157 results about "Transition matrices" patented technology

The present invention relates to the related technical field of natural language inference. Disclosed is a large language model joint inference method based on a knowledge graph-enhanced chain-of-thought prompt, comprising: constructing a local knowledge subgraph; decomposing an original question text into S sub-question texts and concatenating the original question text and the S sub-question texts; inputting a concatenated question text into a graph inference model to obtain a weighted entity distribution; from the weighted entity distribution, extracting first G answer entities having the highest confidence level, using an inference transition matrix to trace an inference process of each answer entity, and generating an inference path from a question entity to the corresponding answer entity; and using the inference path to assist a large language model in predicting an answer to the original question text. By means of the method, the large language model can quickly and accurately find an answer to a question text.

The invention belongs to the field of unmanned aerial vehicle control, and particularly relates to a complex task-oriented unmanned aerial vehicle modular integrated control platform, which comprises an acquisition module for acquiring and processing information such as a real-time flight state of an unmanned aerial vehicle group, a complex task text, meteorological data and the like; the task analysis module decomposes a complex task into a single subtask set and an incidence matrix, and evaluates the task complexity for resource allocation; the atmospheric analysis module is used for inverting a real-time flight association adjustment graph based on the corrected meteorological data, flight states and the like in combination with a target optimization algorithm and an optimization function taking resource quantity minimization and task completion probability maximization as targets; the record analysis module generates a performance transfer matrix by using a hidden Markov algorithm; and the flight control module responds to the adjustment graph to correct the flight state, and ensures that the task completion probability meets a preset safety threshold. According to the platform, efficient cooperative control and resource optimization allocation of the unmanned aerial vehicle group under complex tasks are realized through modular integration.

The invention discloses a self-adaptive integrated navigation method based on geometric accuracy factors, which belongs to the technical field of underwater navigation and positioning, is used for underwater navigation and positioning, and comprises the following steps: initializing an inertial navigation system and an integrated navigation filter, and setting a state vector and a noise covariance matrix; predicted navigation parameters of the recursive carrier are mechanically arranged through inertial navigation, and the state transition matrix is used for state prediction; and dynamically adjusting an observation noise covariance matrix according to the geometric precision factor value, further calculating a Kalman gain, performing optimal estimation and correction on a prediction state by fusing acoustic observation information, and finally outputting a high-precision carrier position, speed and attitude. According to the method, the statistical characteristics of observation noise are dynamically remodeled by calculating and feeding back geometric precision factor values in real time, so that the Kalman filter has the capabilities of knowing the own geometric situation and adjusting the trust degree of each information source, and the global optimal navigation precision and reliability are realized under any motion track.

The invention relates to a target track association method in a multi-target environment. The method comprises the following steps: firstly, preprocessing a new received trace point; secondly, predicting the next position of each track by using the existing track historical information and adopting a Kalman filtering algorithm, meanwhile, adjusting a state transition matrix and a process noise covariance matrix in real time by considering factors such as ocean current and wind direction in an offshore environment, and adjusting an observation noise covariance matrix according to meteorological information; then comprehensively considering distance, speed difference and motion direction difference to calculate a correlation degree magnitude between a new receiving trace point and each predicted track position; and finally, judging whether the newly received trace point belongs to the existing target track or forms a new target by adopting a threshold judgment method according to the correlation degree value, and updating the state of the track. The method can effectively cope with multi-target intersection, observation loss and clutter interference, improves the stability of a matching result, can realize continuous tracking of a marine target, can adjust parameters according to the environment and target characteristics, and has high adaptability.

This invention relates to the field of data prediction technology, specifically to a collaborative control method for the opening degree of hydropower station gates based on BeiDou technology. This method constructs state parameters and analyzes the transition of these parameters over historical periods to obtain the transition probability for each state parameter transition. By combining the correlation between meteorological data and opening data, as well as the correlation between historical meteorological data and predicted meteorological data, the transition probabilities are continuously weighted to obtain a second weighted transition probability. The quantitative characteristics of the state transition of the flood discharge channel gate under its predicted state transition are statistically analyzed for the non-flood discharge channel gate, thereby obtaining the upper-level influence weight, obtaining the second state transition matrix of the flood discharge channel gate, and performing state prediction. This embodiment of the invention, by combining the influence of historical meteorological data and further combining the control influence between gates, obtains accurate gate state prediction results.

