149 results about "Fisher information" patented technology

The invention provides an intelligent query semantic understanding method based on information geometry and Riemannian manifold, and the method comprises the steps: constructing a query semantic space based on Riemannian measurement, extracting query geometric features through a Fisher information matrix, and achieving the precise recognition of a query intention in combination with Wasserstein distance constraint and geodesic optimization. According to the method, an adaptive optimization mechanism is established, a parallelization calculation strategy is adopted, and multi-modal feature fusion and robustness enhancement are achieved. Experiments show that compared with the prior art, the query understanding accuracy, the processing time delay and the semantic representation compression rate of the method are well improved, and the semantic understanding performance in a complex query scene is remarkably improved. The method can be widely applied to the fields of search engines, intelligent questions and answers, knowledge maps and the like, and has important practical value.

The invention belongs to the technical field of federated learning privacy protection, and particularly relates to a heterogeneous federated learning adaptive privacy protection method and device based on momentum optimization. The method comprises the steps that feature extractor parameters broadcasted by a central server are received, the local gradient of a current iteration round is calculated based on a local data set, and a momentum mechanism is introduced into local gradient updating for a classifier part; a gradient descent algorithm is applied, and local model parameters are updated through the local gradient obtained through calculation; for the feature extractor part, calculating a Fisher information matrix; performing normalization processing on the Fisher information of each layer, and screening strategies; and then differential privacy processing is carried out, the global model parameters are uploaded to a central server for parameter aggregation, and updated global model parameters are broadcasted back to each local client for next round of training. According to the method, the problems of negative influence of a differential privacy mechanism on model training, slow model convergence and poor privacy protection effect are solved.

The invention provides a digital twin model incremental learning method and system and an equipment state prediction method.The digital twin model incremental learning method introduces an incremental learning mechanism based on elastic weight merging to train an industrial equipment state parameter prediction model; the current working condition samples are stored in an experience playback buffer area, the experience playback buffer area adopts a circular queue structure, the samples stored in the buffer area are used in a mixed mode for training when a new working condition is learned, and mixed loss optimization is executed on a current data batch; after batch training is completed, Fisher information matrix diagonal terms, about historical data, of current parameter configuration are calculated firstly, then global parameter importance is updated until Neural ODE incremental training converges, and finally a complete model state tuple fusing multi-working-condition knowledge is output; the incremental Neural ODE can continue to learn knowledge of new working conditions on the premise that knowledge of old working conditions is not forgotten, and the sustainable evolution ability of the digital twin under the limited resource condition is achieved.

The invention provides a rolling bearing digital twinning dynamic evolution method and system based on continuous learning, and belongs to the technical field of bearing life prediction. The method comprises the following steps: processing a bearing monitoring signal through short-time Fourier transform to generate standardized time-frequency data; constructing health indexes based on the index degeneration function and dividing health levels; training a life prediction model by using the extended LSTM network and taking the time-frequency data as input; real-time data is collected through a fixed time window to predict the life, and when the root-mean-square error of a predicted value and an actual value exceeds the limit, the edge device is triggered to upload new data; evaluating parameter importance in combination with a Fisher information matrix, and dynamically adjusting a regularization intensity updating model; and monitoring the data standard deviation in real time, triggering shutdown when the data standard deviation exceeds the limit, otherwise, predicting the remaining life by updating the model, and stopping when the remaining life reaches the threshold value. According to the method, dynamic evolution of the digital twin model is realized through continuous learning, and the industrial equipment state monitoring and predictive maintenance capability is effectively improved.

The invention discloses a lithium battery health state prediction method based on a feature screening technology. The lithium battery health state prediction method based on the feature screening technology comprises the following steps: obtaining a data set; extraction and primary processing of initial features; screening a feature set; preliminarily training a machine learning model; evaluating and optimizing a machine learning model; and predicting and verifying the health state of the lithium battery. The method indirectly reveals the complex relationship between the features and the target output by evaluating the uncertainty of the parameter estimation value by using the effective Fisher information matrix-based feature screening technology, not only can effectively capture the linear and nonlinear interaction between the features, but also can reflect the combined influence of each feature in the multi-parameter model, and can effectively improve the target output. A more comprehensive and accurate feature selection basis is provided; besides, the Fisher information matrix-based feature screening technology is suitable for linear and nonlinear models, and particularly shows relatively high robustness and precision when processing a multi-parameter model with complex interaction.

