16 results about "Orthogonal complement" patented technology

The invention relates to a trajectory prediction model robustness enhancement method based on dynamic subspace projection decomposition. Comprising the following steps: firstly, extracting hidden layer semantic features containing historical tracks and map topology through a multi-modal feature encoder; secondly, constructing a dynamic routing mechanism based on scene self-adaption, and calculating projection weights of input features on a plurality of expert subspaces; then, executing truncation projection operation based on orthogonal decomposition, retaining core semantics located in a low-dimensional space, and filtering out adversarial disturbance located in an orthogonal complementary space; and finally, introducing a feature consistency constraint training mechanism, taking the reconstructed features of the clean sample as anchor points, and compulsively aligning the purified features of the confrontation sample with the anchor points. Compared with the prior art, the method has the advantages that the robustness of the model in white box gradient attack, black box query attack and physical semantic deception scenes is remarkably improved through feature purification of a physical level and structured consistency constraint, and the prediction reliability of the automatic driving system is ensured.

The invention provides a single-bit digital beam forming method and device, a terminal and a medium, and relates to the technical field of radar detection and wireless communication, and the method comprises the steps: constructing a target single-bit low-rank matrix recovery model of a uniform linear array comprising a plurality of sensors, and an optimization problem of the target single-bit low-rank matrix recovery model, and converting the target single-bit low-rank matrix recovery model into a processable problem, solving and determining a low-rank matrix by using a near-end block coordinate descent algorithm and a gradient descent algorithm, calculating an estimated covariance matrix of a target and an interference signal based on the matrix, performing eigendecomposition to calculate an interference subspace, and modeling a single-bit digital beam forming problem according to a base vector of the interference subspace to obtain a single-bit digital beam forming matrix. And calculating a closed-form solution based on the principle that the optimal solution is located in the orthogonal complement space of the interference subspace. According to the method, the resolution requirement of the analog-to-digital converter can be reduced by fully utilizing the single-bit quantization characteristic, and the beamforming performance of single-bit digital beamforming is close to that of high-bit digital beamforming when the hardware complexity and the power consumption are reduced.

The application discloses a kind of double-channel collaborative backdoor defense methods and systems for longitudinal federal learning, comprising: active party equipment is according to the difference between local prediction result and true label Complementary supervision target for constraining passive party aggregation output is built, the sum of each element of complementary supervision target is zero;Passive party uploads local output, and active party calculates complementary supervision loss after weighted aggregation;Active party calculates original gradient according to the loss, and after grouping statistics according to true label, category discriminant gradient subspace is extracted;In subspace and its orthogonal complement space, respectively set perturbation intensity and generate structured perturbation, superimposed to original gradient after return passive party, and passive party updates local model accordingly, and active party synchronously updates local model and aggregation weight.The application weakens the ability of association establishment of trigger and target category by controlling learning target channel and gradient feedback channel, from the source of backdoor formation, while maintaining main task performance, effectively defending training stage backdoor attack.

The invention provides a security alignment method based on neural network parameter subspace restoration, which is applied to a federated learning server side, does not access private training data of each participant in the execution process of the method, and executes the following steps in each round of distributed cooperative training: based on the global neural network model parameters of the current round, performing the subspace restoration on the basis of the global neural network model parameters of the current round; respectively obtaining a security parameter increment vector and an unsecurity parameter increment vector; calculating to obtain a difference parameter vector, and decomposing to obtain an orthogonal basis matrix of a danger parameter subspace; projecting the to-be-repaired global parameter updating vector to an orthogonal complementary space of the danger parameter subspace based on the orthogonal basis matrix so as to remove a projection component of the to-be-repaired global parameter updating vector on the danger subspace; and fusing the weighted safety parameter increment vector and the projected parameter updating vector to obtain a repaired global parameter updating vector so as to update the parameters of the global neural network model.

A computer-implemented method includes receiving at a processor an input set of multiway data. The method determines a set of persistent homology barcodes based on the multiway data and identifies at least a first significant persistent homology barcode in the set of persistent homology barcodes. A representative cycle of the first significant persistent homology is returned and an orthonormal basis for a boundary of the multiway data is computed. The method obtains a harmonic representative by computing a projection of the representative cycle to an orthogonal complement and generates a set of feature vectors using the harmonic representative. Each feature vector has a magnitude corresponding to an impact that feature has on a given condition. The method then ranks the feature vectors by feature vector magnitude.

