50 results about "Adaptive regularization" patented technology

The invention discloses a data deep learning and intelligent analysis method based on an AI artificial intelligence technology, and relates to the technical field of basic AI models, and the method comprises the steps: employing a multi-modal data preprocessing module to carry out the expansion of small sample data through a generative model, and combining with meta-learning to extract prototype features, meanwhile, an epsilon-differential privacy budget is dynamically allocated based on the data sensitivity level so as to inject dynamic noise; establishing a layered federated learning architecture, training a model by local training nodes through a loss function containing a self-adaptive regularization item, and performing sparse processing and gradient disturbance before uploading parameters; the global aggregation node adopts a weighted federated average algorithm to aggregate parameters, and dynamically adjusts the communication frequency according to the loss convergence speed; and a target model is obtained through iterative training, and a decision interpretation report containing the attention thermodynamic diagram and the desensitization identifier is generated when a result is output. According to the method, the problems of small sample overfitting, data islands and privacy disclosure are effectively solved, and the accuracy and practicability of the model are improved.

The invention relates to a fixed parameter optimization cubature Kalman filtering tracking method for a high-speed high-maneuvering target, belongs to the technical field of signal processing and target tracking, and aims to solve the problem of insufficient tracking performance caused by difficulty in parameter tuning and isolation of an improved mechanism when an existing method is used for coexistence of model mismatch, noise time variation and outlier interference. According to the scheme, a framework integrating off-line multi-parameter collaborative optimization and on-line multi-mechanism adaptive filtering is constructed; in the off-line stage, an optimal combination of key parameters such as covariance adjustment factors is determined through a grid search system; in the online stage, the combination is loaded, and a complete filtering process including innovation feedback type dynamic covariance adjustment, sliding window type noise estimation and outlier suppression and trace-related adaptive regularization is executed, so that an enhanced tracking method with active pre-judgment and closed-loop learning capabilities is formed. The method is mainly used for carrying out high-precision and high-robustness real-time state estimation on the high-speed high-maneuvering target.

The invention discloses a calibration method and system based on multi-scale feature fusion and an adaptive optimization algorithm, and belongs to the technical field of computer vision. The method comprises the following steps: collecting a plurality of images containing calibration modes; detecting and positioning feature points by adopting a multi-scale feature fusion technology, and distributing a confidence score for each feature point; carrying out feature point matching and removing abnormal values by using an improved RANSAC algorithm based on the prior geometrical relationship and confidence score; preliminarily calculating camera parameters; and a nonlinear optimization model fusing confidence weighting and adaptive regularization is constructed, joint optimization is performed on camera parameters, and a high-precision calibration result is output. According to the method, through an optimization mechanism of multi-scale feature fusion and confidence guidance, the problems of unstable feature point detection, abnormal value sensitivity and insufficient precision of a traditional method in a complex environment are effectively solved, and the accuracy, robustness and automation degree of image calibration are remarkably improved.

The invention discloses an atmospheric ocean multi-source data high-precision assimilation method and device, and relates to the technical field of ocean multi-source data assimilation, and the assimilation method comprises the steps: collecting multi-source observation data of a typhoon region of the South China Sea; preprocessing the data; decomposing the observation field into different scales by using wavelet transform; acquiring a background field from the numerical model, interpolating the background field to a grid consistent with observation, projecting the background field to an observation space, and performing multi-scale decomposition to enable the background field to correspond to an observation scale; calculating a difference vector of observation and background fields on each scale, and dynamically calculating an adaptive weight of each point on each scale according to an observation error, a background error and a current deviation; and constructing a cost function, solving through an optimization algorithm to obtain an analysis field, and outputting a final analysis field. According to the method, a multi-scale decomposition and self-adaptive regularization cooperation mechanism is introduced, so that high-precision dynamic fusion of the multi-source observation data and the numerical model can be realized.

The invention provides a small sample tunnel low-temperature asphalt optimization design method based on machine learning. According to the method, an asphalt sample is prepared through orthogonal test design, performance indexes are measured, and an initial small sample data set is established; the SMOTER technology is used for data enhancement, and the sample scale is effectively expanded; constructing an AR-CatBoost machine learning model, and introducing an adaptive regularization and gradient weighting mechanism to enhance the learning ability of the nonlinear relationship between the asphalt component mixing amount and the performance; carrying out hyper-parameter automatic optimization by adopting an improved reptile search algorithm fusing Cauchy variation and dynamic boundary adjustment; based on the optimization model, large-scale virtual ratio performance prediction is generated in a component mixing amount range, and the optimal ratio is accurately screened by integrating a scoring function. The method effectively solves the problem of insufficient model generalization ability under the small sample condition, realizes high-precision and low-cost automatic design of the tunnel low-temperature asphalt mix proportion, and is suitable for rapid research and development and performance optimization of tunnel asphalt materials in cold regions.