The invention discloses a model fine tuning method and system based on task priority and elastic weight solidification, belongs to the technical field of machine learning, and aims to solve the technical problem of how to provide a stable, efficient and low-power silk screen manufacturing model which can adapt to various complex production scenes. Comprising the following steps: collecting sample data, and dividing the sample data into a plurality of subsets; service weights and priority coefficients of all new tasks in the silk screen manufacturing process are analyzed based on the sorted subsets; constructing a knowledge graph based on the entities and the relationship between the entities; calculating the importance degree of the model parameters to the old task based on the Fisher information matrix, and adjusting the priority level of the model parameters based on the knowledge graph; and on the basis of the Fisher information matrix, updating a loss function of the silk screen manufacturing model under the new task through EWC, and correcting the current silk screen manufacturing model on the basis of entities and relationships in the knowledge graph, so that multi-task model fine tuning is realized.

The invention discloses an intelligent power grid equipment fault monitoring system and method based on the Internet of Things, and relates to the technical field of the Internet of Things. Discrete Fourier transform and a Fisher information processing method are adopted to fuse multi-source data, meteorological interference is corrected, and accurate collection of electrical quantity and quantification of environmental influence are achieved; and setting a fault monitoring sliding window and a harmonic analysis time interval, and obtaining harmonic data of different time scales. Extracting harmonic features through slope calculation, and identifying voltage abnormal features; historical data is processed according to a time sequence prediction algorithm to construct a prediction set, voltage abnormity is judged based on a historical data threshold value, intelligent early warning of the voltage state is achieved, the abnormity capturing precision is improved, and when the system detects the voltage abnormity, information of strain section equipment sending out a fault early warning signal is obtained and transmitted to an OTA upgrading terminal. And enabling the OTA upgrading terminal to carry out state identification and data interaction on the upgrading process of the intelligent power grid equipment based on the equipment information.

The invention discloses a sensor optimization arrangement method for vibration monitoring of a T-shaped bridge, which comprises the following steps of: firstly, preliminarily drawing up a measuring point group and determining an interested modal order through a finite element model and / or a modal result of the T-shaped bridge obtained by field measurement; then, based on a modal shape matrix of a bridge in a finite element model result, deriving a maximum diagonal element value of a corresponding MAC matrix and a trace of a Fisher information matrix, and increasing and decreasing the degree of freedom of the MAC matrix and the Fisher information matrix through a correction matrix so that the MAC matrix and the Fisher information matrix meet monitoring requirements; comprehensively considering whether the values of the two parameters are in the optimal range or not based on actual measurement and historical data; the measurement point group is the optimal arrangement mode of the vibration sensor if the measurement point group is in the optimal range, and the measurement point group is the optimal arrangement mode of the vibration sensor if the measurement point group is not in the optimal range, the correction matrix acts on the MAC matrix and the Fisher information matrix respectively, the corrected parameter values are compared, and the measurement point group under the optimal parameter value is taken as the optimal arrangement of the vibration sensor.

The invention discloses an anti-shielding three-dimensional scanning measurement field construction method for large-size measurement, and belongs to the technical field of optical three-dimensional measurement. The method comprises the following steps: firstly, performing adaptive sampling based on geometric features of a workpiece to generate a candidate mark point set; secondly, constructing a virtual measurement scene, simulating a dynamic scanning process by using a ray tracing technology, and calculating the visibility and observation quality of each point under multiple viewpoints; then measuring precision is quantified by using a Fisher information matrix, and a layout optimization model which takes maximization of information gain as a target and meets coverage rate and engineering constraint at the same time is established; and finally, solving the model to obtain an optimal mark point subset, and verifying and outputting the optimal mark point subset. According to the method, the problems of measurement interruption, low point distribution efficiency and difficulty in guaranteeing precision caused by shielding of a large-size complex component in scanning are solved, and automatic construction of a high-robustness and high-precision measurement field is realized.

