Industrial internet of things optimization system and method fusing digital twin and federated learning

Abstract

The application discloses an industrial internet of things optimization system and method fusing digital twinning and federated learning, belongs to the technical field of industrial internet of things, and aims to solve the problems of physical mapping inconsistency, model drift and low collaborative efficiency under resource limitation. The system comprises a physical entity layer, a digital twinning mapping layer, a distributed federated learning layer and a cross-domain resource scheduling layer. The digital twinning layer ensures consistency between virtual and real; the federated learning layer realizes model collaboration by using physical constraint alignment and adaptive aggregation; and the scheduling layer optimizes calculation, communication and energy resources. The application realizes deep fusion of mechanism and data driving, suppresses model drift, improves resource utilization and system adaptive capacity while ensuring privacy.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) technology, and in particular to an industrial IoT optimization system and method that integrates digital twins and federated learning. Background Technology

[0002] As the global industrial sector deepens its transformation towards digitalization, networking, and intelligence, the Industrial Internet of Things (IIoT) has become a key infrastructure supporting the restructuring of production paradigms and the enhancement of core competitiveness in modern manufacturing. In complex industrial production environments, thousands of sensors, actuators, and smart terminals interact frequently, constructing a complex mega-system deeply coupled with the information space. To achieve ultimate optimization of production efficiency, accurate prediction of equipment health status, and adaptive adjustment of production processes, various advanced data processing and model evolution technologies have been widely introduced into the industrial field. These technologies aim to extract knowledge models that drive production decisions through in-depth mining of massive amounts of industrial data.

[0003] In this technological evolution, digital twin technology, with its ability to perform high-fidelity mapping of physical entities in virtual space, has become a core tool for industrial system modeling and real-time monitoring. By integrating multi-scale, multi-physics, and multi-probability simulation processes, digital twins can synchronize the operating status, environmental parameters, and workload of physical entities in real time. Based on historically accumulated operating data and mechanistic models, it can perform pre-simulation, diagnosis, and evaluation of production processes. This "virtual-real mapping, two-way closed-loop" mechanism greatly enhances the observability and controllability of industrial systems in dynamic environments, providing a solid theoretical foundation for achieving precise industrial control.

[0004] Meanwhile, because industrial data often involves core production processes, trade secrets, and national security, data privacy protection and the "data silo" effect have become key bottlenecks restricting the implementation of large-scale artificial intelligence algorithms in the Industrial Internet of Things (IIoT). Against this backdrop, Federated Learning, as an emerging distributed machine learning framework, allows industrial nodes to collaboratively train a globally optimal model by exchanging model parameters or gradient information without disclosing the original data, based on the core idea of ​​"the model moves while the data remains stationary." This mechanism demonstrates significant technical advantages in addressing issues such as limited sample size, fragmented data distribution, and high privacy compliance requirements in the industrial sector, opening up new paths for collaborative optimization across equipment and factories.

[0005] However, as industrial applications place increasingly stringent demands on real-time performance, reliability, and sophisticated management, relying solely on the independent operation of digital twins or federated learning, or simply loosely coupling the two, is revealing deeper technical contradictions in practical engineering applications. Specifically, the effectiveness of a digital twin system highly depends on the synchronization accuracy between the virtual model and the physical entity. However, in a distributed industrial IoT environment, there is significant asymmetry in the computing resources, storage capacity, and communication bandwidth of different terminal devices. This heterogeneity leads to substantial communication overhead and processing latency when the digital twin model performs high-frequency state synchronization and parameter updates, making it difficult to guarantee the real-time mapping between the virtual and physical worlds.

[0006] Furthermore, traditional federated learning mechanisms are highly susceptible to severe model drift when processing non-independent, identically distributed (Non-IID) data in industrial scenarios. Due to differences in operating conditions, environmental disturbances, and sensor accuracy among industrial equipment at different workstations and in different batches, the model parameters generated through local training often possess strong individual characteristics. Directly aggregating these significantly differentiated parameters leads to slow or even divergent global model convergence, resulting in a substantial decrease in the accuracy of collaborative optimization. The root cause lies in the fact that existing collaborative frameworks often overlook the deep interaction logic between the physical constraints provided by digital twins and the statistical learning process of federated learning. In dynamically changing production processes, the degradation of physical entities, random environmental disturbances, and instantaneous switching of production tasks all cause non-stationary evolution of data distribution. Without a mechanism that can perceive changes in physical state in real time and dynamically adjust the learning strategy accordingly, federated learning will struggle to capture the physical evolution patterns inherent in the digital twin model, resulting in a lack of physical consistency in the trained model and making reliable deployment in harsh industrial environments.

[0007] Furthermore, edge nodes in the Industrial Internet of Things (IIoT) typically face stringent energy efficiency constraints. Achieving the optimal balance between overall system energy efficiency and performance—while ensuring model training accuracy and digital twin synchronization quality—by dynamically optimizing computational task offloading strategies and communication resource allocation weights, constitutes a highly coupled nonlinear optimization challenge. This trade-off between privacy protection strength, model training accuracy, system response latency, and resource consumption makes existing single optimization methods insufficient to cope with the complex and ever-changing industrial implementation conditions.

