49results about How to "Effective decoupling" patented technology

The invention discloses a smoke scene three-dimensional reconstruction method and system based on double-branch physical knowledge embedding, and the method comprises the steps: firstly collecting data, calculating a fluid perception parameter under each observation view angle, calculating a state estimation value of each pixel, taking the state estimation value as priori supervision, pre-training a state estimation module, and carrying out the reconstruction of a smoke scene through the pre-training state estimation module; therefore, dynamic objects and static objects in the smoke scene are distinguished. The static object is sent to a static branch, the dynamic object is sent to a dynamic branch to be processed respectively, the volume density and appearance color of the static object are estimated, and the volume density, appearance color and velocity field of the dynamic object are estimated; updating the velocity field and the volume density of the dynamic object by using a fluid solving module; output results of the static branches and the dynamic branches are respectively rendered and synthesized, multi-view image information and physical knowledge priori are used as supervision constraints, module parameters are continuously iteratively optimized, finally, a high-fidelity three-dimensional dynamic smoke scene is reconstructed, and image generation of any view angle in the scene is supported.

The invention provides a dynamic nerve radiation field anti-aliasing reconstruction method based on three planes and multistage textures, which comprises the following steps: inputting a training video frame to be reconstructed, projecting a light cone to each pixel of the training video frame in an observation space, and sampling an inscribed ball on the axis of the light cone to obtain a sampling ball; mapping a sampling ball of the observation space to a standard space by using a dynamic deformation network; the standard space is represented by using multi-level three-plane textures, sampling balls in the standard space are respectively projected to three planes to obtain three corresponding discs, features on corresponding multi-level texture planes are respectively extracted from each disc and are spliced, and multi-level three-plane texture features are obtained; and inputting a to-be-reconstructed video frame into the dynamic deformation network and the multi-level texture three-plane, inputting the obtained multi-level three-plane texture features and the specified view angle into the static neural radiation field network for neural rendering, and reconstructing a rendered image of the to-be-reconstructed video frame under the specified view angle.

The invention provides a dynamic double-closed-loop control method for a magnetic suspension energy storage flywheel rotor based on LADRC. The dynamic double-closed-loop control method is used for realizing stable suspension and dynamic control of the magnetic suspension energy storage flywheel rotor. Firstly, a mathematical model of each degree of freedom of a flywheel rotor and a linearization model of a power amplifier-electromagnet coil under a magnetic bearing coordinate system are established; then, defining an electromagnetic coupling term, a gyroscope coupling term and an unbalance term in the flywheel rotor mathematical model as external disturbance, deducing a flywheel rotor and power amplifier-electromagnet coil mathematical model for controller design, and designing a position loop second-order LADRC controller and a current loop first-order LADRC controller; and finally, setting controller parameters of a position loop and a current loop by adopting a pole assignment method. Simulation results show that the double-closed-loop control strategy can realize effective decoupling among degrees of freedom under static and high-speed suspension working conditions of the flywheel rotor, and compared with PID control, the double-closed-loop control strategy is good in dynamic characteristic and strong in anti-interference capability. The method is of great significance in guaranteeing stable and reliable operation of the magnetic suspension energy storage flywheel rotor.

The present invention discloses a method for fine-grained control of bridge construction progress based on a digital twin multidimensional model, comprising the following steps: analyzing the composition of the bridge from the perspective of physical entities, establishing a physical entity decomposition structure of the bridge facility in combination with the fine-grained control objectives; constructing a digital twin geometric model of the bridge; further constructing a digital twin physical model of the bridge; further constructing a digital twin behavior model of the bridge; further constructing a digital twin rule model of the bridge; and using the digital twin multidimensional model of the bridge to fine-grained control of the construction progress. The present invention proposes a method for establishing a digital twin model for bridge construction progress control from four dimensions: geometry, physics, behavior, and rules, and innovatively proposes a method for calculating the number of days of construction progress delay based on "time" and "number of components" in the dimension of the rule model, which effectively improves the degree of refinement of the whole process control of the bridge construction progress.

The invention relates to the technical field of voice processing, in particular to an emotion voice conversion method and device from source emotion to target emotion. The method comprises the following steps of: 1, acquiring voice data Voice1 with source emotion and voice data Voice2 with target emotion; step 2, carrying out feature processing on the Voice1 and the Voice2 respectively; step 3, inputting a processing result of the step 2 into a trained emotional speech conversion model for processing to obtain a reconstructed Mel spectrum NewMel3; and step 4, carrying out sound code reduction on the NewMel3 to obtain voice data Voice3 with target emotion. Through the designed emotional voice conversion model, effective decoupling of emotional information and non-emotional information can be realized, dependence on target emotional information of a converted speaker is not needed, and emotional voice with a better conversion effect can be obtained.

