Open domain adaptive generalization method and system based on meta-learning and multi-source data enhancement

Abstract

The invention relates to an open domain adaptive generalization method and system based on meta-learning and multi-source data enhancement. The method comprises the steps of designing a meta-learning task and a loss function, simulating data distribution of an unknown target domain, enabling a model to be trained on a plurality of source domains, and learning cross-domain universal feature representation; in a feature level, simulating data distribution differences among different source domains through a feature mixing technology; on the label level, a soft label distillation method is adopted, and knowledge is transferred from other source domains; feature mapping between a source domain and an unknown target domain is learned, feature conversion is carried out, an alignment strategy is set, and a feature extraction mechanism of the model is adjusted, so that the model is more adaptive to data characteristics of the target domain. Through meta-learning and multi-source data enhancement, the model can enhance the generalization ability of the model by simulating the data distribution difference between different source domains under the condition that only single source domain data exists.

Description

Technical Field

[0001] The present invention relates to the field of domain adaptation technology, and in particular to an open domain adaptation generalization method and system based on meta-learning and multi-source data enhancement. Background Art

[0002] In the field of machine learning, especially when facing intelligent recognition tasks, the generalization ability of the model is particularly critical. Generalization ability ensures that the model not only performs well on training data, but also makes accurate predictions on new data that has not been seen. However, most existing technologies are limited to the training and testing of a single data set. Once faced with a new domain that is inconsistent with the distribution of training data, the model performance will drop significantly. This phenomenon is widely recognized as domain shift. When dealing with generalization problems, existing machine learning models usually rely on a large amount of labeled data for training. However, in practical applications, obtaining this data is both expensive and time-consuming. In addition, existing technologies are also unable to quickly adapt to unknown domains. For example, in the sentiment analysis task of natural language processing, if the model is only trained on a specific domain (such as hotel reviews), when it is applied to another domain (such as restaurant reviews), the performance will be greatly reduced.

[0003] To address this challenge, a variety of methods have been proposed, including domain adaptation, domain generalization, and adversarial training. Domain adaptation methods attempt to use labeled data from the source domain and unlabeled data from the target domain to adjust the model, but their application is limited by access to target domain data. Domain generalization methods attempt to learn from multiple source domains to improve the performance of the model in an unknown target domain, but their effectiveness is limited when the source and target domains are very different. Adversarial training can enhance the robustness of the model against attacks, but it focuses mainly on model security rather than generalization capabilities in open domains.

[0004] Therefore, traditional domain adaptation methods often have problems such as poor generalization ability, low model adaptability, and strong data dependence. Summary of the invention

[0005] Based on this, in order to solve the above technical problems, an open domain adaptation generalization method and system based on meta-learning and multi-source data enhancement is provided, which can quickly adapt to new fields under limited data conditions, effectively deal with category imbalance problems, and improve the generalization ability of the model in unknown domains.

[0006] An open domain adaptation generalization method based on meta-learning and multi-source data enhancement, the method comprising:

[0007] Collecting raw data, and preprocessing the raw data to obtain processed data;

[0008] Acquire an initial model, design a meta-learning task for the initial model based on the processed data, and construct a loss function corresponding to the initial model based on the meta-learning task, formulate a meta-learning strategy, and formulate a parameter update strategy to complete the meta-learning process construction of the initial model, and obtain an initial model based on meta-learning;

[0009] Introducing multi-source data enhancement technology into the initial model based on meta-learning, simulating data distribution differences between different source domains through feature mixing technology at the feature level of the initial model based on meta-learning, and adopting soft label distillation method at the label level of the initial model based on meta-learning to transfer knowledge from other source domains to obtain the initial model after data enhancement;

[0010] In the initial model after data enhancement, the feature representation on the unknown domain is optimized, the feature mapping between the source domain and the unknown target domain is learned, a feature alignment strategy is adopted to ensure the consistency of the source domain and the unknown target domain features, the source domain features are adjusted by learning linear transformation to achieve feature conversion, and the feature extraction mechanism is adjusted to generate a target model;

[0011] Based on the target model, corresponding auxiliary decisions are made according to the original data.

[0012] In one embodiment, the meta-learning task includes a base learning task and a meta-learning subtask;

[0013] The base learning task is used to train the basic classification ability of the initial model on each source domain;

[0014] The meta-learning subtask is used to optimize the adaptability of the initial model to the new domain.

[0015] In one embodiment, constructing a loss function corresponding to the initial model based on the meta-learning task includes:

[0016] Determine a base learning task loss based on the base learning task, determine a meta-learning sub-task loss based on the meta-learning sub-task, and obtain a regularization term;

[0017] A composite loss function corresponding to the initial model is constructed according to the base learning task loss, the meta-learning subtask loss, and the weight of the regularization term.

[0018] In one embodiment, formulating a meta-learning strategy includes:

[0019] A hybrid strategy based on Dirichlet distribution is used to generate new samples in the feature space;

[0020] For the generated new sample, a weight vector is extracted from the Dirichlet distribution to calculate the weighted sum.

[0021] In one embodiment, the method further comprises:

[0022] Combine the mixed features and the features transferred by the soft label distillation method for optimization;

[0023] The diversity of data from different source domains is measured, and enhanced ensemble learning is performed on data from different source domains.

