Source model adaptive fusion method and system for multi-source heterogeneous data

Abstract

The invention discloses a source model adaptive fusion method and system for multi-source heterogeneous data, and the method comprises the steps: carrying out the parameter optimization of each pre-training source model through employing the unlabeled data of a target domain, and obtaining a plurality of target source models matched with the target domain; splitting each target source model into a backbone network and a classifier, and averaging parameters of a plurality of backbone networks according to layers to generate a fused backbone network; calculating the weight of each classifier according to the fused backbone network and the label-free data; and obtaining a source fusion model of the target domain according to the fusion backbone network, each classifier and the corresponding weight. A plurality of independent backbone networks are integrated into a unified backbone, and meanwhile, on the basis of the fusion backbone and the target domain label-free data statistical classifier weight, an additional weight model does not need to be trained, so that the cross-domain multi-source heterogeneous data model adaptive scene of the source domain data cannot be directly obtained, the computing resources are reduced, and the data processing efficiency is improved. And the target domain model performance is maintained and even improved.

Description

Technical Field

[0001] This invention relates to the field of data migration technology, and in particular to an adaptive fusion method and system for source models of multi-source heterogeneous data. Background Technology

[0002] In multi-source heterogeneous data scenarios, there is a domain gap (data distribution difference) between different source domains and the target domain, and the cost of multi-source data annotation is high. Therefore, unsupervised domain adaptation technology has become the core means. Traditional unsupervised domain adaptation methods assume that source domain data can be directly accessed. By aligning the data distribution of source domain and target domain, source domain knowledge is transferred to the target domain task to improve the performance of the target domain model.

[0003] However, in practical applications, in privacy-sensitive fields such as healthcare and finance, due to data privacy protection policies (such as patient medical record privacy and user financial data confidentiality) or storage resource limitations, source domain data cannot be directly obtained. Only source models pre-trained based on source domain data can be used. When multiple source domain data cannot be directly obtained, how can we effectively integrate the knowledge of multiple source models to adapt to the target domain task and avoid the high memory and computational overhead caused by parallel computing of multiple models?

[0004] Existing methods primarily rely on model ensemble and model parameter merging. Model ensemble combines the predictions of multiple source models by running them simultaneously and dynamically adjusting their weights. However, this approach requires storing and computing multiple models, leading to a significant increase in resource consumption and making it difficult to deploy on resource-constrained devices such as mobile devices and embedded systems. Model parameter merging reduces the number of models by arithmetically averaging all the parameters of all models, but it typically requires additional training steps, increasing the computational burden, and its performance is often inferior to model ensemble. Therefore, in cross-domain, multi-source, heterogeneous data model adaptation scenarios where source domain data cannot be directly obtained, how can we reduce computational resources while maintaining or even improving the performance of the target domain model? Summary of the Invention

[0005] This invention provides a source model adaptive fusion method and system for multi-source heterogeneous data, which can reduce computing resources while maintaining or even improving the performance of the target domain model in cross-domain multi-source heterogeneous data model adaptation scenarios where source domain data cannot be directly obtained.

[0006] To achieve the above objectives, embodiments of the present invention provide an adaptive fusion method for source models of multi-source heterogeneous data, comprising: Obtain several pre-trained source models in the source domain and unlabeled data in the target domain. Use the unlabeled data to optimize the parameters of each pre-trained source model to obtain several target source models adapted to the target domain. Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are averaged layer by layer to generate a fused backbone network. The weights of each classifier are calculated based on the fused backbone network and the unlabeled data. Based on the fusion backbone network, each classifier, and its corresponding weights, the source fusion model for the target domain is obtained.

[0007] As an improvement to the above scheme, the step of splitting each target source model into a backbone network and a classifier, and generating a fused backbone network by averaging the parameters of several backbone networks layer by layer, includes: Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are aligned according to network layers. The parameters of the same network layer are arithmetically averaged, and the average parameters of all network layers are combined to generate a fused backbone network.

[0008] As an improvement to the above scheme, the step of calculating the weight of each classifier based on the fused backbone network and the unlabeled data includes: Based on the fused backbone network and the unlabeled data, the discriminability of each classifier is calculated to obtain the discriminability score of each classifier; Based on the fused backbone network and the unlabeled data, the transferability of each classifier is calculated to obtain the transferability score of each classifier; The discriminability and transferability scores are converted into weights for each classifier using a softmax-like function and a temperature coefficient.

