34 results about "Orthogonal subspace" patented technology

The invention relates to the technical field of image data processing, and discloses a few-sample scene-oriented industrial product surface defect image analysis method, which comprises the following steps of: obtaining surface gray level image data of a product to be analyzed, calculating a structure tensor matrix and generating an anisotropy degree graph; searching similar blocks in a preset search neighborhood, and constructing a local texture data matrix; performing singular value decomposition on the local texture data matrix to extract a main subspace; constructing a projection operator and utilizing the projection operator to carry out orthogonal projection reconstruction on the local image block to generate a reconstructed background image block; according to the method, through an orthogonal subspace projection mechanism, good product textures and defect signals are separated, random noise is removed, meanwhile, high-frequency structural features are completely reserved, and the method is high in robustness, high in robustness and high in robustness. And the defect detection precision of a complex texture surface in a few-sample scene is improved.

The invention relates to the field of rotating machinery fault diagnosis, and discloses a bearing fault diagnosis method based on a physical perception KAM network, and the method comprises the steps: carrying out the discretization of a bearing vibration signal through a Gabor filter group based on physical prior initialization, and generating modal feature lexical elements with physical frequency band meanings; and inputting the lexical elements into a PC-KAM backbone network, calculating a hidden state vector by using a state space model branch, and dynamically adjusting the position of a primary function node of a Kolmogov-Arnod network branch to realize collaborative dynamic feature extraction. In the training stage, an orthogonal subspace constraint and physical perception low-rank adaptation fine tuning mechanism is introduced. And finally, searching a historical fault case, performing multi-modal fusion on the historical fault case and the deep feature sequence, mapping a fusion representation into a soft prompt, and inputting the soft prompt into a large language model to generate a diagnosis report. According to the method, the problems of poor physical interpretability of characteristics and few-sample diagnosis under variable working conditions are effectively solved, and the generalization and decision-making ability of a diagnosis system are improved.

The invention discloses a self-supervised tomographic SAR (Synthetic Aperture Radar) reconstruction method based on orthogonal subspace decomposition. According to the method, a traditional single original space is reconstructed and expanded into a Range-Null-Raw triple space collaborative learning framework, an R2R and N2N principle-based self-supervised loss function is constructed by using an orthogonal subspace decomposition characteristic of a linear measurement operator, direct mapping learning from noise measurement data to a high-quality reconstruction result is realized, and the real-time performance of a real-time reconstruction system is greatly improved, so that the real-time performance of the real-time performance of the real-time performance of the real-time performance of the real-time performance of the real-time performance of the real-time performance is improved. Observable components are learned through Range space, unobservable components are learned through Null space, a cross constraint strategy of global consistency is ensured through Raw space, and the adaptability of the model to real measurement conditions is enhanced. The performance gap between simulation training and actual deployment in deep learning TomoSAR reconstruction is effectively solved, and a new solution is provided for reliable application of an unsupervised deep reconstruction network in a real TomoSAR imaging task.

The invention discloses a modal modeling method for a tendon-driven continuum robot. The method comprises the following steps: carrying out differential modeling on a configuration under a SE (3) Lie group-se (3) Lie algebra framework; the strain field is subjected to spectrum parameterization and is decomposed into active / passive orthogonal subspaces, and an inner product adopts energy weighted Hilbert measurement containing material and section parameters. The active mode is constructed by a linear continuous Jacobian of a tendon path; legendre polynomial expansion is adopted under the collinear layout, and complex Fourier harmonic expansion is adopted under the non-collinear layout; the passive mode is selected according to a spectrum slot position mutual exclusion principle so as to ensure that the passive mode is orthogonal to the active mode in energy and suppress spectrum leakage. Lie group integration and spectral integration are adopted for numerical value realization; newton-Rafson iteration is used in the quasi-static state, and generalized-alpha time integration or Newmark-beta integration is used in the dynamics. The method has high precision and robustness under nonlinear large deformation and multi-tendon coupling conditions, and is suitable for scenes such as software operation, minimally invasive intervention and precise detection.