A hydrogen energy storage power station control system adopts a layered structure, comprehensively judges from electricity price information, renewable energy output, environment temperature and hydrogen storage tank state data through a reinforcement learning model, selects a hydrogen production, power generation or standby mode, and sets power targets of an electrolytic cell and a fuel cell; the middle layer uses an optimization algorithm to refine power distribution, the lower layer executes a power instruction through an industrial control algorithm, and an independent safety guarantee layer is set to monitor indexes such as hydrogen concentration, pressure and temperature. In the optimization process of the system, energy consumption and equipment life loss of an electrolytic bath and auxiliary equipment are considered at the same time, full-life-cycle economy is achieved, the economy of a hydrogen energy storage power station is improved, and the service life of key equipment is prolonged.

The invention relates to a high-precision steel surface defect detection method suitable for a complex industrial environment, and aims to solve the problems of feature coupling and background interference caused by data scarcity, long tail distribution and multi-defect coexistence. According to the method, a conditional generative adversarial network (CGAN) and a convolutional neural network (CNN) are combined, an SE-Net channel attention mechanism is integrated, and the method is used for automatic detection and classification of multiple defects on the steel surface. By introducing a physical constraint CGAN data generation method, a multi-defect coexistence composite image conforming to industrial reality can be generated, so that the diversity and accuracy of a training data set are effectively enhanced. The model optimizes the expression of defect features through multi-level feature extraction and an attention mechanism, and enhances the distinguishing ability and recognition precision of multiple defects. Experiments show that the method is excellent in performance in a multi-defect identification task, the accuracy rate of 98.89% and the F1 score of 99.72% are achieved, the method is remarkably superior to a traditional detection method and other deep learning models, and the method has high robustness and practical value in a complex industrial environment.

The invention discloses an operation and maintenance release management system, method, device and equipment and a storage medium. The system comprises a system application layer, an API adaptation layer, a tool layer and an instance layer. The system application layer is used for managing at least one management application system and issuing operation and maintenance release management tasks generated by the management application systems to the API adaptation layer; the API adaptation layer is used for determining tool execution instructions corresponding to the API execution instructions according to the task execution related information and issuing the task execution related information and the tool execution instructions to the tool layer; the tool layer is used for scheduling a task execution tool to execute a tool execution instruction, generating a machine execution script, determining an instance server according to task execution related information, and issuing the machine execution script to the instance server in the instance layer; and the instance layer is used for managing at least one instance server to execute a corresponding machine execution script, generating and feeding back a script execution result, and managing an operation and maintenance release state of the system instance according to the script execution result.

The invention discloses a vehicle-mounted voice assistant application module, equipment and a vehicle. The vehicle-mounted voice assistant application module comprises a voice acquisition and recognition sub-module and a data output sub-module, the voice acquisition and recognition sub-module is used for acquiring user voice, and generating a voice request according to the user voice if the user voice comprises a wake-up word; sending the voice request to a corresponding request processing end, and sending response data fed back by the request processing end to the data output sub-module through a target protocol; wherein in the target protocol, the voice acquisition and recognition sub-module is used as a communication client, and the data output sub-module is used as a communication server; and the data output sub-module is used for receiving the response data sent by the voice acquisition and recognition sub-module and outputting the response data. Therefore, through module division, effective decoupling can be realized, the development cost and the maintenance cost are reduced, and the stability and the expandability of voice assistant application are also provided.

The invention discloses a small-scale deformation InSAR time sequence inversion method and system based on multiple stable points, and the method comprises the following steps: S1, carrying out the small baseline interference processing of multi-temporal synthetic aperture radar data, and screening high-coherence pixel points, S2, recognizing a stable point candidate set with dense spatial distribution; s3, constructing an initial atmospheric delay phase model covering a full interferogram range; s4, obtaining an interferometric phase after atmospheric correction; and S5, inverting the deformation rate and the deformation time sequence parameters of the InSAR again. According to the method, dense stable points are screened based on a deformation rate threshold value, and a space atmosphere model is constructed, so that the problem that atmosphere delay and nonlinear deformation are difficult to separate in InSAR time sequence analysis is effectively solved, and high-signal-to-noise-ratio and high-precision time sequence monitoring of small-scale weak deformation of the earth surface is realized.