The invention relates to the technical field of attitude estimation, in particular to an attitude estimation method in a magnetic interference environment, which comprises the following steps of: obtaining a conversion relation between an Euler angle and a quaternion; calculating the quaternion at the next sampling moment; respectively solving a first state transition matrix, a first process noise matrix and a first prior covariance according to the first state equation; respectively solving a second state transition matrix, a second process noise matrix and a second prior covariance according to a second state equation; elements in the nb-series rotation matrix are used as observed quantities of state variables in the parameter-series rotation matrix, a corresponding Kalman gain is calculated, and a Jacobian matrix is designed by using the observed quantities; calculating an observation noise covariance matrix by using the Jacobian matrix; updating the Kalman gain by using the observation noise covariance matrix; and updating the state equation and the posterior covariance by using the updated Kalman gain and the observed quantity. The method solves the problem of how to prevent the calculation of the roll angle and the pitch angle from being influenced by magnetic interference when the magnetic interference exists.

The invention provides a river trajectory data protection method based on differential privacy, and belongs to the field of environmental data privacy protection. According to the method, river flow velocity and flow direction data are discretized into a two-dimensional grid, spatial-temporal characteristics are extracted by using a graph convolutional network (GCN), and a Markov transfer matrix is established; and according to the grid access frequency and the spatial density, privacy budget is adaptively allocated through a dual attenuation factor model, and differential privacy protection is realized. In the trajectory generation stage, dynamic privacy budget is combined, noise is injected into direction and time features by adopting an index mechanism and a Laplace mechanism, and a synthetic trajectory conforming to physical constraints is generated; and finally, evaluating and optimizing the track quality by using Savitzky-Golay filtering and multiple indexes (such as Frechet distance, access frequency error and KL divergence). According to the method, the privacy leakage risk is effectively reduced while the space-time continuity of the trajectory is kept, the data privacy and availability are considered, and the method is suitable for hydrological monitoring, ecological analysis and environmental data sharing.

The invention discloses an electric reactor mechanical vibration diagnosis method, device and equipment and a storage medium, and relates to the technical field of fault diagnosis, and the method comprises the steps: obtaining N window vectors obtained by dividing an electric reactor vibration signal; performing binning according to the amplitude in the window vector to obtain a Markov transfer matrix; generating a Markov transfer field according to the Markov transfer matrix; taking each row of the Markov transfer field as an initial node feature to obtain an initial node feature matrix; calculating a correlation coefficient between every two window vectors according to the N window vectors so as to construct an adjacent matrix; constructing graph structure data according to the initial node feature matrix and the adjacent matrix; inputting the graph structure data into the graph neural network model to obtain a fault classification probability of each window vector; and generating a fault diagnosis conclusion of the whole reactor according to the fault classification probabilities of all window vectors. The method can improve the diagnosis accuracy of the whole reactor.

The invention relates to a river channel sand body random model generation method and device based on hidden Markov model data filling, a medium and equipment. The method comprises the steps that logging data of a target area are collected and arranged; dividing logging data according to a logging gas-bearing interpretation conclusion, and carrying out analytical statistics on the velocity and density of longitudinal and transverse waves in the logging data; initializing an initial state probability, a state transition matrix, and a mean value and a covariance of Gaussian distribution in the hidden Markov model; inputting a three-dimensional observation vector composed of the velocity and density of the longitudinal and transverse waves into the initialized hidden Markov model for iterative training; performing down-sampling on the lithology data, then constructing a state transition probability matrix, and further constructing a lithology sequence conforming to a geological law through simulation; according to the lithologic sequence, calling a hidden Markov model to generate longitudinal wave velocity, transverse wave velocity and density corresponding to the lithologic sequence; and calculating a reflection coefficient of the synthetic stratum, setting a Ricker wavelet dominant frequency, and obtaining synthetic seismic response data through convolution.

The invention provides a task-based dialogue data generation method and device, a medium and electronic equipment, and the method comprises the steps: obtaining a state transition matrix which represents the probability of transferring from any state node of a task-based dialogue to another state node; performing sampling processing on the state transition matrix to obtain a state sequence of a plurality of sampling paths; processing each state sequence by using a large language model and corresponding cue words to generate a plurality of dialogue contents, and scoring each dialogue content to obtain a plurality of dialogue content scores, the dialogue contents being in one-to-one correspondence with the dialogue content scores; and performing multiple iterations on the state transition matrix based on the dialogue content score, and generating task-type dialogue data by adopting the iterated state transition matrix meeting the iteration stop condition, thereby solving the problem of insufficient training caused by lack of high-quality data of the task-type dialogue system.

