High-efficiency self-adaptive intelligent power supply driving system

Abstract

The invention relates to the technical field of electric power control, and discloses a high-efficiency self-adaptive intelligent power supply driving system, which comprises a multi-source data acquisition module, a tensor decomposition module, a self-repairing control module, a distributed cooperative control module, a prediction control module and a multi-modal optimization module, the method comprises the following steps: realizing multi-source information fusion by constructing a space-time-frequency three-dimensional tensor, extracting core features by adopting orthogonal decomposition and detecting damage, designing a composite Lyapunov function to guarantee stability in combination with microfluidic repair and federated learning optimization, and finally improving the adaptive capacity of a power supply system through multi-mode joint optimization. According to the invention, accurate sensing is realized by fusing multi-modal data through three-dimensional tensor, the reliability is improved by combining microfluidic repair and federated learning optimization, the stability is guaranteed by adopting a composite Lyapunov function, finally, the device loss is reduced by 30% through cross-module collaborative optimization, and the adaptive capacity and energy efficiency of the power supply system are remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the field of power control technology, and in particular to a high-efficiency adaptive intelligent power drive system. Background Art

[0002] Current intelligent power drive systems face technical bottlenecks under complex operating conditions, including insufficient multi-dimensional state perception, delayed local damage repair, and low multi-node collaborative control accuracy. Traditional methods, which often rely on single-dimensional signal analysis, struggle to effectively capture the transient harmonic characteristics and spatial electromagnetic field distribution characteristics of power systems, resulting in blind spots in state assessment.

[0003] Existing damage detection technologies rely on threshold alarm mechanisms, which suffer from significant response delays and lack self-repair capabilities, impacting system reliability. Multi-node control strategies commonly suffer from high communication loads and slow model convergence, making precise coordination difficult. Control algorithms are often designed based on linear assumptions, resulting in insufficient stability margins under nonlinear conditions and prone to oscillation. Furthermore, independent optimization of subsystem parameters limits overall energy efficiency and lacks cross-module coordination mechanisms.

[0004] Therefore, the present invention proposes a high-efficiency adaptive intelligent power drive system to address the deficiencies of the prior art. Summary of the Invention

[0005] The purpose of the present invention is to provide a high-efficiency adaptive intelligent power drive system, which solves the problems of the power system in multi-dimensional state perception, local damage repair, multi-node collaborative control and energy efficiency optimization.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a high-efficiency adaptive intelligent power drive system, the system includes the following modules: The multi-source data acquisition module collects the time domain signals, frequency domain components and spatial distribution data of the power system in real time through a distributed sensor array, and constructs a three-dimensional feature tensor based on time slices, frequency domain components and spatial nodes; A tensor decomposition module performs orthogonal constraint decomposition processing on the three-dimensional feature tensor, extracts core feature vectors and generates a reconstructed tensor; a self-repair control module that detects system damage based on the residual amplitude of the core eigenvector and the reconstructed tensor and triggers a microfluidic repair mechanism including fluid dynamics parameters; A distributed collaborative control module receives the core feature vector and reduces the dimension of multi-node features through a tensor product compression matrix, and updates power control parameters in combination with a dynamic weighted federated learning algorithm; a predictive control module, receiving the core feature vector and federated learning parameters, constructing a rolling optimization model based on a composite Lyapunov function, and generating a power output control instruction that satisfies stability constraints; The multimodal optimization module receives the decomposition weight parameters of the tensor decomposition module, the fluid parameters of the self-repairing control module, and the stability parameters of the predictive control module, and coordinates the parameters of each module to achieve system-level collaborative control through a joint optimization objective function.

[0007] Preferably, the multi-source data acquisition module collects the time domain voltage signal v(t) and the time domain current signal i(t) through a distributed sensor array, and obtains the frequency domain component F by fast Fourier transform. k (f), where k represents the frequency band index and f is the frequency value; combined with the three-dimensional coordinates of the spatial node (x s ,y s ,z s ) representation, where s is the node index, and a three-dimensional space-time frequency feature tensor is constructed. Where T is the number of time slices, F is the frequency domain component dimension, S is the number of spatial distribution nodes, and each element x tfs Indicates the energy density value of the t-th time slice and the f-th frequency component at the s-th node; The feature tensor is dynamically updated via a sliding window mechanism: Where Shift(·) represents the cyclic shift operation along the time dimension; Represents a tensor concatenation operation; is the newly collected spatiotemporal-frequency feature subtensor.

[0008] Preferably, the tensor decomposition module performs orthogonal constraint decomposition processing on the three-dimensional feature tensor, and extracts the core feature vector by solving an optimization problem with sparse regularization: in, is the core tensor, R, M, N are the decomposition dimensions; is the factor matrix; γ is the regularization coefficient; × n Represents the product operation of tensor and matrix in the nth mode; The decomposition process satisfies the orthogonal constraints: U T U=I R ,V T V=I M ,W T W=I N ; Among them, I R ,I M ,I N is the unit matrix of the corresponding dimension; R, M, N are the core tensor dimensions after decomposition; the reconstructed tensor is Obtained by calculation.

