Distributed model training method based on edge network equipment

Abstract

The invention discloses a distributed model training method based on edge network equipment, and relates to the technical field of edge computing distributed model training, and the method comprises the steps: carrying out the logic hierarchical aggregation operation of an intermediate parameter set, forming a multi-layer aggregation structure, and generating a global topology parameter set through topology perception mapping; performing unified fusion calculation on the global topological parameter set to obtain a fusion result set, and performing sample vector extraction and collaborative feature aggregation to obtain a global parameter set; and based on the global parameter set and the structured data set, performing parameter adaptation optimization on the distributed model to obtain a trained distributed model. According to the method, local training data in an edge network is cleaned, formatted and subjected to feature conversion, nonlinear mapping and fractional aggregation of local training results are combined, a multi-level aggregation structure is formed, a global topology parameter set is generated, and global sensing and coding of the network topology structure and the hierarchical relation between nodes are achieved.

Description

Technical Field

[0001] This invention relates to the field of distributed model training technology in edge computing, and in particular to a distributed model training method based on edge network devices. Background Technology

[0002] With the rapid development of artificial intelligence technology, especially the application of deep learning and large-scale data processing technologies, traditional centralized model training methods face performance bottlenecks in scenarios with large data volumes and widely distributed nodes. In recent years, the widespread deployment of edge computing and edge network devices has provided a new technological foundation for distributed model training. Existing distributed training methods typically rely on cloud-based central nodes to aggregate and uniformly compute distributed data, achieving global model updates through parameter synchronization or gradient aggregation. When processing locally generated data, these methods can complete unified training of model parameters through network communication, improving training speed and resource utilization efficiency.

[0003] While existing distributed model training methods can alleviate cloud computing pressure to some extent, there is still room for optimization in edge environments. Current methods rely heavily on a unified central node for parameter aggregation, particularly for cleaning local data, feature transformation, and structured aggregation of local training results. This can lead to high network bandwidth consumption, insufficient topology awareness, and inadequate exploitation of local feature synergies. In practical edge network deployments, the heterogeneity of data and the complexity of topological distribution among nodes limit the efficiency of global parameter fusion and optimization. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a distributed model training method based on edge network devices to solve the problems of high data heterogeneity at edge nodes and low efficiency of global parameter fusion.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a distributed model training method based on edge network devices, which includes collecting local training data in a distributed edge network and cleaning, formatting and feature transformation to generate a structured dataset;

[0008] Perform local training operations on the structured dataset to obtain local update results;

[0009] The local update results are extracted and formatted to generate an intermediate parameter set;

[0010] The intermediate parameter set is logically hierarchically aggregated to form a multi-level aggregation structure, and a global topology parameter set is generated through topology-aware mapping.

[0011] A unified fusion calculation is performed on the global topology parameter set to obtain a fusion result set, and sample vector extraction and collaborative feature aggregation are performed to obtain the global parameter set;

[0012] Based on the global parameter set and structured dataset, the distributed model is optimized for parameter adaptation to obtain the trained distributed model.

[0013] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for generating the structured dataset are as follows:

[0014] Collect local training data from distributed edge networks, perform time-series synchronization processing to generate the original training data set, and perform outlier detection and cleaning to obtain a standardized dataset;

[0015] The standardized dataset is mapped to a unified feature space, and a formatted dataset is generated through nonlinear transformation and standardization.

[0016] The system performs feature interaction and mapping on formatted datasets to generate structured datasets.

[0017] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for obtaining the local update result are as follows:

[0018] The structured dataset is computed through nonlinear mapping and feature interaction to obtain local model parameter vectors, and local optimization values ​​of the samples are formed by aggregated gradient updates and nonlinear fractional calculations.

[0019] The local optimization values ​​of the samples are aggregated and trained to generate local update results.

[0020] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for generating the intermediate parameter set are as follows:

[0021] Extract sample feature values ​​from the local update results, perform nonlinear combination mapping to obtain a set of mapped features, and perform fractional aggregation operation to generate an intermediate parameter vector;

[0022] The intermediate parameter vector is structured and formatted to form an intermediate parameter set.

[0023] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for forming a multi-level aggregation structure are as follows:

[0024] The intermediate parameter set is processed through hierarchical mapping and inter-layer interaction to obtain the hierarchical mapping feature set;

[0025] The hierarchical mapping feature set is normalized and grouped within the layer to obtain the hierarchical aggregated value set, and then cross-layer combination and nonlinear mapping are performed to generate a multi-level aggregated structure vector.

