CT induction power supply system based on cooperative control and distributed optimization and control method

Abstract

The invention relates to the technical field of CT induction power supplies, in particular to a CT induction power supply system based on cooperative control and distributed optimization and a control method. According to the technical scheme, the CT induction power supply control method based on cooperative control and distributed optimization comprises the following steps: S1, deploying a multi-mode sensing array, and synchronously collecting a three-phase current harmonic spectrum of a power supply node, an iron core magnetostriction vibration waveform and a radiator surface infrared thermal image; generating a digital twinborn body with space-time relevance through multi-physics field coupling modeling; and S2, constructing a data fusion engine driven by a space-time attention mechanism at an edge computing node, and carrying out LSTM feature extraction on the output of the step S1 and a historical operation log under physical constraint. According to the method, through collaborative innovation of multi-mode federated learning and physical embedded edge calculation, a multi-physical field dynamic modeling mechanism under privacy protection is constructed, and the correlation precision and information security of electromagnetic-vibration-thermodynamic data are effectively balanced.

Description

Technical Field

[0001] The present invention relates to the technical field of CT induction power supplies, and particularly to a CT induction power supply system and a control method based on cooperative control and distributed optimization. Background Art

[0002] A CT induction power supply is a device that, based on the principle of electromagnetic induction, uses a current transformer (CT for short) to induce electrical energy from a power transmission wire and provides a stable and reliable power supply for various intelligent electronic devices.

[0003] In the prior art, there is a contradiction between the external multi-physical field coupling modeling and real-time optimization of the CT induction power supply system. Traditional methods cannot achieve the dynamic correlation analysis of electromagnetic-vibration-thermal multi-source heterogeneous data on the premise of ensuring device data privacy (such as the risk of gradient leakage), and lack the ability to embed physical laws such as electromagnetic-thermal coupling equations into the edge computing framework, resulting in a lag in the adjustment of energy entropy weight behind the power grid transient process.

[0004] In addition, in the face of external disturbances such as sudden changes in the grid topology impedance, the existing control architecture has problems of centralized strategy lag and cross-domain response mismatch; on the one hand, the blockchain consensus delay restricts the millisecond-level synchronous issuance of transient instructions; on the other hand, traditional PID parameters and anti-resonant compensation networks are difficult to dynamically cooperate, and the real-time adaptive tuning of compensation gain cannot be achieved through reinforcement learning, resulting in insufficient anti-disturbance ability of the system.

[0005] The patent with the publication number CN110034589B discloses an inductive energy harvesting AC-DC combined power supply system and a control method, including a main circuit and a power supply control and energy management system; the main circuit includes an AC power acquisition circuit, a main power supply circuit, a charging circuit, a backup power supply circuit, and a support capacitor C1; the positive terminal of the support capacitor C1 is connected to the rectifier bus; the power supply control and energy management system includes a DC-DC main controller and a rectifier bus voltage sampling circuit; the rectifier bus voltage can reflect the energy balance relationship between the AC input power of the power acquisition CT and the power consumption of the load and the power supply system; the DC-DC main controller can judge the energy balance relationship according to the rectifier bus voltage, select corresponding control strategies according to different energy balance relationships, and generate control signals that can meet the corresponding control strategies. It also discloses an AC-DC combined power supply method, which uses the inductive energy harvesting AC-DC combined power supply system of the present invention and performs zone control according to the rectifier bus voltage, but it does not involve the dynamic correlation analysis of multi-source heterogeneous data and has limitations in use. Summary of the Invention

[0006] The present invention proposes a CT induction power supply system and control method based on cooperative control and distributed optimization, which solves the contradiction between the external multi-physical field coupling modeling and real-time optimization of the CT induction power supply system in the prior art. The traditional method cannot realize the dynamic correlation analysis of electromagnetic-vibration-thermal multi-source heterogeneous data on the premise of ensuring the device data privacy.

[0007] To solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0008] A CT induction power supply control method based on cooperative control and distributed optimization includes:

[0009] Step S1: Deploy a multi-modal sensing array to synchronously collect the three-phase current harmonic spectrum of the power supply node, the core magnetostrictive vibration waveform, and the infrared thermal imaging on the surface of the radiator. Generate a digital twin with spatio-temporal correlation through multi-physical field coupling modeling, where the local model parameter encryption sharing of distributed nodes is realized by using a federated learning framework;

[0010] Step S2: Build a data fusion engine driven by a spatio-temporal attention mechanism at the edge computing node. Perform LSTM feature extraction under physical constraints on the output of Step S1 and the historical operation log to generate a health assessment matrix containing load fluctuation sensitive factors, and dynamically correct the energy entropy weight coefficient through a quantum hybrid optimization algorithm;

[0011] Step S3: Based on the health assessment matrix in Step S2, construct a multi-party game optimization model, define power supply stability, equipment life balance degree, and minimum reactive power loss as game objectives, use an improved NSGA-III algorithm to solve the Pareto optimal strategy set, and introduce a dynamic reward and punishment mechanism to realize the autonomous evolution of the strategy space;

[0012] Step S4: Execute policy consensus verification through a heterogeneous blockchain architecture. The main chain deploys a global policy verification based on Proof of Contribution (PoC), and the side chain implements a local policy fast synchronization with a Verifiable Delay Function (VDF). Use homomorphic encryption technology to achieve data privacy protection during cross-chain verification;

[0013] Step S5: Based on the consensus result in Step S4, construct an adaptive control law driven by a Lyapunov energy function. By real-time monitoring the change of the singular value of the topological impedance matrix, dynamically switch the anti-resonant compensation network parameters, and fuse a deep reinforcement learning algorithm to online optimize the gain scheduling strategy of the PID controller.

[0014] Further, the deployment of the multi-modal sensing array in Step S1 specifically includes:

[0015] Set Rogowski coil sensors with a bandwidth ≥ 100 kHz in the three-phase current acquisition channels, and set the sampling rate to 512 points per cycle;

[0016] The magnetostrictive vibration sensor uses an optical fiber Bragg grating array, which is arranged at the four corners and the center of the iron core, and the sampling frequency is 20 kHz;

[0017] The infrared thermal imager is equipped with a dual-band temperature measurement module, 3-5 μm and 8-12 μm, and the spatial resolution is not less than 640×480 pixels.