The invention discloses a Fisher information matrix-based deformation monitoring sensor optimization arrangement method. The method comprises the steps of defining observation variables and unknown parameters; establishing a relation matrix; calculating a probability density function and a quality function of the observation variable; calculating a Fisher information matrix according to the deformation reconstruction equation; the traces of the Fisher information matrix are maximized to achieve optimal arrangement of the sensors. According to the method, the relation between the Fisher information matrix and the deformation reconstruction precision is deduced from the basic principle of an inverse finite element, optimal arrangement of the sensor is completed by maximizing the trace of the Fisher information matrix represented by the overall stiffness matrix, and efficient optimal arrangement of the sensor is achieved.

The invention discloses a profile optimization design method in a positioning system based on an RIS, and the method comprises the steps: building a system model, determining the system composition elements, namely an access point (AP), the RIS, UE and the position relation of the AP, the RIS and the UE, and building a signal model and a measurement model. Fisher information analysis is carried out, a correlation matrix and parameters are calculated, and a position error boundary expression is deduced. And then an optimal RIS profile design strategy is provided, an optimization problem is constructed, and an optimal covariance matrix and RIS profile design are obtained through matrix analysis and derivation. Meanwhile, hardware limitation is considered, constrained RIS profile design is proposed, an optimization problem with constraints is constructed, and an RIS profile meeting actual hardware constraints is obtained through solving by an alternating minimization technology. The present invention improves the positioning accuracy of user equipment (UE).

The invention discloses a joint communication and perception method based on FIM-CRB optimization. The joint communication and perception method comprises the following steps: establishing an RIS-assisted joint communication and perception system model; constructing an echo signal model of a sensing link in the system, and deducing vectorization expression of an echo signal; constructing a Fisher information matrix FIM, and further deducing a lower limit CRB of perception parameter estimation precision; constructing a joint optimization problem, taking the minimum CRB as a target function, introducing the minimum signal to interference plus noise ratio constraint and the total transmitting power constraint of a communication user, and jointly optimizing a transmitting beam matrix of the BS and a reflection matrix of the RIS; and equivalently reconstructing the joint optimization problem into a solvable convex optimization problem by using a positive semidefinite programming and fractional programming method, and iteratively updating the transmitted beam, the RIS reflection matrix, the sensing beam matrix and the auxiliary variables by using an alternate optimization algorithm to finally obtain an optimization result. The method can effectively improve the sensing precision and robustness of the system, and is suitable for a 6G sensing communication network and an intelligent environment construction scene.

The invention relates to the technical field of large language model privacy protection, in particular to a large language model black box forgetting method based on a double-auxiliary model, which comprises the following steps of: constructing the double-auxiliary model and generating output difference characteristics, analyzing parameter criticality through a Fisher information matrix to carry out self-matching fine tuning, calculating a weighted forgetting relevance score, and carrying out self-matching fine tuning on the weighted forgetting relevance score. And a forgetting response is generated based on score matching target model output. According to the method, the forgetting demand is judged through the output difference of the double-auxiliary model, the self-matching fine-tuning optimization parameter updating strategy is utilized, the forgetting relevance is calculated in combination with the weighted cross entropy and the target model loss, the target information is accurately forgotten in the black box environment, meanwhile, the performance of the model on the non-forgetting task is maintained, and the forgetting efficiency is improved. And the practicability and efficiency of privacy protection of the large language model are improved.

The invention discloses a load prediction method and system based on error division multiple tasks, a medium and equipment, and the method comprises the steps: combining an error division task with a continuous learning mechanism, firstly calculating a sample relative error in a training process, classifying low-error data into a data set corresponding to a learned load feature mode, and carrying out the calculation of a data set corresponding to the learned load feature mode; the high-error data is delimited as a new task to be learned; measuring model parameter importance through a Fisher information matrix of EWC, introducing an L2 regularization item to constrain core parameter updating, and avoiding forgetting a learned old mode; and circularly executing task division and model training, so that a single model gradually masters various load characteristic modes in continuous learning. A division basis does not need to be set manually, a multi-mode learning process is simplified, and prediction accuracy and robustness in a complex load scene are improved.

The invention relates to the technical field of computers, and discloses an industrial object-oriented Gaussian approximation filtering method for a nonlinear system. Mathematical modeling is carried out on a physical system to be subjected to state estimation, and a nonlinear state space model corresponding to the physical system is obtained; matching the first Gaussian distribution of the system state at the time t to a given distribution through a moment matching operation based on the non-linear state-space model to determine a predictive estimate and a predictive estimate error of the system state at the time t; acquiring a system observation value at the moment t; iteratively updating the optimal estimation and the optimal estimation error of the system state at the moment t based on a natural gradient descent method and a Fisher information matrix associated with the second Gaussian distribution of the system state at the moment t, and obtaining the optimal estimation and the optimal estimation error at the moment t until an iteration stop condition is met; linearization errors can be avoided, and more accurate system state estimation of a nonlinear system is provided.