[0009] As an improvement to the above scheme, the step of calculating the discriminability of each classifier based on the fused backbone network and the unlabeled data to obtain the discriminability score of each classifier includes: The fusion backbone network is used to extract features from the unlabeled data to obtain sample features of the target domain; Calculate the prediction entropy and feature neighborhood density of each classifier based on the sample features and the corresponding nearest neighbor features; The discriminability of each classifier is calculated based on the predicted entropy and feature neighborhood density to obtain the discriminability score of each classifier.

[0010] As an improvement to the above scheme, the step of calculating the transferability of each classifier based on the fused backbone network and the unlabeled data to obtain the transferability score of each classifier includes: A random perturbation is applied to the fused backbone network to generate a perturbed backbone network; The perturbated backbone network is used to extract features from the unlabeled data to obtain the perturbated sample features of the target domain; The channel variance of the feature output of each classifier is calculated based on the features of the perturbation samples. The transferability of each classifier is calculated based on the channel variance to obtain the transferability score of each classifier.

[0011] As an improvement to the above scheme, after obtaining the source fusion model of the target domain, the method further includes: The source fusion model is used to classify and predict the new unlabeled data in the target domain, and the final classification result of the new unlabeled data is obtained.

[0012] As an improvement to the above scheme, the step of using the source fusion model to classify and predict the new unlabeled data in the target domain to obtain the final classification result of the new unlabeled data includes: The fusion backbone network in the source fusion model is used to extract features from the new unlabeled data of the target domain to obtain the target features of the target domain. The target features are input into each classifier in the source fusion model to obtain the classification prediction probability of each classifier. The classification prediction probabilities of each classifier are weighted and summed to obtain the final prediction probability of the target feature; The category with the highest probability among the final predicted probabilities is selected as the final classification result of the new unlabeled data.

[0013] As an improvement to the above scheme, the step of obtaining several pre-trained source models in the source domain and unlabeled data in the target domain, and using the unlabeled data to optimize the parameters of each pre-trained source model to obtain several target source models adapted to the target domain, includes: Obtain pre-trained source models from several source domains and unlabeled data from the target domain; The unlabeled data is input into each pre-trained source model, and the model parameters of each pre-trained source model are adjusted by the SHOT++ algorithm to obtain several target source models adapted to the target domain.

[0014] To achieve the above objectives, embodiments of the present invention provide a source model adaptive fusion system for multi-source heterogeneous data, comprising: The model parameter optimization module is used to obtain pre-trained source models in several source domains and unlabeled data in the target domain, and to optimize the parameters of each pre-trained source model using the unlabeled data to obtain several target source models adapted to the target domain. The backbone network generation module is used to split each target source model into a backbone network and a classifier, and to generate a fused backbone network by averaging the parameters of several backbone networks layer by layer. The classification weight calculation module is used to calculate the weight of each classifier based on the fused backbone network and the unlabeled data. The fusion model generation module is used to obtain the source fusion model of the target domain based on the fusion backbone network, each classifier and its corresponding weight.

[0015] As an improvement to the above solution, the backbone network generation module is used for: Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are aligned according to network layers. The parameters of the same network layer are arithmetically averaged, and the average parameters of all network layers are combined to generate a fused backbone network.

[0016] Compared with existing technologies, the present invention discloses a source model adaptive fusion method and system for multi-source heterogeneous data. This method involves acquiring pre-trained source models from several source domains and unlabeled data from a target domain. The unlabeled data is used to optimize the parameters of each pre-trained source model, resulting in several target source models adapted to the target domain. Each target source model is then split into a backbone network and a classifier. The parameters of the backbone networks are averaged layer-by-layer to generate a fused backbone network. The weights of each classifier are calculated based on the fused backbone network and the unlabeled data. Finally, the source fusion model for the target domain is obtained based on the fused backbone network, each classifier, and its corresponding weights. By integrating multiple independent backbone networks into a unified backbone, cross-source domain common knowledge can be effectively integrated through average, avoiding interference from classifier heterogeneity on backbone features, and only one set of backbone parameters needs to be stored, significantly reducing memory consumption. At the same time, based on the fusion backbone and unsupervised statistical classifier weights of unlabeled data in the target domain, it does not need to rely on source domain data and train additional weight models, meeting data security requirements and reducing computational overhead. In cross-domain multi-source heterogeneous data model adaptation scenarios where source domain data cannot be directly obtained, it can reduce computational resources while maintaining or even improving the performance of the target domain model. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an adaptive fusion method for source models of multi-source heterogeneous data provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a source model adaptive fusion system for multi-source heterogeneous data provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that the terms "comprising" and "specific" in this invention, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0020] Please see Figure 1 , Figure 1 This is a flowchart illustrating an adaptive fusion method for source models of multi-source heterogeneous data provided in an embodiment of the present invention. The adaptive fusion method for source models of multi-source heterogeneous data includes: S1, obtain several pre-trained source models in the source domain and unlabeled data in the target domain, and use the unlabeled data to optimize the parameters of each pre-trained source model to obtain several target source models adapted to the target domain. S2, each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are averaged layer by layer to generate a fused backbone network. S3, calculate the weight of each classifier based on the fused backbone network and the unlabeled data; S4. Based on the fusion backbone network, each classifier, and the corresponding weights, the source fusion model of the target domain is obtained.