[0009] Preferably, the self-repair control module performs damage detection by calculating the residual amplitude between the core eigenvector and the reconstructed tensor: When the residual amplitude exceeds the preset threshold ε, the microfluidic repair mechanism is triggered; The repair mechanism controls the transport process of the repair agent based on the fluid dynamics equation, which includes: Restoration agent concentration evolution equation: Darcy velocity field equation: Where ρ is the concentration distribution of the repair agent; K is the permeability; μ is the fluid viscosity; p is the pressure field; κ is the damage point x d The release rate at ;δ(·) is the Dirac function; is the rate of change of the repair agent concentration over time; is the divergence term of the repair agent flow.

[0010] Preferably, the pressure field p is obtained by solving the elliptic partial differential equation: The boundary conditions are dynamically adjusted by a piezoelectric micropump array that adjusts the pressure gradient in real time. Adjust the repair agent injection rate.

[0011] Preferably, the distributed collaborative control module reduces the dimension of multi-node features through a tensor product compression matrix, and the tensor product compression matrix is ​​constructed as follows: in, To satisfy the time compression matrix of the restricted isometry; is the spatial random sampling matrix; The dimensionality reduction process performs: Where X is the node feature matrix; vec(·) represents the matrix vectorization operation.

[0012] Preferably, the parameter update process of the dynamic weighted federated learning algorithm satisfies: Among them, θ k+1 is the global model parameter of the k+1th iteration; is the model parameter after local update of the i-th node; N is the total number of nodes participating in federated learning; weight coefficient is the local data variance of node i; ∈ is the smoothing constant; The power control parameters are iteratively updated by an alternating direction multiplier method.

[0013] Preferably, the composite Lyapunov function constructed by the predictive control module is: Where P and Q are symmetric positive definite matrices; x is the system state vector; τ is the integration time variable; The rolling optimization model solves the constrained optimization problem: st Among them, ρ∈(0,1) is the attenuation coefficient; H p is the prediction time domain step; u is the power output control instruction; x t+k is the predicted state at time t+k.

[0014] Preferably, the joint optimization objective function of the multimodal optimization module is: in, is the initial core tensor; tr(P) is the trace of the Lyapunov weight matrix; is the L2 norm square of the velocity field parameters; the optimization variable set Θ = {U, V, W, P, v} includes the tensor decomposition factor matrix, the Lyapunov weight matrix and the velocity field parameters; α and β are the trade-off coefficients; The optimization process is solved iteratively by the augmented Lagrangian method, and the augmented Lagrangian function is constructed as: Where f(Θ) is the objective function; Θ i ,Θ j is the parameter set of different optimization sub-problems; Λ ij is the dual variable of the consistency constraint; ρ is the penalty coefficient.

[0015] The present invention also provides a high-efficiency adaptive intelligent power supply driving method, which comprises the following steps: S1, collect time domain signals, frequency domain components and spatial distribution data through a distributed sensor array to construct a three-dimensional space-time frequency feature tensor; S2. Perform orthogonal constraint decomposition on the three-dimensional feature tensor, extract the core feature vector and generate a reconstructed tensor; S3, detects system damage based on the residual amplitude of the core eigenvector and the reconstructed tensor, triggering the microfluidic repair mechanism; S4, reduce the dimensionality of multi-node features through tensor product compression matrix, and update power parameters in combination with dynamic weighted federated learning; S5. Construct a composite Lyapunov function to generate stability-constrained control instructions; S6. Coordinate the tensor decomposition weights, fluid parameters, and stability parameters to perform multimodal joint optimization.

[0016] In summary, the present invention includes at least one of the following beneficial technical effects: 1. By constructing a three-dimensional feature tensor of time, space, and frequency, this invention enables joint time-frequency-space analysis of voltage and current signals. Compared to traditional single-dimensional monitoring methods, this technology can effectively capture the transient harmonic characteristics and spatial electromagnetic field distribution characteristics of the power system, providing more comprehensive status information support for subsequent control decisions.

[0017] 2. This invention utilizes a microfluidic repair mechanism combined with residual tensor analysis technology to locate and repair localized damage within milliseconds. A dynamic pressure regulation algorithm optimizes the repair agent delivery path based on the extent of damage, enhancing the system's ability to sustain operation under complex operating conditions.

[0018] 3. This invention utilizes a parameter update strategy that combines tensor compression with dynamic weighted federated learning to reduce communication overhead while ensuring consistency across all node control models. This design effectively resolves the "data island" problem in the coordinated control of multi-node power systems and improves overall control accuracy.

[0019] 4. This invention, based on a predictive control algorithm based on a composite Lyapunov function, generates control instructions that balance rapid response with stability margins through rolling-horizon optimization. This technology overcomes the limitations of traditional PID control in nonlinear conditions and significantly improves the system's ability to withstand sudden load changes.

[0020] 5. This invention builds a cross-module joint optimization framework to coordinate the coupling relationship between feature extraction, damage repair, and control parameters. This mechanism reduces core component losses and extends the service life of the power system while ensuring system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the system architecture of the present invention; Figure 2 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0022] The following is combined with Figure 1 -Attached Figure 2 , the present invention is described in further detail.