The invention relates to the technical field of automobile frame devices, and particularly discloses an auxiliary frame structure which comprises a front cross beam, a rear cross beam and two longitudinal beams, the two longitudinal beams are arranged in a spaced mode, the front cross beam and the rear cross beam are connected between the two longitudinal beams respectively, the longitudinal beams are provided with safety collision force transmission induction structures, and the safety collision force transmission induction structures are connected between the front cross beam and the rear cross beam. The safety collision force transmission induction structure comprises a first induction part, a second induction part and a third induction part, the longitudinal beam is provided with a first end face and a second end face, the first induction part is formed at the end, close to the front cross beam, of the first end face, the second induction part is formed at the end, close to the front cross beam, of the second end face, and the third induction part is formed at the end, close to the front cross beam, of the second end face. The third induction part is formed on the second end face, and the third induction part is farther away from the front cross beam relative to the second induction part. The invention further discloses a vehicle. According to the auxiliary frame structure, effective decoupling of collision force can be achieved, the collision prevention performance of the auxiliary frame structure is improved, and the impact absorption capacity of a whole vehicle is improved.

The invention relates to an engineering calculation-oriented algorithm editing system and method.An algorithm editing instruction received by the system comprises a storage path, an arrangement relation and related parameter information of operators needing to be called, and an operator library is set for calling and arrangement during engineering algorithm editing; and processing the arranged operator combination in combination with related parameter information in the algorithm editing instruction to generate a to-be-edited engineering algorithm, and calling the generated engineering algorithm by the system according to user requirements to calculate the target calculation object. According to the method, an algorithm programming function and a software programming function can be effectively decoupled, the algorithm development efficiency is improved, meanwhile, an algorithm running function is designed, an engineering algorithm engineer can conveniently develop and apply the algorithm, whether problems occur in the application process of the engineering algorithm or not can be conveniently known in time, and the engineering algorithm is better perfected.

This invention provides a fault detection method and system based on low-rank decomposition and common trend decoupling. The method first acquires historical process data under normal operating conditions of an industrial process as training data. Then, it decomposes the training data into clean data representing the main structural information and sparse noise data representing noise interference using low-rank decomposition. Next, it constructs a common trend learning model based on the clean data, reconstructs the non-stationary common trend, and uses the residual between the clean data and the non-stationary common trend as a stationary feature. Decoupling is achieved by applying stationarity constraints. Further, a fault detection model is obtained through joint optimization. In the detection stage, the process data to be detected is input into the model, and monitoring statistics based on the reconstruction error of the stationary feature and the common trend are constructed respectively. The result of the control limit comparison determines whether a fault has occurred, thereby improving the robustness of fault detection in non-stationary industrial processes.

The application provides a security-constrained economic dispatch method based on a dual Halpern iteration, and belongs to the technical field of power system optimization dispatch and artificial intelligence computing infrastructure. The method comprises the following steps: constructing an original model of a security-constrained economic dispatch problem SCED; modeling the renewable energy output uncertainty by using a Gaussian mixture model, converting the chance constraint in the original model into a deterministic linear constraint, and thus converting the original model into a convex quadratic programming problem; constructing a limited Wolfe dual model for the convex quadratic programming problem, solving the dual model by using a Halpern iteration, and obtaining a dispatch optimization result. The application can solve the problems of low calculation efficiency, large memory overhead, low GPU utilization rate and difficulty in stable solution in a super-large-scale scenario in large-scale SCED, and can quickly generate an executable dispatch scheme, so that the calculation efficiency and engineering applicability of online dispatch and rolling optimization are improved.

The invention discloses a satellite technology-based surface temperature and surface emissivity synchronous inversion method. The method comprises the following steps of: acquiring adjacent observation data; atmospheric parameter initial values are calculated; performing estimation based on the atmospheric parameter initial value and an I-DN algorithm to obtain a first surface temperature, a first surface emissivity, a temperature profile correction value and a humidity profile correction coefficient; the initial atmospheric temperature profile and the initial humidity profile are corrected, and corrected atmospheric parameters are calculated through an RTTOV model; estimating based on the first surface temperature, the first surface emissivity, the corrected atmospheric parameters and an I-TES algorithm to obtain a second surface temperature and a second surface emissivity; calculating weighted surface emissivity; estimating based on the weighted surface emissivity and an I-SW algorithm to obtain a third surface temperature; calculating a weighted surface temperature; and completing integrated inversion of the surface temperature and the surface emissivity of the target area. According to the method, the accuracy and the stability of synchronous inversion of the surface temperature and the emissivity are improved.