This invention discloses a method and system for fitting safety monitoring data based on adaptive hybrid basis functions, belonging to the field of computer-aided data analysis technology. The method achieves fully automatic high-precision curve fitting through a six-layer adaptive architecture: it automatically analyzes the trend complexity, periodicity, noise level, and nonlinearity of the data using multi-dimensional feature recognition technology; it calculates the data complexity score based on the feature analysis results and intelligently selects basis function combinations from an extended basis function library containing 32 sub-functions across 8 categories; it constructs a dynamic weighted hybrid basis function model, achieving adaptive adjustment of model parameters through time-varying weight functions and coupling correction terms; and it employs a multi-model dynamic fusion mechanism, integrating multiple candidate models based on six-dimensional confidence evaluation. This invention innovatively introduces adaptive regularization technology and a hierarchical optimization strategy, effectively balancing fitting accuracy and generalization ability, and possesses advantages such as full automation, strong robustness, and complete uncertainty quantification.

The application discloses an infrared image segmentation method based on a local region self-organizing mapping algorithm, which comprises the following steps: firstly, a local region self-organizing mapping algorithm is designed, a plurality of feature bias fields are extracted by moving a local sliding window before iterative evolution of a level set, and a plurality of feature local data driving terms are constructed under a multiplicative bias field model framework; secondly, a plurality of feature global data driving terms are calculated, and a plurality of feature hybrid data driving terms are constructed through an adaptive weight function; then, an adaptive regularization function is used to regularize an energy value range of the plurality of feature hybrid data driving terms, and an initial level set is driven to perform iterative evolution; then, in the iterative process of the level set, a hyperbolic tangent function is used to keep the level set in the iterative evolution, and a symbol rule feature of positive outside and negative inside and a distance rule feature with a modulus value of 1 are kept; meanwhile, a mean filter template is used to continuously smooth the level set and eliminate redundant non-boundary contour lines; finally, a gradient descent method is used to continuously iteratively solve a minimum value of an energy function until a convergence criterion of the level set is reached or a maximum iteration number is reached, and then the method is stopped, at this time, a position of a zero level set is output, and image segmentation is completed. The local region self-organizing mapping algorithm can accurately mine a foreground contour from an infrared image, and has good segmentation speed and precision.

The application discloses a substation equipment anomaly detection method based on a confidence meta-learning framework, relates to the technical field of substation equipment anomaly detection, and solves the problems that the existing method is susceptible to abnormal sample interference and poor in adaptability. The steps are as follows: a visible light image training set of normal substation equipment is constructed, an anomaly detection model is trained to evaluate the equipment state, meanwhile, a reconstruction, density and equipment key area abnormal loss function is constructed, the sample is converted into an abnormal score; the abnormal score distribution algorithm is used to identify a threshold and assign a sample weight, the training loss is fused, the covariance matrix is built by training and verifying the loss to evaluate the model uncertainty, the adaptive regularization is introduced to reduce the uncertainty, and then the model parameters are updated by using the meta-learning; the model with the updated parameters is used to process visible light images of equipment to be detected and output results. The application quantifies two types of uncertainty in a label-free unsupervised framework, reduces abnormal sample interference, improves the adaptability of the model by combining the meta-learning, and improves the detection accuracy.

The invention provides a flight parameter identification method for a multi-aircraft system, which relates to the field of aircraft dynamics, and comprises the following steps: modeling a dynamic model: in a multi-aircraft cooperative communication network, for each node in the network, modeling based on an ARMAX model with external input and a colored noise structure, each node constructs a local dynamic model according to local observation information; preliminary parameter estimation: performing preliminary parameter estimation on the local dynamic model of each node by adopting a distributed extended least square algorithm; and correcting the initial parameter estimation value based on an adaptive Lasso regularization mechanism to obtain a final parameter estimation result under sparsity constraint. The parameter identification method provided by the invention has excellent modeling precision and engineering adaptability, and is suitable for typical scenes of online modeling, cooperative control, health state monitoring and the like in a multi-aircraft system.

