technologies: INT8 quantization, layer fusion, and memory layout optimization; integrate the SHAP interpretable analysis engine, and introduce hierarchical grouping SHAP and Monte Carlo sampling strategies to ensure real-time analysis efficiency. Interpretable industrial controllers: It integrates a deep reinforcement learning policy network with a SHAP value calculation engine, and outputs control decisions and their attribution explanations simultaneously, achieving synchronous output of "decision-explanation"; Constructing a Markov decision process: The state space of an industrial control system is defined as follows: (29) in, Dimensions representing state characteristics; At any decision moment The system's state vector Specifically defined as: , This refers to the level of production capacity and inventory. To maintain the level of technological reserves, Environmental risk level, This represents the environmental damping value. For the first A historical state characteristic.

[0046] The output motion space of the controller is defined as the servo motor power coefficient as follows: (30) in, The actuator is in a completely shut-down state; The maximum power output of the actuator under rated conditions; action value. This directly determines the instantaneous output power ratio of the servo system.

[0047] To guide the agent in learning a control strategy that balances efficiency, economy, and safety, this invention designs a composite reward function that integrates four key objectives, as shown in the formula: (31) in, These are non-negative weighting coefficients used to balance the relative importance of different objectives; the generation efficiency reward is... , Physical gain calibrated for federated learning The instantaneous throughput of raw materials; energy consumption penalty is , Energy consumption cost coefficient; safety constraint reward is , and These represent the lower limit of safe production capacity and inventory, and the upper limit of safe risk level; the smoothing reward is... .

[0048] To achieve optimal control in a continuous action space, the Deep Deterministic Policy Gradient (DDPG) algorithm is employed, with its core consisting of parallel Actor and Critic networks. Actor Network It is a deterministic policy network that maps the current state to a specific action, with the input being a state vector. The output is Network parameters are .

[0049] Critic Network Evaluate the long-term expected reward of performing a specific action in a given state, with the state as the input. and actions The concatenation; the output is a state-action value estimate. Network parameters are .

[0050] Algorithm training is accomplished by alternately optimizing the Actor and Critic networks; the Critic network is updated by minimizing the temporal difference error to update its parameters. The loss function is defined as: (32) in, This represents the batch size of transferred samples sampled from the experience replay pool; the target value is... , Indicates the discount factor; , They are respectively and Target network.

[0051] The Actor network is updated using the policy gradient method, updating the network parameters in the direction that increases the Q-value. The policy gradient is approximated as: (33) Real-time inference optimization: To meet industrial real-time requirements (<10ms response time), three optimization techniques are employed.

[0052] 1. INT8 quantization linearly quantizes network weights and activation values ​​from a 32-bit floating-point (FP32) dynamic range to an 8-bit integer (INT8), significantly reducing memory usage and computational bandwidth. The quantization process is as follows: (34) in, This is the original FP32 weight matrix.

[0053] 2. Layer fusion optimization: The BatchNorm layers are fused into the convolutional layers, using the following formula: , (35) in, This is the scaling factor for the BatchNorm layer. The variance of batch data is calculated for the BatchNorm layer. It is a constant. The average value of batch data calculated by the BatchNorm layer. These are the bias parameters of the original convolutional layer. This is the offset coefficient for the BatchNorm layer.

[0054] 3. Memory layout optimization: The default NCHW (batch-channel-height-width) memory layout is converted to NHWC (batch-height-width-channel) layout.

[0055] SHAP value interpretability analysis is employed: SHAP values ​​are based on Shapley values ​​in cooperative game theory, assigning a fair contribution value to each feature; for features containing... Model prediction of features ,feature SHAP value The calculation formula is as follows: (36) in, It is the set of all features; yes The middle does not contain features any subset of; It is a characteristic function, indicating that only a subset is used. The features in the model are the expected values ​​of the predictions. Let be the input state vector; the decision function is . The characteristic function is .