[0008] In summary, existing technical solutions, while integrating the real-time sensing capabilities of digital twins with the distributed collaborative advantages of federated learning, still suffer from core bottlenecks such as insufficient integration of physical mechanisms and data-driven approaches, lack of flexibility in heterogeneous resource scheduling, and poor convergence due to non-independent and identically distributed data. Therefore, how to construct an industrial IoT collaborative optimization system that deeply integrates digital twins and federated learning, achieving global collaborative optimization of model training accuracy, system real-time response, and resource utilization while ensuring data privacy and physical consistency, has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0009] This invention provides an industrial Internet of Things (IIoT) collaborative optimization system and method that integrates digital twins and federated learning, aiming to solve the technical problems of existing IIoT in distributed environments, such as inconsistent physical mapping, model drift caused by non-independent and identically distributed data, and low collaborative efficiency under heterogeneous resource constraints.

[0010] According to a first aspect of the present invention, an industrial Internet of Things collaborative optimization system integrating digital twins and federated learning is provided, the system comprising a physical entity layer, a digital twin mapping layer, a distributed federated learning layer, and a cross-domain resource scheduling layer.

[0011] The physical entity layer consists of multiple heterogeneous industrial IoT nodes deployed in the industrial field. Each heterogeneous industrial IoT node integrates a sensing module, an execution module, and an edge computing module. The sensing module is responsible for collecting multi-source heterogeneous data of the physical entity in real time, including but not limited to equipment operating parameters, environmental state variables, and process flow data. The execution module adjusts the physical production process in real time according to received control commands. The edge computing module provides the necessary computing and storage resources for local data processing and model training.

[0012] The digital twin mapping layer achieves high-fidelity mapping of the entire lifecycle of physical entities by establishing virtual models that correspond one-to-one with the physical entity layer. The digital twin mapping layer includes a state observer, a mechanism evolution model, and a real-time synchronization unit. The state observer is used to preprocess and extract features from the collected multi-source heterogeneous data to construct the real-time state vector of the physical entity. The mechanism evolution model predicts the future state of the physical entity based on preset physical laws and dynamic equations. The real-time synchronization unit dynamically adjusts the parameters of the virtual model by calculating the mapping deviation between the physical entity and the virtual model to ensure real-time consistency between the virtual and real spaces.

[0013] The distributed federated learning layer is built upon the digital twin mapping layer. It achieves collaborative training of the globally optimal model by exchanging model parameters among nodes instead of raw data. The distributed federated learning layer includes a local training module, a physical constraint alignment module, and an adaptive aggregation module. The local training module utilizes local node data for model evolution. The physical constraint alignment module transforms the physical mechanism information provided by the digital twin mapping layer into constraint terms, which are then incorporated into the loss function of the local training. The adaptive aggregation module dynamically adjusts the aggregation weights based on the twin fidelity and data quality of each node to generate a global model.

[0014] The cross-domain resource scheduling layer is responsible for the global collaborative optimization of the computing, communication, and energy resources of the entire system. This layer includes a task offloading unit, a bandwidth allocation unit, and an energy efficiency management unit. The task offloading unit dynamically determines the execution location of digital twin synchronization and federated learning training tasks based on the real-time load and computing capabilities of each node. The bandwidth allocation unit optimizes the spectrum resource usage of each node during model parameter uploading and state synchronization. The energy efficiency management unit minimizes the total system power consumption while meeting real-time requirements by adjusting the processor's operating frequency and transmit power.

[0015] Furthermore, the state observer in the digital twin mapping layer describes the evolution process of physical entities by constructing a state-space model, and its state transition logic is as follows:

[0016]

[0017] In the above formula, Indicates the first A physical entity in The state vector at any given time; The state transition matrix represents the evolution of the physical entity's state over time. Indicates in The control input vector applied to the entity at any given time; This represents the input matrix, which defines the weights of the control inputs on the state. This represents the process noise vector, used to simulate the impact of random environmental disturbances on physical entities.

[0018] Furthermore, the real-time synchronization unit in the digital twin mapping layer updates the virtual model parameters by minimizing the residual between the observed data and the predicted data, and its synchronization criterion satisfies the following mathematical expression:

[0019]

[0020] In the above formula, Indicates in The actual observation vectors collected at each moment; This represents the observation mapping function, used to map virtual states to the observation space; Indicates virtual model in The parameter vector at time step; This represents the regularization coefficient, used to control the smoothness of model parameter updates and avoid drastic model fluctuations caused by transient noise.

[0021] Furthermore, the physical constraint alignment module in the distributed federated learning layer ensures that the data-driven model training process conforms to the inherent mechanisms of physical entities by introducing a physical consistency loss term. The improved loss function for local training is defined as:

[0022]

[0023] In the above formula, Indicates the first Local model parameters for each node; Indicates based on local dataset Experience risk loss; The physical consistency loss is represented by measuring the deviation between the model's predictions and the predictions of the digital twin mechanism model. This represents the physical constraint weighting factor, used to adjust the intensity of the influence of physical mechanisms on the learning process.