[0024] In one embodiment, optimizing the feature representation on the unknown domain in the data-augmented initial model and learning the feature mapping between the source domain and the unknown target domain includes:

[0025] Determine a mapping function based on the initial model after data enhancement;

[0026] The source domain features are mapped to the target domain feature space according to the mapping function, and the feature mapping between the source domain and the unknown target domain is estimated and learned through supervised learning.

[0027] In one embodiment, a feature alignment strategy is used to ensure the consistency of features between the source domain and the unknown target domain, including:

[0028] The feature covariance of the source domain and the unknown target domain is calculated, and the feature covariance is maximized until the features of the source domain and the unknown target domain are consistent.

[0029] In one embodiment, feature conversion is achieved by adjusting source domain features through learning linear transformation, including:

[0030] A conversion loss function is defined, and whether the features of the source domain and the unknown target domain are consistent is determined by the conversion loss function to obtain a judgment result;

[0031] If the judgment result is that the source domain and the unknown target domain features are consistent, a conversion matrix and a bias vector are calculated according to the source domain features and the target domain features;

[0032] The transformation matrix and the bias vector are used to control the difference between the transformed features and the unknown target domain features to be minimal.

[0033] In one embodiment, the method further comprises:

[0034] According to the position target domain, the target model is evaluated for performance, and it is determined whether the target model can meet the generalization requirements, and a performance evaluation result is obtained;

[0035] According to the performance evaluation result, the target model is optimized or the learning is terminated.

[0036] An open domain adaptation generalization system based on meta-learning and multi-source data enhancement, the system comprising:

[0037] A data acquisition module is used to collect raw data and pre-process the raw data to obtain processed data;

[0038] A meta-learning framework module, used to obtain an initial model, design a meta-learning task for the initial model based on the processed data, and construct a loss function corresponding to the initial model based on the meta-learning task, formulate a meta-learning strategy, formulate a parameter update strategy, complete the meta-learning process construction of the initial model, and obtain an initial model based on meta-learning;

[0039] A multi-source data enhancement module, used to introduce multi-source data enhancement technology into the initial model based on meta-learning, simulate the data distribution difference between different source domains by feature mixing technology at the feature level of the initial model based on meta-learning, and adopt a soft label distillation method at the label level of the initial model based on meta-learning to transfer knowledge from other source domains to obtain the initial model after data enhancement;

[0040] A feature calibration module is used to optimize the feature representation on the unknown domain in the initial model after data enhancement, learn the feature mapping between the source domain and the unknown target domain, adopt a feature alignment strategy to ensure the consistency of the source domain and the unknown target domain features, adjust the source domain features by learning linear transformation to achieve feature conversion, and adjust the feature extraction mechanism to generate a target model;

[0041] A decision module is used to make corresponding auxiliary decisions according to the original data based on the target model.

[0042] The above-mentioned open domain adaptation generalization method and system based on meta-learning and multi-source data enhancement, through meta-learning and multi-source data enhancement, enables the model to learn more generalized feature representations when only a single source domain data is available; through feature calibration, it can quickly adjust the feature distribution of the unknown target domain, effectively narrowing the distribution difference between the source domain and the target domain; through model rapid adaptation and evaluation, it can achieve rapid adaptation and performance evaluation of unknown domains, ensuring the stability and accuracy of the model in the unknown domain; multi-source data enhancement technology can enhance the generalization ability of the model by simulating the data distribution differences between different source domains under limited labeled data conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is an application environment diagram of an open domain adaptation generalization method based on meta-learning and multi-source data enhancement in one embodiment;

[0044] Figure 2 1 is a flow chart of an open domain adaptation generalization method based on meta-learning and multi-source data enhancement in one embodiment;

[0045] Figure 3 is a structural block diagram of an open domain adaptation generalization system based on meta-learning and multi-source data enhancement in one embodiment;

[0046] Figure 4 A schematic diagram of the structure of a meta-learning framework module in an embodiment;

[0047] Figure 5 is a schematic diagram of the structure of a multi-source data enhancement module in one embodiment;

[0048] Figure 6 is a schematic diagram of the structure of a feature calibration module in one embodiment;

[0049] Figure 7 A state transition diagram for model learning in one embodiment;

[0050] Figure 8 It is a schematic diagram of a solution flow of an open domain adaptation generalization method based on meta-learning and multi-source data enhancement in one embodiment;

[0051] Fig. 9 is a schematic diagram of a process of environmental adaptability learning applied to autonomous driving in one embodiment;

[0052] Fig.10 A schematic diagram of a process for product quality control applied in intelligent manufacturing in one embodiment;

[0053] Fig.11 A schematic diagram of a process for abnormal behavior detection applied to smart security in one embodiment;

[0054] Fig.12 is a schematic diagram of a process for disease diagnosis applied to medical image analysis in one embodiment;