[0056] Calculating the Shapley value precisely has exponential complexity. Kernel SHAP is an efficient approximation method based on a local surrogate model, which can be used to approximate the value of the sample to be explained. Nearby, using a simple linear model To approximate complex black box models Defined as: (37) in, A simplified input indicating whether a feature exists; It is the prediction of the model when all features take the baseline value; Features to be solved The SHAP value.

[0057] Using SHAP kernel weights to define the sample weight kernel function requires defining a kernel function for each simplified input. Assign a weight The formula is: (38) The optimization problem is solved by minimizing the weighted squared loss to obtain the SHAP value, as shown in the formula: (39) in, Represents the set of all possible subsets of features; This means that the simplified input is mapped to the original feature space.

[0058] Given that the original computational cost of Kernel SHAP may still exceed real-time limitations, two further real-time optimization strategies are introduced: hierarchical grouping SHAP and Monte Carlo sampling with gradient integration. Hierarchical grouping SHAP groups industrial state characteristics according to their physical meaning, and first calculates the between-group SHAP values. Then, calculate the relative contribution of each feature within the group. The calculation formula is: (40) in, It is a fast approximation based on gradient.

[0059] Monte Carlo sampling and gradient integration, Monte Carlo approximation through a sampled feature subset To approximate the integral of the SHAP value, the formula is as follows: (41) in, Indicates the number of randomly selected feature subsets; Indicates the first The feature subset selected in the second sampling; Representing a subset Add features The subsequent feature set; Indicates that only a subset is retained. The features in the data are replaced with baseline values, while other features are replaced with baseline values.

[0060] Step 4: Generate cloud environment damping signal flow, set up normal and adversarial test dual modes to evaluate system robustness; input the disturbed signal and real-time status into the programmable logic controller, and simultaneously execute servo power decision and interpretable attribution analysis; advance the simulation with a fixed step size and output full-cycle timing data; by comparing key performance indicators in the dual modes, realize automated quantitative evaluation of robustness and provide a basis for controller tuning. Physical execution verification: By automatically generating diverse interference signals through generative adversarial networks, the closed-loop execution of control strategies and the dynamic response of physical systems are realized in a hardware-in-the-loop (HIL) simulation environment. The robustness of the system is verified by injecting adversarial interference.

[0061] 1. Generation of cloud-based environmental damping signal flow: Each verification cycle begins with acquiring the environmental damping signal sent by the cloud semantic server. The signal is synthesized from basic damping, semantic event increments, and random noise, as shown in the formula: (42) in, As the reference damping; For semantic event increments, by analyzing discrete event sequences The formula is obtained by modeling exponential decay: (43) in, Represents a set of events; The time when the event occurred; The event amplitude; The decay coefficient is used to simulate the time-varying decay of information influence; Gaussian white noise is used to simulate environmental uncertainty.

[0062] 2. Adversarial interference injection and test mode switching: To evaluate the system's robustness under unexpected disturbances, two simulation modes were set up: normal and adversarial testing. In normal mode, the system was directly used... As an environmental input; when in adversarial testing mode, Superimposed with composite interference inspired by Generative Adversarial Networks (GANs) The formula is: (44) The interference signal is designed as follows: Among them, setting Simulate high-frequency vibration interference; set The simulation of periodic pulse disturbances aims to test the stability and recovery capability of the control system in the face of unexpected, strong dynamic disturbances.

[0063] 3. Explainable intelligent control decisions and real-time attribution: Damped signal after injection of interference With the current system state vector The input is sent to the industrial programmable logic controller (PLC); among which, For production capacity inventory, For technological reserves.

[0064] The controller executes servo power decisions based on a deep reinforcement learning strategy, using the following formula: (45) in, To normalize the risk level, the decision simultaneously triggers the Explainable Artificial Intelligence (XAI) module, which calculates the contribution of each state variable to the decision in real time based on the SHAP framework. The formula is as follows: (46) (47) (48) in, The SHAP contribution value of environmental risk state variables to decision-making. The SHAP contribution of capacity and inventory status variables to decision-making. The SHAP contribution of the technology reserve variable to the decision.