[0024] Furthermore, the physical consistency loss The specific construction method is as follows:

[0025]

[0026] In the above formula, This represents the predicted state provided by the digital twin model; This represents the differential equation operator followed by the physical system; this term ensures that the model generated by federated learning strictly follows the dynamic constraints of the physical entity in terms of parameter evolution.

[0027] Furthermore, the adaptive aggregation module in the distributed federated learning layer employs a weighted algorithm based on Siamese fidelity. The update logic for the global model parameters is as follows:

[0028]

[0029] In the above formula, This represents the aggregated global model parameters; This represents the total number of nodes participating in collaborative training; Indicates the first The twin fidelity score for each node. It is calculated based on the reciprocal of the average mapping deviation between physical entities and virtual models within a preset time window. The smaller the deviation, the higher the fidelity, and the greater its weight in the global model aggregation.

[0030] Furthermore, the cross-domain resource scheduling layer achieves Pareto optimality of system performance by establishing a multi-objective optimization model. The optimization objective function... Defined as:

[0031]

[0032] In the above formula, Indicates the first The computation frequency allocation for each node; Indicates the transmission power allocation; This indicates the proportion of communication bandwidth allocated to this node; This indicates the total time taken for the node to complete the predetermined rounds of collaborative tasks, including local training time, twin synchronization time, and parameter upload time. Indicates total energy consumption; This represents the energy efficiency regulation coefficient, used to balance time response speed and energy consumption.

[0033] Furthermore, the total time consumed The specific components satisfy the following logic:

[0034]

[0035] In the above formula, This indicates the number of computation cycles required to process a single sample. This represents the total number of samples used in local training. Indicates the amount of model parameter data that needs to be transmitted; Indicates the total available bandwidth; Indicates channel gain; This represents the power of Gaussian white noise.

[0036] Furthermore, the system addresses the dynamic stochasticity in the Industrial Internet of Things (IIoT) by introducing a Lyapunov optimization framework, ensuring system stability during long-term operation. The Lyapunov function is defined. for:

[0037]

[0038] In the above formula, Indicates the first Each node The task queue backlog length at any given time. By minimizing the drift-plus-penalty term, system energy efficiency is optimized while ensuring queue stability.

[0039] According to a second aspect of the present invention, an industrial Internet of Things (IoT) collaborative optimization method integrating digital twins and federated learning is provided. The method is implemented based on the aforementioned system and specifically includes the following steps:

[0040] Step 1: Initialization Phase. The system control center initializes the global model parameters and distributes them to each heterogeneous industrial IoT node. Simultaneously, each node initiates the digital twin mapping program to establish an initial virtual entity based on the preset mechanistic model parameters.

[0041] Step Two: Virtual-Real Synchronization and Perception Phase. Sensors in the physical entity layer collect production data in real time. The digital twin mapping layer uses a state observer to clean and extract features from the data, and uses a real-time synchronization unit to update the virtual model parameters. If the mapping deviation between the virtual model and the physical entity exceeds a preset threshold, a model reconstruction mechanism is triggered to ensure the fidelity of the virtual space.

[0042] Step 3: Local Training Phase under Physical Constraints. Each node initiates a distributed federated learning layer in its local edge computing module. The local training module uses real-time acquired physical data to calculate gradients. The physical constraint alignment module obtains the current physical evolution pattern from the digital twin mapping layer and introduces it as a penalty term into the loss function. By solving the optimization problem with physical constraints, the local model parameters are updated to ensure that the model not only fits the data distribution but also conforms to the physical mechanism.

[0043] Step 4: Adaptive Weight Calculation and Parameter Upload Phase. Each node calculates its twin fidelity score for the current moment. Specifically, the system compares the predicted output of the virtual model with the actual output of the physical entity within a preset period and calculates a preset deviation metric. The adaptive aggregation module determines the node's contribution to global aggregation based on the reciprocal of this error. Subsequently, the node uploads its local model parameters and corresponding fidelity score to the cross-domain resource scheduling layer.

[0044] Step 5: Cross-Domain Resource Cooperative Scheduling Phase. The cross-domain resource scheduling layer senses the real-time channel status, remaining power, and computing load of each node. Using a preset optimization algorithm, it dynamically allocates the bandwidth, computing frequency, and transmit power required for the next round of training, and decides whether to offload some high-complexity computing tasks to the cloud processing center to optimize the overall system response latency and energy efficiency.

[0045] Step Six: Global Model Aggregation and Distribution Phase. The system's central server receives parameters uploaded from each node and performs global aggregation according to adaptive weight logic. The aggregated global model is then fed back to each node as the base model for the next round of training.

[0046] Step Seven: Closed-Loop Optimization and Adaptive Adjustment Phase. The system monitors the convergence of the global model and the synchronization accuracy of the digital twin in real time. If model drift is detected due to non-independent and identically distributed data, knowledge distillation and alignment are performed on the local model using simulation data generated from the digital twin model. This process continues iteratively until the global model accuracy reaches the preset target and the system operation stabilizes.