[0018] Further, the encrypted sharing of the federated learning framework in step S1 includes:

[0019] Gradient noise injection: When training the local model, differential privacy noise satisfying ε = 0.5 and δ = 1e-5, Gaussian noise is added to the backpropagation gradient where g is the original gradient vector, N(0,σ 2 I) is Gaussian noise with a mean of 0 and a variance of σ 2 and is the standard deviation of the noise;

[0020] Parameter aggregation protocol: The Paillier homomorphic encryption algorithm is adopted, and the key length is set to 2048 bits to realize the secure weighted average of the global model parameters;

[0021] Trusted verification mechanism: Verify the integrity of the TEE environment through remote attestation technology to ensure memory encryption and process isolation during model update.

[0022] Further, the LSTM feature extraction under physical constraints in step S2 is specifically as follows:

[0023] Construct a physical knowledge module containing electromagnetic-thermal coupling equations, and embed the Maxwell equations as constraint conditions into the LSTM network through the Lagrange multiplier method;

[0024] Design an improved LSTM unit with a time-varying forget gate, and its forget gate coefficient is dynamically adjusted by the temperature gradient of the radiator;

[0025] Introduce spectral normalization processing in the feature output layer to limit the L2 norm of the health assessment matrix not to exceed a preset threshold.

[0026] Further, the dynamic correction process of the quantum hybrid optimization algorithm in step S2 includes:

[0027] Initialization stage: Map the energy entropy weight coefficient to the superposition state of qubits, and each weight corresponds to the entangled state of 3 qubits;

[0028] Iterative update: Introduce the quantum rotation gate operation into the velocity update equation of the classical particle swarm, and the rotation angle is determined by the difference between the current particle fitness value and the global optimal value;

[0029] Constraint handling: When the fitness change rate for 10 consecutive iterations is lower than 0.1%, start the quantum annealing process to accelerate convergence.

[0030] Furthermore, the optimization strategy of the improved NSGA-III algorithm in step S3 includes:

[0031] Reference point dynamic adjustment: Calculate the reference point offset according to the Mahalanobis distance of the population distribution, and the offset step size is proportional to the curvature of the Pareto front;

[0032] Chaotic initialization: Use the Logistic chaotic map to generate the initial population, with the chaotic parameter μ = 4.0 and the initial value

[0033] X0 = 0.31;

[0034] Elite screening: Calculate the Shannon entropy value of the solutions in the reserve pool, and preferentially retain the solutions with an entropy value higher than 0.8 to maintain diversity.

[0035] Furthermore, the PoC mechanism for proving the contribution degree of the main chain in step S4 specifically includes:

[0036] The contribution degree evaluation index consists of the following parameters:

[0037] Historical verification accuracy rate, with a weight of 40%, denoted as Hi;

[0038] Computing resource input amount, with a weight of 30%, denoted as Ri;

[0039] Network latency performance, with a weight of 30%, denoted as Li;

[0040] The consensus weight calculation formula is:

[0041] wi = (0.4Hi + 0.3Ri) / (0.3Li)

[0042] Where Hi, Ri, and Li respectively represent the historical verification accuracy rate, computing resource input amount, and network latency value of node i, and wi represents the consensus weight value of the i-th stage;

[0043] Adopt the BLS threshold signature scheme, requiring at least 7 signatures out of 11 nodes to pass the verification.

[0044] Furthermore, the triggering conditions for singular value monitoring in step S5 include:

[0045] Set the dynamic threshold κ_max = 100×N, where N is the number of currently running nodes, and trigger compensation when the condition number of the impedance matrix exceeds κ_max;

[0046] The singular value spectrum entropy calculation adopts a sliding window mechanism, and the window length is set to 10 power frequency cycles;

[0047] Stability margin assessment: When the real part of the Lyapunov exponent is greater than 0 for 5 consecutive iterations, start the emergency control strategy.

[0048] Furthermore, the optimization of the PID gain scheduling strategy in step S5 specifically includes: constructing a Twin Delayed Deep Deterministic Policy Gradient (TD3) framework, setting the learning rate of the actor network to 3×10-4, and the learning rate of the critic network to 1×10-3; the reward function is designed as follows: R = 0.5(1 - Vd / Vmax) + 0.3(1 - Es / Emax) - 0.2|Ct - Ct_prev|;

[0049] Where: Vd: the real-time deviation value of the output voltage, Vmax is the maximum allowable voltage deviation; Es: the system steady-state error value, Emax is the maximum allowable steady-state error; Ct, Ct_prev: the control quantity output values at the current moment and the previous moment; R represents the system dynamic performance evaluation index; amplitude constraints need to be imposed on the PID gain parameters in the exploration strategy.

[0050] Furthermore, a CT induction power supply system includes:

[0051] Multimodal sensing module: integrating three-phase current, vibration, and infrared sensing units, and configuring a synchronous data acquisition architecture;

[0052] Edge computing module: built-in encrypted computing environment to achieve fast iteration of local models;

[0053] Intelligent decision-making module: integrating a physical constraint prediction model and a multi-objective optimization algorithm to support dynamic resource scheduling;

[0054] Blockchain network module: adopting a main-side chain architecture and a hardware-accelerated consensus mechanism to ensure cross-chain interaction efficiency;

[0055] Dynamic control module: including an adaptive compensation network and a reinforcement learning controller, and configuring high-frequency power devices and a real-time control interface;

[0056] Digital twin module: realizing synchronous interaction between the physical system and the virtual model and multi-field coupling visualization.

[0057] The positive effects of the present invention are as follows: Through the collaborative innovation of multimodal federated learning and physics-embedded edge computing, breakthroughs are achieved in the data perception and optimization levels: constructing a multi-physical field dynamic modeling mechanism under privacy protection, effectively balancing the correlation accuracy and information security of electromagnetic-vibration-thermal data; introducing a weight regulation system driven by quantum optimization, significantly improving the real-time matching ability of energy entropy adjustment and the power grid transient process; the feature extraction module constrained by physical laws greatly reduces the computational complexity, supporting the efficient generation of multi-objective optimization strategies on the edge side.