[0026] The multi-level aggregated structure vector is obtained by summarizing and integrating the full sample.

[0027] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for generating the global topology parameter set are as follows:

[0028] The multi-level aggregation structure is combined with the hierarchical mapping feature set, and topology-aware mapping is performed to generate topology-aware vectors.

[0029] The topology-aware vectors are fused with inter-layer features through nonlinear mapping to obtain a set of sample topology fusion values.

[0030] The sample topology fusion value set is aggregated and standardized to generate a global topology parameter set.

[0031] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for generating the global topology parameter set are as follows:

[0032] The operational availability, bandwidth normalization score, and synchronization success rate are extracted from the sample topology fusion value set and then calculated using a geometric mean to obtain the sample reliability score.

[0033] The topological importance coefficient is obtained by combining the sample reliability score with the sample topology fusion value set for correlation mapping and weight normalization.

[0034] Multidimensional parameter mapping and tensor aggregation operations are performed on the topological importance coefficients to generate dimensional aggregation values. Multidimensional orthogonal projection and normalization synthesis operations are then performed to obtain the global topological parameter set.

[0035] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for obtaining the fusion result set are as follows:

[0036] The global topology parameter vector is extracted from the global topology parameter set, and then nonlinear mapping and weight interaction fusion are performed to obtain the comprehensive topology feature vector.

[0037] The comprehensive topological feature vector is subjected to fractional fusion calculation to generate a fusion result vector, which is then summarized and aggregated to form a fusion result set.

[0038] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for obtaining the global parameter set are as follows:

[0039] Extract fused sample vectors from the fusion result set, and perform cross-sample similarity mapping and asymmetric interactive combination operations to generate a collaborative feature set;

[0040] Perform fractional nonlinear aggregation operations on the collaborative feature set to generate a global parameter vector, and then perform set-based aggregation processing to generate a global parameter set.

[0041] As a preferred embodiment of the distributed model training method based on edge network devices described in this invention, the specific steps for obtaining the trained distributed model are as follows:

[0042] The global parameter set and the structured dataset are fused and integrated according to the sample correspondence to generate an updated distributed model parameter vector.

[0043] The updated distributed model parameter vector is combined with the structured dataset and the global parameter set, and then input into the distributed model for forward inference and error calculation. The internal parameters of the distributed model are iteratively optimized through backpropagation to obtain the trained distributed model.

[0044] The beneficial effects of this invention are as follows: by cleaning, formatting, and transforming the local training data in the edge network, and combining the nonlinear mapping and fractional aggregation of the local training results, a multi-level aggregation structure is formed and a global topology parameter set is generated, realizing the global perception and encoding of the network topology and the hierarchical relationship between nodes; by extracting sample vectors and generating a global parameter set through cross-sample collaborative feature aggregation, the deep fusion of information between different nodes and samples is realized, thereby enhancing the training accuracy, robustness, and scalability of the distributed model in the heterogeneous edge network environment, while optimizing the model parameters as a whole, so that the final distributed model has high performance and adaptability. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a distributed model training method based on edge network devices.

[0047] Figure 2 Generate flowcharts for structured datasets.

[0048] Figure 3 Generate a flowchart for the partial update results.

[0049] Figure 4 Generate a flowchart for the set of intermediate parameters. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0053] Reference Figures 1-4 This is one embodiment of the present invention, which provides a distributed model training method based on edge network devices, including the following steps:

[0054] S1. Collect local training data from the distributed edge network and perform cleaning, formatting, and feature transformation to generate a structured dataset.

[0055] S1.1 Collect local training data in the distributed edge network, perform time series synchronization processing, generate the original training data set, and perform outlier detection and cleaning to obtain a standardized dataset.

[0056] Furthermore, local training data is obtained from distributed edge network nodes, and data streams from different sources are synchronized and aligned based on a unified time benchmark to ensure temporal consistency between samples. Outlier detection is performed on the synchronized data, and invalid or erroneous data is removed through statistical discrimination and pattern bias identification. The local training data is also cleaned to remove noise and redundant content, so that the data remains complete and usable, resulting in a standardized dataset.

[0057] S1.2 Map the standardized dataset to a unified feature space, and generate a formatted dataset through nonlinear transformation and standardization.