[0065] 4. Dynamic evolution of physical systems: Control Action When the system is controlled via a servo motor model, the system state evolves according to the following discrete-time equation: 41. Calculation of raw material throughput: ; 42. Capacity and Inventory Updates: ; 43. Technological Reserve Updates: ; in, Physical gain as defined by the federal government; ; ; ; .

[0066] This is the damping loss coefficient; and These are the natural depreciation rates for production capacity and technology, respectively. This is the coefficient for technological accumulation.

[0067] 5. Full-cycle data recording and robustness assessment: The entire verification process uses a fixed step size. Iterative progress, with each step recorded including: timestamp System status Controlling actions Environmental damping and SHAP attribution vector .

[0068] Output the complete time series dataset after the simulation. By comparing key indicators such as production capacity stability, control action smoothness, and system recovery time under normal and adversarial modes, the robustness of the control system can be automatically and quantitatively evaluated, providing a basis for controller parameter tuning and anti-disturbance strategy design.

[0069] An industrial adaptive digital twin system based on cloud-edge collaboration and explainable AI includes: The cloud-based semantic reasoning and nonlinear damping mapping module is configured to execute the following process: Real-time collection of unstructured text data from multiple channels; extraction of key semantic events using a pre-trained NLP model; construction of an attention-based event-damping mapping model; setting of basic environmental damping based on industry benchmarks and historical data; simulation of the timeliness of event impact using an exponential decay model; and finally, synthesis of a damping signal, which is then transmitted to the edge computing gateway via an encrypted channel. This achieves the mapping from unstructured semantic information to continuously adjustable physical signals; such as... Figure 2 As shown, on the timeline At this point, the system recognizes the positive event of "subsidy disbursement," and the damping coefficient decreases accordingly, providing support for subsequent proactive control strategies; At this point, the system detects the sudden negative event of "tariff increase", and the damping coefficient rises rapidly to its peak value; this damping signal is sent to the edge in real time via the MQTT protocol as a feedforward control signal to trigger the system's preventive adjustment and robust response; The edge federation parameter collaborative calibration and secure aggregation module is configured to perform the following processes: constructing an industrial process physical model and performing parameter calibration locally on each edge computing gateway using a two-stage least squares method based on instrumental variables; encrypting the parameter gradients generated by calibration using the Paillier homomorphic encryption algorithm and uploading them to the cloud federation server, where secure aggregation is performed under encrypted conditions; the server performs statistical inference and confidence interval analysis on the aggregated estimated physical gain parameters, while introducing differential privacy and secure multi-party computation verification mechanisms to enhance system privacy protection and security; technological synergies such as... Figure 3 As shown, the gray and light blue scatter dots represent the local production data (raw material input and yield output) of two physically isolated factory nodes (node ​​A and node B), respectively. Due to the heterogeneity of data distribution among nodes and the inability to share data directly due to privacy constraints, the local fitting model of a single node (as shown by the blue and green dashed lines) is prone to bias or overfitting. Through encrypted gradient uploading and federated averaging aggregation, the system finally obtains a global physical gain model (red solid line). This model is located at the trend center of the two sets of data distributions, and its calibrated physical gain parameters effectively integrate the features of multiple nodes, achieving high-precision collaborative modeling under the premise of data privacy protection.