[0021] For example, the source model adaptive fusion method for multi-source heterogeneous data is implemented by a fusion server, which can interact with the target user. The fusion server acquires pre-trained source models from several source domains (e.g., art, clip art, and products) and unlabeled data (e.g., unlabeled images) from the target domain (real world). It uses a single-source unsupervised domain adaptation algorithm (SHOT++ algorithm) to optimize the parameters of each pre-trained source model using the unlabeled data, resulting in several target source models adapted to the target domain. Through unsupervised distribution alignment (single-source unsupervised domain adaptation algorithm), the domain gap between a single source model and the target domain is reduced, avoiding performance degradation due to excessive domain differences during subsequent fusion. Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are... The parameters are averaged layer by layer to generate a fusion backbone network. No additional training is required. It directly integrates general domain-invariant knowledge (such as basic features like image edges and textures) across source domains through parameter averaging, while reducing the number of backbone networks to one, significantly reducing memory usage. The weight of each classifier is calculated based on the fusion backbone network and the unlabeled data. This weight essentially reflects the adaptability of each classifier to the target domain, allowing classifiers with strong adaptability to contribute higher weights and suppressing classifiers with weak adaptability, thereby avoiding performance degradation and improving classification accuracy. Based on the fusion backbone network, each classifier, and its corresponding weight, the source fusion model for the target domain is obtained. This invention integrates multiple independent backbone networks into a unified backbone. Cross-source domain common knowledge can be effectively integrated through average, avoiding interference from classifier heterogeneity on backbone features. Moreover, only one set of backbone parameters needs to be stored, significantly reducing memory usage. At the same time, based on the fusion of the backbone and the unsupervised statistical classifier weights of unlabeled data in the target domain, it does not require relying on source domain data and training additional weight models, meeting data security requirements, reducing computational overhead, and can be deployed on resource-constrained devices such as mobile devices and embedded systems. It can adapt to cross-domain multi-source heterogeneous data model adaptation scenarios where source domain data cannot be directly obtained, thereby reducing computational resources while maintaining or even improving the performance of the target domain model.

[0022] Specifically, step S2 includes: S21, each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are aligned according to network layers. S22, perform an arithmetic average of the parameters of the same network layer, and combine the average parameters of all network layers to generate a fused backbone network.

[0023] For example, each target source model Split into backbone network and classifier ,right The backbone network parameters of each target source model are averaged layer by layer to generate a fused backbone network, retaining the classifiers from all source models. This forms a lightweight structure with a single backbone and multiple classifiers; the parameter expression for the fused backbone network is: , In the formula, To integrate the backbone network Layer parameters; For the first The backbone network of the target source model is the first... Layer parameters.

[0024] Specifically, step S3 includes: S31, Perform discriminability calculation on each classifier based on the fused backbone network and the unlabeled data to obtain the discriminability score of each classifier; S32, perform transferability calculation on each classifier based on the fused backbone network and the unlabeled data to obtain the transferability score of each classifier; S33, using a Softmax-like function and a temperature coefficient, transforms the discriminability score and transferability score into weights for each classifier.

[0025] For example, the formula for calculating the transferability score of each classifier is as follows: , The formula for calculating the transferability score of each classifier is as follows: , The formula for calculating the weight of each classifier is: , In the formula, For the first Discriminability score of the classifier for each target source model; For the first The classifier of the target source model for the first target source model Feature neighborhood density of each sample feature; For the first The classifier of the target source model for the first target source model Predictive entropy of each sample feature; For the first The transferability score of the classifier for each target source model; The total number of sample features; For the first The classifier of the target source model for the first target source model The channel variance of the feature output of each perturbation sample feature; For the first The weights of the classifiers for each target source model; and This is a temperature coefficient used to adjust the smoothness of the weight distribution.