The invention provides an attack detection method and system based on Bayesian incremental learning and a storage medium, and the method comprises the steps: 1, collecting a data set, and dividing the data set into a plurality of tasks according to years; 2, using a Bayesian continuous learning framework, taking posterior distribution obtained by learning of a previous task as prior distribution of a current task, and adopting a gradient projection method to project a gradient of the current task to an orthogonal subspace of an old task feature space to obtain projection parameters; step 3, performing label deviation and noise processing on the task; 4, minimizing new task loss to obtain parameters, and adopting a training strategy according to a label deviation and noise processing result; 5, finding an optimal combined solution; and step 6, taking the combined model parameters as initialization parameters of the next task, returning to the step 2, and entering the next round of iteration until training of all tasks is completed. The method has the beneficial effects that the knowledge retention capability can be remarkably improved, and the problem of disastrous forgetting is effectively solved.

This document describes techniques and systems of multiple-target, simultaneous beamforming for four-dimensional (4D) radar systems for efficient angle estimation in two dimensions with a high dynamic range. For example, a processor can use electromagnetic (EM) energy received by a two-dimensional (2D) array to determine first angles in a first dimension associated with one or more objects. The processor can then determine a subspace projection matrix using the first angles without an estimate of the power of noise or interference signals in the received EM energy. Using the subspace projection matrix, the processor can determine an interference-orthogonal subspace projection-based beamformer. With the interference-orthogonal subspace projection-based beamformer, the processor can determine the desired signal output from an adaptive beamformer for the EM energy and second angles corresponding to respective first angles for the objects.

The application provides a power system core data quantum-resistant encryption method and system based on lattice cryptography, relates to the technical field of power system information security, and comprises the following steps: decomposing core data into orthogonal subspace components, constructing a quantum-resistant key pair based on a lattice difficult problem, performing state migration through multiple rounds of lattice basis transformation and noise injection, and finally assembling a ciphertext-related hash chain. The application can resist quantum computing attacks, improve the security of power system core data, realize efficient encryption processing, and guarantee the structural integrity and traceability of data.

The invention provides a spectral data cube reconstruction method, device and equipment based on compressed physical prior, a storage medium and a program product, and relates to the technical field of spectral reconstruction, and the method comprises the steps: obtaining a two-dimensional measurement image formed after a to-be-observed target is coded and modulated by a spectral imaging chip; obtaining an original transmission spectrum matrix corresponding to each pixel in the spectral imaging chip; performing principal component analysis on the original transmission spectrum matrix in a spectrum dimension, and extracting a principal component feature vector through a singular value decomposition method; projecting the original transmission spectrum matrix into an orthogonal subspace formed by the principal component feature vectors to obtain a compressed physical prior vector; and inputting the two-dimensional measurement image and the compressed physical prior vector into a deep expansion neural network, and outputting a high-dimensional spectral data cube of the to-be-observed target. The reconstruction efficiency of the spectral data cube can be remarkably improved.

The application relates to the technical field of industrial equipment state monitoring, and particularly discloses a fault prediction and health management system based on an industrial robot, which collects robot benchmark operation data and real-time operation data, and constructs a benchmark data set containing individual identity labels; multi-domain feature extraction and weighted fusion are performed on the data, a high-dimensional feature space is obtained through phase space reconstruction; the high-dimensional features are decomposed into mutually orthogonal individual attribute subspaces and degradation state subspaces by using an orthogonal subspace learning algorithm, and pure benchmark degradation features are obtained; a health index is constructed based on the benchmark degradation features, and a segmented continuous degradation model is established; real-time features are projected into the degradation state subspace to obtain real-time degradation features, which are input into the degradation model to invert the remaining service life and output graded early warning information; the application effectively suppresses false abnormal alarms by stripping individual difference interference through orthogonal decomposition of the feature space, and improves the cross-device generalization capability and prediction accuracy.