[0070] The interpretable intelligent control and real-time decision attribution module is configured to execute the following processes: constructing a Markov decision process for the industrial control system; employing a deep deterministic policy gradient algorithm to achieve optimal control in the continuous action space; and completing policy training through alternating optimization of Actor and Critic networks, implementing three inference optimization techniques: INT8 quantization of network weights and activation values, fusion of BatchNorm layers and convolutional layers, and NHWC memory layout transformation; integrating the SHAP interpretability analysis engine and introducing two acceleration strategies—hierarchical grouping SHAP and Monte Carlo sampling combined with gradient integration—to ensure real-time analysis efficiency; the output results of the interpretable analysis unit running inside the industrial controller (PLC) are as follows: Figure 4 As shown, a color-stacked bar chart visually presents the SHAP contribution value of each state characteristic to the controller's decision; in At this point, the red bars representing risk factors show significant negative values ​​and extend downwards considerably; this indicates that although the technology reserves (green bars) and production capacity status (blue bars) at that time still provided positive support for production, the risk factor weights increased sharply due to the external "tariff increase" event. This can explain why the AI ​​controller made a defensive decision to proactively reduce servo power based on a comprehensive assessment of the contributions of multiple features.

[0071] The hardware-in-the-loop adversarial testing and physical system evolution verification module is configured to execute the following process: generate a cloud-based environmental damping signal stream; set two simulation modes, normal and adversarial testing, to evaluate system robustness; input the disturbed signal and real-time system state into an industrial programmable logic controller (PLC); and simultaneously execute servo power decision-making based on deep reinforcement learning (DRL) and explainable artificial intelligence (XAI) attribution analysis; the simulation process iterates with a fixed step size and outputs a time-series dataset covering the entire lifecycle; by comparing key performance indicators such as production stability, control smoothness, and dynamic recovery time under the two modes, an automated quantitative evaluation of system robustness is completed, thus providing a reliable basis for controller parameter tuning and anti-interference strategy optimization. The system state evolution of the physical actuator in response to the above PLC control commands is as follows: Figure 5 As shown, in the Hardware-in-the-Loop (HIL) test mode, when the environmental damping is... When drastic fluctuations occur and servo power adjusts accordingly, the system's capacity inventory (solid green line) does not experience a catastrophic drop, but rather achieves a smooth transition to a new steady state; the trajectories of the system control variable (servo power) and state variable (capacity inventory) in the phase plane are as follows: Figure 6 As shown, its convergence shape indicates that the system has not diverged or oscillated continuously, verifying that the differential topology network constructed based on the instrumental variable (IV) calibration parameters has Lyapunov stability, which meets the requirements of the industrial control system for dynamic safety.

[0072] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for constructing a cloud-edge collaborative and industrial adaptive digital twin, characterized in that, Includes the following steps: Step 1: Extract key semantic events and construct an event-damping mapping model based on an attention mechanism; An exponential decay model is used to simulate the timeliness of the event's impact, and the final damping signal is synthesized. Step 2: Construct a physical model of the industrial process and perform two-stage least squares parameter calibration of instrumental variables; The local gradients are encrypted and securely aggregated; statistical inference and confidence evaluation are performed on the physical gain parameters, and differential privacy and security verification mechanisms are introduced. Step 3: Construct a Markov decision process and use a deep deterministic policy gradient algorithm to achieve optimal control of the continuous action space; perform quantization, layer fusion and memory layout optimization; introduce hierarchical grouping SHAP and Monte Carlo sampling strategies to ensure real-time analysis efficiency; Step 4: Generate cloud-based environmental damping signal stream; input the disturbed signal and real-time status into the controller, and synchronously execute servo power decision and interpretable attribution analysis.

2. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 1, characterized in that, Step one specifically includes: Step 11: Collect unstructured text data; Step 12: Use a pre-trained NLP model to extract key events; Step 13: Map the semantic vectors to damping increments using an attention mechanism network: And train the mapping network; Step 14: Set the reference environmental damping , This represents the industry average capacity utilization rate. This represents the theoretical maximum production capacity. This is an industry adjustment coefficient; Step 15: Use the exponential decay model ; The attenuation coefficient; Indicates an indicator function; The timestamp of the event; Step 16: Introduce Gaussian noise into the damping signal; Step 17: Send the synthesized damping signal to the edge computing gateway through an encrypted channel.

3. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 2, characterized in that, The formula for the upper and lower bound constraints of the synthesized damping signal is as follows: , This is the synthesized damping signal.

4. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 1, characterized in that, Step two specifically includes: Step 21: Construct a physical model of the industrial process ; For output variables; For input variables; For environmental damping variables; This represents the physical gain parameter to be calibrated; This is the damping effect coefficient; Represents the random error term; Step 22, Build For instrumental variables and linear regression model ; For the intercept term; Instrumental variable right The influence coefficient; For environmental damping right The influence coefficient; The parameters are estimated using the least squares method based on the first-stage regression residuals. , to obtain the predicted value ; Step 23, Construction right and regression model ; For the second-stage regression residuals; Step 24, let the first... Local datasets of edge nodes; Step 25: Perform parameter estimation in both stages. , ; Step 26: The edge nodes encrypt the local parameter gradients using the Paillier public key encryption algorithm.

5. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 4, characterized in that, Step 26 specifically includes: Randomly select two prime numbers and ,satisfy , The greatest common divisor; calculate pq modulus and Carmichael function value ; Randomly select an integer ,satisfy ; Calculate the modular inverse ; Set the public key to The private key is ; For plaintext messages Random selection : ; Decryption of ciphertext ,have: ; Construct the first The parameter gradient of each edge node is calculated based on the local loss function. ; Perform gradient calculation ; The gradient is encrypted using Paillier's public key. ; The cloud-based federated server receives encryption gradients from all edge nodes. ; Decrypt the aggregation result using the private key. ; Global parameters are updated using the federated average algorithm. ; Indicates the learning rate. This is the global physical gain parameter.

6. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 1, characterized in that, Constructing a Markov decision process and employing a deep deterministic policy gradient algorithm to achieve optimal control in the continuous action space specifically includes: Step 31: Construct a Markov decision process ; Step 32: Define the system's state vector , , This refers to the level of production capacity and inventory. To maintain the level of technological reserves, Environmental risk level, This represents the environmental damping value. For the first A historical state characteristic; Step 33: Set the controller's output motion space to the servo motor power coefficient; Step 34: Construct a composite reward function ;, The weighting coefficient is non-negative; the generation efficiency reward is... , Physical gain calibrated for federated learning The instantaneous throughput of raw materials; energy consumption penalty is , Energy consumption cost coefficient; safety constraint reward is , and These represent the lower limit of safe production capacity and inventory, and the upper limit of safe risk level; the smoothing reward is... ; Step 35: Use the DDPG algorithm to control the motion space.

7. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 1, characterized in that, Quantization, layer blending, and memory layout optimization include: INT8 quantization is used; BatchNorm layers are merged into convolutional layers; NCHW memory layout is converted to NHWC layout.

8. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 1, characterized in that, Hierarchical grouping SHAP includes: Build includes Model prediction of features ,feature SHAP value ; Constructing a linear model ; A simplified input indicating whether a feature exists; Construct weights ; By minimizing the weighted squared loss Solve for the SHAP value.

9. The method for constructing a cloud-edge collaborative and industrial adaptive digital twin according to claim 1, characterized in that, Step four specifically includes: Step 41: Obtain environmental damping signal ; As the reference damping, For semantic event increments, It is Gaussian white noise; Step 42, in Superimposed composite interference inspired by GAN networks : ; Step 43: The controller executes servo power decisions based on a deep reinforcement learning strategy. , For technological reserves; Step 44: Calculate the contribution of each state variable to the decision in real time based on the SHAP framework. ; ; ; Step 45: Calculate raw material throughput, update production capacity inventory and technology reserves.

10. A cloud-edge collaborative and industrial adaptive digital twin construction system, characterized in that: include: Memory is used to store instructions that can be executed by the processor; A processor for executing instructions to implement the cloud-edge collaborative and industrial adaptive digital twin construction method as described in any one of claims 1-9.

Images