The invention discloses a multi-modal ultrasonic Lamb wave imaging spatial resolution quantification method, and relates to the technical field of ultrasonic Lamb wave imaging, and the method specifically comprises the steps: 1, employing an array transducer to carry out point-by-point excitation and receiving on the basis of a full-focus method data collection mode, collecting complete full-matrix data, building a multi-modal Lamb wave propagation model, and carrying out the calculation of the multi-modal Lamb wave propagation model; the method comprises the following steps: step 1, constructing a corresponding Fisher information matrix based on partial derivative information of each channel in combination with observation noise variance, and step 2, carrying out CRLB numerical calculation by adopting an MATLAB platform based on a derivation result of the Fisher information matrix. Physical relations among the defect imaging precision, the wavelength and the modal characteristics are established theoretically, and a quantitative basis is provided for multi-modal selection and system design.

The invention relates to the technical field of smart power grids and communication, and discloses a power utilization information acquisition method and system based on dual-mode communication, and the method comprises the steps: 1, initializing a configuration network, building the neighborhood of each acquisition terminal node, carrying out the probability distribution parameterization of the channel states of a power line carrier channel and a wireless radio frequency channel, and carrying out the network configuration; according to the probability distribution parameterization, calculating a Fishi information matrix used for quantifying the information distance between different channel states; and step 2, defining a channel state geometrical relationship based on the Fishi information matrix. The probability distribution parameter space of the channel state is constructed into the information geometric manifold, the manifold Kalman filtering algorithm is adopted for state prediction, and compared with an existing channel prediction scheme adopting standard Kalman filtering or a simple time sequence model, the problem that the statistical characteristics of the channel cannot be accurately described due to model mismatch is solved, and the channel prediction accuracy is improved. Therefore, the problems of low prediction precision and incapability of effectively supporting a prospective routing decision are solved.

The invention discloses a three-dimensional sensor optimal arrangement method and system based on a genetic algorithm, and relates to the technical field of three-dimensional sensor optimal arrangement, and the method comprises the following steps: building a finite element model; selecting a proper number of candidate measuring points by combining the size of a grid unit in the finite element model, and determining a to-be-selected measuring point set of the sensor; the vibration mode extraction number is determined, and modal information extraction is carried out; based on the mode number selection of the Fisher information matrix, the Fisher information matrix of the structure is constructed by utilizing the mode information of each measuring point of the structure obtained through mode analysis calculation; constructing a three-dimensional sensor optimization arrangement model, and establishing the sensor optimization arrangement model based on a three-dimensional modal confidence criterion; and solving the three-dimensional sensor optimization arrangement model by using a genetic algorithm to obtain a sensor arrangement scheme. According to the method, the arrangement position of the sensor is optimized, so that the recognition precision of the structural modal parameters is remarkably improved, and accurate capture of key modal information is ensured.

The invention discloses a method for optimizing a three-dimensional sensor in an operation mode, which combines an effective independence method with redundancy, sorts candidate sensors according to contribution of candidate sensor positions to linear independence of a target mode, and removes an unimportant position through iteration each time. The final sensor configuration tends to maximize the determinant of the Fisher information matrix, so that the corresponding target modal shape is linearly independent as far as possible, and meanwhile, the signal strength of the target modal response in the test data is maximized, so that the modals are distinguished as far as possible, the signal-to-noise ratio is improved, and the influence of noise on the test is reduced. According to the invention, the number of required sensors can be minimized while the accuracy of data collection is ensured, so that the potential influence of the sensors on the structure function is reduced, and the test cost is reduced.