This application belongs to the interdisciplinary field of artificial intelligence, computer vision, and network information security. It discloses a deepfake detection method and system based on orthogonal subspace decomposition and hyperspherical metric. By applying singular value decomposition to the weight matrix of a pre-trained visual model, it explicitly constructs a frozen principal subspace that preserves general semantic knowledge and a trainable orthogonal residual subspace that captures specific forgery traces, achieving orthogonal isolation of the parameter space. Simultaneously, hyperspherical metric learning is introduced into the feature space, performing L2 normalization on the features and applying alignment and uniformity losses. Combined with spherical linear interpolation, latent space data augmentation is performed while preserving the Riemannian geometric structure. Through the synergistic constraints of the parameter and feature spaces, this application can reduce the interference of fine-tuning on pre-trained general visual knowledge and improve the feature discrimination stability and cross-forgery generalization ability in deepfake detection tasks.

The invention discloses a multi-harmonic injection-based dual three-phase permanent magnet synchronous motor torque increasing method, which comprises the following steps of: establishing a mathematical model of a dual three-phase permanent magnet synchronous motor in a six-dimensional natural coordinate system, and decoupling the six-dimensional natural coordinate system into three orthogonal subspaces; in the subspace, the fifth harmonic current and the seventh harmonic current are mapped into sixth harmonic current, and a phase current model containing fundamental waves, fifth harmonic waves and seventh harmonic waves is constructed; taking a phase current peak value as a constraint, carrying out collaborative optimization on fifth and seventh harmonics by adopting a genetic algorithm, and obtaining an optimal harmonic gain coefficient by taking maximization of a fundamental current as a target; generating a harmonic current reference sequence of the future step based on the rotor position, the angular velocity, the sampling period, the rated current and the optimal harmonic gain coefficient; and establishing a discrete prediction model of the dual three-phase permanent magnet synchronous motor, and carrying out multi-step tracking on the harmonic current reference sequence by adopting generalized prediction control. According to the invention, harmonic collaborative optimization can be realized, and the dynamic tracking performance is improved.

The invention relates to the technical field of vision measurement, in particular to a bridge cable bent tower measuring device and method based on orthogonal subspace decomposition. The method comprises the following steps: S1, decomposing a rotation matrix and optimizing an objective function; s2, optimizing a solution algorithm, and verifying convergence; s3, the anti-noise capability is enhanced through projective invariance constraint and quadratic distance constraint; and S4, predicting a target shielding pose based on the hybrid predictor. According to the bridge cable bent tower measuring device and method based on orthogonal subspace decomposition, a rotation matrix is decomposed into three orthogonal projection subspaces, and the problem that a traditional iteration method is sensitive to an initial value is avoided; secondly, on the basis of Grassmann popularity, a convex relaxation optimization method is provided, and global convergence is guaranteed; and finally, introducing algebraic geometric constraint to enhance the anti-noise capability of the algorithm.

The invention relates to a large-scale power system discrete eigenvalue parallel computing method based on surrounding channel integration. Comprising the following steps: performing contour integral spectrum transformation on a system state matrix, and determining characteristic value dominance; differentiated integral curves are designed based on characteristic value dominance, the number of characteristic values in each integral curve is estimated, and the initial dimension of the corresponding characteristic subspace is determined; determining the number of initial integral points of each integral curve by adopting a self-adaptive method, and performing self-adaptive integral point configuration on each integral curve; determining an initial subspace dimension of each block based on the number of feature values and the number of parallel integral blocks; performing singular value decomposition on a block basis matrix to construct a standard orthogonal basis; combining the orthogonal basis of each block to form a global orthogonal subspace basis; performing Rayleigh-Ritz projection in a global space to solve a dimension-reduced generalized feature value; and outputting the discrete eigenvalue and the eigenvector. And accurate and efficient calculation of discrete characteristic values of a large-scale power system is realized.