This invention belongs to the field of electronic digital data processing technology, specifically relating to a low-altitude communication data anomaly detection system and method based on edge intelligence. The system parses protocol fields from the raw bitstream of the low-altitude communication link layer, extracts type identifiers and payload lengths to construct a protocol state machine transition matrix, and utilizes a Long Short-Term Memory (LSTM) network to extract state transition probability sequence features. Geographically adjacent edge nodes are set as micro-federated learning groups. Each node, after training using local features, only uploads network gating weight parameters to its neighboring nodes for weighted aggregation to update its local model. An alarm is triggered when the real-time state transition probability sequence deviates from the normal transition matrix and exceeds a preset topology threshold. This invention can identify protocol state machine logic errors and transition anomalies, reducing the amount of communication data required for collaborative model updates between edge nodes.

The application discloses a method and system for quantitatively analyzing self-insulation resistance shunt effect of a self-powered detector, and the design steps are as follows: a circuit model of the self-powered detector is established according to a working principle of the detector; specific parameters of each element in the circuit are calculated according to geometric structure and physical parameters of the detector; a state space equation in a matrix form is established according to a current generation mechanism of the self-powered detector in a radiation field and the circuit model, and a system state transition matrix of the self-powered detector system is established by solving a matrix index; and influences of geometric dimensions and physical parameters of the detector on the insulation resistance and the shunt effect thereof are analyzed according to the system state transition matrix. The method can quantitatively calculate the shunt effect of the insulation layer resistance of the self-powered detector with different structures and parameters, and can accurately calculate a change of a capacitive current with time, thereby providing guidance for design and application of the detector.

The invention provides a graph neural network training method and device, a medium, equipment and a program product, and the method comprises the steps: obtaining sample graph data, and enabling a plurality of nodes of the sample graph data to comprise a plurality of nodes with noise marks; iteratively training the graph neural network and the matrix estimator based on the sample graph data and the noise marks; in any round of iteration, the sample graph data and the noise marks are input into the graph neural network, so that the graph neural network outputs original prediction marks of all the nodes in the multiple nodes, the original prediction marks are corrected based on the current noise transfer matrix, and corrected prediction marks are obtained; performing parameter adjustment on the graph neural network based on the corrected prediction mark and the noise mark; and after parameter adjustment of the graph neural network is completed, inputting the original prediction mark and the noise mark into a matrix estimator, so that the matrix estimator outputs a noise transfer matrix, and performing parameter adjustment on the matrix estimator based on the noise transfer matrix of the adjacent nodes.

The application provides a watermark adding method, a watermark detection method, a watermark processing device and a computer storage medium. The watermark adding method comprises the following steps: inputting user interaction content and historical output text into a text generation model to obtain an original text probability distribution; obtaining a preset probability transition matrix; correcting the original text probability distribution by using the probability transition matrix to obtain a corrected text probability distribution; and taking a text corresponding to a maximum probability value in the corrected text probability distribution as current output text after watermark adding. By using the watermark adding method, the accuracy and usability of generated content are not damaged by directly setting a red-green set for a single character by using a hash function or other methods, and the output of the large model is corrected by using the probability transition matrix, so that the accuracy of watermark adding is improved.

The application provides an acquisition method and device of a push strategy, equipment and a storage medium. The acquisition method of the push strategy comprises the following steps: obtaining historical feedback data of a target user for different push strategies; firstly, determining a state transition matrix of the target user according to the historical feedback data, wherein the state transition matrix is used for indicating the probability value of state change of the target user under different push strategies; and secondly, determining a target push strategy with the maximum long-term and short-term value from a push strategy set according to the state transition matrix of the target user, the preset short-term value corresponding to the state change under different push strategies and the benchmark long-term value. Compared with the prior art, the reliability of the user value evaluation is higher and the selected push strategy is better because the value of the user in the short term and the long term is considered simultaneously.