[0026] Specifically, step S31 includes: S311, The fusion backbone network is used to extract features from the unlabeled data to obtain sample features of the target domain; S312, calculate the prediction entropy and feature neighborhood density of each classifier based on the sample features and the corresponding nearest neighbor features; S313, the discriminability of each classifier is calculated based on the predicted entropy and the feature neighborhood density to obtain the discriminability score of each classifier.

[0027] For example, the formula for calculating the prediction entropy of each classifier is as follows: , The formula for calculating the feature neighborhood density of each classifier is: , In the formula, For the first Individual sample features; For the first A classifier for a target source model; is the entropy function, used to quantify the uncertainty of the classifier's prediction results; The number of nearest neighbor features; For the first The first sample feature The nearest neighbor features can be obtained using the K-Nearest Neighbors (KNN) algorithm; is the cosine distance function, used to quantify the similarity between two feature vectors. Lower entropy results in higher prediction confidence; higher density indicates a tighter feature distribution and more stable classification; higher discriminability scores indicate stronger classifier reliability.

[0028] Specifically, step S32 includes: S321, Apply random perturbation to the fused backbone network to generate the perturbed backbone network; S322, The perturbed backbone network is used to extract features from the unlabeled data to obtain the perturbed sample features of the target domain; S323, Calculate the channel variance of the feature output of each classifier based on the features of the perturbed samples; S324, perform transferability calculation on each classifier based on the channel variance to obtain the transferability score of each classifier.

[0029] For example, applying random perturbations to the converged backbone network Calculate the channel variance of the feature output. : , In the formula, The number of channels for the output features of the perturbed backbone network; is the variance function, used to quantify the degree of fluctuation in the output characteristics of the backbone network after perturbation; For the first The perturbation backbone set of the backbone network of a target source model; For the first The backbone network of a target source model; For the first The backbone network of the target source model for the first target source model Unlabeled data The extracted features The characteristic values ​​of each channel; For random disturbance coefficients.

[0030] Furthermore, after obtaining the source fusion model of the target domain, the method further includes: S5, the source fusion model is used to classify and predict the new unlabeled data in the target domain to obtain the final classification result of the new unlabeled data.

[0031] Specifically, step S5 includes: S51, the fusion backbone network in the source fusion model is used to extract features from the new unlabeled data of the target domain to obtain the target features of the target domain; S52, Input the target features into each classifier in the source fusion model to obtain the classification prediction probability of each classifier; S53, the classification prediction probabilities of each classifier are weighted and summed to obtain the final prediction probability of the target feature; S54, Select the category with the highest probability among the final predicted probabilities as the final classification result of the new unlabeled data.

[0032] For example, the expression for the classification prediction probability is: , In the formula, Represents new unlabeled data in a given target domain. In this case, the new unlabeled data belongs to the category The probability of; To integrate the backbone network.

[0033] Specifically, step S1 includes: S11, obtain pre-trained source models from several source domains and unlabeled data from the target domain; S12, input the unlabeled data into each pre-trained source model, and adjust the model parameters of each pre-trained source model through the SHOT++ algorithm to obtain several target source models adapted to the target domain.

[0034] For example, adjusting the model parameters of each pre-trained source model is mainly used to optimize the fit between the feature extractor and the classifier; by aligning unsupervised distributions, the domain gap between a single source model and the target domain is reduced, avoiding performance degradation due to excessive domain differences during subsequent fusion.