Embodiments of the present application disclose a fluid steady-state characteristic determination method and related device, the method comprising: obtaining physical properties and state parameters of a fluid in a target system; constructing a fluid dynamics equation according to the physical properties and the state parameters, and discretizing the fluid dynamics equation into a linear equation set, the linear equation set being expressed as a product of a coefficient matrix and a vector corresponding to unknown quantities being equal to a constant vector; constructing a quantum circuit to solve the linear equation set according to a target unitary operator, and obtaining steady-state characteristics of the fluid, the target unitary operator acting on two registers, when a quantum state of one of the registers represents an arbitrary row of the coefficient matrix, the target unitary operator performs a NOT operation on the other register, and when the quantum state of one of the registers represents an orthogonal complement space of an arbitrary row of the coefficient matrix, the target unitary operator does not perform any operation on the other register. By using the embodiments of the present application, the steady-state characteristics of the fluid can be quickly determined.

The application provides a secure alignment method based on a neural network parameter subspace repair, which is applied to a federal learning server side, and private training data of each participant is not accessed in a method execution process. In each round of distributed collaborative training, the following steps are performed: based on a current round global neural network model parameter, a secure parameter increment vector and an unsafe parameter increment vector are respectively obtained; a difference parameter vector is calculated and decomposed to obtain an orthogonal basis matrix of a dangerous parameter subspace; based on the orthogonal basis matrix, a to-be-repaired global parameter update vector is projected to an orthogonal complement space of the dangerous parameter subspace to remove a projection component of the to-be-repaired global parameter update vector on the dangerous subspace; the weighted secure parameter increment vector and the projected parameter update vector are fused to obtain a repaired global parameter update vector to update a parameter of the global neural network model.

The invention relates to the technical field of image data processing, and discloses a Chinese ancient book text restoration method and system based on big data, and the method comprises the steps: constructing a homologous font skeleton index database, and extracting the topological nodes and curvature features of the same version stroke skeleton; recognizing a damaged area and extracting a residual skeleton, and matching an optimal homologous skeleton; constructing a thin-plate spline interpolation function and executing non-rigid geometric deformation to generate a target guide skeleton; calculating a damaged region neighborhood background image gradient field to construct a background interference tensor field, and extracting an interference direction vector; constructing an interference vector orthogonal complement space and projecting the skeleton direction vector to the interference vector orthogonal complement space to obtain a net skeleton direction vector; a structure tensor field is reconstructed based on a net skeleton direction vector, texture mapping is guided, through topological isomorphic mapping and orthogonal tensor field constraint, the coupling problem of stroke structure distortion and background noise interference in ancient book repair is solved, and high-fidelity structure reconstruction and texture fusion under a complex engraving style are achieved.

Embodiments of the present disclosure disclose an intelligent scheduling control method in a complex environment based on imitation learning. The specific implementation of the method comprises: calling a large language model to analyze text and convert it into a structured data set, and switching to a conditional random field model when it fails; calculating the minimum Euclidean distance between the environment context vector and the historical expert trajectory matrix to establish the state trajectory envelope drift distance, and when it is greater than a threshold, reconstructing the local expert trajectory segment and refreshing the weight; projecting the policy update gradient to the orthogonal complement space of the expert behavior cloning gradient, calculating the actual update gradient and refreshing the policy parameters to output instructions. This implementation suppresses parameter oscillation caused by state deviation, restricts the update direction using gradient space orthogonality, avoids weight rewriting, and ensures stable scheduling control flow.

The invention discloses an iterative generalization remote sensing recognition model construction method and device and a medium, and the method comprises the steps: carrying out the feature extraction of an initial remote sensing image sample, constructing a semantic topological graph, and constructing an initial sample library through combining the importance of the sample and the similarity between the samples; an orthogonal constraint loss function is adopted, and correlation between features extracted by a target sensing head and a background sensing head of the base model is optimized based on the initial sample library; the optimized base model is used for monitoring new scene remote sensing image samples, and the prediction entropy and the feature drift distance of each sample are calculated to establish a domain adaptation data set training adaptation module; and projecting the parameter variation trained by the adaptive module to an orthogonal complementary space of a first feature subspace determined based on the optimized base model and the initial sample library, and constructing an iterative generalization remote sensing recognition model through parameter fusion. According to the method, efficient adaptation to a new scene is realized, learned old knowledge is not disturbed, and the problem that the old knowledge is forgotten is solved.