The invention discloses a Hilbert space guided physical consistency dynamic scene three-dimensional reconstruction and rendering method and system, and the method comprises the steps: 1, receiving video data of a dynamic scene, constructing a three-dimensional Gaussian point set, and initializing the static attributes of Gaussian points; selecting part of Gaussian points as control points; 2, constructing an orthogonal subspace used for describing the motion trail of the Gaussian point in the Hilbert space; 3, each Gaussian point forms a Gaussian point track according to a time sequence; learning a group of subspace coupling coefficients for each Gaussian point trajectory, and representing each Gaussian point trajectory as a linear combination of primary functions of orthogonal subspaces; 4, calculating a residual vector of each Gaussian point track; 5, updating the central position of each Gaussian point; and 6, inputting the updated Gaussian points into a rendering pipeline, and generating a rendering image corresponding to the visual angle and the time. According to the method, on the premise of ensuring physical consistency, the dynamic scene rendering precision is remarkably improved.

The application provides a spectral data cube reconstruction method and device based on compressed physical prior, equipment, storage medium and program product, relating to the technical field of spectral reconstruction, and the method comprises the following steps: obtaining a two-dimensional measurement image formed by encoding and modulation of a to-be-observed target through a spectral imaging chip; obtaining an original transmission spectrum matrix corresponding to each pixel in the spectral imaging chip; performing principal component analysis on the original transmission spectrum matrix in the spectral dimension, and extracting a principal component feature vector by a singular value decomposition method; projecting the original transmission spectrum matrix into an orthogonal subspace formed by the principal component feature vector to obtain a compressed physical prior vector; inputting the two-dimensional measurement image and the compressed physical prior vector into a deep unfolding neural network to output a high-dimensional spectral data cube of the to-be-observed target. The application can significantly improve the reconstruction efficiency of the spectral data cube.

The invention relates to the field of computer vision and camouflage target segmentation, in particular to a camouflage target segmentation method of a reversible expansion network based on SAM guidance, which comprises the following steps of: firstly, constructing a foreground space priori graph, a background space priori graph and a high-quality SAM pseudo mask by utilizing a segmentation cutting model SAM; the prior redundancy is eliminated through low-dimensional orthogonal subspace projection, and the separability of the foreground and the background is enhanced; pixel-level and gradient-level feature fitting items and SAM subspace priori constraint items are fused to construct an overall objective function, the objective function is expanded into a multi-stage alternating iteration process of a foreground optimization submodule SFOS and a background optimization submodule SBOS, and a foreground feature map and a background feature map are refined step by step; and finally, generating a camouflage target segmentation mask according to the iteratively optimized foreground feature map. Through large model prior guidance, two-stage feature modeling and multi-stage expansion optimization, the integrity and accuracy of camouflage target segmentation are significantly improved.

This application relates to the field of carpooling demand prediction technology, and particularly to a method, device, electronic device, and storage medium for predicting carpooling demand. The method includes: constructing a feature tensor set based on historical order data, meteorological parameters, and regional static feature data; transforming the target low-dimensional statistical features into target high-dimensional semantic features; and mapping the target high-dimensional semantic features to a target orthogonal subspace to generate a feature set that eliminates redundant temporal correlations. The feature set is then used to optimize the hyperparameters of a pre-constructed ensemble learning gradient boosting tree model until an iteration stopping condition is met, thus constructing a carpooling demand prediction model. This model outputs carpooling demand. This solves the problem that related technologies fail to couple the temporal and spatial dependencies of passenger travel demand, and that the prediction models are sensitive to hyperparameters, making it difficult to adapt to unconventional scenarios and output accurate carpooling demand.