This invention discloses a method for calculating the dynamic reliability of electromechanical systems based on inter-component interoperability, comprising: (1) constructing a hidden Markov model for system risk transmission and an inter-component risk state transmission matrix; (2) combining load effects to calculate inter-component interoperability and resistance coefficients, quantifying relevant parameters of the state transition matrix, and calculating the probability that the risk of a component will not escalate after risk transmission; (3) applying inter-component interoperability, determining the final state sequence of the system based on the idea of ​​stochastic computation, and calculating the dynamic reliability of the system. For complex electromechanical systems, the interoperability calculation method of this invention can make the system reliability calculation closer to the actual situation, and based on this, an improved maintenance strategy is proposed. Compared with the traditional maintenance strategy, the maintenance strategy considering inter-component interoperability proposed in this invention can maintain the system at a higher reliability and have a lower maintenance cost in the long term.

A parameter continuous creep SWAT time-varying water and sediment simulation method, system and medium, the method comprises the following steps: constructing a SWAT database for land use scenarios before and after the yield calibration parameter set; 24 types of parameters are divided into continuous creep and fixed parameters, and an expression is embedded in the main loop of the SWAT source code; update parameter execution boundary, change rate, rollback control; output runoff and sediment results. Based on the transition matrix, the HRU land class transition path and intensity are identified, and the HRU differentiated creep coefficient and slope-river transmission lag factor are determined by combining slope, distance from river, confluence, hydrological connectivity and area; in the update, the runoff, surface erosion and sediment transport parameter coupling are established, and the water conservation, sediment transport continuity, peak suppression and abnormal rollback are executed. The present application embeds the transfer heterogeneous response and channel transmission into dynamic simulation, reduces the risk of parameter mutation, sediment shock and non-physical combination, and improves the continuity, stability and physical reasonableness in the gradual change period.

The application discloses a rotating machinery fault diagnosis method based on delay-weighted visual graph and Netpro2vec, and belongs to the technical field of mechanical state monitoring and fault diagnosis. The method comprises the following steps: collecting a rotating machinery vibration signal; converting the signal into a delay-weighted visual graph, wherein, when judging whether two data points are visual, adjacent nodes of the two data points are ignored, and a weight is assigned to a visual edge, and the weight is defined as the minimum vertical distance between the two points and all data points therebetween; based on the graph, a low-dimensional feature vector is extracted by using the Netpro2vec method, specifically, a word sequence is generated by calculating a node distance distribution and a transition matrix, and a document is formed, and then a vector representation is obtained by learning through a document embedding model; and finally, the feature vector is input into a classifier to obtain a diagnosis result. The application enhances the noise resistance through a delay mechanism, and finely depicts the signal structure through a weighting mechanism, so that the precision and robustness of fault diagnosis are significantly improved.

The invention discloses a T cell receptor sequence design method based on a discrete diffusion model, and relates to a T cell receptor sequence design method. The invention aims to solve the problems that the traditional TCR discovery method depends on experimental screening and has high cost, long period and limited target coverage. According to the method, it is ensured that the generated sequence conforms to the TCR natural evolution law through a BLOSUM transfer matrix, the epitope binding specificity is improved by means of EPredactor gradient guidance, and through experimental verification, the AUC value of the generated sequence on prediction tools such as PISTE reaches 0.925 or above, the KL divergence of bit-score distribution and the real TCR is smaller than or equal to 0.006, the average value of the accessible surface area change (delta SASA) of a binding interface solvent is higher than that of the real TCR, and the method has the advantages that the method is simple and convenient to implement, and the method is suitable for popularization and application. The TCR sequence meeting the immunotherapy application requirement can be efficiently generated, and a key technical support is provided for personalized immunotherapy target development. The invention belongs to the technical field of bioinformatics, computational biology and immunotherapy.

The invention discloses an extreme rainfall flood risk propagation path identification method based on hydrological connectivity. The method comprises the following steps: constructing a hydrological topology network; calculating the runoff yield and hydrological potential energy of each grid unit; constructing a connectivity weight matrix according to the potential energy difference and the Manning roughness coefficient; a dynamic connectivity adjacency matrix is generated through a ponding depth triggering threshold value; constructing a risk propagation probability transfer matrix; and identifying a risk propagation main path and a key conduction node. According to the method, hydrological connectivity changes under extreme rainfall conditions can be dynamically represented, flood risk propagation paths and key nodes are effectively identified, and technical support is provided for urban flood disaster risk early warning and emergency management.

This invention relates to the field of biosensor technology, specifically to a sleep staging method and system based on the fusion of label calibration and inference calibration. The method includes the following steps: constructing a sleep stage transition matrix M driven by both data statistics and domain knowledge; performing a light calibration on the original sleep staging labels of the training set based on the transition matrix M to obtain calibration labels; extracting the time-frequency features of the electroencephalogram (EEG) signal, and training an LSTM model using the calibration labels to obtain a base model; and using the LSTM model to perform preliminary staging on a fixed-length EEG signal sequence to obtain an initial staging sequence. This invention employs a fusion scheme of a dual-driven transition matrix, light label calibration, dynamic programming global calibration, and iterative feedback.