[0035] In one specific embodiment, the source domains include art, clip art, and products, while the target domain is the real world. Under unsupervised conditions, knowledge from the three source domains is transferred to the target domain to improve classification accuracy. The feature extractor (Tiny Swin Transformer) and classifier (fully connected layer) for each source domain are independently pre-trained to obtain a pre-trained source model for each source domain. The parameters of each pre-trained source model are optimized using unlabeled data from the target domain, resulting in several target source models adapted to the target domain. The parameters of each feature extractor (backbone network) are averaged and merged; and each classifier is weighted and integrated. For example, inputting source domain data (annotated images of art, clip art, and products), Tiny Swin Transformer models (feature extractor + classifier) ​​are pre-trained on the three source domains respectively; SHOT++ is used to individually adapt each source model to the target domain, resulting in three target source models adapted to the target domain; the backbone network parameters of the three models are averaged layer by layer to output a single fused feature extractor. For the target domain data (unlabeled images of the real-world domain), the prediction entropy and neighborhood density of each classifier are calculated to obtain a discriminability score. Simultaneously, Gaussian noise perturbation is applied to the backbone network, and the feature output variance is calculated to obtain a transferability score. A softmax-like function and a temperature coefficient are used to convert the discriminability and transferability scores into weights for each classifier. The target domain image is input into a single fusion feature extractor to extract features. The extracted features are then input into three classifiers to obtain multiple prediction probabilities. These multiple prediction probabilities are weighted and fused to obtain the final prediction probability. The class with the highest probability among the final prediction probabilities is selected as the final classification result for the target domain image.

[0036] This invention discloses an adaptive fusion method for source models of multi-source heterogeneous data. The method involves acquiring pre-trained source models from several source domains and unlabeled data from a target domain. The parameters of each pre-trained source model are optimized using the unlabeled data to obtain several target source models adapted to the target domain. Each target source model is then split into a backbone network and a classifier. The parameters of the backbone networks are averaged layer-by-layer to generate a fused backbone network. The weights of each classifier are calculated based on the fused backbone network and the unlabeled data. Finally, the source fusion model for the target domain is obtained based on the fused backbone network, each classifier, and its corresponding weights. By integrating multiple independent backbone networks into a unified backbone, cross-source domain common knowledge can be effectively integrated through average, avoiding interference from classifier heterogeneity on backbone features, and only one set of backbone parameters needs to be stored, significantly reducing memory consumption. At the same time, based on the fusion backbone and unsupervised statistical classifier weights of unlabeled data in the target domain, it does not need to rely on source domain data and train additional weight models, meeting data security requirements and reducing computational overhead. In cross-domain multi-source heterogeneous data model adaptation scenarios where source domain data cannot be directly obtained, it can reduce computational resources while maintaining or even improving the performance of the target domain model.

[0037] See Figure 2 , Figure 2 This is a schematic diagram of the structure of a source model adaptive fusion system 10 for multi-source heterogeneous data provided in an embodiment of the present invention. The source model adaptive fusion system 10 for multi-source heterogeneous data includes: The model parameter optimization module 11 is used to obtain several pre-trained source models in the source domain and unlabeled data in the target domain, and to optimize the parameters of each pre-trained source model using the unlabeled data to obtain several target source models adapted to the target domain. The backbone network generation module 12 is used to split each target source model into a backbone network and a classifier, and to generate a fused backbone network by averaging the parameters of several backbone networks layer by layer. The classification weight calculation module 13 is used to calculate the weight of each classifier based on the fused backbone network and the unlabeled data. The fusion model generation module 14 is used to obtain the source fusion model of the target domain based on the fusion backbone network, each classifier and its corresponding weight.

[0038] Specifically, the backbone network generation module 12 is used for: Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are aligned according to network layers. The parameters of the same network layer are arithmetically averaged, and the average parameters of all network layers are combined to generate a fused backbone network.

[0039] Furthermore, the source model adaptive fusion system 10 for multi-source heterogeneous data also includes: The classification result prediction module is used to perform classification prediction on the new unlabeled data of the target domain using the source fusion model, so as to obtain the final classification result of the new unlabeled data.

[0040] The source model adaptive fusion system 10 for multi-source heterogeneous data provided in this embodiment of the invention can realize all the processes of the source model adaptive fusion method for multi-source heterogeneous data in the above embodiment. The functions and technical effects of each module in the system are the same as those of the source model adaptive fusion method for multi-source heterogeneous data in the above embodiment, and will not be repeated here.

[0041] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A source model adaptive fusion method for multi-source heterogeneous data, characterized in that, include: Obtain several pre-trained source models in the source domain and unlabeled data in the target domain. Use the unlabeled data to optimize the parameters of each pre-trained source model to obtain several target source models adapted to the target domain. Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are averaged layer by layer to generate a fused backbone network. The weights of each classifier are calculated based on the fused backbone network and the unlabeled data. Based on the fusion backbone network, each classifier, and its corresponding weights, the source fusion model for the target domain is obtained.