[0058] The innovation of the cross - domain control architecture provides the core guarantee for the system's anti - disturbance and adaptive capabilities: optimizing the consensus mechanism to achieve millisecond - level synchronous distribution of control instructions, ensuring rapid response during the power grid's transient process; integrating reinforcement learning and anti - resonance compensation network to break through the dynamic adaptation bottleneck of traditional PID parameters in scenarios of impedance mutation and realizing intelligent tuning of compensation gain; the digital - twin - driven strategy evolution mechanism significantly enhances the system's stability under complex disturbances. Brief Description of the Drawings

[0059] Figure 1 This is the schematic block diagram of the CT induction power supply system in the present invention. Detailed Embodiments

[0060] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present invention.

[0061] Embodiment 1

[0062] A CT induction power supply control method based on cooperative control and distributed optimization includes:

[0063] Step S1: Deploy a multi - modal sensing array to synchronously collect the three - phase current harmonic spectrum, core magnetostrictive vibration waveform, and infrared thermal imaging on the surface of the radiator of the power supply node, and generate a digital twin with spatio - temporal correlation through multi - physical - field coupling modeling, where a federated learning framework is used to achieve encrypted sharing of local model parameters of distributed nodes;

[0064] Step S2: Build a data fusion engine driven by a spatio - temporal attention mechanism on the edge computing node, perform LSTM feature extraction under physical constraints on the output of Step S1 and the historical operation log, generate a health assessment matrix containing load - fluctuation - sensitive factors, and dynamically correct the energy entropy weight coefficient through a quantum hybrid optimization algorithm;

[0065] Step S3: Based on the health assessment matrix in Step S2, build a multi - party game optimization model, define power supply stability, equipment - life balance, and minimum reactive power loss as game objectives, use an improved NSGA - III algorithm to solve the Pareto - optimal strategy set, and introduce a dynamic reward - punishment mechanism to achieve autonomous evolution of the strategy space;

[0066] Step S4: Execute strategy consensus verification through a heterogeneous blockchain architecture. The main chain deploys a global strategy verification based on Proof - of - Contribution (PoC), and the side chain implements fast synchronization of local strategies with a Verifiable Delay Function (VDF). Homomorphic encryption technology is used to achieve data privacy protection during cross - chain verification.

[0067] Step S5: Based on the consensus result of Step S4, construct an adaptive control law driven by the Lyapunov energy function. By real-time monitoring the change of the singular value of the topological impedance matrix, dynamically switch the anti-resonant compensation network parameters, and fuse the deep reinforcement learning algorithm to online optimize the gain scheduling strategy of the PID controller.

[0068] Multi-modal perception and modeling: Deploy high-precision sensing arrays at key nodes of the power cabinet. For three-phase current harmonic spectrum acquisition, use broadband Rogowski coils, for vibration monitoring, select distributed fiber Bragg grating arrays, and for the thermal imaging module, configure dual-band infrared detectors. After the data is preprocessed by the edge gateway, through the federated learning framework, realize the secure interaction of cross-node model parameters, and establish a digital twin model of electromagnetic-mechanical-thermal multi-field coupling.

[0069] Edge feature optimization: Deploy a physically constrained LSTM network on the edge server, conduct spatio-temporal correlation analysis on the real-time data stream and historical operation logs, and embed Maxwell's equations as hard constraints during the feature extraction process. Dynamically adjust the energy entropy coefficient through quantum hybrid optimization to ensure that the health assessment matrix accurately reflects the transient characteristics of the device.

[0070] Multi-objective decision-making: Construct a three-party game model, and define power supply quality, life balance, and energy consumption economy as optimization objectives; use the chaotic initialization NSGA-III algorithm to generate the Pareto front solution set, combine the dynamic reward and punishment mechanism to realize the online evolution of the strategy set, and maintain decision diversity through entropy value screening.

[0071] Trusted consensus: The main chain uses the PoC consensus mechanism to evaluate the contribution degree of nodes, and the side chain realizes second-level strategy synchronization through VDF; use homomorphic encryption to protect data privacy during cross-chain verification to ensure the secure and trusted transmission of control instructions among distributed nodes.

[0072] Dynamic control: Based on the Lyapunov energy function, construct an adaptive control law, and real-time monitor the singular value spectral entropy of the power grid impedance matrix. When an impedance mutation is detected, through the switching of anti-resonant network parameters and the PID gain scheduling optimized by the TD3 algorithm, realize millisecond-level dynamic compensation.

[0073] In this embodiment, specifically:

[0074] Multi-modal data acquisition

[0075] Three-phase harmonic monitoring: Deploy high-frequency Rogowski coils and high-precision analog-to-digital converters at the power supply node, and use programmable logic devices to analyze the current harmonic spectrum in real time to dynamically capture harmonic components from 3 to 50 times.

[0076] Core vibration perception: An optical fiber grating sensor array is attached to the surface of the core, and the vibration signal and temperature drift interference are separated through wavelength demodulation technology to achieve sub-millimeter vibration positioning.

[0077] Thermal imaging acquisition: A dual-band infrared camera is configured, and the environmental light reflection is eliminated by combining a dynamic filtering algorithm to generate a temperature field distribution map of the radiator surface.

[0078] Multi-physical field modeling

[0079] Establish an electromagnetic-mechanical-thermal three-field coupling model: The magnetic field change is correlated with the mechanical deformation through the magnetostrictive constitutive equation, and the Joule heat equation is introduced to describe the evolution process of the temperature field, and a digital twin with spatio-temporal correlation characteristics is constructed.

[0080] Federated learning implementation: After each node locally trains the physical field model, the homomorphic encryption technology is used to upload the model parameters to the central aggregator, and the gradient masking mechanism is used to prevent privacy leakage.

[0081] Spatio-temporal feature extraction

[0082] Construct a two-channel deep learning engine: The spatial channel uses a three-dimensional convolutional network to extract the spatial distribution characteristics of the sensor array, and the temporal channel analyzes the historical data trend through a gated recurrent unit, and the spatio-temporal characteristics are dynamically fused using attention weights.