The invention belongs to the technical field of network security, and particularly relates to a continuous learning network intrusion detection method and system based on gradient projection and concept drift detection. The problems of serious concept drift, disastrous forgetting, new and old knowledge gradient interference and the like of an existing intrusion detection system in a dynamic network environment are solved, network flow data flow is monitored in real time, variance fluctuation of feature dimensions is monitored through a drift detection module based on principal component analysis (PCA), and unsupervised identification of distribution offset is achieved; a strategic sample selection and forgetting mechanism is adopted, a memory buffer area is optimized based on KL divergence, and balanced learning of normal traffic and an evolution abnormal mode is achieved under limited resources; a gradient projection mechanism (GPM) is introduced in a model fine tuning stage, and a parameter updating direction of a new task is limited in an orthogonal complement space of historical knowledge by constructing an orthogonal basis of a historical feature subspace, so that destructive interference between tasks is eliminated on a gradient level.

The invention discloses a video diffusion model concept erasing method and system based on two-stage adjustment, and belongs to the technical field of artificial intelligence generated content security. Based on prompt embedding and target concept embedding, respectively generating three groups of projection matrixes of a prompt subspace, a target concept subspace and an orthogonal complement space of the prompt subspace, the target concept subspace and the orthogonal complement space; analyzing and screening trigger words through an orthogonal projection distance, and performing space projection correction on prompt embedding; a dynamic scaling factor is calculated by fusing correction prompt and target concept embedding, and the denoising direction is adjusted; and carrying out multi-step denoising under the corrected noise estimation, and outputting a high-quality video without a target concept. According to the method, a diffusion model does not need to be trained, a mainstream text is compatible to a video diffusion model in a plug-and-play mode, various tasks such as object deletion, artistic style stripping, celebrity portrait hiding and sensitive content erasing are supported, the generation probability of a target concept can still be remarkably reduced in a countermeasure attack scene, and the method is high in practicability. And non-target content generation quality and video high fidelity are compatible.

The invention discloses a power distribution network fault multi-source heterogeneous feature recognition method, system and device and a medium, and the method comprises the steps: constructing a multi-source heterogeneous data base covering internal electrical quantity and state quantity and external weather and geographical environment, and improving the quality through employing a data management technology; performing time-frequency localization feature extraction on the transient electrical waveform by using wavelet transform, and constructing a Bayesian network model to process logic uncertainty to realize orthogonal complementation of physical waveform and logic state features; mapping the heterogeneous data to a unified space-time coordinate system through a three-level space-time grid mechanism; constructing a multi-time scale fault prediction model system comprising SVR, LSTM, VAE and GCN; and realizing differentiated operation and maintenance decision recommendation through the power distribution network fault multi-mode knowledge graph. According to the method, the fault identification accuracy and the emergency response efficiency of the power distribution network in a complex environment are remarkably improved.

A computer-implemented method includes receiving at a processor an input set of multiway data. The method determines a set of persistent homology barcodes based on the multiway data and identifies at least a first significant persistent homology barcode in the set of persistent homology barcodes. A representative cycle of the first significant persistent homology is returned and an orthonormal basis for a boundary of the multiway data is computed. The method obtains a harmonic representative by computing a projection of the representative cycle to an orthogonal complement and generates a set of feature vectors using the harmonic representative. Each feature vector has a magnitude corresponding to an impact that feature has on a given condition. The method then ranks the feature vectors by feature vector magnitude.

This invention discloses a method for obtaining beamforming weight vectors in an integrated radar-communication system, within the field of array signal processing technology. The method comprises the following steps: S1: Establishing a receiver signal model for a uniform linear array antenna in the integrated radar-communication system; S2: Introducing orthogonal polynomials and the Gauss-Legendal numerical integration method to perform a low-complexity, fast fitting solution for the interference subspace; S3: Introducing the Schmitt orthogonalization method to recursively solve the orthogonal complement projection operator of the interference subspace and project it to obtain the beamforming weight vector; S4: Introducing an equivalent representation of the beamforming weight vector under a dual-phase shifter architecture to obtain the updated beamforming weight vector value under the dual-phase shifter architecture. This beamforming weight vector acquisition method solves the problems of high computational complexity, poor performance in low-quantization-bit hardware scenarios, and desired signal suppression under high signal-to-interference-plus-noise ratios in existing methods, achieving low-complexity and highly robust beamforming.