[0055] Fig.13 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0057] The open domain adaptation generalization method based on meta-learning and multi-source data enhancement provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Figure 1As shown, the application environment includes a computer device 110 . The computer device 110 can collect raw data and preprocess the raw data to obtain processed data; the computer device 110 can obtain an initial model, design a meta-learning task for the initial model based on the processed data, and construct a loss function corresponding to the initial model based on the meta-learning task, formulate a meta-learning strategy, and formulate a parameter update strategy to complete the meta-learning process construction of the initial model to obtain an initial model based on meta-learning; the computer device 110 can introduce multi-source data enhancement technology in the initial model based on meta-learning, simulate the data distribution differences between different source domains through feature mixing technology at the feature level of the initial model based on meta-learning, and use a soft label distillation method at the label level of the initial model based on meta-learning to transfer knowledge from other source domains to obtain an initial model after data enhancement; the computer device 110 can optimize the feature representation on the unknown domain in the initial model after data enhancement, learn the feature mapping between the source domain and the unknown target domain, use a feature alignment strategy to ensure the consistency of the source domain and the unknown target domain features, adjust the source domain features through learning linear transformation to achieve feature conversion, and adjust the feature extraction mechanism to generate a target model; the computer device 110 can make corresponding auxiliary decisions based on the original data based on the target model. The computer device 110 may be, but is not limited to, various personal computers, laptop computers, robots, unmanned aerial vehicles, and other devices.

[0058] In one embodiment, Figure 2 As shown, an open domain adaptation generalization method based on meta-learning and multi-source data enhancement is provided, including the following steps:

[0059] Step 202: collect original data and pre-process the original data to obtain processed data.

[0060] Among them, preprocessing can include data cleaning, standardization and denoising. By preprocessing the original data, subsequent model optimization operations are facilitated.

[0061] Step 204, obtain the initial model, design a meta-learning task for the initial model based on the processed data, and construct a loss function corresponding to the initial model based on the meta-learning task, formulate a meta-learning strategy, and formulate a parameter update strategy to complete the meta-learning process construction of the initial model and obtain an initial model based on meta-learning.

[0062] Specifically, the computer device can construct and optimize the meta-learning process of the model. Through the meta-learning tasks and loss functions, the initial model can be trained on multiple source domains and learn universal feature representations across domains. By simulating the data distribution of the unknown target domain, the initial model can still maintain stable performance when facing new and unknown domains.

[0063] In one embodiment, a provided open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a meta-learning task design process, wherein the meta-learning task includes a base learning task and a meta-learning sub-task; the base learning task is used to train the basic classification ability of the initial model on each source domain; the meta-learning sub-task is used to optimize the adaptability of the initial model to the new domain.

[0064] Among them, the design of the meta-learning task aims to improve the generalization ability of the model by simulating the data distribution of the unknown target domain. The meta-learning task consists of two parts: the base learning task and meta-learning subtasks

[0065] The base learning task is used to train the basic classification ability of the model in each source domain. The base learning task is defined as: Where N is the total number of training samples, x i and i are the features and labels of the i-th sample, p(y i |x i ; θ) is the predicted probability under the model parameters θ.

[0066] The meta-learning subtask is used to optimize the model's ability to quickly adapt to new fields. The meta-learning subtask is defined as: Where M is the number of samples in the meta-learning subtask, and are samples and labels generated by the meta-learning strategy, and θ' is the model parameter after fast adaptation.

[0067] In one embodiment, a provided open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a process of constructing a loss function, the specific process including: determining a base learning task loss based on a base learning task, determining a meta-learning subtask loss based on a meta-learning subtask, and obtaining a regularization term; constructing a composite loss function corresponding to the initial model according to the weights of the base learning task loss, the meta-learning subtask loss, and the regularization term.

[0068] The loss function is a key factor in guiding model training. In this embodiment, a composite loss function is designed, which combines the base learning task loss and the meta-learning subtask loss, as well as a regularization term to control the complexity of the model and avoid overfitting. The composite loss function is defined as: Among them, λ and γ are the weights of the meta-learning task loss and regularization term, respectively.

[0069] In one embodiment, a provided open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a process of formulating a meta-learning strategy, the specific process including: using a hybrid strategy based on Dirichlet distribution to generate new samples in the feature space; for the generated new samples, extracting a weight vector from the Dirichlet distribution to calculate a weighted sum.

[0070] Among them, the meta-learning strategy is responsible for generating samples in the meta-learning task. These samples are generated by a special data augmentation technique to simulate the data distribution of the unknown target domain. Specifically, this embodiment adopts a hybrid strategy based on Dirichlet distribution to generate new samples in the feature space. Let α be a hyperparameter used to control the concentration of the Dirichlet distribution, and K is the number of source domains. For each newly generated sample, a weight vector λ~Dir(α) is first extracted from the Dirichlet distribution, and then the weighted sum is calculated: Among them, x k is a sample in the kth source domain.

[0071] The open domain adaptation generalization method based on meta-learning and multi-source data enhancement provided in this embodiment may also include a process of formulating a parameter update strategy. Specifically, in the meta-learning framework, the update strategy of the model parameters is crucial to the generalization ability and adaptability of the model. In this embodiment, an update strategy based on model parameter differences is adopted, which encourages the model to share knowledge between different tasks while maintaining task-specific feature representations: Among them, Δθ k It is task T k The model parameter update amount is ρ, and ρ is the weight of shared knowledge, which controls the degree of sharing of model parameters between different tasks. The meta-learning framework module fully considers the particularity of the open domain generalization problem, enabling the model to adapt to the new data distribution through rapid parameter adjustment when facing unknown domains, thereby maintaining stable performance.