[0083] Physical constraint embedding: The energy conservation equation is introduced as a regularization constraint in the LSTM network to ensure that the feature extraction conforms to the actual physical laws.

[0084] Hybrid optimization strategy

[0085] Quantum-classical collaborative architecture: The energy entropy weight is encoded as a quantum bit state vector, and the solution space is quickly traversed through quantum gate operations, and the classical optimizer makes fine adjustments to the continuous parameters to form a two-layer optimization loop.

[0086] Multi-objective optimization modeling

[0087] Define the three-party game players: The power supply unit takes voltage stability as the optimization goal, the device management module focuses on the life balance, and the energy efficiency controller is oriented towards minimizing the reactive power loss.

[0088] Dynamic policy library construction: The multi-objective genetic algorithm is used to generate the Pareto optimal policy set, and the inefficient policies are eliminated through the reward and punishment mechanism and the random mutation strategy is introduced to maintain the diversity.

[0089] Heterogeneous blockchain architecture

[0090] Main chain global verification: The proof-of-contribution mechanism is adopted to evaluate the effectiveness and response speed of the node's historical strategies, and only high-reputation nodes are allowed to participate in the global consensus.

[0091] Side-chain fast synchronization: Deploy verifiable delay functions in a local network to ensure that policy updates are verified by multiple nodes within a limited time, and achieve cross-chain data privacy protection through homomorphic encryption.

[0092] Real-time control optimization

[0093] Impedance monitoring and compensation: Continuously detect the impedance characteristics of the power grid topology, and automatically switch the anti-resonant compensation network configuration when the singular value exceeds the threshold.

[0094] Intelligent parameter tuning mechanism: The deep reinforcement learning module dynamically adjusts the proportional, integral, and differential coefficients of the PID controller through online trial-and-error learning to form a closed-loop optimization system.

[0095] Multimodal sensing and digital twin technology achieve comprehensive monitoring of electromagnetic, mechanical, and thermal fields, solving the blind spot problem of traditional single-dimensional detection; The combination of spatio-temporal attention mechanism and quantum hybrid optimization breaks through the convergence speed bottleneck of traditional algorithms in complex multi-objective scenarios.

[0096] The blockchain architecture ensures the consistency of policies while ensuring the secure transfer of industrial data among distributed nodes through encryption means; The integration of Lyapunov stability theory and deep reinforcement learning enables the system to maintain the optimal control state under complex working conditions such as sudden load changes. The health assessment matrix and the equipment life balance optimization model greatly extend the service life of key components and significantly reduce maintenance costs.

[0097] Through the collaborative innovation of multimodal federated learning and physically embedded edge computing, breakthroughs are achieved at the data perception and optimization levels: Construct a dynamic modeling mechanism for multi-physical fields under privacy protection to effectively balance the correlation accuracy and information security of electromagnetic-vibration-thermal data; Introduce a weight regulation system driven by quantum optimization to significantly improve the real-time matching ability of energy entropy adjustment and the power grid transient process; The feature extraction module constrained by physical laws greatly reduces the computational complexity and supports the efficient generation of multi-objective optimization strategies on the edge side.

[0098] The innovation of the cross-domain control architecture provides the core guarantee for the system's anti-interference and adaptive capabilities: Optimize the consensus mechanism to achieve millisecond-level synchronous issuance of control instructions, ensuring a rapid response during the power grid transient process; Integrate reinforcement learning and anti-resonant compensation network to break through the dynamic adaptation bottleneck of traditional PID parameters in the impedance mutation scenario and achieve intelligent tuning of compensation gain; The digital twin-driven strategy evolution mechanism significantly enhances the stability of the system under complex disturbances

[0099] Example 2

[0100] The deployment of the multimodal sensing array in step S1 specifically includes:

[0101] In the three-phase current acquisition channel, a Rogowski coil sensor with a bandwidth ≥ 100 kHz is set, and the sampling rate is set to 512 points per cycle;

[0102] The magnetostrictive vibration sensor uses a fiber Bragg grating array, which is arranged at the four corners and the center of the iron core, and the sampling frequency is 20 kHz;

[0103] The infrared thermal imager is configured with a dual-band temperature measurement module, 3 - 5 μm and 8 - 12 μm, and the spatial resolution is not less than 640 × 480 pixels.

[0104] Among them, for current acquisition: A Rogowski coil with a bandwidth of 100 kHz is installed on the three-phase busbar, and it is sampled at 512 points per cycle in cooperation with a 16-bit ADC to effectively capture ≥ 50 harmonic components.

[0105] Vibration monitoring: An 8-channel FBG array is arranged at the four corners and the center of the iron core, and the mechanical vibration modes caused by magnetostriction in the range of 10 - 500 Hz are accurately identified at a sampling frequency of 20 kHz.

[0106] Thermal imaging: A dual-band infrared camera with a resolution of 640 × 480 is used. Local overheating is identified through the 3 - 5 μm band, and the overall temperature rise distribution is monitored through the 8 - 12 μm band.

[0107] In the three-phase current acquisition channel, a Rogowski coil sensor with a bandwidth ≥ 100 kHz is configured, and a high-density sampling mode of 512 points per cycle is adopted to ensure the complete capture of more than 50 harmonic components; An optical fiber Bragg grating array is arranged at the four corners and the center of the iron core, and the magnetostrictive vibration frequency spectrum characteristics are analyzed in real time through 20 kHz high-frequency sampling; A dual-band (3 - 5 μm and 8 - 12 μm) infrared thermal imager is configured to construct a three-dimensional distribution map of the surface temperature field of the radiator at a pixel-level resolution of 640 × 480, and the three-dimensional perception of the equipment state is realized by combining multi-physical field data fusion.

[0108] Through the coordinated configuration of high-frequency sampling and multi-band detection, the synchronous and accurate perception of electromagnetic - mechanical - thermal parameters is realized, the early recognition ability of abnormal harmonic components and local overheating is improved, and multi-dimensional data support is provided for equipment health assessment.