The invention provides a Laplace-DLTS defect parameter extraction method and a Laplace-DLTS defect parameter extraction system. The Laplace-DLTS defect parameter extraction method comprises the following steps: acquiring capacitance transient data obtained by a deep energy level transient spectrum test; carrying out preprocessing on the capacitance transient data; constructing a Laplacian transformation kernel function according to the capacitance transient data, and performing discretization processing to obtain a discrete Laplacian transformation model; regularization constraint parameters are introduced to construct a target function, a corresponding emissivity spectrum is obtained through numerical solution, and the regularization constraint parameters are adaptively determined according to signal characteristics of capacitance transient data; performing automatic peak identification on the emissivity spectrum based on a preset criterion; and according to the identified effective defect peak, automatically extracting and outputting physical parameters of the corresponding defect. By introducing the adaptive regularization constraint parameter, the regularization degree can be dynamically adjusted according to the actual signal characteristics of the capacitance transient data, the instability problem of noise amplification and solution is effectively suppressed, and the resolution capability and reliability of the emissivity spectrum analysis result are improved.

The application discloses a magnetic resonance image registration method based on adaptive regularization, constructs a training set and a test set; an adaptive regularization registration network is built; the floating image and the fixed image in the training set are spliced in the channel dimension and then input into the student registration network for bidirectional registration operation, the appearance disturbance of the floating image domain is calculated, the spatial transformation uncertainty map is calculated according to the predicted deformation field output by the teacher registration network; and the adaptive regularization registration network is trained. The appearance disturbance of the floating image domain is added in the teacher registration network to obtain the spatial transformation uncertainty map, the weight of the regularization term in the loss function is automatically adjusted by utilizing the spatial transformation uncertainty, adaptive regularization is realized, and the magnetic resonance image registration precision is improved.

The present application provides a kind of geologic multi-source constraint adaptive full waveform inversion method and system, constructs multiple complementary constraint norms, organically fuses multiple modal prior information, forms unified multi-source constraint framework, compared with traditional single type prior constraint, the present application can more comprehensively utilize available prior information, guide inversion to the direction that both conforms to geophysical data and geological understanding converges;While introducing adaptive regularization parameter strategy based on Bayes theory, realize the adaptive update of weight factor, avoid the subjectivity and randomness caused by the dependence of regularization parameter on experience selection in traditional method, make the inversion result more stable, and do not need tedious manual parameter adjustment, can dynamically adjust the constraint strength according to the data noise level, enhance the guiding role of prior constraint in low signal-to-noise ratio, more rely on observation data in high signal-to-noise ratio;At the same time, through the combination of model compression and forward simulation calculation, the calculation efficiency is greatly improved.

The invention relates to the technical field of electrical digital data processing, and discloses a radiology department image focus prediction system based on electrical digital data processing, which comprises the following steps: generating a local computing environment fingerprint representing a current computing environment before processing image data, and carrying out association binding on the fingerprint and a generated prediction result; and the system also stores a golden fingerprint of a standard environment, and automatically executes an adjustment operation based on the difference between the local computing environment fingerprint and the golden fingerprint when the local computing environment fingerprint and the golden fingerprint are inconsistent, so that the computing environment identity is anchored for each data processing result, and a set of environment baseline self-adaptive regularization and self-healing mechanism is established; therefore, the problems of data comparability loss and systematic trust crisis caused by dynamic evolution of a computing environment are solved, and the context credibility and long-term value of a data processing result in the whole life cycle are ensured.

The present invention relates to an active contour model image segmentation method based on K-medoids clustering, comprising: using a K-medoids clustering algorithm to perform binarization processing on an image to be segmented to obtain a fitting result, and acquiring a foreground image and a background image; constructing a pre-segment fitting function based on the binarization fitting result; using a zero level set that meets the Lipschitz condition to replace the contour curve of the pre-segment fitting function to obtain a KM pre-segment fitting function; using a gradient descent method to minimize the energy functional of the KM pre-segment fitting function to obtain a gradient flow equation; using an adaptive regularization function to regularize a data-driven term of the gradient flow equation; using a rulesig function to regularize the zero level set function, using a kernel function to smooth and shorten the curve of the regularized zero level set function, and outputting a level set function as a segmentation result of the image to be segmented.