[0072] Step 206, introduce multi-source data enhancement technology into the initial model based on meta-learning, simulate the data distribution differences between different source domains through feature mixing technology at the feature level of the initial model based on meta-learning, and use soft label distillation method at the label level of the initial model based on meta-learning to transfer knowledge from other source domains to obtain the initial model after data enhancement.

[0073] In order to enhance the adaptability of the model to different distributions, this embodiment introduces innovative multi-source data enhancement technology. At the feature level, feature mixing technology is used to simulate the data distribution differences between different source domains; at the label level, soft label distillation method is used to transfer knowledge from other source domains. By increasing the diversity and complexity of data, the generalization ability of the model can be effectively improved.

[0074] Among them, the feature-level mixing technology aims to enhance the adaptability of the model to the new domain by mixing features from different source domains. s (x) is the feature representation of sample x in the source domain s, α s is the corresponding weight, then the feature mixture can be expressed as: Where S is the number of source domains, and Weight α s It can be determined by an optimization problem so that the mixed feature representation can maximize the performance of the model on the target task.

[0075] The soft label distillation method at the label level is used to transfer knowledge from other source domains to compensate for the class imbalance problem that may exist in some source domains. s (y|x) is the predicted probability distribution of sample x belonging to category y in the source domain s. The goal of soft label distillation is to learn a mapping T such that: in, is the soft label of sample x in target domain t, β st are the transfer weights from the source domain s to the target domain t. These weights can be learned by minimizing the Kullback-Leibler divergence between the source and target domains:

[0076] In one embodiment, an open domain adaptation generalization method based on meta-learning and multi-source data enhancement is provided, which may also include an optimization process for multi-source data enhancement, the specific process including: optimizing the features after joint mixing and the features after transfer by the soft label distillation method; measuring the diversity of data from different source domains, and performing enhanced integrated learning on data from different source domains.

[0077] Among them, the optimization of multi-source data enhancement involves the joint optimization of feature mixing and soft label distillation. The optimization of multi-source data enhancement aims to find a balance point so that the model can learn cross-domain universal representations from feature mixing and obtain category diversity from soft label distillation. The optimization problem can be expressed as: in, It is a multi-source domain loss function that combines classification loss and distribution consistency loss, ensuring that the enhanced data not only helps improve the generalization ability of the model but also maintains the distribution characteristics of the source domain data.

[0078] Step 208, optimize the feature representation on the unknown domain in the initial model after data enhancement, learn the feature mapping between the source domain and the unknown target domain, adopt a feature alignment strategy to ensure the consistency of the source domain and the unknown target domain features, adjust the source domain features by learning linear transformation to achieve feature conversion, and adjust the feature extraction mechanism to generate a target model.

[0079] In this embodiment, the computer can also perform feature calibration operations on the model, the purpose of which is to optimize the feature representation of the model in the unknown domain. By learning the feature mapping between the source domain and the unknown target domain, the feature extraction mechanism of the model can be adjusted to make it more suitable for the data characteristics of the target domain. This process involves sophisticated feature conversion and alignment strategies to ensure the classification accuracy of the model in the unknown domain.

[0080] In one embodiment, a provided open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a process of performing feature mapping learning, the specific process including: determining a mapping function based on an initial model after data enhancement; mapping source domain features to the space of target domain features according to the mapping function, and estimating and learning the feature mapping between the source domain and the unknown target domain through supervised learning.

[0081] Among them, feature mapping learning is the basis of feature calibration, which aims to find a mapping function h that converts the source domain feature f s Mapped to the target domain feature f t The mapping can be achieved through supervised learning. Perform estimation learning, where N is the number of samples, and are the features of the i-th source domain and target domain samples respectively.

[0082] In one embodiment, a provided open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a process of ensuring the consistency of source domain and unknown target domain features, the specific process including: calculating the feature covariance of the source domain and the unknown target domain, and maximizing the feature covariance until the source domain and the unknown target domain features are consistent.

[0083] In this embodiment, the feature alignment strategy is used to ensure the consistency of source domain and target domain features. The alignment method is achieved by maximizing the feature covariance of the two domains: Among them, φ is the alignment parameter to be learned, and λ is the regularization coefficient used to prevent overfitting.

[0084] In one embodiment, a provided open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a process of performing feature conversion, and the specific process includes: defining a conversion loss function, judging whether the source domain and unknown target domain features are consistent through the conversion loss function, and obtaining a judgment result; if the judgment result is that the source domain and unknown target domain features are consistent, calculating a conversion matrix and a bias vector based on the source domain features and the target domain features; and controlling the difference between the converted features and the unknown target domain features to be minimized through the conversion matrix and the bias vector.

[0085] The core of feature conversion technology is to adjust the source domain features by learning linear transformation to better adapt to the data distribution of the target domain. This can be optimized by defining a loss function, which measures the consistency between the converted features and the target domain features. and target domain features The goal of feature transformation is to find a transformation matrix W and a bias vector b so that the transformed features Minimize the difference between the target domain features, which can be achieved by minimizing the following loss function To achieve: Where N is the number of samples, are the features after transformation.

[0086] In one embodiment, an open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a process of constructing a loss function for feature calibration.