The invention discloses a streaming federal learning method, system and device based on human-like key learning period plasticity, and the method comprises the steps: carrying out the online training of each local deep learning model through a client, and calculating a Fischer information matrix trace of a model parameter; the trained local deep learning models and the corresponding Fischer information matrix traces are uploaded; the server carries out asynchronous aggregation on each local deep learning model, and carries out weighted aggregation based on lag attenuation on all Fischer information matrix traces to obtain a federated Fischer information trace of an asynchronous streaming federated model; the method comprises the steps of determining whether an asynchronous streaming federated model is in a key learning period or not on the basis of a federated Fischer information trace, dynamically adjusting training hyper-parameters of the asynchronous streaming federated model on the basis of the federated Fischer information trace, and training the asynchronous streaming federated model. The model is enabled to fully learn diversified category knowledge, and the problem of time distribution over-fitting in streaming data is effectively relieved.

The invention discloses an industrial mechanical arm joint rigidity online identification method based on an optimal excitation track, and belongs to the field of mechanical arm parameter identification. The method comprises the following steps: firstly, establishing a mechanical arm rigid-flexible coupling dynamic model, and constructing a linear regression equation taking joint rigidity and viscous damping as parameters; a Fisher information matrix is derived based on the regression model to quantify the amount of information of the excitation trajectory. And further, parameterizing the excitation trajectory by using finite term Fourier series, and solving the optimal excitation trajectory by using a scalar index of a Fisher information matrix and a sustainable excitation condition as optimization targets under the condition of meeting joint motion constraints. And finally, the mechanical arm is driven to execute the track, motion data are collected and input into parameter identification algorithms such as unscented Kalman filtering and recursive least square, and online and high-precision estimation of the time-varying joint stiffness is achieved. According to the method, the excitation track is designed directly from the information theory, and the precision, robustness and real-time performance of joint rigidity identification are effectively improved.

The application relates to the technical field of intelligent identification, and discloses a gesture recognition method based on adaptive uncertainty-aware meta learning, which formalizes a gesture small sample regression framework, extracts features by using a linearized neural network and a neural tangent kernel, establishes a Gaussian process regression model by combining Bayesian inference, projects a covariance matrix in a reduced dimension by using a Fisher information matrix in view of high-dimensional characteristics, simultaneously constructs an adaptive weight generator, generates a dynamic threshold value by means of historical loss moving average, training progress and an uncertainty correction term, and assigns specific weights to tasks; a posterior mean value is output as a predicted angle in a meta test stage, and the posterior covariance is used to quantize uncertainty. The application can effectively cope with object symmetry ambiguity and feature loss, focuses the model on difficult tasks through an adaptive mechanism, significantly improves prediction accuracy and reliability in a small sample scene, and reduces computational complexity.

A learning device includes processing circuitry configured to acquire data of which a label is predicted, and reduce, in a model representing a probability distribution of the label of the acquired data, a rank of a Fisher information matrix for the data to a value less than a predetermined value and learn the model.

The invention relates to the technical field of artificial intelligence and privacy protection, and discloses a vertical field large model privacy forgetting method and terminal based on double Fisher matrix guidance, and the method comprises the steps: separating two text sequence data subsets from an original training corpus of a vertical field large language model, namely a reserved data set and a forgetting data set; the reserved data set contains vertical field professional data which does not contain privacy information, and the forgotten data set contains privacy sensitive data to be eliminated. Based on an information geometry theory, calculating a reserved Fisher information matrix and a forgotten Fisher information matrix of model parameters on the reserved data set and the forgotten data set respectively; based on collaborative analysis of a reserved Fisher information matrix and a forgotten Fisher information matrix, a parameter forgotten guide function is constructed, guide function values of all parameters of a model are calculated, and a standardized guide vector is constructed according to the guide function values. According to the method, refined balance of privacy forgetting of the large language model in two levels of privacy elimination and capability keeping is realized.

The invention discloses a neural network model fusion method based on robust interpolation and Bayesian optimization, and belongs to the technical field of artificial intelligence and deep learning model safety protection. The method aims at solving the problems that robustness and accuracy of a deep neural network are difficult to consider under white-box adversarial attacks, and an existing method has the defects in the aspects of interpolation parameter selection, Fisher information estimation and iterative optimization. According to the method, efficient fusion of the neural network model is realized through data division and adversarial sample generation, baseline model normal training and adversarial training, interpolation model generation and experience Fisher information matrix calculation, Bayesian optimization dynamic search of optimal interpolation parameters, and return retraining and iterative optimization. According to the method, the confrontation robustness is remarkably improved while the standard accuracy of the model is kept, and the method can be widely applied to safety key scenes such as a face recognition access control system, financial voiceprint authentication and Internet of Things malicious software detection.