The invention relates to the technical field of natural gas load prediction, and particularly provides a natural gas load prediction method and device. The method comprises the following steps: acquiring historical data and exogenous variable data, and performing dynamic characteristic decomposition to generate a multi-dimensional characteristic matrix; obtaining a preliminary predicted value through an improved hybrid neural network model, wherein the model is composed of a CNN, a BiLSTM and an adaptive regularization layer; a disturbance curve is generated by combining extreme weather and an emergency simulator to calculate and correct a predicted value, and the two predicted values are fused by adopting a dynamic weight distribution mechanism. Through exogenous variable decomposition, extreme event response and adaptive regularization technologies, the prediction accuracy and robustness are improved, and effective support is provided for natural gas supply optimization and emergency scheduling.

The application discloses a network slice resource allocation method based on Tsallis-MAAC, comprising: acquiring real-time state characteristics and global resource states of each network slice; calculating dynamic priority weights of each slice; constructing a Tsallis divergence regularization term with asymmetric geometric morphology based on the dynamic priority weights and a reference strategy, wherein the priority weights are mapped as geometric morphology parameters to regulate the sensitivity of the strategy constraint; predicting SLA violation risk probabilities of each slice, combining the priority weights with the resource states to generate adaptive regularization coefficients; coupling the regularization term with the adaptive coefficients to obtain a total regularization loss, combining a strategy gradient target to construct a total strategy loss function, and updating and executing resource allocation strategies of each slice according to the total strategy loss function. The application realizes fine network slice resource allocation under differentiated SLA guarantee through the cooperation of the priority-driven geometric constraint and the risk-perceived adaptive adjustment double channels.

The invention provides a reservoir production dynamic optimization decision method and system, and relates to the fields of physics, machine learning and data mining. According to the global quantum state characterization method based on quantum reservoir parametric modeling and weak Hopf tensor product fusion, a quantum interference fringe-based adaptive regularization parameter dynamic adjustment mechanism, a quantum gradient descent algorithm and a distributed collaborative architecture are combined; the technical bottlenecks of traditional static parameter setting, classical optimization dimension limitation and multi-well-site isolated optimization are broken through, and dynamic global optimal balance of physical fidelity and economic benefits is achieved for the first time in reservoir production optimization.

The invention discloses a gravity inversion method and system based on cluster number self-optimization and sub-domain adaptive regularization, a storage medium and equipment, belongs to the technical field of gravity inversion, and aims to solve the problems that regularization parameters are difficult to adaptively adjust, the clustering number depends on manual setting, a partition structure is not matched with geologic features and the like in gravity inversion. The optimal clustering cluster number is automatically determined by calculating the marginal benefit decreasing rule of the intra-cluster quadratic sum error; and constructing a global reference regularization factor based on the spatial density characteristics of each sub-region, and introducing a normalized spatial density exponential attenuation strategy to realize sub-region differential dynamic updating of the regularization factor. According to the method, automatic determination of the cluster number in gravity inversion is realized, dependence on manual setting and fixed prior is avoided, stability, spatial resolution and imaging precision of gravity inversion are improved, spatial adaptive adjustment of regularization constraint parameters is realized, and the method is suitable for mineral resource exploration and geological structure identification under complex geological conditions.

The invention discloses a fracture pressure data collection system based on a coalbed methane reservoir, and the system comprises a data collection layer which collects pressure, temperature, strain and sound wave signals through a fiber bragg grating array and a piezoelectric sensor, carries out the local filtering, and then uploads the signals through a LoRa network in the form of a structured data packet; in the data transmission layer, edge nodes clean and compress data in real time, add metadata labels, and preferentially transmit high-priority alarm signals through a 5G channel and a QoS route; and the data processing layer is used for fusing real-time data and geological information based on an ST-CNN double-flow architecture, carrying out transfer learning to initialize a model and dynamically updating parameters, and carrying out adaptive regularization balance generalization. Extracting spatio-temporal features and geological topological features, generating a prediction map, a trend and a risk identifier, and triggering early warning to output visual data, decision suggestions and digital twinning instructions to an application layer; and the application layer integrates an intelligent decision-making platform and a digital twin module, and supports dynamic threshold setting, strategy optimization and reverse adjustment of model parameters.

The invention discloses a substation equipment anomaly detection method based on a confidence element learning framework, relates to the technical field of substation equipment anomaly detection, and solves the problems that an existing method is easily interfered by abnormal samples and is poor in adaptability. The method comprises the following steps: constructing a transformer substation normal equipment visible light image training set, training an anomaly detection model to evaluate an equipment state, constructing reconstruction, density and equipment key area anomaly loss functions, and converting a sample into an anomaly score; calculating an identification threshold according to abnormal score distribution, distributing sample weights, fusing training loss, establishing a covariance matrix through training and verification loss to evaluate the uncertainty of the model, introducing adaptive regularization to reduce the uncertainty, and updating model parameters by meta learning; and processing the visible light image of the to-be-detected equipment by using the model with updated parameters, and outputting a result. According to the method, two types of uncertainty are quantified in a label-free and supervision-free framework, abnormal sample interference is reduced, the model adaptability is improved in combination with meta-learning, and the detection accuracy is improved.