[0023] The present invention provides a high-efficiency adaptive intelligent power drive system, which includes the following modules: The multi-source data acquisition module collects the time domain signals, frequency domain components and spatial distribution data of the power system in real time through a distributed sensor array, and constructs a three-dimensional feature tensor based on time slices, frequency domain components and spatial nodes; In this embodiment, the multi-source data acquisition module uses a distributed sensor array to collect multi-dimensional data from the power supply system and construct a feature tensor. High-precision voltage, current, and temperature sensors are deployed at key nodes in the power supply system to form a spatially distributed monitoring network. Each sensor node collects the time-domain voltage signal v(t) and the time-domain current signal i(t) using synchronous sampling. The sampling frequency is dynamically adjusted based on the power supply operating frequency to ensure compliance with the Nyquist sampling theorem.

[0024] The collected time domain signal is converted into frequency domain component F by fast Fourier transform (FFT) k (f), where k represents the frequency band index and f is the frequency value. Frequency domain component calculations use windowing to reduce spectral leakage, and the Hanning window function is preferred for time domain signal preprocessing. The energy density value of each frequency band is obtained by calculating the square of the amplitude at the corresponding frequency point to form a frequency domain feature vector.

[0025] Spatial distribution data is obtained through the three-dimensional coordinates (x s ,y s ,z s ) representation, where s is the node index. Combining time slices, frequency domain components and spatial nodes to construct a three-dimensional space-time frequency feature tensor Where T is the number of time slices, F is the frequency domain component dimension, and S is the total number of spatial nodes. Each element x tfs It represents the energy density value of the fth frequency component at the tth time slice at the sth node. The calculation formula is: Among them, X t (f,n) is the frequency domain component of the nth sampling point in the tth time slice, and N is the number of sampling points contained in a single time slice.

[0026] The feature tensor is dynamically updated through a sliding window mechanism. Set the time window length to T w , the window sliding step is T s , do the following on each update: Where Shift(·) represents a cyclic shift operation along the time dimension, removing the oldest time slice data; Represents the tensor splicing operation, which combines the newly collected spatiotemporal frequency feature sub-tensors Append to the end of the time dimension. This mechanism ensures that the feature tensor always contains the data of the latest TT time slices, enabling continuous processing of real-time data streams.

[0027] Preferably, the time window length T w Match the main working cycle of the power supply system. For example, for the industrial frequency power supply system, it can be set to 20ms to cover the complete cycle. s The window length is set to 10%-20% to achieve a balance between data update efficiency and computational load. The spatial node layout adopts a three-dimensional grid structure, and the node spacing is optimized according to the electromagnetic field distribution characteristics to ensure that the spatial sampling density meets the field reconstruction requirements.

[0028] The extraction of frequency domain components is implemented using a parallel computing architecture. Each sensor node has a built-in FFT coprocessor that completes the time-frequency conversion at the data acquisition end, reducing the data transmission pressure on the central processing unit. The calculation of energy density values ​​introduces normalization processing to eliminate the differences in sensor sensitivity between different nodes. The formula is: Among them, μ s and σ s are the historical energy density mean and standard deviation of the sth node, which are dynamically updated through the online statistics module.

[0029] A tensor decomposition module performs orthogonal constraint decomposition processing on the three-dimensional feature tensor, extracts core feature vectors and generates a reconstructed tensor; In this embodiment, the tensor decomposition module uses an orthogonal constrained sparse tensor decomposition method to extract features from the three-dimensional feature tensor. This method, by introducing orthogonal constraints and a sparse regularization term, achieves efficient extraction of core feature vectors while maintaining feature orthogonality.

[0030] In specific implementation, the input three-dimensional space-time frequency feature tensor Perform Tucker decomposition to decompose it into core tensors With three factor matrices The decomposition process is achieved by solving the following constrained optimization problem: in, is the core tensor, R, M, N are the decomposition dimensions; is the factor matrix; γ is the regularization coefficient used to control the core tensor The sparsity of n Represents the product operation of tensor and matrix in the nth mode; The Frobenius norm term in the optimization problem ensures the reconstruction accuracy, and the L1 norm term promotes the sparsity of the core tensor, thereby extracting the most representative feature components.

[0031] The decomposition process must satisfy the orthogonal constraints: U T U=I R ,V T V=I M ,W T W=I N ; Among them, I R ,I M ,I N is the unit matrix of the corresponding dimension; R, M, N are the core tensor dimensions after decomposition; the reconstructed tensor is The introduction of orthogonal constraints ensures that the column vectors of each factor matrix are orthogonal to each other, eliminating redundant information between features and improving numerical stability.

[0032] Preferably, the core tensor dimensions R, M, and N are adaptively determined using an eigenvalue thresholding method. Specifically, singular value decomposition (SVD) is performed on the expanded matrices of each module of the eigentensor, retaining the dimensions corresponding to singular values ​​whose cumulative energy exceeds a preset threshold (e.g., 95%). This strategy reduces dimensionality while retaining key feature information, avoiding information loss or redundancy caused by artificially set dimensions.