[0047] Furthermore, in step three, to address the challenges posed by the non-independent and identically distributed (Non-IID) nature of industrial data, this invention employs a local alignment strategy based on digital twin enhancement. The logic is as follows:

[0048]

[0049] In the above formula, Indicates the learning rate; This represents the proximal term coefficient, used to limit the degree to which the local model deviates from the global model; This represents the step size factor of the physical constraints. By introducing a global model bias term and a physical constraint gradient term, model divergence caused by differences in local operating conditions can be effectively suppressed.

[0050] Furthermore, in step five, the cross-domain resource scheduling layer employs the Lagrange multiplier method to handle the constrained optimization problem. The Lagrange function is constructed as follows:

[0051]

[0052] In the above formula, Represents the Lagrange multiplier vector; This represents the system's preset maximum allowable latency threshold. Optimal resource allocation under latency constraints is achieved through iterative optimization of variables and multipliers.

[0053] Furthermore, in step seven, when a sudden change in the physical environment of a specific node is detected (such as parameter drift caused by increased equipment wear), the digital twin mapping layer adjusts the parameters of its internal mechanistic evolution model. and This captures the non-stationary evolution characteristics of physical entities in real time. Subsequently, this change is processed through a physical consistency loss term. Rapid feedback to the federated learning process enables the model to quickly adapt to new physical conditions, thereby eliminating the lag between physical evolution and statistical learning.

[0054] The beneficial effects of this invention are:

[0055] First, it achieves a deep integration of physical mechanisms and data-driven approaches. By introducing physical consistency constraints provided by digital twins into the loss function of federated learning, this invention overcomes the shortcomings of traditional machine learning models in terms of lack of physical interpretability and robustness in harsh industrial environments. Even in the presence of noise interference and sensor failures, the model constructed in this invention can consistently maintain consistency with physical laws, significantly improving the reliability of control decisions.

[0056] Second, it effectively solves the model drift problem caused by the non-independent and identically distributed nature of industrial data. Utilizing the precise characterization of the physical state of each node through a digital twin model, this invention designs an adaptive aggregation mechanism based on twin fidelity. This mechanism can automatically identify and reduce the impact of low-quality, high-biased nodes on the global model, accelerating the convergence speed of federated learning and achieving higher global prediction accuracy under heterogeneous operating conditions.

[0057] Third, it significantly improves the system's operating efficiency in resource-constrained environments. Through joint optimization of computing, communication, and energy resources via a cross-domain resource scheduling layer, this invention can dynamically balance the load of each node while meeting industrial real-time requirements. In particular, the introduction of Lyapunov optimization and the Lagrange multiplier method enables the system to adaptively respond to dynamically changing channel environments and computing demands, significantly reducing the overall energy consumption of the system.

[0058] Fourth, it safeguards the privacy and security of industrial data. This invention adheres to a federated learning architecture of "data remains stationary while the model moves," combined with the virtual mapping capabilities of digital twins, enabling enterprises or factories to achieve collaborative evolution without sharing sensitive production process data. This not only meets data compliance requirements but also provides a secure technological foundation for building a cross-industry industrial internet ecosystem.

[0059] Fifth, it enhances the self-adaptive and self-evolving capabilities of industrial systems. The real-time synchronization and mechanism prediction functions provided by the digital twin mapping layer enable the system to have a keen perception of physical entity degradation, environmental disturbances, and task switching. By transforming this perception into dynamic adjustments to learning strategies, the collaborative optimization system implemented in this invention can autonomously update itself as the physical environment evolves, achieving truly intelligent closed-loop management. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall structure of the system of the present invention;

[0061] Figure 2 This is a flowchart illustrating the method of the present invention.

[0062] The attached diagram is labeled as follows: 1. Physical Entity Layer; 2. Digital Twin Mapping Layer; 3. Distributed Federated Learning Layer; 4. Cross-Domain Resource Scheduling Layer; 11. Sensing Module; 12. Execution Module; 13. Edge Computing Module; 21. State Observer; 22. Mechanism Evolution Model; 23. Real-Time Synchronization Unit; 31. Local Training Module; 32. Physical Constraint Alignment Module; 33. Adaptive Aggregation Module; 41. Task Offloading Unit; 42. Bandwidth Allocation Unit; 43. Energy Efficiency Management Unit. Detailed Implementation

[0063] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0064] like Figure 1 As shown, the overall architecture of this system consists of four layers: Physical Entity Layer 1, Digital Twin Mapping Layer 2, Distributed Federated Learning Layer 3, and Cross-Domain Resource Scheduling Layer 4. These layers interact with each other through standardized interfaces for data exchange and control command transmission, forming a closed-loop collaborative optimization system. Physical Entity Layer 1 is deployed in actual industrial settings and includes multiple heterogeneous industrial IoT nodes. Each node integrates a sensing module 11, an execution module 12, and an edge computing module 13. The sensing module 11 is equipped with a high-precision multimodal sensor array, including but not limited to vibration accelerometers (range ±50g, sampling rate 10kHz), temperature thermocouples (accuracy ±0.5℃), pressure transmitters (range 0–10MPa, resolution 0.01MPa), and process vision cameras (resolution 2048×2048, frame rate 30fps), used for real-time acquisition of multi-source heterogeneous data such as equipment operating parameters, environmental state variables, and process flow images. The execution module 12 consists of a programmable logic controller (PLC) and a servo driver. After receiving control commands from the upper layer, it adjusts the physical production process in milliseconds by regulating motor speed, valve opening, or heating power. The edge computing module 13 uses an embedded computing unit based on the ARM Cortex-A72 architecture with a main frequency of 2.0GHz, equipped with 4GB LPDDR4 memory and 64GB eMMC storage, and supports the TensorFlow Lite Micro inference framework, providing the necessary computing and storage resources for local data preprocessing, digital twin synchronization, and federated learning training.