[0109] The encrypted sharing in the federated learning framework in step S1 includes:

[0110] Gradient noise injection: During the local model training, differential privacy noise satisfying ε = 0.5 and δ = 1e-5, Gaussian noise is added to the backpropagation gradient where g is the original gradient vector, N(0,σ 2 I) is Gaussian noise with a mean of 0 and a variance of σ 2 and is the standard deviation of the noise, and I is the identity matrix;

[0111] After substituting ε = 0.5 and δ = 1e - 5, it is calculated that σ = 9.68. Therefore, the standard deviation of the Gaussian noise is σ ≈ 9.68, and the final Gaussian noise is sampled from a Gaussian distribution with a mean of 0 and a variance of σ 2 = 9.68 2 ≈ 93.70, that is: N(0, 93.70I);

[0112] The standard deviation of the noise σ ≈ 9.68

[0113] The variance of the noise: σ 2 ≈ 93.70;

[0114] Parameter aggregation protocol: The Paillier homomorphic encryption algorithm is adopted, and the key length is set to 2048 bits to achieve the secure weighted average of the global model parameters;

[0115] Trusted verification mechanism: Verify the integrity of the TEE environment through remote attestation technology to ensure memory encryption and process isolation during model updates.

[0116] In the federated learning framework, Laplace noise that satisfies (ε = 0.5, δ = 1e - 5) differential privacy is added to the local model backpropagation gradient to form privacy - protected gradients; the 2048 - bit Paillier homomorphic encryption algorithm is used to securely aggregate the model parameters to ensure that the data in the weighted average process is not decrypted; the integrity certificate of the trusted execution environment (TEE) is verified through SGX remote attestation technology, and memory encryption and process isolation are enforced to prevent the model parameters from being maliciously stolen during training.

[0117] Combining differential privacy and homomorphic encryption technologies, global model optimization is achieved on the premise of protecting the data privacy of distributed nodes. Effective defense against side - channel attacks is achieved through TEE environment verification, balancing the security of data sharing and the efficiency of model training.

[0118] The physical - constraint - based LSTM feature extraction in step S2 is specifically as follows:

[0119] Construct a physical knowledge module containing electromagnetic - thermal coupling equations, and embed the Maxwell's equations as constraint conditions into the LSTM network through the Lagrange multiplier method;

[0120] Design an improved LSTM cell with a time - varying forget gate, and its forget - gate coefficient is dynamically adjusted by the radiator temperature gradient;

[0121] Introduce spectral normalization processing in the feature output layer to limit the L2 norm of the health assessment matrix not to exceed a preset threshold.

[0122] Technical details of feature extraction:

[0123] Constrained embedding: The residual term of the electromagnetic-thermal coupling equation is introduced in the LSTM hidden layer, and the network output is forced to satisfy the physical conservation law through the Lagrange multiplier.

[0124] Dynamic forgetting: The forgetting gate coefficient is controlled by the temperature gradient. When the radiator temperature difference exceeds 5°C, the memory weight of the historical vibration features is automatically enhanced.

[0125] Spectral normalization: impose L2 norm constraints on the output features, limit the numerical range of the health matrix to the interval [0,1], and avoid divergence of the evaluation value.

[0126] The Maxwell equations of electromagnetic field and the heat conduction equation are combined to construct a physical constraint module, and the partial differential equations are embedded in the LSTM network loss function through Lagrange multipliers. A time-varying forget gate sensitive to temperature gradient is designed so that the forgetting coefficient β=1 / (1+e^(-ΔT)), where ΔT is the temperature difference of the radiator. Spectral normalization is implemented in the LSTM output layer to force the L2 norm of the health matrix not to exceed the preset threshold θ=10, so as to prevent overfitting while maintaining feature stability.

[0127] Physical constraints are embedded in the deep learning network to enhance the physical interpretability of feature extraction. The time-varying forget gate design improves the model's adaptability to temperature drift, and spectral normalization processing effectively suppresses the risk of overfitting.

[0128] The dynamic correction process of the quantum hybrid optimization algorithm in step S2 includes:

[0129] Initialization phase: map the energy entropy weight coefficient to the superposition state of quantum bits, and each weight corresponds to the entangled state of three quantum bits;

[0130] Iterative update: Introduce quantum rotating door operation into the velocity update equation of the classical particle swarm, and the rotation angle is determined by the difference between the current particle fitness value and the global optimal value;

[0131] Constraint processing: When the fitness change rate for 10 consecutive iterations is less than 0.1%, the quantum annealing process is started to accelerate convergence.

[0132] Energy entropy weight correction steps:

[0133] Quantum coding: Map the weight coefficients into 3-qubit entangled states and prepare 256 sets of superposition states to form the initial population.

[0134] Rotation update: Calculate the quantum gate rotation angle θ=arctan(Δf / 10) based on the fitness difference, and adjust the quantum state phase synchronously when updating the particle position.

[0135] Annealing convergence: When the iteration falls into the local optimum, a transverse magnetic field is applied for quantum tunneling to complete the global optimization of the weight coefficient within 20ms.

[0136] Map the energy entropy weight coefficient to the 3-qubit entangled state |ψ> = α|000> + β|111>, and explore the parameter space through the quantum rotation gate; superimpose the quantum rotation amount Δθ = k * (f_global - f_current) on the classical particle swarm velocity update term, where k is the fitness proportionality factor; when the fitness change rate is continuously lower than 0.1% for 10 times, start the quantum annealing process: perform annealing evolution with the Hamiltonian H(t) = (1 - t / T)H_0 + (t / T)H_p, T = 100 ms, to break through the local optimum.

[0137] The quantum superposition state mapping expands the search dimension of the solution space, the quantum rotation gate operation enhances the global optimization ability of the classical algorithm, and the annealing mechanism dynamically balances local exploitation and global exploration, significantly improving the convergence speed and accuracy of the weight coefficient correction.

[0138] The optimization strategy of the improved NSGA-III algorithm in step S3 includes:

[0139] Dynamic adjustment of reference points: Calculate the reference point offset according to the Mahalanobis distance of the population distribution, and the offset step size is proportional to the curvature of the Pareto front;

[0140] Chaotic initialization: Use the Logistic chaotic mapping to generate the initial population, with the chaotic parameter μ = 4.0 and the initial value

[0141] X0 = 0.31;

[0142] Elite screening: Calculate the Shannon entropy value of the solutions in the retention pool, and preferentially retain the solutions with an entropy value higher than 0.8 to maintain diversity.