The application discloses a personalized federated learning method for cloud-edge collaboration of Internet of Things, which is characterized by resource limitation and high heterogeneity in the cloud-edge collaborative Internet of Things scene. Dynamic parameter decomposition and integration are introduced into personalized federated learning. In view of the potential risks of catastrophic forgetting and bias accumulation in the parameter transmission and aggregation process of federated learning, an adaptive regularization optimization method based on continuous learning is adopted. To solve the above problems, firstly, the global shared parameter and the client-specific parameter are decoupled by using the parameter decomposition strategy. Then, the local optimization process performed by each client is compared to sequential multitasking, and the online importance score and sensitivity in the optimization path are calculated. Finally, an elastic regularization term based on the learning trajectory is introduced to constrain the parameter update, and the learning rate of the next round is dynamically adjusted according to the importance weight of each parameter in the last round of global tasks. This method effectively balances the global sharing and local individualization requirements, and is particularly suitable for cloud-edge collaborative Internet of Things environments with limited resources, frequent communication and strong data heterogeneity. It not only improves the accuracy and robustness of the model on the edge side, but also significantly reduces the communication and computing overhead, and has good practical deployment prospects and expansion capabilities.

The invention discloses a dynamic response reconstruction-oriented wireless sensor dimensionality reduction sparse configuration evolution method, and belongs to the technical field of wireless sensor networks. The method comprises the following steps: constructing a feature selection driving sparse optimization model based on compressed sensing, and optimizing a sensor configuration weight by taking an aggregation function of minimizing residual adaptive regularization and a quadratic form accumulation term as a target; a dimension reduction differential evolution algorithm is provided, the dimension of a decision vector is dynamically reduced through an adaptive restart strategy, and the convergence efficiency and the global search capability are improved; and finally, the number of sensors is remarkably reduced while high-precision prediction of unreachable node response is ensured. The method is suitable for health monitoring in complex projects such as spacecraft structures, and has the advantages of low configuration cost, high prediction precision and high applicability.

The embodiment of the invention provides a target detection model training method and system based on model transfer learning. The method is applied to the technical field of transfer learning and comprises the steps of obtaining a target detection model and target domain data, determining strategies of a frozen layer and a fine tuning layer, training the model and calculating parameter differences between a source domain and a target domain, and adjusting regularization and negative transfer parameters. And integrating the self-adaptive regularization formula to a loss function, and continuously optimizing training. And updating the loss function after each round of training, evaluating the performance of the source domain and the target domain, detecting negative migration and taking mitigation measures until the model reaches a preset standard. According to the scheme, the performance of the model on the target domain task can be improved through an accurate strategy and dynamic adjustment, overfitting or negative migration is avoided, and finally a high-quality model which can reserve source domain knowledge and can adapt to the target domain task is obtained.

The application relates to a data prediction evaluation method and system for CAE simulation, and belongs to the technical field of CAE simulation. The method solves the problem of low prediction result precision caused by insufficient model parameter optimization capability. The method comprises the following steps: based on a historical CAE simulation data set, taking CAE simulation parameters as input and corresponding simulation responses as output, a probability generation model based on a Gaussian process is constructed; a multiple start mechanism is adopted, and an adaptive regularization term is introduced in the maximization of the marginal likelihood function to obtain a target function; based on the historical CAE simulation data set and the target function, the hyperparameters of the probability generation model are optimized to obtain an optimal hyperparameter set, and the training of the probability generation model is completed; and new CAE simulation parameters are input into the trained probability generation model to output simulation response prediction values and uncertainty quantification indexes. The precision of the prediction result is improved.

The application discloses a campus public opinion monitoring-oriented incremental learning type neural network model continuous optimization method, and relates to the technical field of neural networks. In the method, the incremental learning type neural network model to be optimized comprises at least four layers, namely an input layer, two hidden layers and an output layer, each hidden layer comprises at least two neurons, and is divided into shallow feature extraction neurons and deep monitoring neurons. In the incremental learning process, a network model for neuron state recognition is established, a fully-connected deep neural network is adopted as a basic framework, and a total loss function of the network model is composed of three parts, namely a regularization loss term, a task loss term and a weight recovery loss term. Through weight reconstruction, bias correction and adaptive regularization constraint, model parameters are iteratively optimized, and the capacity protection and self-healing ability of the network topology structure are realized.