2. The source model adaptive fusion method for multi-source heterogeneous data as described in claim 1, characterized in that, The step of splitting each target source model into a backbone network and a classifier, and generating a fused backbone network by averaging the parameters of several backbone networks layer by layer, includes: Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are aligned according to network layers. The parameters of the same network layer are arithmetically averaged, and the average parameters of all network layers are combined to generate a fused backbone network.

3. The source model adaptive fusion method for multi-source heterogeneous data as described in claim 1, characterized in that, The step of calculating the weight of each classifier based on the fused backbone network and the unlabeled data includes: Based on the fused backbone network and the unlabeled data, the discriminability of each classifier is calculated to obtain the discriminability score of each classifier; Based on the fused backbone network and the unlabeled data, the transferability of each classifier is calculated to obtain the transferability score of each classifier; The discriminability and transferability scores are converted into weights for each classifier using a softmax-like function and a temperature coefficient.

4. The source model adaptive fusion method for multi-source heterogeneous data as described in claim 3, characterized in that, The discriminability calculation for each classifier based on the fused backbone network and the unlabeled data, to obtain a discriminability score for each classifier, includes: The fusion backbone network is used to extract features from the unlabeled data to obtain sample features of the target domain; Calculate the prediction entropy and feature neighborhood density of each classifier based on the sample features and the corresponding nearest neighbor features; The discriminability of each classifier is calculated based on the predicted entropy and feature neighborhood density to obtain the discriminability score of each classifier.

5. The source model adaptive fusion method for multi-source heterogeneous data as described in claim 3, characterized in that, The transferability calculation for each classifier based on the fused backbone network and the unlabeled data, to obtain a transferability score for each classifier, includes: A random perturbation is applied to the fused backbone network to generate a perturbed backbone network; The perturbated backbone network is used to extract features from the unlabeled data to obtain the perturbated sample features of the target domain; The channel variance of the feature output of each classifier is calculated based on the features of the perturbation samples. The transferability of each classifier is calculated based on the channel variance to obtain the transferability score of each classifier.

6. The source model adaptive fusion method for multi-source heterogeneous data as described in claim 1, characterized in that, After obtaining the source fusion model of the target domain, the method further includes: The source fusion model is used to classify and predict the new unlabeled data in the target domain, and the final classification result of the new unlabeled data is obtained.

7. The source model adaptive fusion method for multi-source heterogeneous data as described in claim 6, characterized in that, The step of using the source fusion model to classify and predict the new unlabeled data in the target domain, and obtaining the final classification result of the new unlabeled data, includes: The fusion backbone network in the source fusion model is used to extract features from the new unlabeled data of the target domain to obtain the target features of the target domain. The target features are input into each classifier in the source fusion model to obtain the classification prediction probability of each classifier. The classification prediction probabilities of each classifier are weighted and summed to obtain the final prediction probability of the target feature; The category with the highest probability among the final predicted probabilities is selected as the final classification result of the new unlabeled data.

8. The source model adaptive fusion method for multi-source heterogeneous data as described in claim 1, characterized in that, The process involves acquiring pre-trained source models for several source domains and unlabeled data for the target domain. The unlabeled data is then used to optimize the parameters of each pre-trained source model, resulting in several target source models adapted to the target domain, including: Obtain pre-trained source models from several source domains and unlabeled data from the target domain; The unlabeled data is input into each pre-trained source model, and the model parameters of each pre-trained source model are adjusted by the SHOT++ algorithm to obtain several target source models adapted to the target domain.

9. A source model adaptive fusion system for multi-source heterogeneous data, characterized in that, include: The model parameter optimization module is used to obtain pre-trained source models in several source domains and unlabeled data in the target domain, and to optimize the parameters of each pre-trained source model using the unlabeled data to obtain several target source models adapted to the target domain. The backbone network generation module is used to split each target source model into a backbone network and a classifier, and to generate a fused backbone network by averaging the parameters of several backbone networks layer by layer. The classification weight calculation module is used to calculate the weight of each classifier based on the fused backbone network and the unlabeled data. The fusion model generation module is used to obtain the source fusion model of the target domain based on the fusion backbone network, each classifier and its corresponding weight.

10. The source model adaptive fusion system for multi-source heterogeneous data as described in claim 9, characterized in that, The backbone network generation module is used for: Each target source model is split into a backbone network and a classifier, and the parameters of several backbone networks are aligned according to network layers. The parameters of the same network layer are arithmetically averaged, and the average parameters of all network layers are combined to generate a fused backbone network.

Images