The application relates to a method for generating wind power output simulation data according to historical data, which comprises the following steps: dividing historical wind power output data into several seasons through different date segmentation modes; calculating the expected value of wind power output of each season formed by each date segmentation mode, selecting a date segmentation mode with the maximum expected value deviation between seasons, and dividing the historical wind power output data into several seasons through the date segmentation mode; statistically obtaining the wind power output probability distribution in each season and the Markov transition matrix at each time; and simulating the wind power output at each date and each time in each season by using a Metropolis-Hastings algorithm to obtain annual simulation data. The simulation data generated by the scheme can better conform to the probability distribution of local historical data and the output characteristics of local wind power, and is suitable for Monte Carlo simulation calculation and other calculation researches which require a large amount of data.

The application discloses an omnidirectional guiding interaction multi-model positioning method and system for a motorized mother ship to recover an underwater robot and a storage medium, the method obtains the state information and distance data of the mother ship by integrating a water acoustic ranging and a water acoustic communication device, constructs an interaction multi-model singular value decomposition unscented Kalman filtering algorithm fusing a uniform straight line and a uniform bow turning model, and realizes real-time estimation of the position of the mother ship under a mouth-shaped maneuvering track. An event-triggered communication mechanism is designed, information is broadcast only when the turning state of the mother ship changes, so as to reduce the communication load and maintain real-time performance; meanwhile, a Markov transition matrix is updated, the turning model weight is enhanced, and the estimation accuracy and model matching are improved. According to the model probability and the estimation result, a LOS guiding law with an adaptive look-ahead distance is constructed, and the autonomous guiding recovery of the underwater robot to the motorized mother ship is realized. The method provided by the application can realize dynamic estimation and guiding based on the measurement data of a single water acoustic device under the conditions of the mother ship maneuvering and communication limitation.

The invention provides a navigation method, system and device for dynamic perception and obstacle avoidance based on an unmanned aerial vehicle, and the method comprises the steps: obtaining a real-time depth map of a to-be-detected region, and determining the instantaneous positions of a dynamic target and a static target in the real-time depth map; establishing a polar coordinate system by taking the instantaneous position of the dynamic target as an original point, discretizing a peripheral space into a polar unit network, constructing a Markov chain based on probability distribution of polar units, and updating a time-varying transfer matrix of the Markov chain based on a transfer probability between the polar units; and selecting a continuous polar unit sequence based on the updated Markov chain, fitting the continuous polar unit sequence by adopting a polynomial trajectory to obtain a prediction path, and proposing dynamic target adaptive trajectory planning. According to the invention, the perception precision and the state estimation stability of the unmanned aerial vehicle to the dynamic and static targets are obviously improved, the real-time response capability of trajectory planning to the motion change of the dynamic targets is enhanced, and the autonomy and the flight safety of the unmanned aerial vehicle in a dynamic environment are ensured.

This invention belongs to the field of equipment control technology, specifically relating to a power equipment control method and system based on the Internet of Things (IoT), comprising the following steps: S1, acquiring real-time operating data of nodes of each power device in the IoT, constructing a second-order state tensor representing the operating state of each node; projecting the second-order state tensor onto a low-dimensional subspace defined by basis vectors to obtain a low-dimensional state descriptor, wherein the basis vectors are adaptively updated based on the historical state covariance matrix of all nodes; calculating the Wasserstein distance of the low-dimensional state descriptor sequences between nodes, and generating a weighted transition matrix for parameter aggregation between nodes by combining the communication link quality; S2, if the condition number of the local state covariance matrix of a node exceeds a first threshold, then local optimization is initiated. This invention improves the convergence speed and accuracy of the optimization algorithm, and achieves more stable collaborative control of power equipment groups through a two-level control mechanism combining local optimization and global correction.