The application relates to the technical field of electronic component reliability, and discloses a flat transformer voltage withstand performance sure reliability modeling and evaluation method, which comprises the following steps: firstly, a constitutive model fusing a physical mechanism is constructed, a time degradation model and a temperature influence model are established based on the geometric size, material properties and environmental stress of an insulating layer, and a process capability parameter is introduced to represent manufacturing fluctuation; secondly, a reliability evaluation model is established, a failure threshold and a margin equation are defined, and physical parameters are identified as random variables to quantify uncertainty; then, Monte Carlo algorithm is used to sample random variables, virtual samples are generated, and voltage withstand margin is calculated on an evaluation time sequence; finally, the number of samples meeting a safety criterion is counted, instantaneous reliability is calculated, and a full-life-cycle reliability curve is generated. The application effectively decouples and quantifies the influence of manufacturing processes, temperature stress and time degradation on voltage withstand performance, and realizes sure reliability evaluation with high confidence.

This invention belongs to the field of vehicle control technology and relates to a vehicle motion control system based on angular module dynamics. The system constructs a two-degree-of-freedom driver control system including steering wheel input and vehicle posture lever input, used to represent path curvature tracking requirements and vehicle posture adjustment requirements, respectively. Based on the driver control input, ideal yaw rate control targets and ideal center-of-gravity sideslip angle control targets are constructed, and the four-wheel steering angle feedforward control quantities are determined through vehicle kinematics methods of pure yaw motion and pure lateral motion. Through feedback optimization control, the additional yaw moment and four-wheel steering angle control quantities are coordinated under composite operating conditions. Based on adaptive adjustment of control weights and constraints under operating conditions, the distribution control of four-wheel drive / braking torque and four-wheel steering angle is achieved. This invention achieves decoupled control of vehicle yaw motion and lateral motion by constructing a two-degree-of-freedom driver control system and combining feedforward kinematics and feedback dynamics control.

The invention relates to the technical field of data analysis, in particular to a palmprint recognition anti-fraud method based on double-branch self-supervised learning, and the method comprises the steps: constructing a double-branch self-supervised reconstruction architecture, determining a first to-be-processed image based on a high-frequency self-supervised branch, and determining a second to-be-processed image based on a chromaticity regularization self-supervised branch. Dividing the two images into patch blocks which are not overlapped with each other, independently generating random mask matrixes with different spatial distributions, determining visible areas, inputting the patch blocks of the visible areas into a shared encoder for feature extraction, outputting latent variable features, and splicing the latent variable features with learnable mask marks to obtain a patch matrix; and respectively sending to a high-frequency decoder and a chroma decoder, determining total loss, carrying out physical decoupling on illumination and material attributes, and carrying out fine adjustment on the architecture to finish convergence of the architecture. According to the method, the palm print features with discrimination are extracted through the shared encoder, and chromaticity distribution consistency constraints are introduced, so that effective decoupling of ambient light and real material attributes is realized.

This invention discloses an underwater wireless optical communication alignment method and system based on spatial coherence adaptive differential observation. The method includes: constructing a photodetector array in the receiver, constructing a normalized differential observation structure, and through differential observation and dynamic spacing adjustment scheme, a laser system transceiver alignment method that can effectively suppress turbulent multiplicative scintillation and adapt to different water dynamic conditions can be developed, significantly improving the alignment accuracy and robustness between the receiver and the transmitter in complex turbulent environments.

The application relates to a city functional area classification method based on a graph adaptive propagation network and a medium, which comprises the following steps: respectively preprocessing original road network data, point of interest (POI) data and street view image (SVI) data to obtain a road network graph and an adjacency matrix thereof, a street POI sequence and a street SVI set; respectively extracting features from the street POI sequence and the street SVI set to obtain semantic feature vectors and visual feature vectors; using two graph adaptive propagation networks with the same structure but independent parameters, performing multi-scale propagation and adaptive aggregation on the semantic feature vectors and the visual feature vectors on the shared road network graph to obtain a POI modal feature matrix enhanced by graph propagation and an SVI modal feature matrix enhanced by graph propagation; fusing the obtained POI modal feature matrix and SVI modal feature matrix through cross-modal gating to obtain fusion features; and inputting the fusion features as an input into a trained multilayer perceptron classifier.