[0143] Key points for the implementation of algorithm optimization:

[0144] Adaptive reference points: Calculate the covariance matrix according to the population distribution. When the Mahalanobis distance exceeds the threshold, offset the reference point by a factor of 1.2 times the curvature of the front.

[0145] Chaotic initialization: Use the Logistic mapping with μ = 4.0 to generate the initial solutions, and inject the chaotic sequence into the population after screening by Fourier transform.

[0146] Entropy value screening: Use the sliding window method when calculating the Shannon entropy of the solutions, and retain the solutions with an entropy value > 0.8 to form an elite pool to ensure the diversity of the Pareto solution set.

[0147] Calculate the Mahalanobis distance \(d = \sqrt{(x - \mu)^T\Sigma^{-1}(x - \mu)}\) based on the covariance matrix \(\Sigma\) of the population distribution, and dynamically adjust the reference point offset \(\Delta r=\eta\cdot\kappa\), where \(\kappa\) is the curvature radius of the Pareto front; generate the initial population using the Logistic chaotic map with \(\mu = 4.0\), and the iterative formula \(x_{n + 1}=\mu x_n(1 - x_n)\), with the initial value \(x_0 = 0.31\); calculate the Shannon entropy \(H=-\sum p_i\log p_i\) of the solution set during elite screening, where \(p_i\) is the proportion of the \(i\)-th solution in the population, and the higher the entropy value, the greater the population diversity;

[0148] By setting the entropy threshold to 0.8, ensure that the population diversity is maintained at a high level and avoid premature convergence to local optimal solutions; high diversity helps to explore more solution spaces and increases the possibility of finding the global optimal solution; retain the solutions with \(H\geq0.8\) to maintain population diversity.

[0149] Dynamic reference point adjustment improves the approximation efficiency of the Pareto front, chaotic initialization enhances the uniformity of the population distribution, and the Shannon entropy screening mechanism ensures the engineering applicability of the solution set, making the optimization strategy more adaptable to multi-objective game scenarios.

[0150] The Proof of Contribution (PoC) mechanism for the contribution degree of the main chain in step S4 specifically includes:

[0151] The contribution degree evaluation index consists of the following parameters:

[0152] Historical verification accuracy rate, with a weight of 40%, denoted as \(H_i\);

[0153] Computing resource input, with a weight of 30%, denoted as \(R_i\);

[0154] Network latency performance, with a weight of 30%, denoted as \(L_i\);

[0155] The consensus weight calculation formula is:

[0156] \(w_i=\frac{0.4H_i + 0.3R_i}{0.3L_i}\)

[0157] Where \(H_i\), \(R_i\), and \(L_i\) respectively represent the historical verification accuracy rate, computing resource input, and network latency value of node \(i\), and \(w_i\) represents the consensus weight value of the \(i\)-th stage;

[0158] Adopt the BLS threshold signature scheme, requiring at least 7 signatures out of 11 nodes to pass the verification.

[0159] The contribution degree of the main-chain nodes is evaluated through three-dimensional metrics: The historical verification accuracy rate Hi is calculated using a sliding window (window size: 100 blocks). The computing resource input Ri = CPU core hours × memory occupancy rate, and the network latency Li = average response time / benchmark value. The consensus weight wi = (0.4Hi + 0.3Ri) / (0.3Li) is used to achieve multi-dimensional measurement. The BLS threshold signature scheme is adopted, with the threshold set as (t, n) = (7, 11), requiring at least 7 nodes to pass the elliptic curve pairing verification: e(g, σ) = e(vk, H(m)), ensuring the security of policy verification.

[0160] Based on the calculation of the consensus weight with multi-dimensional contribution metrics, the optimal allocation of node resources is realized. The BLS threshold signature reduces the communication overhead while ensuring the Byzantine fault tolerance ability, and a highly reliable decentralized policy verification system is constructed.

[0161] The triggering conditions for singular value monitoring in step S5 include:

[0162] Set the dynamic threshold κ_max = 100 × N, where N is the number of currently running nodes. When the condition number of the impedance matrix exceeds κ_max, compensation is triggered.

[0163] The calculation of the singular value spectrum entropy adopts a sliding window mechanism, and the window length is set to 10 power frequency cycles.

[0164] Stability margin evaluation: When the real part of the Lyapunov exponent is greater than 0 for 5 consecutive iterations, the emergency control strategy is started.

[0165] Set the dynamic threshold κ_max = 100N (N is the number of online nodes). When the condition number of the impedance matrix cond(Z) = σ_max / σ_min > κ_max, compensation is triggered. The sliding window mechanism is used to calculate the singular value spectrum entropy H_sv = -Σ(σ_i / Σσ_j)log(σ_i / Σσ_j), and the window length covers 10 power frequency cycles (200 ms @ 50 Hz). When the real part of the Lyapunov exponent Re(λ) is greater than 0 for 5 consecutive times, the emergency control strategy is started: Cut off 50% of the compensation branches and switch in the standby damping resistors.

[0166] Among them, the singular value dynamic monitoring system operates according to the following mechanism:

[0167] Threshold triggering mechanism:

[0168] The dynamic threshold is set as κ_max = 100 × N (N is the number of currently online nodes). When the condition number of the impedance matrix exceeds the threshold, the anti-resonance compensation network is reorganized and triggered.

[0169] The sliding recursive method is used to calculate the condition number in real time: The singular value decomposition result of the impedance matrix is updated every 10 ms.

[0170] Spectrum entropy monitoring module:

[0171] Configure a sliding window synchronized with the power frequency period (window length = 10 × 20 ms = 200 ms)

[0172] Normalization is adopted when calculating the singular value spectrum entropy:

[0173] [H = -\sum_{i = 1}^nσ_i / |σ|_2\cdot\log(σ_i / |σ|_2)]

[0174] When the spectrum entropy value exceeds 0.85 three times in a row, activate the warning mechanism;

[0175] Stability margin control:

[0176] Calculate the real part of the Lyapunov exponent in real time and adopt moving average filtering (window length = 5 control cycles);

[0177] Emergency strategy activation condition: the real part of the exponent > 0 and the duration ≥ 100 ms, immediately switch to the preset safety control mode.