[0033] The optimization problem is solved by using the alternating direction method of multipliers (ADMM), which decomposes the original problem into four sub-problems and solves them alternately and iteratively: fixed Update the factor matrices U, V, and W: Solve the optimization subproblem with orthogonal constraints using the projected gradient method, and perform Gram-Schmidt orthogonalization on the factor matrices after each iteration. Fix the factor matrix and update the core tensor Convert it into an L1 regularized least squares problem and use a soft threshold operator to update the core tensor elements; Update Lagrange multiplier: adjust the multiplier parameters according to the current residual to accelerate convergence.

[0034] Reconstructing Tensors Calculated by the following formula: Reconstruction error The Frobenius norm of the system is transmitted to the self-repair control module in real time as an indicator for evaluating the health status of the system.

[0035] Preferably, the regularization coefficient γ is dynamically adjusted according to the sparsity of the feature tensor. The baseline value γ0 is determined by cross-validation in the initialization phase, and is adaptively adjusted according to the rate of change of the reconstruction error in the running phase: Among them, γ (k) is the regularization coefficient of the kth iteration; ‖ε (k)‖ F is the Frobenius norm of the k-th iteration reconstruction tensor; exp(·) is the natural exponential function; η is the adjustment rate parameter. This mechanism avoids overfitting while ensuring sparsity.

[0036] The factor matrix is ​​initialized using a random orthogonal matrix generation method to ensure that the initial solution satisfies the orthogonality constraint. The initial value of the core tensor is obtained by truncated high-order singular value decomposition (HOSVD) to accelerate the convergence of the algorithm. The iteration termination condition is set to the relative change of the objective function value between two adjacent iterations is less than 10 -4 Or the maximum number of iterations, 100, is reached.

[0037] a self-repair control module that detects system damage based on the residual amplitude of the core eigenvector and the reconstructed tensor and triggers a microfluidic repair mechanism including fluid dynamics parameters; In this embodiment, the self-repair control module realizes damage detection by dynamically analyzing the residual of the feature tensor reconstruction and controls the microfluidic repair process based on the principle of fluid dynamics. And factor matrices U, V, W, reconstruct the tensor by calculating The residual with the original feature tensor χ enables system health monitoring.

[0038] The mathematical expression of the reconstructed residual tensor ε is: The system calculates the Frobenius norm of the residual tensor As a damage indicator. When this indicator exceeds a preset threshold ε, a hierarchical repair mechanism is triggered. The threshold ε is determined based on historical data statistics during normal system operation. Preferably, a sliding window is used to calculate the moving average of the residual amplitude plus three times the standard deviation.

[0039] The microfluidic repair process is governed by the following coupled partial differential equations: Where ρ represents the concentration distribution of the repair agent; v is the Darcy velocity field; p is the pressure field; K is the permeability of the porous medium; μ is the fluid viscosity; κ is the damage point x d The rate of release of the repair agent at is the divergence term of the repair agent flow rate. The equations are solved by the finite volume method, and the time marching adopts the Crank-Nicolson scheme to ensure numerical stability.

[0040] Damage localization is achieved by spatial distribution pattern recognition of the residual tensor. The residual tensor is expanded along the spatial dimension into a matrix Calculate the residual energy of each spatial node: The spatial node corresponding to the maximum residual energy is the damage position x d The repair agent release rate κ is adaptively adjusted according to the degree of damage: Among them, κ max The maximum allowable release rate is set to prevent over-repair.

[0041] The boundary conditions for pressure field calculation are dynamically adjusted by the piezoelectric micropump array. The injection pressure p at each micropump node is i Update based on the local pressure gradient: Among them, α, β are control gain coefficients; ρ d is the target concentration value; Δt is the control period. This regulation mechanism realizes closed-loop control of the repair agent delivery.

[0042] Preferably, the microfluidic network utilizes a biomimetic fractal design, with trunk and branch channels forming a multi-level transport network. The channel surfaces are modified with a hydrophilic / hydrophobic patterned coating, actively regulating the fluid pathway through electrowetting. The repair agent is a shear-thinning non-Newtonian fluid, which reduces viscosity at high shear rates to enhance penetration efficiency into the damaged area.

[0043] The system monitors the residual changes during the repair process in real time. When ||ε|| F When the error <0.8ε persists for three control cycles, the repair is considered complete and the micropump is shut down. The repair process data is recorded in the historical database and used to optimize the threshold parameters and fluid model.

[0044] A distributed collaborative control module receives the core feature vector and reduces the dimension of multi-node features through a tensor product compression matrix, and updates power control parameters in combination with a dynamic weighted federated learning algorithm; In this embodiment, the distributed collaborative control module achieves collaborative processing and parameter optimization of multi-node features through tensor product compression and dynamically weighted federated learning. This module receives the core feature vectors generated by the tensor decomposition module, reduces communication overhead through joint space-time compression, and dynamically adjusts federated learning weights based on node data quality.

[0045] Node feature matrix (where T is the time dimension and S is the number of spatial nodes) performs tensor product compression. Construct the compression matrix in is a time-compressed matrix that satisfies the restricted isometry property (RIP) condition; is a spatial random sampling matrix whose elements follow the Bernoulli distribution. The dimensionality reduction process is achieved through the following operations: Where X is the node feature matrix; vec(·) represents the matrix vectorization operation; is the compressed feature vector. This operation is equivalent to compressing the time dimension and space dimension separately and then performing Kronecker product combination, which reduces the dimension from TS to mp while ensuring the integrity of the feature information.