[0058] Furthermore, a one-to-one mapping relationship is established with the unified feature space through feature selection, category encoding, and numerical scale alignment operations. During this mapping process, one-hot encoding or lookup table embedding is performed on discrete attributes in the standardized dataset, while continuous attributes in the standardized dataset are normalized or scaled according to the dimensional requirements of the unified feature space. This ensures that each sample in the standardized dataset has consistent feature dimensions and semantic representations within the unified feature space. After mapping, nonlinear transformations are applied to the representations in the unified feature space to enhance feature expressive power. These nonlinear transformations can employ polynomial feature expansion, kernel function mapping, or piecewise mapping to generate new higher-order features or nonlinear combination features. The output of the nonlinear transformation is then standardized to eliminate dimensional differences and stabilize the numerical distribution. Standardization can employ known methods such as zero-mean unit variance standardization or min-max scaling. After standardization, the results are validated for consistency and missing values ​​are filled in to form a formatted dataset.

[0059] It should also be noted that discrete attributes refer to categorical features with finite and enumerable values ​​in a standardized dataset. Each sample can only belong to one or several specific categories in terms of discrete attributes, and there is no continuous quantitative relationship between these values.

[0060] S1.3 Perform feature interaction and mapping processing on the formatted data set to generate a structured data set.

[0061] Furthermore, pairwise interactive computations are performed on various features in the formatted dataset. These computations can include feature multiplication, feature ratios, and cross-combinations, forming interactive feature vectors that reflect the relationships between features. Nonlinear mapping operations are then applied to the interactive feature vectors. These nonlinear mappings can employ kernel function mapping, polynomial expansion, or piecewise function mapping to project the interactive features into a higher-dimensional feature space to enhance their expressive power. After the nonlinear mapping is complete, the mapping results are numerically stabilized and dimensionally aligned to ensure consistency across all samples in terms of feature dimensions. Finally, the results of feature interaction and nonlinear mapping are integrated in a unified format to output a structured dataset.

[0062] S2. Perform local training operations on the structured dataset to obtain local update results.

[0063] S2.1. The structured dataset is processed through nonlinear mapping and feature interaction to obtain local model parameter vectors, and local optimization values ​​of the samples are formed through aggregated gradient updates and nonlinear fraction calculations.

[0064] Furthermore, the structured dataset is input into a nonlinear mapping process. This process can utilize methods such as polynomial expansion, kernel function transformation, or piecewise mapping to enhance the nonlinear representation of the structured dataset in the feature space. Feature interaction computation is then performed on the nonlinearly mapped features, including operations such as product interaction and proportional mapping across different feature dimensions, generating local feature combinations that reflect the complex relationships between features. After obtaining these local feature combinations, parameterization operations are performed, and local weight factors are extracted to form a local model parameter vector. This local model parameter vector is then applied to the gradient update process, which incorporates nonlinear fractional computation. After both gradient update and nonlinear fractional computation are completed, the update results for each sample are processed, and the local optimization values ​​for each sample are output, expressed as:

[0065]

[0066] in, Indicates the first Local optimization values ​​for each sample Indicates the sample index. Indicates the first The local feature combination vector of each sample Indicates the first The local weight factor vector of each sample, Indicates the first Loss function for each sample Regarding distributed model parameter vectors gradient vector, Represents the parameter vector within the distributed model. Indicates the first The nonlinear fractional calculation results for each sample Represents the regularization coefficient;

[0067] S2.2 Perform aggregation training operations on the local optimization values ​​of the samples to generate local update results.

[0068] Furthermore, the local optimization values ​​of the samples are organized and grouped according to the sample index order to ensure the consistency of the same batch of samples during the calculation process. Aggregation operations such as summation, mean calculation, or cumulative calculation are performed on the organized local optimization values ​​to integrate the optimization information from different samples during training. During the aggregation process, normalization or scaling methods are used to control the stability of the numerical distribution to avoid excessive deviations affecting the overall results. After aggregation, multiple rounds of iterative training calculations are performed on the aggregation results to gradually improve the convergence and stability of the local parameter representation through iterative updates. After all iterative training calculations are completed, the final results are uniformly organized into local parameter update outputs, and the local update results are output.

[0069] It should also be noted that the “sample index order” is determined by the unique identifier (e.g., sample ID or record number) of each sample in the structured dataset. By arranging all samples in ascending order of the unique identifier or the original record order, all samples are numbered sequentially, thereby ensuring that the position and order of each sample are consistent during the sorting, grouping and aggregation process, so as to maintain the consistency and traceability of the same batch of samples in the calculation.