[0087] Among them, the loss function of feature calibration combines the goals of mapping learning, feature alignment and conversion to ensure the generalization ability of feature representation: Among them, ν and v are weights used to balance different loss terms. Feature calibration technology takes into account the possible feature distribution differences between the source domain and the target domain in the open domain generalization problem. Through feature mapping learning and alignment strategies, it provides an effective feature adaptation mechanism for the model, which can effectively improve the feature representation ability of the model in the unknown target domain. In summary, the total loss function of the model is

[0088] Step 210, making corresponding auxiliary decisions based on the target model according to the original data.

[0089] In one embodiment, a provided open domain adaptation generalization method based on meta-learning and multi-source data enhancement may also include a process of evaluating the model, and the specific process includes: performing performance evaluation on the target model according to the position target domain, judging whether the target model can meet the generalization requirements, and obtaining a performance evaluation result; according to the performance evaluation result, optimizing the target model or ending learning.

[0090] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0091] In one embodiment, Figure 3 As shown, a system for an open domain adaptation generalization method based on meta-learning and multi-source data enhancement is provided, including: a data acquisition module 310, a meta-learning framework module 320, a multi-source data enhancement module 330, a feature calibration module 340 and a decision module 350, wherein:

[0092] The data acquisition module 310 is used to acquire raw data and pre-process the raw data to obtain processed data;

[0093] The meta-learning framework module 320 is used to obtain an initial model, design a meta-learning task for the initial model based on the processed data, and construct a loss function corresponding to the initial model based on the meta-learning task, formulate a meta-learning strategy, formulate a parameter update strategy, complete the meta-learning process construction of the initial model, and obtain an initial model based on meta-learning;

[0094] A multi-source data enhancement module 330 is used to introduce multi-source data enhancement technology into the initial model based on meta-learning, simulate the data distribution difference between different source domains through feature mixing technology at the feature level of the initial model based on meta-learning, and adopt a soft label distillation method at the label level of the initial model based on meta-learning to transfer knowledge from other source domains to obtain the initial model after data enhancement;

[0095] A feature calibration module 340 is used to optimize the feature representation on the unknown domain in the initial model after data enhancement, learn the feature mapping between the source domain and the unknown target domain, adopt a feature alignment strategy to ensure the consistency of the source domain and the unknown target domain features, adjust the source domain features by learning linear transformation to achieve feature conversion, and adjust the feature extraction mechanism to generate a target model;

[0096] The decision module 350 is used to make corresponding auxiliary decisions according to the original data based on the target model.

[0097] In one embodiment, the meta-learning task includes a base learning task and a meta-learning sub-task; the base learning task is used to train the basic classification ability of the initial model on each source domain; the meta-learning sub-task is used to optimize the adaptability of the initial model to the new domain.

[0098] In one embodiment, the meta-learning framework module 320 is also used to determine the base learning task loss based on the base learning task, determine the meta-learning subtask loss based on the meta-learning subtask, and obtain the regularization term; and construct a composite loss function corresponding to the initial model according to the base learning task loss, the meta-learning subtask loss, and the weight of the regularization term.

[0099] In one embodiment, the meta-learning framework module 320 is further used to adopt a hybrid strategy based on Dirichlet distribution to generate new samples in the feature space; for the generated new samples, a weight vector is extracted from the Dirichlet distribution to calculate a weighted sum.

[0100] In one embodiment, the structure of the meta-learning framework module is as follows: Figure 4 As shown, it mainly includes:

[0101] Meta-learning task design unit: responsible for generating meta-learning tasks that simulate the data distribution of unknown target domains, providing training signals for the model to adapt to new domains;

[0102] Loss function construction unit: constructs a composite loss function for training, combining the base learning task loss, meta-learning task loss, and regularization term to optimize the generalization ability of the model;

[0103] Meta-learning strategy unit: A hybrid strategy based on Dirichlet distribution is used to generate samples for training, enhancing the model's adaptability to new domain data;

[0104] Parameter update strategy unit: updates the model according to the model parameter differences and adopts a weight adjustment strategy based on shared knowledge to adapt to different tasks.

[0105] exist Figure 4 In the figure, arrows are used to represent data flow and information transmission. The meta-learning task designed by the meta-learning task design unit drives the loss function construction unit to build the loss function. The samples generated by the meta-learning strategy unit are input to the loss function construction unit to calculate the loss and guide the parameter update strategy unit to update the model parameters. The parameter update strategy of the parameter update strategy unit is fed back to the meta-learning task design unit and the meta-learning strategy unit to optimize the meta-learning task and sample generation strategy.

[0106] In one embodiment, the multi-source data enhancement module 330 is also used to optimize the combined mixed features and the features transferred by the soft label distillation method; measure the diversity of data from different source domains, and perform enhanced integrated learning on data from different source domains.

[0107] In one embodiment, the structure of the multi-source data enhancement module is as follows: Figure 5 As shown, it mainly includes:

[0108] Feature mixing technology unit: responsible for mixing feature representations from different source domains to enhance the model's generalization ability to new domains;

[0109] Soft label distillation unit: transfers knowledge from other source domains to generate soft labels for the target domain to compensate for the class imbalance problem;

[0110] Optimization strategy unit: jointly optimizes feature mixing and soft label distillation to enhance the generalization ability of the model;

[0111] Data diversity measurement unit: measures the diversity of data from different source domains and guides the selection and combination of data enhancement;

[0112] Enhanced data ensemble learning unit: Integrate enhanced data from different source domains to improve the generalization ability of the model.