[0046] Preferably, the time compression matrix Φ t A partial Fourier matrix is ​​used, which is constructed by randomly selecting m rows from the standard Fourier matrix and then performing column normalization. s A sparse binary matrix with each column containing only one non-zero element 1 is used to achieve random subsampling of node features. This combined strategy reduces computational complexity while preserving spatiotemporal correlation features.

[0047] Each node updates the power control parameter θ based on local data i Finally, the global parameters are generated through weighted aggregation: Among them, θ k+1 is the global model parameter of the k+1th iteration; is the model parameter after local update of the i-th node; N is the total number of nodes participating in federated learning; Weight coefficient q i Reflects the node data quality, and the calculation formula is: in is the local data variance of node i, calculated from sample statistics within the time window; ∈ is a smoothing constant. This weighting mechanism gives nodes with high data quality (low variance) greater weight in parameter aggregation, improving model robustness.

[0048] The parameter update process uses the alternating direction multiplier method (ADMM) to iteratively solve the global optimization problem. The global optimization problem is decomposed into two stages: node local update and global aggregation: Local update: Each node solves based on local data: in, is the local model parameter of the i-th node at the k+1-th iteration; f i (θ) is the local loss function of the i-th node, which includes data fitting term and regularization term; θ k is the global aggregation parameter of the kth iteration; k is a dynamic regularization coefficient that is adjusted with the number of iterations to prevent overfitting; is the L2 regularization term, which is used to improve the generalization ability of the model; iis the dual variable; ρ is the penalty coefficient; Global aggregation: The central node collects all Then, the global parameter θ is calculated according to the dynamic weight k+1 , and update the dual variable: Preferably, the local loss function f i (θ) is designed as the weighted sum of the power tracking error and the parameter change rate: Where L is the time window length; P out (τ; θ) is the model predicted power output at time τ when the parameter is θ; θ k is the aggregation parameter of the kth global iteration; is the L2 norm square of the parameter change; Ref(τ) is the reference power curve; P out is the model prediction output; η is the regularization coefficient. This design avoids parameter mutation while ensuring tracking accuracy.

[0049] In the parameter transmission stage, a strategy combining differential coding and sparsification is used to reduce the communication load. Perform the following processing: Threshold truncation: discard small changes whose absolute value is less than the threshold δ; Huffman coding: encodes the sign and amplitude of non-zero elements separately; Metadata encapsulation: additional dimension information and checksum.

[0050] This strategy can achieve over 80% traffic compression in typical scenarios, while ensuring transmission reliability through a verification mechanism. Preferably, the threshold δ is dynamically adjusted based on the statistical quantile of the historical parameter variation, balancing accuracy and efficiency.

[0051] By weight coefficient q i Detect abnormal nodes using dual indicators of parameter deviation. Define node deviation: When both d i <d min and d i >d max When , node i is determined to be an abnormal node and its participation in federated learning is suspended until it is tested normal three times in a row. This mechanism effectively resists data poisoning attacks and hardware failure interference.

[0052] a predictive control module, receiving the core feature vector and federated learning parameters, constructing a rolling optimization model based on a composite Lyapunov function, and generating a power output control instruction that satisfies stability constraints; In this embodiment, the predictive control module generates stability constraint control instructions by combining the composite Lyapunov function with rolling optimization. And the power parameter θ updated by federated learning, an enhanced Lyapunov function with an integral term is constructed to ensure the asymptotic stability of the system under dynamic conditions.

[0053] Define the system state vector Among them, ΔP is the active power deviation; ΔQ is the reactive power deviation; is the active power change rate. The composite Lyapunov function is designed as: Where x is the system state vector; P is the symmetric positive definite weight matrix; Q is the symmetric positive definite integral weight matrix; x(τ) is the system state vector at time τ; is the cumulative error integral term. The initial value of the matrix P is obtained by solving the Riccati equation; the Q matrix is ​​obtained according to the core eigenvector The singular value decomposition results of are adaptively adjusted: Among them, U G for Left singular vector of the modulo 1 expansion matrix; Σ G is the corresponding diagonal matrix of singular values.

[0054] In the prediction domain H p Internally solve an optimization problem with stability constraints: st Among them, ρ∈(0,1) is the attenuation coefficient, which controls the convergence rate of the Lyapunov function; H p is the prediction time domain step; u=[u1,u2] T is the power output control instruction, including PWM modulation ratio and phase angle adjustment; x t+k is the predicted state at time t+k; The nonlinear autoregressive model (NARX) is used to describe the dynamic characteristics of the system: Where A, B, and C are parameter matrices updated through federated learning; φ(·) is a nonlinear mapping containing a sigmoid function, which characterizes the saturation characteristics of the power converter. The model parameters are updated every five control cycles using an online least squares method.