[0065] The digital twin mapping layer 2 is built on the edge computing module 13 or a nearby edge server. Its core function is to establish a high-fidelity virtual model corresponding one-to-one with each industrial node in the physical entity layer 1. This layer includes a state observer 21, a mechanism evolution model 22, and a real-time synchronization unit 23. The state observer 21 first performs timestamp alignment, outlier removal, and missing value interpolation on the received raw sensor data. Specifically, for vibration signals, wavelet threshold denoising is used to eliminate high-frequency electromagnetic interference; for temperature sequences, cubic spline interpolation is used to fill gaps caused by communication packet loss; and for visual images, the contours of moving targets are extracted using background subtraction. Subsequently, the state observer 21 extracts key features from the cleaned data to construct the real-time state vector of the physical entity. In a specific embodiment, for a CNC machine tool spindle, its state vector is defined as follows: ,in Main spindle speed (rpm). Bearing temperature (°C). The radial vibration acceleration is (m / s²). The main axis angular displacement (rad).

[0066] The state-space model constructed by state observer 21 follows the following state transition logic:

[0067]

[0068] in, It is a 4×4 state transition matrix, whose elements are obtained offline from historical operating data using system identification methods (such as the N4SID subspace identification algorithm). For example, in the application of a certain machining center, Typical values ​​are:

[0069]

[0070] It is a 4×2 input matrix, corresponding to the control input vector. , representing the speed setting value and the coolant flow rate setting value, respectively. Let Gaussian white noise be a vector with zero mean, and its covariance matrix be... The maximum likelihood estimation method is used to simulate the impact of random factors such as cutting force fluctuations and grid voltage disturbances on the system state.

[0071] Mechanism evolution model 22 constructs a system of differential equations based on classical physical laws to predict the future state of physical entities. Taking the thermal deformation of the principal shaft as an example, its heat conduction equation can be expressed as:

[0072]

[0073] in For material density, For specific heat capacity, Thermal conductivity, This is the term for frictional heat generation. After being discretized using the finite element method, this partial differential equation is embedded into the mechanism evolution model 22, enabling it to predict the temperature field distribution and the resulting geometric deformation within the next 5–10 seconds.

[0074] The real-time synchronization unit 23 is responsible for maintaining the consistency between the virtual model and the physical entity. Its working principle is to dynamically update the virtual model parameters by minimizing the residual between the observed data and the model predictions. The synchronization criterion satisfies the following optimization problem:

[0075]

[0076] In this formula, These are actual observation vectors, such as direct measured values ​​including spindle end face runout and tool wear. The observation mapping function, typically a nonlinear neural network, is used to map the internal states. Mapped to the observable output space; The regularization coefficient is set to 0.01 in a preferred embodiment to prevent the parameters from oscillating drastically due to instantaneous sensor noise. This optimization problem is solved online using stochastic gradient descent, with parameter updates performed every 100ms.

[0077] The distributed federated learning layer 3 is built on top of the digital twin mapping layer 2. Its core components include a local training module 31, a physical constraint alignment module 32, and an adaptive aggregation module 33. The local training module 31 independently runs a lightweight neural network model on each node, such as a two-layer fully connected network (128-dimensional input layer, 64-dimensional hidden layer, and 10-dimensional output layer), for tasks such as device fault prediction, energy efficiency assessment, or quality classification. The training data comes entirely from the historical running records cached by the local edge computing module 13, without involving any external transmission of raw data.

[0078] The physical constraint alignment module 32 transforms the physical mechanism information provided by the digital twin mapping layer 2 into mathematical constraints and integrates them into the local training process. Its core lies in constructing an improved loss function:

[0079]

[0080] in, The loss can be either standard cross-entropy loss or mean squared error loss, depending on the specific task type; The physical constraint weighting factor is determined to be 0.5 in this embodiment through grid search; The physical consistency loss is specifically constructed as follows:

[0081]

[0082] Here, For digital twin models in The predicted state provided at any given moment; Operators for the differential equations obeyed by a physical system. For example, if the physical system obeys Newton's second law. ,but This forces the output acceleration of the neural network to match the force analysis results. During the training process, It is estimated from a continuous sequence of states by numerical differentiation (such as the five-point stencil formula).

[0083] The adaptive aggregation module 33, located on the system's central server, is responsible for integrating local model parameters uploaded from various nodes. Its innovation lies in the introduction of twin fidelity scoring. As aggregate weights, the global model parameter update logic is as follows:

[0084]

[0085] Twin fidelity rating Defined as the reciprocal of the average mapping deviation within a preset time window (e.g., the most recent 100 synchronization cycles). The mapping deviation is calculated using a weighted Euclidean distance.