The invention discloses an underwater vegetation classification method based on unsupervised learning and feature fusion. The method comprises the following steps: S1, collecting an original image; s2, inputting the original image into a deep convolutional neural network, removing redundant information, and generating a low-dimensional visual feature vector; s3, performing quantification processing on the original image through a multi-modal large model, performing ID processing on the unstructured graph according to the biological cue word, inputting the processed unstructured graph into an EAPCR-AE model, and extracting a low-dimensional high-density semantic feature vector; s4, performing L2 normalization on the low-dimensional visual feature vector and the low-dimensional high-density semantic feature vector respectively, and then performing splicing; constructing an orthogonal subspace by using a principal component analysis technology, extracting a principal component of which the cumulative variance contribution rate reaches a preset threshold value, and outputting a fusion feature vector; and S5, performing clustering and voting judgment on the fused feature vector to form a classification result. According to the underwater vegetation classification method, the feature fusion barrier of visual features and word meaning features in label classification is overcome, and the classification accuracy is high.

The invention discloses a dual three-phase permanent magnet synchronous motor model predictive current control method based on an optimized duty ratio error, and belongs to the technical field of motor control. The method comprises the following steps: firstly, constructing a six-phase mathematical model of a motor, and converting the six-phase mathematical model into an orthogonal subspace model through vector space decoupling and Parker transformation; then, establishing a cost function based on a current tracking error, and solving an analytical solution of an unconstrained modulation voltage of the cost function; on the basis, a quadratic programming problem with the minimum modulation voltage error as the target is constructed, and the optimal modulation voltage corresponding to the optimal duty ratio is solved under the linear boundary constraint. In order to efficiently solve, the optimization problem is processed by adopting a primal-dual interior point algorithm. And finally, generating modulation voltage according to the solved optimal duty ratio to drive the inverter to output. According to the invention, the dynamic response speed of the system is effectively improved, the steady-state current error and the harmonic component are obviously reduced, and the method is suitable for a high-performance motor driving system.

The invention discloses a machine forgetting method based on gradient decomposition, which comprises the following steps of: determining a forgetting sample set and a reserved sample set from a training sample set of a trained model, and respectively inputting training samples in the forgetting sample set and the reserved sample set into the trained model; respectively calculating to obtain a forgetting gradient and a retention gradient corresponding to the forgetting loss and the retention loss, carrying out orthogonal decomposition of the forgetting gradient on the retention gradient, extracting components of the forgetting gradient in an orthogonal subspace, and then carrying out gradient symbol space decomposition to obtain a positive space and a negative space of dimensions in the gradient; and fusing the forgetting gradient and the reserved gradient in different modes according to the space to which the gradient dimension belongs to obtain a final gradient, updating parameters of the trained model according to the final gradient, and repeating the process until an end condition is reached. According to the method, forgetting information and reserved knowledge are explicitly decoupled through gradient decomposition, and the interpretability and robustness of machine forgetting are improved.

This invention relates to the field of undisturbed sample transportation. To achieve hazardous transportation assessment of undisturbed samples, this application provides a method and system for assessing the transportation risks of undisturbed samples based on vibration reduction theory. The method involves acquiring the vibration of the transport container, the force and displacement of the vibration-damping support, and the constraint changes of the encapsulated soil sample cylinder to form a transportation response segment sequence. This sequence is then input into MTS-JEPA, where joint embedding mapping is performed on short-term impact and long-term cumulative scales to obtain a transportation potential state sequence. An orthogonal subspace state codeword is constructed using AMP in a soft codebook to form an orthogonal subspace codebook. The transportation potential state sequences are then categorized to form hazardous state groups and hazardous state change sequences. Finally, a hazard assessment result is generated under the structural instability threshold. This method achieves hazardous transportation assessment of undisturbed samples with high accuracy.