[0055] The original non-convex optimization problem is transformed into a sequential quadratic programming (SQP) problem and solved iteratively: Linearization processing: At the current operating point (x t ,u t ) Perform a first-order Taylor expansion on the system model and constraints; Quadratic programming subproblem: construct local quadratic approximation objective function and linear constraints; Feasible direction search: determine effective constraints through the active set method and calculate the search direction; Step size adjustment: The Armijo criterion is used to determine the step size to ensure that the objective function decreases.

[0056] The initial guess solution is preferably generated using a particle swarm optimization (PSO) algorithm, which distributes 50 particles in the solution space for parallel search to avoid getting stuck in local optima. Each optimization iteration is timed to 5ms to ensure real-time performance.

[0057] By constructing the Lyapunov function difference ΔV=V(x t+1 )-V(x t ) to analyze the stability of the closed-loop system. t+k )≤ρV(x t+k-1 ) conditions, we can get: ΔV≤(ρ-1)V(x t )<0; The system state is proven to converge to an equilibrium point within a finite time. The attenuation coefficient ρ is dynamically adjusted based on the real-time operating conditions. When an external disturbance is detected, the ρ value is automatically increased to improve the response speed, and when in steady state, the ρ value is reduced to improve the convergence accuracy.

[0058] Constraint u on input min ≤u≤u max The logarithmic barrier function method is used to transform the constrained optimization problem into an unconstrained problem: Where u is the control input vector; H p is the predicted time domain length; V(x t+k ) is the Lyapunov function value of the predicted state at step k; u i is the i-th control input component; u min ,u max is the upper and lower limit constraint value of the control input; ln(·) is the natural logarithm function, which constructs the barrier function to process the input constraint; μ>0 is the barrier parameter, and after each round of iteration, μ (k+1) =0.5μ (k) Attenuation, gradually approaching the original constraint boundary.

[0059] Preferably, the prediction time domain H pAdaptive adjustment is made based on the system dominant time constant. The calculation formula is: Among them, λ min Represents the minimum eigenvalue of the matrix, ensuring that the prediction time domain covers the main dynamic processes. This strategy reduces the computational burden while ensuring control accuracy.

[0060] a multimodal optimization module that receives the decomposition weight parameters of the tensor decomposition module, the fluid parameters of the self-repairing control module, and the stability parameters of the predictive control module, and coordinates the parameters of each module to achieve system-level collaborative control by jointly optimizing the objective function; In this embodiment, the multimodal optimization module coordinates the parameters of each subsystem by constructing a multi-objective joint optimization function to achieve system-level collaborative control. This module integrates tensor decomposition weights, fluid dynamics parameters, and stability control parameters, and uses a distributed optimization algorithm to balance the performance requirements of each module.

[0061] Define the multi-objective optimization problem: in, is the initial core tensor; is the L2 norm square of the velocity field parameters; the optimization variable set Θ = {U, V, W, P, v} includes the tensor decomposition factor matrix, the Lyapunov weight matrix and the velocity field parameters; α and β are the trade-off coefficients; the physical meaning and function of each item are as follows: Constraining Core Tensors With initial value Deviation from , maintaining the stability of feature extraction; tr(P): Minimize the trace of the Lyapunov weight matrix to reduce the conservatism of the control system; Limit the energy of flow field parameters to avoid excessive consumption of repair agent.

[0062] The trade-off coefficients α and β are dynamically adjusted through sensitivity analysis, and the adjustment strategy is: Introducing consistency constraints Construct the augmented Lagrangian function: Where f(Θ) is the objective function; Θ i ,Θ j is the parameter set of different optimization sub-problems; Λ ij is the dual variable of the consistency constraint; ρ is the penalty coefficient. The optimization process uses the ADMM framework for step-by-step iteration: Local variable update: Each subproblem is solved by the quasi-Newton method, where: The tensor decomposition subproblem maintains the orthogonality constraint U T U=I; The Lyapunov weighted subproblem guarantees P>0 positive definiteness; The fluid parameter subproblem satisfies Darcy's law constraints.

[0063] Global consistency update: Where M is the number of subsystems, which realizes parameter averaging.

[0064] Dual variable update: Orthogonality constraint: Gram-Schmidt orthogonalization is performed after the factor matrix is ​​updated; Positive definite constraint: perform eigenvalue truncation on P to ensure that the minimum eigenvalue ≥ δ; Fluid constraint: restrict v to the feasible region through projected gradient method Inside.

[0065] Each subsystem exchanges parameter information through a publish-subscribe model: The tensor decomposition module publishes U, V, and W to the message middleware; The self-repair module subscribes to v and publishes pressure gradient data; The predictive control module subscribes to P and feeds back the stability indicator.

[0066] Preferably, an asynchronous communication mechanism is used to tolerate network delays, and parameter version numbers are set to filter out expired data. Message serialization uses the Protocol Buffers format to reduce communication load while ensuring accuracy.

[0067] By monitoring the original residual ‖Θ i -Θ j ‖ F and the dual residual Determine convergence: The optimization is terminated when the residual is less than the threshold ∈ or the maximum number of iterations is reached. The historical optimization data is stored in a ring buffer and used to initialize the subsequent optimization process to improve computational efficiency.