This application relates to the field of rail transit technology and provides a method and system for deploying metamaterial tags in a railway scenario. The method includes: constructing an observation equation with the Euclidean distance between the train and the metamaterial tag as the observation distance, and constructing observation vectors for N metamaterial tags based on the observation equation; constructing a Fisher information matrix about the motion state vector based on the observation vectors and the train's motion state vector, and using the trace of the inverse of the Fisher information matrix as the Cramer-Rhodes lower bound; using the spatial coordinates of the N metamaterial tags as optimization variables, minimizing the Cramer-Rhodes lower bound as the objective, and introducing dynamic environmental constraints to form a constrained optimization problem; solving the optimization problem using a particle swarm optimization algorithm to obtain a set of target tag position coordinates; and deploying metamaterial tags at corresponding positions along the track based on the target tag position coordinates. This improves motion compensation accuracy.

The invention discloses a disease and pest prediction method based on adaptive federated optimization and dynamic knowledge distillation, and the method comprises the steps: carrying out the training of a round of a local model through a loaded local training sample for each client, obtaining a client model parameter, calculating a Fischer information matrix of the client model parameter, and carrying out the calculation of the Fischer information matrix; the method comprises the following steps of: calculating weighted model parameters of a client by combining a Fischer information matrix and a loss function gradient of the client, then calculating a basic weight of the client based on a data volume weight, a data distribution weight and a calculation capability weight, and aggregating the weighted model parameters of the clients by adopting the obtained basic weight to obtain parameters of a global model; and after the client obtains the parameters of the global model, taking the global model as a teacher model, taking the local model of the client as a student model, and optimizing the model parameters of the local model of the client through knowledge distillation. According to the method, the convergence speed and performance of the global model are improved, and the adaptability of the model to different regions and different data distributions is improved.

The application relates to an image classification model optimization training method based on Nesterov momentum and block curvature perception, which comprises the following steps: performing weight attenuation update on the model parameters of the current iteration, constructing a first-order momentum estimation, introducing a Nesterov accelerated gradient method to construct a Nesterov modified momentum term, constructing a second-order moment estimation, constructing a block curvature perception proxy term, jointly constructing a parameter update equation of the image classification model based on the Nesterov modified momentum term, the second-order moment estimation of the current iteration and the block curvature perception proxy term, and realizing adaptive optimization of the model parameters. In the application, the gradient coupling relationship in the parameter block is perceived without explicitly calculating the Hessian matrix or the Fisher information matrix, so that the convergence speed, the generalization performance and the training stability of the image classification model training process are improved.

This invention belongs to the field of intelligent manufacturing equipment industry and industrial automation control technology, and discloses a data-driven gasification combustion fault diagnosis method and a gasification burner. This invention compensates for sensor drift using an adaptive extended Kalman filter, aligns dynamic transmission delays based on the fluid dynamics continuity equation, extracts spatiotemporal features using a graph attention network with injected thermodynamic priors and a gated recurrent unit, outputs the fault probability through a focus loss function, and finally introduces a Fisher information matrix to perform incremental fine-tuning of the neural network. This invention solves the false alarm problem caused by time-varying delays and sensor degradation under variable load conditions, improves diagnostic robustness and adaptability, and can be widely applied to gasification combustion systems in the intelligent manufacturing equipment industry.

The application discloses a Vision Transformer model quantization method based on attention head importance guidance, which adopts a PTQ method based on weight disturbance and output reconstruction. Firstly, a full-precision pre-training model is prepared, and then the weights of the full-precision output part of each attention head are saved, and the Fisher information of each attention head is calculated. Then, the full-precision output is obtained as the benchmark output by inputting the calibration data set into the full-precision model, and the quantization parameters are initialized. According to the Fisher information score of each attention head, the loss before and after quantization is weighted. The AdaRound strategy is adopted to optimize the rounding strategy by minimizing the quantization loss. The application effectively utilizes the characteristics of the multi-head attention mechanism, preferentially optimizes the important attention head in the quantization process of the Vision Transformer model, and retains more effective information, and further improves the model performance in the case of low bits, so that the model has more possibilities in the deployment of edge devices.