[0086]

[0087] in These are the weighting coefficients for each observation dimension, reflecting their criticality to system performance. Ultimately... This ensures that nodes with smaller deviations receive higher aggregation weights.

[0088] The cross-domain resource scheduling layer 4 includes a task offloading unit 41, a bandwidth allocation unit 42, and an energy efficiency management unit 43, which together achieve joint optimization of computing, communication, and energy resources. This layer makes decisions by solving the following multi-objective optimization problem:

[0089]

[0090] in, The energy efficiency adjustment coefficient is set to 0.1 in scenarios with stringent real-time requirements (such as high-speed stamping), and to 10 in scenarios where energy efficiency is prioritized (such as warehousing and logistics). Total time consumed. It consists of computation delay and communication delay:

[0091]

[0092] Here, This represents the number of computation cycles required to process a single sample. For the ResNet-18 model, the measured number of cycles on a Cortex-A72 is approximately 2.5 × 10⁻⁶. 9 cycles / sample; This is the number of local training samples, usually set to 512; The model parameter data size is approximately 50KB for the two-layer network described above. The total available bandwidth is 100MHz in a 5G private network; The channel gain is obtained through channel state information (CSI) feedback. The noise power is set to -174dBm / Hz.

[0093] Energy Efficiency Management Unit 43 further introduces the Lyapunov optimization framework to handle dynamic stochasticity. It defines the task queue backlog. The Lyapunov function is given by the number of twin synchronization and federated training tasks to be processed:

[0094]

[0095] By minimizing the drift penalty term To ensure queue stability (i.e.) Optimize energy efficiency under the premise of ).

[0096] Based on the above system architecture, the collaborative optimization method of the present invention is implemented according to the following steps:

[0097] Step 1: Initialization Phase. The system control center generates the initial global model. (For example, randomly initializing or loading pre-trained weights) and distributing them to all N=20 industrial nodes via a secure channel. Simultaneously, each node loads the corresponding mechanism model template from its local knowledge base based on the equipment model and initializes it. Parameters such as these.

[0098] Step Two: Virtual-Real Synchronization and Sensing Phase. Sensing module 11 acquires data at a frequency of 100Hz, and state observer 21 generates a state vector every 10ms. The real-time synchronization unit 23 continuously monitors the mapping deviation. When the deviation exceeds the threshold for five consecutive times, it will detect the deviation. When the normalized mean squared error (MSE) is reached, the model reconstruction mechanism is triggered: the current parameters are frozen, and the online system identification program is started to re-estimate the parameters. and .

[0099] Step 3: Local Training Phase under Physical Constraints. At the start of each round of federated learning, each node extracts K=512 samples from its local database. The local training module (31) calculates the data loss. The gradient. The physical constraint alignment module 32 obtains the gradient of the most recent 10 time steps from the digital twin mapping layer 2. Sequence, calculation And its gradient. Finally, the local model parameters are updated according to the following rules:

[0100]

[0101] Among them, learning rate =0.01, coefficient of proximal term =0.1 (used to mitigate the Non-IID problem), physics step size factor =0.5. This update process is performed on the edge device for 10 local epochs.

[0102] Step 4: Adaptive Weight Calculation and Parameter Upload Phase. Each node calculates the average mapping deviation over the past 100 synchronization cycles, and obtains... Subsequently, ( , The data is uploaded to the central server via the 5G uplink. Before uploading, the task offloading unit 41 decides whether to offload some computations (such as gradient compression) to the MEC server based on the current channel conditions.

[0103] Step 5: Cross-domain resource collaborative scheduling phase. The cross-domain resource scheduling layer 4 collects data from each node. Remaining battery power and channel gain Construct the Lagrange function:

[0104]

[0105] in The maximum allowable delay is determined. The optimal resource allocation scheme is obtained by solving this constrained optimization problem using the Alternating Direction Multiplier Method (ADMM). And distribute it to each node.

[0106] Step Six: Global Model Aggregation and Distribution Phase. After receiving parameters from all valid nodes, the central server aggregates and generates the model according to the adaptive weight formula. And it is distributed to the entire network via multicast.

[0107] Step 7: Closed-Loop Optimization and Adaptive Adjustment Phase. The system continuously monitors the accuracy of the global model on the validation set and the twin synchronization error of each node. If the mapping deviation of a certain node (e.g., number #7) continuously increases within 30 minutes (slope > 0.001 / minute), it is determined to be equipment degradation (e.g., bearing wear). At this time, the digital twin mapping layer 2 automatically adjusts its... Damping terms in the matrix (e.g., reducing the [3,3] element from 0.90 to 0.85), and passing this change through... This allows for rapid propagation to the federated learning process, enabling the model to adaptively adjust within 2–3 rounds.

[0108] To verify the effectiveness of the present invention, the following embodiments were designed and compared with comparative examples for testing.