[0113] exist Figure 5 In the figure, arrows represent data flow and information transfer. The feature mixing technology unit and the soft label distillation unit independently generate enhanced features and labels, which are then input into the optimization strategy unit for joint optimization. The data diversity measurement unit provides the optimization strategy unit with a measure of data diversity to guide data enhancement decisions. Finally, the enhanced data integration learning unit integrates enhanced data to provide the model with richer learning materials.

[0114] In one embodiment, the feature calibration module 340 is also used to determine a mapping function based on the initial model after data enhancement; map the source domain features to the space of target domain features according to the mapping function, and estimate and learn the feature mapping between the source domain and the unknown target domain through supervised learning.

[0115] In one embodiment, the feature calibration module 340 is further used to calculate the feature covariance of the source domain and the unknown target domain, and maximize the feature covariance until the features of the source domain and the unknown target domain are consistent.

[0116] In one embodiment, the feature calibration module 340 is also used to define a conversion loss function, and to determine whether the source domain and unknown target domain features are consistent through the conversion loss function to obtain a judgment result; if the judgment result is that the source domain and unknown target domain features are consistent, a conversion matrix and a bias vector are calculated based on the source domain features and the target domain features; the difference between the converted features and the unknown target domain features is controlled to be minimized through the conversion matrix and the bias vector.

[0117] In one embodiment, the feature calibration module 340 is also used to perform performance evaluation on the target model according to the position target domain, determine whether the target model can meet the generalization requirements, and obtain a performance evaluation result; based on the performance evaluation result, the target model is optimized or learning is terminated.

[0118] In one embodiment, the structure of the feature calibration module is as follows Figure 6 As shown, it mainly includes:

[0119] Feature space mapping learning unit: responsible for learning the mapping function from source domain features to target domain features, estimating through supervised learning to ensure the consistency of feature representation;

[0120] Feature alignment strategy unit: adopts feature alignment strategy to achieve consistency in feature space by maximizing the covariance of source domain and target domain features;

[0121] Feature transformation technology unit: applies feature transformation technology, such as linear transformation, to adjust the source domain features to better adapt them to the target domain;

[0122] Feature calibration loss function construction unit: Construct a feature calibration loss function that combines the goals of mapping learning, feature alignment, and conversion to ensure the generalization ability of feature representation.

[0123] exist Figure 6 In the figure, arrows are used to represent data flow and information transmission in each part. The feature space mapping learning of the feature space mapping learning unit drives the feature alignment strategy of the feature alignment strategy unit, and the feature conversion technology of the feature conversion technology unit is combined to jointly optimize the feature calibration loss function of the feature calibration loss function construction unit. The loss function result of the feature calibration loss function construction unit is fed back to the feature alignment strategy unit and the feature conversion technology unit to guide further feature space adjustment.

[0124] The open domain adaptation generalization method and system provided in the present application based on meta-learning and multi-source data enhancement are applied in the model. The model can perform state transitions at different learning stages, including the execution of base learning tasks, the adaptation of meta-learning tasks, and parameter updates. The states at different stages may include:

[0125] Initial state: model parameters θ are initialized, ready to start the learning process;

[0126] Base learning task execution status: The model performs base learning tasks on each source domain to obtain basic feature representations;

[0127] Meta-learning task adaptation state: the model adapts to new tasks generated by the meta-learning strategy and learns common features across domains;

[0128] Parameter optimization state: Based on the loss of base learning and meta-learning tasks, the model parameters θ are optimized;

[0129] Performance evaluation status: Evaluate the model performance on the unknown target domain to determine whether it meets the generalization requirements;

[0130] Decision state: Based on the performance evaluation results, the model decides to continue optimization or end learning;

[0131] End state: If the model performance meets the requirements, the learning process ends.

[0132] like Figure 7 As shown in the figure, the arrows represent the transitions and decision logic between states. Starting from the initial state, the model enters the base learning task execution state; after completing the base learning, the model adapts to the meta-learning task; then, the model updates the parameters in the parameter optimization state; after that, the model performs performance evaluation; based on the evaluation results, the model makes a decision to continue the optimization or end the learning process.

[0133] In one embodiment, an open domain adaptation generalization system based on meta-learning and multi-source data enhancement is provided, in which key modules may include a meta-learning framework module, a multi-source data enhancement module, and a feature calibration module, which can achieve efficient migration of source domain knowledge and rapid adaptation to unknown target domains. Specifically:

[0134] Meta-learning framework module: responsible for building and optimizing the meta-learning process of the model. By designing meta-learning tasks and loss functions, the model can be trained on multiple source domains and learn universal feature representations across domains.

[0135] Multi-source data enhancement module: Introduces innovative multi-source data enhancement technology, including hybrid technology at the feature level and soft label distillation at the label level, to enhance the model's adaptability to different distributions.

[0136] Feature calibration module: Through feature mapping learning and alignment strategies, the feature representation of the model in the unknown domain is optimized to ensure the classification accuracy of the model in the unknown domain.

[0137] like Figure 8 As shown in the figure, each module uses arrows to represent data flow and information transmission. The output of the meta-learning framework module points to the multi-source data enhancement module through an arrow, indicating the process of cross-domain feature learning; the output of the multi-source data enhancement module is passed to the feature calibration module through an arrow, indicating the process of enhancing generalization ability; the output of the feature calibration module points to the end through an arrow, indicating the ultimate goal of optimizing feature representation.