This invention discloses an AI-powered intelligent monitoring method for enterprise business stability based on multimodal data fusion, relating to the field of enterprise business stability monitoring. The method includes: collecting and preprocessing multimodal data to obtain an aligned multimodal dataset; extracting topological features from the call chain data in the aligned multimodal dataset and mapping them to a hypergraph adjacency matrix to reconstruct the service dependency hypergraph; converting the indicator data, log data, and call chain data in the aligned multimodal dataset into discrete features of symbolic semantic tokens, performing metric orthogonal subspace alignment under topological constraints to obtain an orthogonally aligned multimodal feature set; and obtaining the current business activity type and priority information, encoding it into a business semantic vector reflecting business stability requirements. This invention achieves enterprise business stability monitoring effects that adapt to business stability requirements and realize accurate anomaly detection and intelligent root cause localization across the entire link dimension.

The invention provides a continuous learning low-rank adaptive method, and belongs to the technical field of continuous learning-oriented image processing. The method comprises the following steps: dividing an obtained image task data set into a plurality of sequential tasks; constructing a low-rank adaptive updating structure, freezing the original weight of the pre-training model, and performing updating modeling on the original weight; constructing a feature matrix and constructing a low-dimensional orthogonal subspace basis matrix; calculating a standard gradient updating direction to obtain a safety gradient direction; on the basis of the safety gradient direction and the current low-rank matrix, updating the low-rank parameter to approach the safety gradient; constructing feature separation loss; in an image processing process for continuous learning, constructing an overall optimization objective function; and performing low-rank self-adaptive fine tuning by using the constructed overall optimization objective function. According to the method, the problem of disastrous forgetting of the model on the prior knowledge of the historical image task when the model learns the new image task is effectively relieved.

The invention discloses a blast furnace air permeability index prediction method based on multi-scale time-delay characteristic mining and application, and belongs to the field of soft measurement modeling in the blast furnace ironmaking process. The invention provides a blast furnace air permeability index prediction framework combining wavelet decomposition, wavelet coherence analysis and a time-frequency fusion model aiming at the characteristics of nonlinearity, unsteady state and large time delay of blast furnace operation data. The method comprises the following steps: firstly, carrying out peak separation and multi-scale decomposition on data, and extracting multi-scale time delay information among variables by utilizing wavelet coherence analysis; and then modeling time domain global features and frequency domain local features by adopting orthogonal subspace analysis and a wavelet neural network, and performing model fusion through Gauss-Markov estimation to realize advanced multi-step prediction of the air permeability index. The dynamic time lag characteristic of the blast furnace process can be effectively captured, the prediction precision and robustness are improved, and technical support is provided for stable operation of the blast furnace.

The application discloses a multi-modal large language model passive forgetting method based on an agent anchor point, and aims at solving the problem that original private data cannot be accessed in a privacy compliance scene. First, a text-guided coarse-to-fine retrieval strategy is adopted, cross-modal feature alignment is utilized to accurately locate an agent anchor point which overlaps with a target semantic from a public data set, and a substitute supervision signal is constructed. Secondly, a double-constraint semantic isolation optimization is implemented. On one hand, a text-anchor semantic repulsion mechanism is introduced to cut off a visual-induced link of a target concept in a feature space, and accurate erasing is realized. On the other hand, a zero-space projection technology is utilized to strictly limit gradient updating in an orthogonal subspace which retains knowledge, and feature isotropic regularization is used to prevent manifold collapse. While completely forgetting sensitive concepts, the method effectively guarantees the general perception and reasoning ability of the model, and significantly reduces the risk of catastrophic forgetting.

This invention relates to a two-stage federated distillation and large model fine-tuning method based on differential privacy, belonging to the fields of artificial intelligence security, federated learning, and differential privacy technology. The method includes: Step 1, federated pre-training based on complexity adaptive budget management and momentum orthogonal subspace directional noise addition; Step 2, local dataset distillation based on frequency domain-aware complex plane anisotropy noise addition; and Step 3, cloud-based large model fine-tuning and end-to-end privacy compliance assessment based on noise prior-driven adaptive rank. This invention achieves precise allocation and tracking of the privacy budget in both the federated pre-training and data distillation stages by introducing mathematically provable RDP differential privacy protection. It completes high-fidelity generation of the distilled dataset and secure large model fine-tuning while strictly protecting the privacy of the client's original data, which is of great significance for promoting the secure and efficient deployment of large models in distributed privacy-sensitive environments.