[0109] Example

[0110] The system and method of this invention are communicatively connected to a management platform and a sensor network platform. The sensor network platform serves as the data transmission hub between the system and the management platform, responsible for high-speed, low-latency data transmission between them, including control, input, management, and route entry. It also includes displaying task status, results, and data, and allowing users to view these information. The system is deployed on 20 robot workstations in an automotive welding workshop, each equipped with the system described herein. The task is to predict weld quality (pass / fail). The local model is a two-layer MLP, with input features including 16-dimensional parameters such as current, voltage, pressure, and time. The digital twin model includes robot dynamics equations and a heat conduction model. Federated learning uses 512 samples for local training in each round, for a total of 20 rounds. The resource scheduling objective is to... Minimize energy consumption.

[0111] Comparative Example 1:

[0112] Traditional Federated Learning (FedAvg) has no digital twin constraint, and the other conditions are the same as in the implementation example.

[0113] Comparative Example 2:

[0114] Digital twin + federated learning, but using uniformly aggregated weights (i.e. ), lacks an adaptive mechanism.

[0115] Comparative Example 3:

[0116] This invention relates to a system that disables cross-domain resource scheduling and uses fixed allocation. .

[0117] The test results are shown in the table below:

[0118] index Example Comparative Example 1 Comparative Example 2 Comparative Example 3 Global model accuracy (%) 96.8 89.2 93.5 95.1 Number of rounds required for convergence 14 28 19 16 Mean synchronization deviation (NMSE) 0.021 0.087 0.035 0.025 Average energy consumption per wheel (J) 186 210 198 245 Task timeout rate (%) 0.3 2.1 0.8 4.7

[0119] Experimental data show that the present invention significantly outperforms the comparative examples in terms of accuracy, convergence speed, synchronization accuracy, and energy efficiency. Particularly in the presence of severe Non-IID data (different robots welding different vehicle models under varying working conditions), the physical constraints and adaptive aggregation mechanism effectively suppress model drift.

[0120] Furthermore, in the embodiment, when the wear of the joint reducer of simulated robot #7 (leading to an increase in positioning error of 0.2 mm), the system detected a sudden increase in its twin deviation (from 0.02 to 0.06) in the 12th round, and then adjusted the mechanistic model parameters, restoring the global accuracy to over 96% in the 14th round. In contrast, the accuracy of Comparative Example 1 continued to drop below 85% under these conditions, verifying the adaptive capability of the present invention.

[0121] In summary, this invention, by deeply integrating the physical mechanism modeling capabilities of digital twins with the distributed collaborative advantages of federated learning, constructs a high-fidelity, highly robust, and energy-efficient collaborative optimization system for the Industrial Internet of Things (IIoT). Its technical solution has undergone end-to-end verification on hardware platforms (NVIDIA Jetson AGX Orin + 5G CPE), software stacks (ROS 2 + PySyft + FMI / FMU), and industrial protocols (OPC UA over TSN), demonstrating clear engineering feasibility and industrial application value.

[0122] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. An industrial IoT optimization system integrating digital twins and federated learning, characterized in that, The system includes a physical entity layer (1), a digital twin mapping layer (2), a distributed federated learning layer (3), and a cross-domain resource scheduling layer (4). The physical entity layer (1) is composed of multiple heterogeneous industrial IoT nodes. Each heterogeneous industrial IoT node integrates a sensing module (11), an execution module (12), and an edge computing module (13). The sensing module (11) is used to collect multi-source heterogeneous data of the physical entity in real time. The execution module (12) is used to adjust the physical production process according to control instructions. The edge computing module (13) is used to provide computing and storage resources required for local data processing and model training. The digital twin mapping layer (2) establishes a virtual model corresponding to the physical entity layer (1), including a state observer (21), a mechanism evolution model (22), and a real-time synchronization unit (23). The state observer (21) is used to preprocess the collected data and construct the real-time state vector of the physical entity. The mechanistic evolution model (22) predicts the future state of physical entities based on physical laws, and the real-time synchronization unit (23) dynamically updates the virtual model parameters by minimizing the residual between the observed data and the model prediction. ; The distributed federated learning layer (3) is built on top of the digital twin mapping layer (2) and includes a local training module (31), a physical constraint alignment module (32), and an adaptive aggregation module (33). The local training module (31) uses local data to train the model. The physical constraint alignment module (32) transforms the physical mechanism information provided by the digital twin into constraint terms in the loss function. The adaptive aggregation module (33) scores the twin fidelity of each node. Dynamically adjust aggregation weights to generate a global model; The cross-domain resource scheduling layer (4) includes a task offloading unit (41), a bandwidth allocation unit (42), and an energy efficiency management unit (43), which are used to jointly optimize computing, communication and energy resources in order to minimize the total energy consumption of the system while meeting the time delay constraints.

2. The industrial IoT optimization system integrating digital twins and federated learning according to claim 1, characterized in that: The state space model constructed by the state observer (21) satisfies the following state transition equation: in, Indicates the first A physical entity in The state vector at any given time; This is the state transition matrix, reflecting the evolution of the physical entity's state over time. for The control input vector at each time step; Given an input matrix, define the weights that control the influence of the input on the state; This is the process noise vector, used to simulate random environmental disturbances.