[0138] In one embodiment, Fig. 9As shown, an open domain adaptation generalization method and system based on meta-learning and multi-source data enhancement provided in this application is applied in the field of autonomous driving to improve the adaptability and generalization ability of the autonomous driving system under different environmental conditions. The specific implementation process is as follows:

[0139] Environmental data collection: Collect environmental data during vehicle driving through on-board sensors, including image, radar and lidar data;

[0140] Data preprocessing: preprocess the collected raw environmental data, including data cleaning, standardization, and denoising;

[0141] Meta-learning framework application: Use the meta-learning framework module to train the autonomous driving model based on multi-source environmental data and learn common feature representations across environments;

[0142] Multi-source data enhancement: Apply the multi-source data enhancement module to enhance the model's adaptability to different environments through feature mixing and soft label distillation technology;

[0143] Feature calibration: Use the feature calibration module to optimize the model's feature representation in specific environments (such as rainy days and nighttime) and improve classification accuracy;

[0144] Driving decision making: Based on the optimized feature representation, the autonomous driving system makes safe and effective driving decisions;

[0145] Performance evaluation and model iteration: Regularly evaluate the performance of the autonomous driving system and iteratively optimize the model based on the evaluation results.

[0146] In one embodiment, Fig.10 As shown, an open domain adaptation generalization method and system based on meta-learning and multi-source data enhancement provided in this application is applied to the field of intelligent manufacturing to improve the accuracy and efficiency of product quality control. The specific implementation process is as follows:

[0147] Manufacturing data collection: Collecting manufacturing data during the production process through sensors and monitoring systems in the factory;

[0148] Data preprocessing: preprocess the collected raw manufacturing data to ensure data quality;

[0149] Meta-learning framework application: Using the meta-learning framework module, the quality control model is trained based on multi-source manufacturing data to learn common feature representations across production links;

[0150] Multi-source data enhancement: Apply the multi-source data enhancement module to enhance the model's adaptability to different production conditions through feature mixing and soft label distillation technology;

[0151] Feature calibration: Use the feature calibration module to optimize the model's feature representation in specific production links (such as welding and assembly) to improve the accuracy of defect detection;

[0152] Quality control decision: Based on the optimized feature representation, the intelligent manufacturing system automatically makes product quality control decisions;

[0153] Performance evaluation and model iteration: Regularly evaluate the performance of the quality control system and iteratively optimize the model based on the evaluation results.

[0154] In one embodiment, Fig.11 As shown, an open domain adaptation generalization method and system based on meta-learning and multi-source data enhancement provided in this application is applied to the abnormal behavior detection system in the field of smart security to improve the ability to identify security events in the monitoring area. The specific implementation process is as follows:

[0155] Video data collection: Video data in the monitoring area is collected in real time through security surveillance cameras;

[0156] Data preprocessing: preprocess the collected video data, including frame extraction, background subtraction, etc.;

[0157] Meta-learning framework application: Using the meta-learning framework module, the model is trained to identify the common feature representations of normal and abnormal behaviors;

[0158] Multi-source data enhancement: Apply the multi-source data enhancement module to enhance the model's adaptability to different monitoring environments through feature mixing and soft label distillation technology;

[0159] Feature calibration: Use the feature calibration module to optimize the model’s feature representation in specific monitoring scenarios and improve the accuracy of abnormal behavior detection;

[0160] Abnormal behavior identification: Based on optimized feature representation, the model automatically identifies and marks abnormal behaviors;

[0161] Security response: The system triggers security responses based on the recognition results, such as sending alarms, dispatching police forces, etc.

[0162] Performance evaluation and model iteration: Regularly evaluate the performance of the abnormal behavior detection system and iteratively optimize the model.

[0163] In one embodiment, Fig.12 As shown, an open domain adaptation generalization method and system based on meta-learning and multi-source data enhancement provided in this application is applied to the field of medical image analysis to improve the accuracy and efficiency of disease diagnosis. The specific implementation process is as follows:

[0164] Medical imaging data collection: Collecting patients’ medical imaging data through medical imaging equipment, such as CT, MRI, etc.;

[0165] Data preprocessing: preprocess the collected medical imaging data, including image denoising, contrast enhancement, etc.;

[0166] Meta-learning framework application: Using the meta-learning framework module, the model is trained to identify the characteristic representations of different types of diseases;

[0167] Multi-source data enhancement: Apply the multi-source data enhancement module to enhance the model's generalization ability for different disease types through feature mixing and soft label distillation technology;

[0168] Feature calibration: Use the feature calibration module to optimize the model’s feature representation for specific disease diagnosis and improve the accuracy of diagnosis;

[0169] Disease diagnosis: Based on optimized feature representation, the model automatically identifies and diagnoses diseases;

[0170] Clinical decision support: The system provides doctors with diagnostic suggestions and treatment options to assist in clinical decision-making;

[0171] Performance evaluation and model iteration: Regularly evaluate the performance of the disease diagnosis system and iteratively optimize the model.