[0178] Based on the dual criteria of matrix condition number and spectrum entropy value, improve the monitoring sensitivity of topological stability. The sliding window mechanism enhances the adaptability to transient processes, and the emergency strategy hierarchical trigger design effectively prevents system-level cascading failures.

[0179] The optimization of the PID gain scheduling strategy in step S5 specifically includes: constructing a twin-delayed deep deterministic policy gradient (TD3) framework, setting the learning rate of the actor network to 3 × 10-4, and the learning rate of the critic network to 1 × 10-3; the reward function is designed as follows: R = 0.5(1 - Vd / Vmax) + 0.3(1 - Es / Emax) - 0.2|Ct - Ct_prev|;

[0180] Where: Vd: the real-time deviation value of the output voltage, Vmax is the maximum allowable voltage deviation; Es: the system steady-state error value, Emax is the maximum allowable steady-state error; Ct, Ct_prev: the control quantity output values at the current moment and the previous moment; R represents the system dynamic performance evaluation index; in the exploration strategy, amplitude constraints need to be imposed on the PID gain parameters.

[0181] Construct the TD3 framework: The actor network uses the Adam optimizer with a learning rate of 3e-4, and the critic network uses a learning rate of 1e-3; The reward function R = 0.5(1 - Vd / Vmax) + 0.3(1 - Es / Emax) - 0.2|Ct - Ct_prev|, where Vmax = 5% and Emax = 3% are thresholds; During the exploration phase, hard constraints are imposed on the PID gains: Kp ∈ [0.1, 10], Ki ∈ [0.01, 5], Kd ∈ [0, 2]; The target network performs soft updates with τ = 0.005 every 2 updates to ensure policy stability.

[0182] The adversarial training mechanism of the TD3 framework improves the self-tuning accuracy of PID parameters. The multi-dimensional reward function design achieves a balance between dynamic performance and steady-state accuracy. The hard constraints on the gains avoid sudden changes in the control quantity and significantly reduce the risk of system overshoot.

[0183] Example 3

[0184] As Figure 1 shown, a CT induction power supply system for implementing a CT induction power supply control method based on cooperative control and distributed optimization includes:

[0185] Multi-modal sensing module: Integrates three-phase current, vibration, and infrared sensing units, and configures a synchronous data acquisition architecture;

[0186] Edge computing module: Builds an encrypted computing environment to achieve fast iteration of local models;

[0187] Intelligent decision-making module: Integrates a physical constraint prediction model and a multi-objective optimization algorithm to support dynamic resource scheduling;

[0188] Blockchain network module: Adopts a main-side chain architecture and a hardware-accelerated consensus mechanism to ensure cross-chain interaction efficiency;

[0189] Dynamic control module: Includes an adaptive compensation network and a reinforcement learning controller, and configures high-frequency power devices and a real-time control interface;

[0190] Digital twin module: Realizes synchronous interaction between the physical system and the virtual model and visualizes multi-field coupling.

[0191] Specifically, the hardware-software cooperation implementation plan of the CT induction power supply system:

[0192] Multi-modal sensing module:

[0193] Current acquisition unit: Configures a 16-bit ADC with a range of ±500A and an inter-channel delay < 10ns;

[0194] Vibration sensing node: The wavelength resolution of the fiber Bragg grating demodulator is 0.1 pm, and the temperature compensation accuracy is ±0.5°C;

[0195] Infrared thermal imager: Integrated TEC cooled detector, NETD < 50mK;

[0196] Edge computing module:

[0197] Equipped with a domestic security chip (encryption computing power ≥ 10Gbps), supporting SM2 / SM9 national cryptographic algorithms;

[0198] Localized model update cycle: 30 minutes per time in normal mode, 5 minutes per time in emergency mode;

[0199] Blockchain network module:

[0200] The main chain adopts the Fabric architecture, and the side chain deploys a fast consensus engine based on Raft;

[0201] The cross-chain communication interface is configured with a hardware acceleration module (FPGA realizes VDF calculation, latency < 5ms);

[0202] The smart contract sets up a two-layer verification mechanism: syntax verification layer + physical rule verification layer;

[0203] Dynamic control module:

[0204] Compensation network topology: Includes 6 groups of LC branches, supporting μs-level solid-state relay switching;

[0205] The power device selects SiC MOSFET, and the switching frequency is designed to be adjustable from 100kHz to 1MHz;

[0206] The real-time interface adopts the TSN network, and the clock synchronization accuracy < 1μs;

[0207] Digital twin module:

[0208] Multi-field coupling visualization engine: Joint solution of electromagnetic field (FEM), temperature field (FVM), and stress field (DEM);

[0209] Virtual-real synchronization mechanism: Perform a full-state check every 50ms, and trigger model reconstruction when the deviation exceeds the limit;

[0210] This implementation method realizes the multi-objective cooperative control of the CT induction power supply system through hardware acceleration, algorithm optimization, and architecture innovation, and improves the dynamic response speed by more than 40% while ensuring safety.

[0211] The above-described embodiments are described in more detail and specifically, expressing the preferred embodiments of the present invention. They are only used to illustrate the technical ideas and features of the present invention. The purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. However, it is not limited to the present invention alone. The patent scope of the present invention cannot be limited only by this embodiment. That is, any equivalent changes or modifications made in accordance with the spirit disclosed by the present invention, for researchers or technicians in the field, within the structure of the present invention, local improvements within the system and changes and transformations between subsystems are still within the patent scope of the present invention.

Claims

1. A CT induction power supply control method based on collaborative control and distributed optimization, characterized in that: include: Step S1: deploy a multimodal sensor array, generate a digital twin with spatiotemporal correlation through multi-physics field coupling modeling, and use a federated learning framework to realize encrypted sharing of local model parameters of distributed nodes; Step S2: Perform LSTM feature extraction under physical constraints on the output of step S1 and the historical operation log to generate a health assessment matrix, and dynamically correct the energy entropy weight coefficient through the quantum hybrid optimization algorithm; Step S3: construct a multi-party game optimization model and use the NSGA-III algorithm to solve the Pareto optimal strategy set; Step S4: Execute policy consensus verification through heterogeneous blockchain architecture, deploy global policy verification based on Proof of Contribution PoC on the main chain, and implement local policy fast synchronization with Verifiable Delay Function VDF on the side chain; Step S5: Based on the consensus result of step S4, an adaptive control law driven by the Lyapunov energy function is constructed, and the gain scheduling strategy of the PID controller is optimized by real-time monitoring of the singular value changes of the topological impedance matrix.

2. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1 is characterized in that: The deployment of the multimodal sensor array in step S1 includes: A Rogowski coil sensor with a bandwidth ≥ 100kHz is set in the three-phase current acquisition channel, and the sampling rate is set to 512 points per cycle; The magnetostrictive vibration sensor uses a fiber Bragg grating array, which is arranged at the four corners and the center of the core, with a sampling frequency of 20kHz; The infrared thermal imager is equipped with a dual-band temperature measurement module, 3-5μm and 8-12μm, and the spatial resolution is not less than 640×480 pixels.

3. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1 is characterized in that: The encrypted sharing of the federated learning framework in step S1 includes: Gradient noise injection: When training a local model, differential privacy noise and Gaussian noise satisfying ε = 0.5 and δ = 1e-5 are added to the back-propagation gradient. Where g is the original gradient vector, N(0,σ 2 I) has a mean of 0 and a variance of σ 2 Gaussian noise, is the standard deviation of the noise; Parameter aggregation protocol: Use the Paillier homomorphic encryption algorithm and set the key length to 2048 bits; Trusted Verification Mechanism: Verify the integrity of the TEE environment through remote attestation technology to ensure memory encryption and process isolation during model updates.

4. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1 is characterized in that: The LSTM feature extraction under physical constraints in step S2 is specifically as follows: Construct a physical knowledge module containing electromagnetic-thermal coupling equations, and embed Maxwell's equations as constraints into the LSTM network through the Lagrange multiplier method; An improved LSTM unit with a time-varying forget gate is designed, and its forget gate coefficient is dynamically adjusted by the radiator temperature gradient; Spectral normalization is introduced in the feature output layer to limit the L2 norm of the health assessment matrix to not exceed the preset threshold.

5. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1 is characterized in that: The dynamic correction process of the quantum hybrid optimization algorithm in step S2 includes: Initialization phase: map the energy entropy weight coefficient to the superposition state of quantum bits, and each weight corresponds to the entangled state of three quantum bits; Iterative update: Introduce quantum rotating door operation into the velocity update equation of the classical particle swarm, and the rotation angle is determined by the difference between the current particle fitness value and the global optimal value; Constraint processing: When the fitness change rate for 10 consecutive iterations is less than 0.1%, the quantum annealing process is started to accelerate convergence.

6. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1 is characterized in that: The optimization strategy of the improved NSGA-III algorithm in step S3 includes: Dynamic adjustment of reference points: The reference point offset is calculated based on the Mahalanobis distance of the population distribution, and the offset step is proportional to the curvature of the Pareto frontier; Chaos initialization: Logistic chaos mapping is used to generate the initial population, with chaos parameter μ = 4.0 and initial value X0＝0.31； Elite screening: Calculate the Shannon entropy of the solutions in the retention pool, and prioritize solutions with entropy values ​​higher than 0.

8.

7. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1 is characterized in that: The contribution proof PoC mechanism of the main chain in step S4 includes: The contribution evaluation index consists of the following parameters: Historical verification accuracy, weighted 40%, denoted as Hi; Calculate the resource input, with a weight of 30%, denoted as Ri; Network delay performance, weight 30%, denoted as Li; The consensus weight calculation formula is: wi=(0.4Hi+0.3Ri) / (0.3Li) Where Hi, Ri, and Li represent the historical verification accuracy, computing resource input, and network delay value of node i, respectively, and wi represents the consensus weight value of the i-th stage; The BLS threshold signature scheme is adopted, which requires at least 7 signatures out of 11 nodes to pass verification.

8. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1 is characterized in that: The triggering conditions for singular value monitoring in step S5 include: Set the dynamic threshold κ_max = 100×N, where N is the number of currently running nodes. When the impedance matrix condition number exceeds κ_max, compensation is triggered. The calculation of singular value spectral entropy adopts a sliding window mechanism, and the window length is set to 10 power frequency cycles; Stability margin assessment: When the real part of the Lyapunov exponent is greater than 0 for five consecutive iterations, the emergency control strategy is initiated.

9. The CT induction power supply control method based on collaborative control and distributed optimization according to claim 1, characterized in that: The PID gain scheduling strategy optimization of step S5 includes: constructing a double-delayed deep deterministic policy gradient (TD3) framework, setting the actor network learning rate to 3×10 -4 , the critic network learning rate is 1×10 -3 ; The reward function is designed as follows: R = 0.5 (1-Vd / Vmax) + 0.3 (1-Es / Emax) - 0.2 | Ct - Ct_prev |; Where: Vd: real-time deviation value of output voltage, Vmax is the maximum allowable voltage deviation; Es: system steady-state error value, Emax is the maximum allowable steady-state error; Ct, Ct_prev: control quantity output value at the current moment and the previous moment; R represents the system dynamic performance evaluation index; the exploration strategy needs to impose amplitude constraints on the PID gain parameters.

10. A CT induction power supply system, used to implement the CT induction power supply control method based on collaborative control and distributed optimization as described in any one of claims 1 to 9, characterized in that: include: Multimodal sensing module: integrates three-phase current, vibration and infrared sensing units, and configures a synchronous data acquisition architecture; Edge computing module: built-in encrypted computing environment to achieve rapid iteration of localized models; Intelligent decision-making module: integrates physical constraint prediction model and multi-objective optimization algorithm to support dynamic resource scheduling; Blockchain network module: adopts the main-side chain architecture and hardware-accelerated consensus mechanism to ensure cross-chain interaction efficiency; Dynamic control module: includes adaptive compensation network and reinforcement learning controller, equipped with high-frequency power devices and real-time control interface; Digital twin module: realizes the synchronous interaction between physical systems and virtual models and the visualization of multi-field coupling.

Images