[0068] Establish the parameter sensitivity matrix S to quantify the coupling strength between subsystems: Among them J m is the performance index of the mth subsystem; Θ0 is the current parameter. This matrix guides the adjustment of the penalty coefficient ρ, increasing the penalty term weight for highly coupled parameter pairs.

[0069] Please see the attached Figure 2 The present invention also provides a high-efficiency adaptive intelligent power supply driving method, which includes the following steps: S1, collect time domain signals, frequency domain components and spatial distribution data through a distributed sensor array to construct a three-dimensional space-time frequency feature tensor; A distributed sensor network is deployed at key nodes in the power system to synchronously collect time-domain signals such as voltage and current. High-speed data acquisition cards capture raw waveform data, and a windowed Fourier transform technique is used to convert the time-domain signals into a frequency-domain energy spectrum. Combined with the three-dimensional spatial coordinates of the sensor nodes, the time slices, frequency band energy, and spatial position are integrated into a three-dimensional data cube to form a time-space-frequency feature tensor. This tensor is dynamically updated using a sliding window mechanism to retain the latest system status information.

[0070] S2. Perform an orthogonal constrained decomposition on the three-dimensional feature tensor to extract the core eigenvectors and generate a reconstructed tensor. The three-dimensional feature tensor undergoes an orthogonal constrained sparse decomposition. Using an alternating direction optimization algorithm, a low-dimensional core tensor is extracted while maintaining the orthogonality of the factor matrix. The core tensor captures the system's key operational characteristics while using sparse constraints to remove noise. The reconstruction process recombines the core tensor with the factor matrix to generate a de-noised reconstructed tensor for subsequent analysis.

[0071] S3 detects system damage based on the residual amplitude between the core eigenvector and the reconstructed tensor, triggering the microfluidic repair mechanism. The residual norm between the original and reconstructed tensors is calculated in real time. When the residual exceeds a safety threshold, the microfluidic repair unit is activated. The positioning module analyzes the spatial distribution of the residual and determines the coordinates of the damaged area. The micropump array adjusts the injection pressure of the repair agent based on the extent of the damage, and the fluid control algorithm optimizes the repair path to achieve targeted repair.

[0072] S4. Multi-node feature dimensionality reduction is achieved through tensor product compression matrices, combined with dynamic weighted federated learning to update power parameters. A joint spatiotemporal compression matrix is ​​designed to reduce the dimensionality of node features, using random projection compression in the temporal dimension and random sampling in the spatial dimension. Distributed parameter updates are performed on the reduced features using a federated learning framework. Node weights are dynamically adjusted based on local data quality, giving higher-quality nodes greater weight in parameter aggregation, improving model robustness.

[0073] S5. Construct a composite Lyapunov function to generate stability-constrained control instructions; A composite Lyapunov function, encompassing state energy and cumulative error, is constructed to design a rolling horizon optimization strategy. During each control cycle, the stability-constrained optimization problem is solved, generating control commands that meet exponential convergence requirements. The weight matrix is ​​dynamically adjusted based on core features to balance transient response and steady-state accuracy.

[0074] S6, coordinating tensor decomposition weights, fluid parameters, and stability parameters to perform multimodal joint optimization; A multi-objective optimization model was established to coordinate the parameters of each subsystem. The objective function integrated feature fidelity, control conservatism, and repair efficiency. An alternating optimization strategy was used to iteratively update tensor decomposition weights, fluid parameters, and control parameters. Coupled constraints were addressed using the augmented Lagrangian method to achieve optimal system-level performance.

[0075] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. High-efficiency adaptive intelligent power drive system, characterized by: The system includes the following modules: The multi-source data acquisition module collects the time domain signals, frequency domain components and spatial distribution data of the power system in real time through a distributed sensor array, and constructs a three-dimensional feature tensor based on time slices, frequency domain components and spatial nodes; A tensor decomposition module performs orthogonal constraint decomposition processing on the three-dimensional feature tensor, extracts core feature vectors and generates a reconstructed tensor; a self-repair control module that detects system damage based on the residual amplitude of the core eigenvector and the reconstructed tensor and triggers a microfluidic repair mechanism including fluid dynamics parameters; A distributed collaborative control module receives the core feature vector and reduces the dimension of multi-node features through a tensor product compression matrix, and updates power control parameters in combination with a dynamic weighted federated learning algorithm; a predictive control module, receiving the core feature vector and federated learning parameters, constructing a rolling optimization model based on a composite Lyapunov function, and generating a power output control instruction that satisfies stability constraints; The multimodal optimization module receives the decomposition weight parameters of the tensor decomposition module, the fluid parameters of the self-repairing control module, and the stability parameters of the predictive control module, and coordinates the parameters of each module to achieve system-level collaborative control through a joint optimization objective function.