The application relates to an air-ground cooperative navigation control method and system based on global priori of an unmanned aerial vehicle and elastic preset performance tracking of an unmanned vehicle, and belongs to the technical field of autonomous navigation and cooperative control of unmanned systems. In view of the problems of a single system, such as a large blind area of sensing, a prominent calculation bottleneck, easy falling into a local minimum value in local obstacle avoidance, weak anti-interference performance in trajectory tracking and waste of communication resources in a complex environment, a layered cooperative architecture is adopted, the unmanned aerial vehicle is used to complete mapping, two-dimensional projection and global path planning and issue a discrete topological path, the unmanned vehicle generates a local trajectory through spatial domain robust differential game, dynamic optimization is realized by combining elastic preset performance control and internal and external double-loop reinforcement learning, high-precision tracking and global optimization of energy consumption are realized, and an event triggering mechanism is used to reduce communication load. The application improves navigation reliability, trajectory stability and resource utilization efficiency in a complex dynamic environment.

The present invention discloses an airport dimming device and method, including a central controller layer, a regional control layer and a unit control layer. The central controller layer is used to receive control requirements for lights from other systems and outputs of system error states; the regional control layer is used to transmit information with the central controller layer and the unit control layer respectively; the unit control layer is responsible for specific lighting control, and the unit control layer includes three functional units at each lighting control node, and each functional unit is at least two. The device of the present invention adopts a three-layer network to effectively decouple the system and effectively ensure the security of the system. The present invention adopts a backup strategy, and in the event of an occasional failure, the failed node can be skipped and the system can continue to work, effectively ensuring the robustness of the system.

This invention discloses a method for measuring Rydberg atoms using micro-Doppler agility under a polarization-multiplexed microwave field, belonging to the field of quantum precision measurement technology. The method includes generating a dynamic coded sequence and an interleaved coded microwave field by aperiodically pairing and jumping the microwave polarization state with the carrier frequency; acquiring the target echo signal and applying it to the Rydberg atom ensemble with a reference microwave field to generate a quantum coherent state; applying a synchronous laser pulse to acquire a time-series optical response signal; obtaining the signal's fundamental frequency, envelope spectrum, and initial phase component through time-frequency transformation and feature separation; and performing parameter mapping and synthesis to generate three-dimensional micro-motion parameters. This invention uses an aperiodically agile polarization-frequency-multiplexed microwave field for active target detection and utilizes the Rydberg atom ensemble for quantum coherent detection and optical readout of the echo signal, enabling high-sensitivity, high-precision, and high-anti-interference measurement of the target's three-dimensional micro-motion parameters.

The invention discloses an anthropomorphic robot multi-part expression cooperative control method, device, system, equipment and product, and relates to the technical field of anthropomorphic humans. The method comprises the steps of firstly obtaining a target expression control signal, then responding to the signal to obtain one-to-one corresponding and mutually independent silence delay parameters for at least two expression execution mechanisms, then synchronously starting a plurality of silence delays based on the parameters, and finally respectively driving the corresponding execution mechanisms when the delays are finished. Therefore, by introducing independent and adjustable silent delay parameters and synchronously starting, the problems of'programmed sequence sense 'and'mechanical cooperation' caused by asynchronous distribution of multi-part action starting instructions in the prior art are solved, and the biological characteristics of'synchronous response and asynchronous execution 'of human facial neural signals are effectively simulated; therefore, the naturalness and life feeling of robot expressions are remarkably improved, and practical application and popularization are facilitated.

This invention discloses a method, device, and medium for querying the capacity of logical units based on pre-allocation, relating to the field of distributed storage technology. By pre-allocating physical storage blocks and establishing mapping relationships for logical storage units that are not bound to physical storage blocks when responding to capacity query commands, the system can accurately calculate the available capacity based on the actual physical attributes of the physical storage block, solving the problem of inaccurate queries. Simultaneously, a pre-allocation state is introduced, distinguishing this state from the conventional allocated or active states. Physical storage blocks in the pre-allocated state are considered idle in the resource reclamation statistics of the storage device. When a logical storage unit receives a data write command, the pre-allocation state is changed back to the allocated state, ensuring the correctness of physical block lifecycle management. While ensuring high accuracy in capacity queries, efficient resource scheduling is achieved, improving the overall reliability and performance stability of the storage system.