3. The industrial IoT optimization system integrating digital twins and federated learning according to claim 1, characterized in that: The real-time synchronization unit (23) updates the virtual model parameters by solving the following optimization problem. : in, for The actual observation vector at time; This is a nonlinear observation mapping function that maps virtual states to the observation space; This is the regularization coefficient, used to control the smoothness of parameter updates.

4. The industrial IoT optimization system integrating digital twins and federated learning according to claim 1, characterized in that: The local training total loss function constructed by the physical constraint alignment module (32) is: in, For the first Local model parameters for each node; For local datasets Experience risk loss; This represents a loss of physical consistency. This is the physical constraint weighting factor.

5. The industrial IoT optimization system integrating digital twins and federated learning according to claim 4, characterized in that: The physical consistency loss Defined as: in, Predicted states provided for digital twin models; Operators for the differential equations followed by a physical system; To predict the derivative of the state with respect to time, it is estimated using numerical differentiation methods.

6. The industrial IoT optimization system integrating digital twins and federated learning according to claim 1, characterized in that: The adaptive aggregation module (33) generates global model parameters. The aggregation logic is as follows: in, The total number of nodes participating in collaborative training; For the first The twin fidelity score of each node is the reciprocal of the average mapping deviation within a preset time window, which is calculated using weighted Euclidean distance.

7. The industrial IoT optimization system integrating digital twins and federated learning according to claim 1, characterized in that: The cross-domain resource scheduling layer (4) determines the resource allocation strategy by solving the following multi-objective optimization problem: in, For the first The computation frequency of each node; This refers to the transmission power. The proportion of communication bandwidth allocated; The total time required to complete the collaborative task; Total energy consumption; This is the energy efficiency adjustment coefficient.

8. The industrial IoT optimization system integrating digital twins and federated learning according to claim 7, characterized in that: The total time It consists of computation delay and communication delay, and satisfies: in, The number of computation cycles required to process a unit sample; This represents the total number of local training samples. The amount of model parameter data to be transmitted; Total available bandwidth; Channel gain; This represents the power of Gaussian white noise.

9. The industrial IoT optimization system integrating digital twins and federated learning according to claim 1, characterized in that: The energy efficiency management unit (43) introduces the Lyapunov optimization framework and defines the Lyapunov function as: in For the first Each node The task queue backlog length at any given time; by minimizing the drift penalty term, the system energy efficiency is optimized while ensuring queue stability, and the maximum delay constraint is handled by combining the Lagrange multiplier method. Construct the Lagrange function: In the formula, This represents the Lagrange multiplier vector.

10. An optimization method for an industrial IoT optimization system integrating digital twins and federated learning as described in any one of claims 1-9, characterized in that, Includes the following steps: Step 1: Initialization Phase: The system control center initializes the global model parameters and distributes them to each heterogeneous industrial IoT node. At the same time, each node starts the digital twin mapping program to establish an initial virtual entity based on the preset mechanism model parameters. Step 2: Virtual-Real Synchronization and Perception Stage: Sensors in the physical entity layer collect production data in real time. The digital twin mapping layer uses a state observer to clean and extract features from the data, and uses a real-time synchronization unit to update the virtual model parameters. If the mapping deviation between the virtual model and the physical entity exceeds a preset threshold, a model reconstruction mechanism is triggered to ensure the fidelity of the virtual space. Step 3: Local training phase under physical constraints: Each node starts a distributed federated learning layer in the local edge computing module. The local training module uses real-time collected physical data to perform gradient calculation. The physical constraint alignment module obtains the current physical evolution law from the digital twin mapping layer and introduces it as a penalty term into the loss function. By solving the optimization problem with physical constraints, the local model parameters are updated to ensure that the model not only fits the data distribution but also conforms to the physical mechanism. Step 4: Adaptive weight calculation and parameter upload stage: Each node calculates the twin fidelity score at the current moment. Specifically, the system compares the predicted output of the virtual model with the actual output of the physical entity within a preset period and calculates its preset deviation metric. The adaptive aggregation module determines the node's contribution to the global aggregation based on the reciprocal of this error. Subsequently, the node uploads the local model parameters and the corresponding fidelity score to the cross-domain resource scheduling layer. Step 5: Cross-domain resource collaborative scheduling phase. The cross-domain resource scheduling layer senses the real-time channel status, remaining power, and computing load of each node. Using a preset optimization algorithm, it dynamically allocates the bandwidth, computing frequency, and transmission power required for the next round of training, and decides whether to offload some high-complexity computing tasks to the cloud processing center to optimize the overall system response latency and energy efficiency. Step Six: Global Model Aggregation and Distribution Phase. The system's central server receives the parameters uploaded by each node, performs global aggregation according to adaptive weight logic, and feeds back the aggregated global model to each node as the base model for the next round of training. Step 7: Closed-loop optimization and adaptive adjustment stage. The system monitors the convergence of the global model and the synchronization accuracy of the digital twin in real time. If model drift is detected due to non-independent and identically distributed data, the local model is aligned with the simulation data generated by the digital twin model through knowledge distillation. This process continues to iterate until the global model accuracy reaches the preset target and the system operation status tends to be stable.

Images