[0172] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig.13 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an open domain adaptation generalization method based on meta-learning and multi-source data enhancement is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0173] Those skilled in the art will understand that Fig.13 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0174] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, steps of an open domain adaptation generalization method based on meta-learning and multi-source data enhancement are implemented.

[0175] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of an open domain adaptation generalization method based on meta-learning and multi-source data enhancement are implemented.

[0176] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0177] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0178] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. An open domain adaptation generalization method based on meta-learning and multi-source data enhancement, characterized in that: The method comprises: Collecting raw data, and preprocessing the raw data to obtain processed data; Acquire an initial model, design a meta-learning task for the initial model based on the processed data, and construct a loss function corresponding to the initial model based on the meta-learning task, formulate a meta-learning strategy, and formulate a parameter update strategy to complete the meta-learning process construction of the initial model, and obtain an initial model based on meta-learning; Introducing multi-source data enhancement technology into the initial model based on meta-learning, simulating data distribution differences between different source domains through feature mixing technology at the feature level of the initial model based on meta-learning, and adopting soft label distillation method at the label level of the initial model based on meta-learning to transfer knowledge from other source domains to obtain the initial model after data enhancement; In the initial model after data enhancement, the feature representation on the unknown domain is optimized, the feature mapping between the source domain and the unknown target domain is learned, a feature alignment strategy is adopted to ensure the consistency of the source domain and the unknown target domain features, the source domain features are adjusted by learning linear transformation to achieve feature conversion, and the feature extraction mechanism is adjusted to generate a target model; Based on the target model, corresponding auxiliary decisions are made according to the original data.

2. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 1, characterized in that: The meta-learning task includes a base learning task and a meta-learning subtask; The base learning task is used to train the basic classification ability of the initial model on each source domain; The meta-learning subtask is used to optimize the adaptability of the initial model to the new domain.

3. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 2, characterized in that: Constructing a loss function corresponding to the initial model based on the meta-learning task, including: Determine a base learning task loss based on the base learning task, determine a meta-learning sub-task loss based on the meta-learning sub-task, and obtain a regularization term; A composite loss function corresponding to the initial model is constructed according to the base learning task loss, the meta-learning subtask loss, and the weight of the regularization term.

4. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 1, characterized in that: Developing a meta-learning strategy involves: A hybrid strategy based on Dirichlet distribution is used to generate new samples in the feature space; For the generated new sample, a weight vector is extracted from the Dirichlet distribution to calculate the weighted sum.

5. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 1, characterized in that: The method further comprises: Combine the mixed features and the features transferred by the soft label distillation method for optimization; The diversity of data from different source domains is measured, and enhanced ensemble learning is performed on data from different source domains.

6. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 1, characterized in that: In the data-augmented initial model, the feature representation on the unknown domain is optimized, and the feature mapping between the source domain and the unknown target domain is learned, including: Determine a mapping function based on the initial model after data enhancement; The source domain features are mapped to the target domain feature space according to the mapping function, and the feature mapping between the source domain and the unknown target domain is estimated and learned through supervised learning.

7. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 1, characterized in that: A feature alignment strategy is used to ensure the consistency of the source domain and unknown target domain features, including: The feature covariance of the source domain and the unknown target domain is calculated, and the feature covariance is maximized until the features of the source domain and the unknown target domain are consistent.

8. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 7, characterized in that: Feature conversion is achieved by adjusting the source domain features through learning linear transformation, including: A conversion loss function is defined, and whether the features of the source domain and the unknown target domain are consistent is determined by the conversion loss function to obtain a judgment result; If the judgment result is that the source domain and the unknown target domain features are consistent, a conversion matrix and a bias vector are calculated according to the source domain features and the target domain features; The transformation matrix and the bias vector are used to control the difference between the transformed features and the unknown target domain features to be minimal.

9. The open domain adaptation generalization method based on meta-learning and multi-source data enhancement according to claim 1, characterized in that: The method further comprises: According to the position target domain, the target model is evaluated for performance, and it is determined whether the target model can meet the generalization requirements, and a performance evaluation result is obtained; According to the performance evaluation result, the target model is optimized or the learning is terminated.

10. An open domain adaptation generalization system based on meta-learning and multi-source data enhancement, characterized in that: The system comprises: A data acquisition module is used to collect raw data and pre-process the raw data to obtain processed data; A meta-learning framework module, used to obtain an initial model, design a meta-learning task for the initial model based on the processed data, and construct a loss function corresponding to the initial model based on the meta-learning task, formulate a meta-learning strategy, formulate a parameter update strategy, complete the meta-learning process construction of the initial model, and obtain an initial model based on meta-learning; A multi-source data enhancement module, used to introduce multi-source data enhancement technology into the initial model based on meta-learning, simulate the data distribution difference between different source domains by feature mixing technology at the feature level of the initial model based on meta-learning, and adopt a soft label distillation method at the label level of the initial model based on meta-learning to transfer knowledge from other source domains to obtain the initial model after data enhancement; A feature calibration module is used to optimize the feature representation on the unknown domain in the initial model after data enhancement, learn the feature mapping between the source domain and the unknown target domain, adopt a feature alignment strategy to ensure the consistency of the source domain and the unknown target domain features, adjust the source domain features by learning linear transformation to achieve feature conversion, and adjust the feature extraction mechanism to generate a target model; A decision module is used to make corresponding auxiliary decisions according to the original data based on the target model.

Images