The present application relates to the technical field of predictive maintenance and fault diagnosis of industrial equipment, and particularly relates to a multi-motor coupled vibration intelligent diagnosis method and system. The technical scheme comprises the following steps: based on the collected vibration data of the multi-motor under the normal operation state and the physical parameters of the motor equipment, a parameterized fault simulation module is used to generate a coupled vibration simulation data set with accurate fault labels. The present application realizes the cold start of the diagnosis system under the condition of no historical fault samples; the deep reinforcement learning agent is used to process the space-time coupled signal, and the layered reward mechanism is combined to significantly improve the decoupling diagnosis accuracy and early warning capability; through the online adaptive learning module, the system can be fine-tuned in real time according to the field feedback, and quickly adapt to the individual equipment state; the simulation-model evolution module uses the accumulated real data to continuously optimize the simulator and the diagnosis model, and forms a self-iterating closed loop.

The present application belongs to the technical field of image processing, and particularly relates to an artistic style transfer method and system based on frequency domain modulation and attention guidance, aiming at the difficulty of effectively separating content structure and style features in the prior art, which easily leads to structural damage or insufficient stylization of the generated results. The present application first initializes the strategy based on structure perception frequency domain modulation, effectively decouples content and style from the frequency domain; constructs a decoupling cross-attention mechanism, processes text, content and style features through parallel channels respectively, thereby avoiding feature interference while precisely regulating the generation process; and uses an attention-driven guidance mechanism to model the style transfer as an iterative optimization problem in the latent space, dynamically calibrates the feature distribution through gradient guidance, and effectively alleviates the feature dilution problem in the deep network.

The composite film is formed by laminating and compounding at least two layers of polymer porous films with anisotropic microporous structures, and the adjacent film layers are stacked after rotating by 45-135 degrees around the axis perpendicular to the film surface, the micropore channels of the upper and lower film layers form a non-through structure in a three-dimensional space, and decoupling of a heat flow transmission direction and a gas diffusion direction is realized. The interlayer dislocation degree is controlled by regulating and controlling the rotation angle, an efficient thermal resistance barrier is formed in the thickness direction, and meanwhile a gas diffusion channel is kept in the plane direction. The composite film has excellent heat insulation performance and air permeability, the heat conductivity coefficient can be as low as 0.028 W / (m.K), the water vapor transmission rate is kept at 980 g / (m.24 h) or above, the preparation process is simple and controllable, and the composite film can be widely applied to the fields of automobile interiors, building heat preservation, garment materials and the like.

The application relates to the technical field of intelligent control of underwater robots, and provides a bionic underwater robot motion control method and system based on asymmetric information. The bionic underwater robot motion control method based on asymmetric information comprises the following steps: first, an underwater robot is abstracted as a double-control system with two sub-controllers of a wave fin and a tail propulsion mechanism. In view of the characteristics that the motion state of the underwater robot cannot be directly measured and the environment is unstructured, the output of acoustic, inertial and other sensors is modeled as a multi-modal linear perception model. Based on the model, corresponding asymmetric measurement signal fusers are designed for each controller. In order to solve the "fusion-control" coupling problem caused by asymmetric information, the random maximum value principle is adopted to realize the decoupling design of the fusers and the controllers, and finally an explicit solution of the optimal decentralized controller is given. The control precision and robustness of the system in a dynamically uncertain environment are significantly improved.

The invention provides a sensitive attribute forgetting method and device for a recommendation system. An embedding layer of the recommendation system comprises a user embedding matrix and an article embedding matrix; the method comprises the following steps: performing matrix decomposition on an article embedding matrix to obtain an independent space orthogonal basis; performing feature decoupling on the user representation vector to obtain a correlation vector and an irrelevant vector; the irrelevant vector is represented as a linear combination of a to-be-optimized parameter and an irrelevant space orthogonal basis; constructing user representation distribution and sensitive attribute distribution, and selecting the optimal bandwidth of the HSIC core; constructing a forgetting loss function used for measuring the dependency between the user representation distribution and the sensitive attribute distribution, and replacing a user representation vector in the forgetting loss function with a combination of a correlation vector and an irrelevant vector to obtain a target function; and under the condition of keeping the correlation vector unchanged, taking a minimization objective function as an objective, obtaining a final user representation vector through iterative optimization, and taking the final user representation vector as a result after the sensitive attribute is forgotten.