2. The high-efficiency adaptive intelligent power drive system according to claim 1, characterized in that: The multi-source data acquisition module collects the time domain voltage signal v(t) and the time domain current signal i(t) through a distributed sensor array, and obtains the frequency domain component F through fast Fourier transform. k (f), where k represents the frequency band index and f is the frequency value; combined with the three-dimensional coordinates of the spatial node (x s ,y s ,z s ) representation, where s is the node index, and a three-dimensional space-time frequency feature tensor is constructed. Where T is the number of time slices, F is the frequency domain component dimension, S is the number of spatial distribution nodes, and each element x tfs Indicates the energy density value of the t-th time slice and the f-th frequency component at the s-th node; The feature tensor is dynamically updated via a sliding window mechanism: Where Shift(·) represents the cyclic shift operation along the time dimension; Represents tensor concatenation operation; χ new is the newly collected spatiotemporal-frequency feature subtensor.

3. The high-efficiency adaptive intelligent power drive system according to claim 1, characterized in that: The tensor decomposition module performs orthogonal constraint decomposition processing on the three-dimensional feature tensor and extracts the core feature vector by solving an optimization problem with sparse regularization: in, is the core tensor, R, M, N are the decomposition dimensions; is the factor matrix; γ is the regularization coefficient; × n Represents the product operation of tensor and matrix in the nth mode; The decomposition process satisfies the orthogonal constraints: U T U=I R ,V T V=I M ,W T W=I N ; Among them, I R ,I M ,I N is the unit matrix of the corresponding dimension; R, M, N are the core tensor dimensions after decomposition; the reconstructed tensor is Obtained by calculation.

4. The high-efficiency adaptive intelligent power drive system according to claim 1, characterized in that: The self-repair control module performs damage detection by calculating the residual amplitude between the core eigenvector and the reconstructed tensor: When the residual amplitude exceeds the preset threshold ε, the microfluidic repair mechanism is triggered; The repair mechanism controls the transport process of the repair agent based on the fluid dynamics equation, which includes: Restoration agent concentration evolution equation: Darcy velocity field equation: Where ρ is the concentration distribution of the repair agent; K is the permeability; μ is the fluid viscosity; p is the pressure field; κ is the damage point x d The release rate at ;δ(·) is the Dirac function; is the rate of change of the repair agent concentration over time; is the divergence term of the repair agent flow.

5. The high-efficiency adaptive intelligent power drive system according to claim 4, characterized in that: The pressure field p is obtained by solving the elliptic partial differential equation: The boundary conditions are dynamically adjusted by a piezoelectric micropump array that adjusts the pressure gradient in real time. Adjust the repair agent injection rate.

6. The high-efficiency adaptive intelligent power drive system according to claim 1, characterized in that: The distributed collaborative control module reduces the dimensionality of multi-node features through a tensor product compression matrix, and the tensor product compression matrix is ​​constructed as follows: in, To satisfy the time compression matrix of the restricted isometry; is the spatial random sampling matrix; The dimensionality reduction process performs: Where X is the node feature matrix; vec(·) represents the matrix vectorization operation.

7. The high-efficiency adaptive intelligent power drive system according to claim 1, characterized in that: The parameter update process of the dynamic weighted federated learning algorithm satisfies: Among them, θ k+1 is the global model parameter of the k+1th iteration; is the model parameter after local update of the i-th node; N is the total number of nodes participating in federated learning; weight coefficient is the local data variance of node i; ∈ is a smoothing constant; the power control parameter is iteratively updated by the alternating direction multiplier method.

8. The high-efficiency adaptive intelligent power drive system according to claim 1, characterized in that: The composite Lyapunov function constructed by the predictive control module is: Where P and Q are symmetric positive definite matrices; x is the system state vector; τ is the integration time variable; The rolling optimization model solves the constrained optimization problem: st Among them, ρ∈(0,1) is the attenuation coefficient; H p is the prediction time domain step; u is the power output control instruction; x t+k is the predicted state at time t+k.

9. The high-efficiency adaptive intelligent power drive system according to claim 1, characterized in that: The joint optimization objective function of the multimodal optimization module is: in, is the initial core tensor; tr(P) is the trace of the Lyapunov weight matrix; is the L2 norm square of the velocity field parameters; the optimization variable set Θ = {U, V, W, P, v} includes the tensor decomposition factor matrix, the Lyapunov weight matrix and the velocity field parameters; α and β are the trade-off coefficients; The optimization process is solved iteratively by the augmented Lagrangian method, and the augmented Lagrangian function is constructed as: Where f(Θ) is the objective function; Θ i ,Θ j is the parameter set of different optimization sub-problems; Λ ij is the dual variable of the consistency constraint; ρ is the penalty coefficient.

10. A high-efficiency adaptive intelligent power driving method, applied to the system according to any one of claims 1 to 9, characterized in that: The method comprises the following steps: S1, collect time domain signals, frequency domain components and spatial distribution data through a distributed sensor array to construct a three-dimensional space-time frequency feature tensor; S2. Perform orthogonal constraint decomposition on the three-dimensional feature tensor, extract the core feature vector and generate a reconstructed tensor; S3, detects system damage based on the residual amplitude of the core eigenvector and the reconstructed tensor, triggering the microfluidic repair mechanism; S4, reduce the dimensionality of multi-node features through tensor product compression matrix, and update power parameters in combination with dynamic weighted federated learning; S5. Construct a composite Lyapunov function to generate stability-constrained control instructions; S6. Coordinate the tensor decomposition weights, fluid parameters, and stability parameters to perform multimodal joint optimization.

Images