[0070] S3. Extract and format the local update results to generate an intermediate parameter set.

[0071] S3.1 Extract sample feature values ​​from the local update results, perform nonlinear combination mapping to obtain the mapping feature set, and perform fractional aggregation operation to generate intermediate parameter vector.

[0072] Furthermore, sample feature values ​​are extracted one by one from the local update results, and organized according to the sample index to ensure consistency of the extraction order. A nonlinear combination mapping operation is performed on the extracted sample feature values. This nonlinear combination mapping can be achieved through methods such as polynomial expansion, kernel function mapping, or piecewise function transformation to generate a mapping feature set that reflects the complex relationships between features. After the mapping feature set is formed, fractional aggregation operations are performed on it. Fractional aggregation operations fuse multiple mapping features in fractional form by performing cumulative summation, proportional calculation, or normalization operations on the numerator and denominator respectively. After the fractional aggregation operation is completed, the output results are uniformly organized and numerically stabilized, and an intermediate parameter vector is output.

[0073] S3.2. The intermediate parameter vector is processed through structured organization and unified formatting to form an intermediate parameter set.

[0074] Furthermore, the intermediate parameter vectors are structured according to sample index order and feature dimensions. Intermediate parameter vectors from different sources are arranged in a fixed order (e.g., according to the order of source nodes or data collection time) and classified into corresponding feature categories (e.g., environmental features, state features, operational features) to ensure the structural integrity of the overall data. The structured intermediate parameter vectors are then uniformly formatted, including steps such as dimension alignment, numerical standardization, and missing value imputation, to ensure consistency in numerical distribution, dimensional structure, and storage format among all intermediate parameter vectors. After the uniform formatting is completed, a consistency check is performed on the results, and the intermediate parameter set is output.

[0075] S4. Perform logical hierarchical aggregation on the intermediate parameter set to form a multi-level aggregation structure, and generate a global topology parameter set through topology-aware mapping.

[0076] S4.1. The intermediate parameter set is processed through hierarchical mapping and inter-layer interaction to obtain the hierarchical mapping feature set.

[0077] Furthermore, the intermediate parameter set is hierarchically divided according to hierarchical criteria (such as node function category, topology depth, and feature type), forming multiple parameter subsets arranged in hierarchical order. A hierarchical mapping operation is performed on each parameter subset, which can be achieved through linear projection transformation or principal component analysis, converting the parameter subset into a reconstructed feature representation within the layer. In the inter-layer interactive processing stage, the mapping representations between different layers are interactively computed. Interactive computation includes element-wise multiplication, vector inner product, and vector concatenation, used to extract correlation features between layers. The interactive computation results are then normalized or scaled to maintain a stable numerical distribution. After completing the interactive computation, the normalized results undergo saliency extraction and redundancy removal, which can be achieved using singular value decomposition or principal component analysis to retain key components. Finally, the results of hierarchical mapping and inter-layer interactive processing are organized according to a unified format and dimensional requirements, outputting a hierarchical mapping feature set.

[0078] S4.2. Perform intra-layer normalization and grouping aggregation on the hierarchical mapping feature set to obtain the hierarchical aggregated value set, and perform cross-layer combination and nonlinear mapping to generate a multi-level aggregated structure vector.

[0079] Furthermore, intra-layer normalization is performed on the mapping representation of each layer in the hierarchical mapping feature set. Normalization can employ zero-mean unit variance standardization or min-max scaling to eliminate dimensional differences between dimensions. After normalization, the normalized mapping representation of each layer is grouped and aggregated according to feature category or index interval. Grouping and aggregation operations include performing statistical aggregations such as summation, mean, and median on vectors within the group, as well as norm calculations or cumulative operations on vectors within the group to summarize intra-group information. After aggregation, the grouping and aggregation results of each layer are summarized in hierarchical order to form a hierarchical aggregated value set. Based on the hierarchical aggregated value set, cross-layer combinations are performed between different layers. Cross-layer combination operations include vector concatenation, element-wise multiplication, and outer product construction to express inter-layer interactions. After cross-layer combination, nonlinear mapping is applied to the intra-group vectors to enhance representational power. Nonlinear mapping can employ methods such as polynomial expansion, kernel function mapping, or piecewise function transformation. Dimension alignment and numerical stabilization are performed on the nonlinear mapping output. The processing results are organized into a multi-level aggregated structure vector according to a unified field order and dimensional requirements, outputting a multi-level aggregated structure vector.