This invention discloses a passive forgetting method for multimodal large language models based on surrogate anchors. Addressing the challenge of inaccessible original private data in privacy-compliant scenarios, this invention first employs a text-guided coarse-to-fine retrieval strategy. It utilizes cross-modal feature alignment to accurately locate surrogate anchors semantically overlapping with the target from public datasets, constructing alternative supervision signals. Secondly, it implements a dual-constraint semantic isolation optimization: on one hand, it introduces a text-anchor semantic exclusion mechanism to sever the visual triggering link of the target concept in the feature space, achieving precise erasure; on the other hand, it uses null space projection technology to strictly restrict gradient updates to an orthogonal subspace that preserves knowledge, and combines this with isotropic feature regularization to prevent manifold collapse. This method effectively safeguards the model's general perception and reasoning capabilities while completely forgetting sensitive concepts, significantly reducing the risk of catastrophic forgetting.

The application provides a kind of attack detection method, system and storage medium based on bayesian incremental learning, method includes: step one: data set is collected, and divided into multiple tasks according to year;Step two: using bayesian continuous learning framework, the posterior distribution obtained by learning last task is used as the prior distribution of current task, gradient projection method is used, the gradient of current task is projected to the orthogonal subspace of old task feature space, and projection parameter is obtained;Step three: label bias and noise processing are carried out on task;Step four: minimize new task loss, obtain parameter, and take training strategy according to the result of label bias and noise processing;Step five: find optimal merging solution;Step six: the model parameter after merging is used as the initialization parameter of next task, returns step two, enters next round iteration, until the training of all tasks is completed.The beneficial effects of the application are: it can significantly improve the knowledge retention ability and effectively overcome the problem of catastrophic forgetting.

This invention relates to a two-stage federated distillation and large model fine-tuning method based on differential privacy, belonging to the fields of artificial intelligence security, federated learning, and differential privacy technology. The method includes: Step 1, federated pre-training based on complexity adaptive budget management and momentum orthogonal subspace directional noise addition; Step 2, local dataset distillation based on frequency domain-aware complex plane anisotropy noise addition; and Step 3, cloud-based large model fine-tuning and end-to-end privacy compliance assessment based on noise prior-driven adaptive rank. This invention, by introducing mathematically provable RDP differential privacy protection in both the federated pre-training and data distillation stages, achieves precise allocation and tracking of the privacy budget in both stages. It completes high-fidelity generation of the distilled dataset and secure fine-tuning of the large model while strictly protecting the privacy of the client's original data, which is of great significance for promoting the secure and efficient deployment of large models in distributed privacy-sensitive environments.

The invention provides a PINN sound field reconstruction coding frequency optimization method based on feature space estimation, and the method comprises the steps: firstly carrying out the feature embedding of a space coordinate through a trigonometric function, calculating a neural tangent kernel on this basis, and carrying out the feature value decomposition of the neural tangent kernel; according to the size of the feature value, the feature vector is divided into two orthogonal subspaces: the feature vector corresponding to the non-zero feature value is stretched into a signal subspace, and the feature vector corresponding to the zero feature value is stretched into a noise subspace; then, correlation analysis is carried out on the noise subspace and a sound pressure vector obtained through sampling, and a function relation with the coding frequency as an independent variable and the orthogonality as a dependent variable is established; when the orthogonality of the function reaches the maximum value, the correlation between the noise subspace and the sampling sound pressure vector is the lowest, and according to the orthogonality of the subspace, the signal subspace and the sampling sound pressure vector have the highest correlation. And the frequency is used as a final feature embedded coding frequency for subsequent PINN training, so that sound field reconstruction is completed.