The invention discloses a medical image segmentation method based on an adaptive regularization network module, and relates to the crossing field of medical image processing and computer vision, and the method comprises the steps: S1, constructing an adaptive regularization network module MARN which comprises an anatomical perception space regularization sub-module and a multi-scale channel modulation MFACM sub-module; the method comprises the steps of S1, preprocessing medical images, S2, inserting MARN into all basic blocks of a deep learning network HRNet, inputting the preprocessed medical images into an ASR sub-module for processing, and outputting regularization features after weight fusion, S3, fusing the regularization features with original features to obtain anatomical region quality scores, S4, inputting the processed medical images into an MFACM sub-module for modulation to obtain modulation output features, and S5, outputting the modulation output features to a deep learning network HRNet. And S5, generating a segmentation result of the medical image through a convolution and activation function according to the task head input to the HRNet. The method can solve the problems that a traditional normalization mechanism is unstable in feature distribution and large in statistical deviation in small-sample, low-contrast and high-noise medical scenes.

The invention discloses a vibration inversion enhancement method and system based on singular value self-adaptive regularization, and the method employs the idea of reverse evolution, carries out the reverse evolution reconstruction of an external signal according to a transmission model and an inversion algorithm, counteracts the attenuation effect of a contact interface in a transmission path, enhances the weak fault features of the external signal, and improves the reliability of the signal. A foundation is laid for state monitoring and fault diagnosis of the gear transmission system, and the operation reliability of the transmission system and high-end equipment is improved.

The application discloses a blast furnace abnormal furnace condition root cause analysis method fusing expert knowledge and Granger causality, which comprises the following steps: integrating furnace condition perception priori knowledge, determining a key variable set according to a current abnormal furnace condition type, and ensuring that the root cause analysis can capture a key target link; based on the Granger causality thought, learning a causality matrix among variables in a variable prediction process, and adopting an adaptive regularization strategy to dynamically adjust the causality matrix; in a time lag process among the variables, firstly performing fast positioning in a coarse granularity, then performing fine search, and further performing global optimization by means of a particle swarm optimization algorithm, and finally adjusting part of the time lag relations in combination with physical priori; constructing a link search and confidence evaluation process, performing propagation path search under the constraints of physical grouping priori, a causality matrix and a time lag matrix, and combining information such as process rules to construct a comprehensive link scoring index, and the application well makes up for the deficiencies of traditional methods in the aspect of blast furnace abnormal furnace condition root cause analysis.

The invention discloses a network slice resource allocation method based on Tsallis-MAAC. The method comprises the following steps: acquiring a real-time state feature and a global resource state of each network slice; calculating the dynamic priority weight of each slice; constructing a Tsallis divergence regularization item with an asymmetric geometrical morphology based on the dynamic priority weight and the reference strategy, wherein the priority weight is mapped into a geometrical morphology parameter to regulate and control the strategy constraint sensitivity; predicting the SLA default risk probability of each slice, and generating an adaptive regularization coefficient in combination with the priority weight and the resource state; and coupling the regularization item and the adaptive coefficient to obtain total regularization loss, constructing a total strategy loss function in combination with a strategy gradient target, and updating and executing a resource allocation strategy of each slice according to the total strategy loss function. According to the method disclosed by the invention, the refined network slice resource allocation under the guarantee of the differentiated SLA is realized through the coordination of the geometric constraint driven by the priority and the self-adaptive adjustment dual-channel of the risk perception.

The present invention discloses an image retrieval method for hash model training and adaptive binary quantization in a noisy environment, including: obtaining multiple sample images, each sample image having its own class label; obtaining an initial model including an initial feature extraction network, an activation function, and an initial classification network: inputting the sample images into the initial model during the z-th training to obtain the z-th hash code value and the z-th prediction value of each sample image; determining the z-th first similarity retention loss according to the number of input sample images, the z-th prediction value, and the class label of each sample image; determining the z-th adaptive regularization loss according to the hyperparameters of the initial model, the number of input sample images, the z-th prediction value, and the (z-1)-th prediction value of each sample image; adjusting the network parameters of the model obtained from the (z-1)-th training according to the two obtained losses until obtaining a first pre-trained feature extraction network, an activation function, and a first pre-trained classification network, thereby obtaining a pre-trained hash model.