[0080] It should also be noted that the unified field order is to arrange the vectors of each layer in a fixed feature order in a multi-level aggregated structure vector. For example, temperature and humidity in the environmental features are arranged first, followed by voltage and power in the state features. This ensures that the position of each feature in the vector is consistent, thereby avoiding dimensional misalignment or confusion.

[0081] Dimensional organization requires that vectors be arranged and aligned according to the hierarchical dimensional structure when combining or splicing vectors across layers. This ensures that the dimensional length, order, and combination method of each layer meet the requirements of subsequent processing, and that all samples and layers maintain dimensional consistency.

[0082] S4.3. The multi-level aggregated structure vector is obtained by summarizing and integrating the full sample.

[0083] Furthermore, element-wise statistical summarization operations are performed on the multi-level aggregated structure vector along the sample dimensions, including commonly used statistics such as element-wise summation, element-wise mean, and element-wise median, to obtain a cross-sample central tendency representation. Robustness processing is applied to the statistical summarization results, using existing robust statistical methods such as truncated mean or absolute deviation of the median to suppress the influence of outliers and stabilize the numerical distribution. The robustly processed summary representation is then scaled and normalized, using known methods such as zero-mean unit variance standardization or min-max scaling to ensure numerical comparability across different dimensions. Dimension compression or principal component extraction is performed on the normalized representation, using methods such as singular value decomposition or principal component analysis to extract key joint components and reduce redundancy. Finally, the dimension-compressed or principal component extracted representation is integrated with the hierarchical identifiers in the multi-level aggregated structure vector, and the output is organized into a multi-level aggregated structure according to a unified field order and dimensionality requirements.

[0084] S4.4 Combine the multi-level aggregation structure with the hierarchical mapping feature set, and perform topology-aware mapping to generate topology-aware vectors.

[0085] Furthermore, field alignment and dimension matching are performed on the multi-level aggregation structure and the hierarchical mapping feature set in both sample and feature dimensions. The hierarchical representation in the multi-level aggregation structure is combined with the corresponding mapping features in the hierarchical mapping feature set according to the sample index and hierarchical index. Topology-aware mapping is performed on the combined representation. Topology-aware mapping projects the combined features into the topological feature space by calculating the similarity, adjacency, and dependency between features of each layer, so as to reflect the topological relationship of samples in the multi-level aggregation structure and the hierarchical mapping feature set. During the topology-aware mapping process, normalization and numerical stabilization processing are performed on the mapping output to ensure the comparability and stability of topological features in different dimensions. The normalized topological feature vectors are uniformly formatted and organized to output the topology-aware vector.

[0086] It should also be noted that similarity refers to the degree of closeness in numerical value or distribution between different levels or different features. It is measured by methods such as calculating Euclidean distance, cosine similarity, or correlation coefficient, and reflects the degree of similarity of the features of samples in multi-level aggregation structures and hierarchical mapping feature sets.

[0087] Adjacency relationships refer to the connections between features or levels in a topological structure, that is, which features or levels are directly related or adjacent to each other. They are represented by adjacency matrices or graph structures and are used to characterize the spatial or logical connections between features within a sample and between layers.

[0088] Dependency refers to the degree to which a change in one feature or level affects another feature or level. It is represented by statistical analysis, correlation analysis, or conditional probability modeling, and reflects the mutual influence and constraint relationships between features in a multi-level aggregate structure and hierarchical mapping feature set.

[0089] S4.5. The topology-aware vector is fused with inter-layer features through nonlinear mapping to obtain a set of sample topology fusion values.

[0090] Furthermore, nonlinear mapping operations are applied to the topology-aware vectors. These nonlinear mappings can employ polynomial expansion, kernel function mapping, or piecewise function transformation methods to enhance the expressive power of the topology-aware vectors in the feature space. Inter-layer feature interaction fusion operations are then performed on the mapped vectors. These operations include element-wise multiplication, vector inner product, and vector concatenation to capture the feature correlations and interdependencies between different layers. During the inter-layer feature interaction fusion process, the fusion results are normalized or scaled to ensure the stability and comparability of the fused values ​​across all dimensions. Finally, the normalized fused features are organized and formatted according to the sample index and hierarchical index, outputting a set of sample topology fusion values.