The invention discloses a unilateral control method and device for an agent in a random game, and relates to the technical field of artificial intelligence and game theory. The problem that an existing random game income control technology cannot support multiple types of control targets is solved. Comprising the following steps: constructing a quintuple random game model, and determining a one-step memory strategy of an action decision basis for controlling and controlled agents; constructing a state space of the Markov chain and a transfer matrix of the Markov chain by using a single-round game result set in the random game process; establishing an identical relationship among stable distribution of the Markov chain, a state transition rule and a strategy for controlling the intelligent agent; and determining the control condition of the long-term expected frequency of the target environment state, the general condition of the zero determinant strategy in the random game, and the control condition of the long-term expected benefit of the controlled agent.

The invention discloses a multi-view fusion cloud native malicious software detection method and system based on high-speed introspection of a virtual machine, and the method comprises the steps: 1), deploying a high-speed introspection module in a monitoring layer of the virtual machine, capturing an API call sequence of an internal process of a target virtual machine from the outside in a safe and low-invasion manner, and structuring the API call sequence into a runtime log; (2) the API calling sequence is regarded as a sentence, and a Word2Vec model is used for training to generate a structure embedding vector of the API; then, constructing a directed heterogeneous graph containing a file, a thread and API calling for each sample, taking the structure embedded vector as an initial feature of an API node, encoding the graph by using a graph attention network, and extracting a structure context feature vector; 3) extracting an API official function description text by utilizing the pre-training language model to generate a semantic embedding vector; constructing a directed heterogeneous graph for each sample, replacing the initial features of the API nodes in the graph with the semantic embedding vector, coding the graph by using the graph attention network again, and extracting a functional semantic feature vector; 4) firstly performing function classification on the APIs, and performing dimensionality reduction on the complete API calling sequence to obtain a limited function state sequence; constructing a Markov transition probability matrix for the state sequence of each sample, and selectively stacking a multi-order transition matrix to form a multi-channel feature tensor; inputting the feature tensor into a convolutional neural network for coding, and extracting a macroscopic behavior evolution feature vector; 5) splicing the structure context feature vector, the function semantic feature vector and the behavior evolution feature vector to form a final comprehensive feature vector; and inputting the comprehensive feature vector into a multi-layer perceptron classifier, and training the classifier in an end-to-end manner to enable an output sample of the classifier to be a classification result of malicious software or benign software.

A method of training, verification, and advanced notice for an autonomous system, comprising: by a machine learning classifier, wherein an adversary vehicle, and a primary vehicle are on the same path; creating a path position probability transition matrix, wherein the path position probability transition matrix comprises an acceleration parameters array of one or more acceleration parameters for a potential position, sorting a sample path, and using Ne elite path; creating a samples array, wherein the sample array is one or more paths; evaluating a sample path; sorting the sample path, based at least in part a custom score of a custom score function and selecting a subset of the sample paths as Ne elite paths; using the Ne elite paths to update the path probability transition matrix and the acceleration parameters array; and repeating until the Ne elite paths stabilizes for a predetermined number of iterations.

The invention discloses a generation method and device of operation and maintenance operation guidance information and a program product, relates to the technical field of operation and maintenance, and is applied to the field of financial science and technology. The method comprises the steps that operation and maintenance operation information of a target object is collected every preset duration, and an operation sequence is generated; state probability distribution is calculated according to a state transition matrix and the operation sequence, and the state transition matrix is determined according to an operation and maintenance operation document and historical operation and maintenance operation; predicting a target operation and maintenance operation executed by the target object in the next step according to the state probability distribution; and according to the operation information of the target operation and maintenance operation, performing retrieval in the operation and maintenance operation document to obtain operation guidance information, and pushing the operation guidance information to the target object. Through the method and the device, the problem that a traditional static document or keyword matching algorithm cannot provide operation guidance for a worker in real time according to the executed operation when the worker executes the operation and maintenance operation in the related technology is solved.

The present application relates to the technical field of image classification, and particularly relates to a training method and device of an image classification model, equipment and a medium. The method obtains an auto-encoding feature vector and a reconstructed image feature vector through an unsupervised image reconstruction model, obtains K feature vector clustering clusters of the auto-encoding feature vector, determines a noise probability transition matrix for representing noise confusion information between various image categories, obtains a category probability estimation vector according to a supervised image classification model, determines model loss by combining the noise probability transition matrix to perform model training, obtains a trained supervised image classification model to perform image classification, measures the distribution of noise labels and non-noise labels through the feature vector clustering clusters, obtains the noise probability transition matrix to represent the noise confusion information between the image categories, and measures the loss of the supervised image classification model, reduces the influence of noise labels on the supervised image classification model, and improves the accuracy of the supervised image classification model.