[0091] S4.6 Extract the operational availability rate, bandwidth normalization score, and synchronization success rate from the sample topology fusion value set, perform geometric mean calculation, and obtain the sample reliability score.

[0092] Furthermore, the availability rate, bandwidth normalization score, and synchronization success rate are extracted one by one according to the sample index, and field alignment and dimension matching are performed. The aligned indicators are then subjected to geometric mean calculation. The geometric mean calculation is performed by taking the logarithm of the availability rate, bandwidth normalization score, and synchronization success rate, summing them, and then taking the exponent to obtain a concentration index reflecting the overall reliability of the sample. After the geometric mean calculation is completed, the calculation results are subjected to numerical stabilization processing, such as truncation of small values ​​or logarithmic shift, to ensure the stability and comparability of the sample reliability scores. The stabilized results are then arranged and output as sample reliability scores according to the sample index order.

[0093] S4.7. Combine the sample reliability score with the sample topology fusion value set to perform correlation mapping and weight normalization operations to obtain the topology importance coefficient.

[0094] Furthermore, the sample reliability scores are aligned and matched with the corresponding sample topology fusion values ​​in the sample topology fusion value set according to the sample index. Correlation mapping is then performed on the aligned data. Existing methods such as Pearson correlation coefficient, cosine similarity, or vector inner product can be used to calculate the correlation between the sample reliability scores and the features of each dimension (such as environmental features, state features, and operational features) in the sample topology fusion value set. After completing the correlation mapping, the mapping results are weighted and normalized. This weighting and normalization includes L1 norm normalization or min-max scaling of the correlation values ​​to generate comparable and normalized topology importance coefficients across different samples and dimensions. The normalized topology importance coefficients are then organized and formatted according to the feature dimensions in the sample index and sample topology fusion value set, and the topology importance coefficients are output.

[0095] S4.8 Perform multidimensional parameter mapping and tensor aggregation operations on the topological importance coefficients to generate dimensional aggregation values, and perform multidimensional orthogonal projection and normalization synthesis operations to obtain the global topological parameter set.

[0096] Furthermore, multidimensional parameter mapping is performed on the topological importance coefficients, mapping them to a higher-dimensional parameter space in each dimension. Mapping methods can employ tensor expansion, feature combination, or multidimensional interpolation to fully express the interrelationships between different dimensions. Tensor quantization aggregation operations are then performed on the mapped multidimensional parameters. These operations include element-wise summation, mean or norm calculation along the sample dimension, feature dimension, and hierarchical dimension to form a dimensional aggregated value reflecting the joint distribution of multidimensional features. After generating the dimensional aggregated value, a multidimensional orthogonal projection operation is performed. This can be achieved using existing methods such as principal component analysis, singular value decomposition, or orthogonal basis projection to project the dimensional aggregated value into an orthogonal space to eliminate redundancy and correlation. Finally, the projected dimensional aggregated value undergoes standardization and synthesis operations, including zero-mean unit variance standardization or min-max normalization, and the global topological parameter set is output according to the feature dimension and sample index.

[0097] S5. Perform unified fusion calculation on the global topology parameter set to obtain the fusion result set, and perform sample vector extraction and collaborative feature aggregation to obtain the global parameter set.

[0098] S5.1 Extract the global topology parameter vector from the global topology parameter set, perform nonlinear mapping and weight interaction fusion to obtain the comprehensive topology feature vector.

[0099] Furthermore, global topological parameter vectors are extracted according to the feature dimensions in the sample index and sample topological fusion value set. The global topological parameter vectors corresponding to each sample are then dimensionally organized and field-aligned. Nonlinear mapping operations are applied to the extracted global topological parameter vectors. Nonlinear mapping can employ polynomial mapping, kernel function mapping, or piecewise function transformation to enhance the expressive power of the global topological parameter vectors in the feature space. After the nonlinear mapping is completed, weighted interactive fusion operations are performed on the mapped global topological parameter vectors. Weighted interactive fusion includes operations such as element-wise multiplication, vector inner product, and vector concatenation to capture the correlation and interdependence between features of different dimensions. The fusion results are then normalized and uniformly formatted to output a comprehensive topological feature vector.

[0100] S5.2 Perform fractional fusion calculation on the comprehensive topological feature vector to generate a fusion result vector, and then summarize and aggregate it to form a fusion result set.

[0101] Furthermore, the comprehensive topological feature vector is organized according to sample index and feature dimension to ensure the alignment of feature dimensions for each sample. Fractional fusion calculation is then performed on the organized comprehensive topological feature vector. This calculation can use fractional operations on the numerator and denominator to perform ratio operations or fractional combinations on feature values ​​of different dimensions, thereby enhancing the expression of relative relationships between features. After the fractional fusion calculation is completed, the resulting fusion vector is aggregated according to sample dimensions. This aggregation can use element-wise summation, mean, or norm calculation to generate a fusion result that reflects the overall distribution of features across all samples. Finally, the aggregated fusion result vector is uniformly organized and formatted, and the fusion result set is output.

[0102] S5.3 Extract the fused sample vector from the fusion result set, and perform cross-sample similarity mapping and asymmetric interactive combination operations to generate a collaborative feature set.

[0103] Furthermore, the fused sample vector corresponding to each sample is extracted, and the fused sample vector is dimensionally organized and uniformly arranged to ensure that the sample vectors can directly correspond in feature dimensions. Cross-sample similarity mapping is performed on the fused sample vector. By comparing the relative differences, feature distribution trends, and inter-sample correlations of each dimension feature in each sample vector, the similarity and differences of features between samples are identified. Asymmetric interactive combination operation is performed on the feature vectors between different samples to integrate similar features and supplement the differences of features with opposite trends or complementary information to preserve the directional features and information completeness between samples, forming a feature representation that can reflect the collaborative relationship between samples. The feature representation is uniformly organized and formatted to output a collaborative feature set.

[0104] S5.4 Perform fractional nonlinear aggregation operation on the collaborative feature set to generate a global parameter vector, and perform set-based aggregation processing to generate a global parameter set.

[0105] Furthermore, each feature vector in the collaborative feature set is organized and uniformly arranged according to its dimension to obtain a collaborative feature vector, ensuring that different collaborative feature vectors can directly correspond in feature dimensions. Fractional nonlinear aggregation operation is performed on the organized collaborative feature vector. Fractional nonlinear aggregation operation can enhance the relative relationship and nonlinear expressive power between features by combining features of different dimensions in the collaborative feature vector in fractional form and applying nonlinear function mapping, such as exponential, logarithmic or power operations, to generate a global parameter vector (including the aggregation feature relationship of each sample, the nonlinear combination value of features of each dimension, and the mutual dependence information between collaborative features). After the global parameter vector is generated, all global parameter vectors are aggregated and processed by set aggregation. By summarizing by element, calculating the mean or norm, etc., the global parameter vectors are integrated into a unified expression to form a global parameter set that reflects the overall sample feature distribution and collaborative feature relationship.

[0106] S6. Based on the global parameter set and structured dataset, perform parameter adaptation and optimization on the distributed model to obtain the trained distributed model.

[0107] S6.1 Unify and integrate the global parameter set and the structured dataset according to the sample correspondence to generate an updated distributed model parameter vector.

[0108] Furthermore, based on the sample correspondence (e.g., encoded by sample index or unique identifier), each global parameter vector is paired with the feature vector of the corresponding sample in the structured dataset, mapping each dimension of the global parameter vector to the corresponding dimension of the feature vector of the sample in the structured dataset. A unified fusion is performed on each pair of paired vectors, combining global parameter information with structured data feature information to form a single vector through vector concatenation, element-wise addition, or multiplication. Simultaneously, parameter integration processing is performed to ensure that the integrated vector maintains consistency and integrity across all dimensions, reflecting both global parameter information and local feature information for each sample. Finally, the fused vectors of all samples are aggregated to form the updated distributed model parameter vector.

[0109] It should also be noted that the sample correspondence refers to the rules or identification methods used to determine the one-to-one pairing of the global parameter vector with specific samples in the structured dataset. For example, the index order of the samples in the structured dataset (sample index) or the unique identifier encoding of each sample can be used to correctly match each global parameter vector with the feature vector of the corresponding sample, thereby ensuring that the fusion operation can accurately reflect the global parameter information and local feature information of each sample.

[0110] S6.2. Combine the updated distributed model parameter vector with the structured dataset and the global parameter set, and input it into the distributed model for forward inference and error calculation. Iteratively optimize the internal parameters of the distributed model through backpropagation to obtain the trained distributed model.

[0111] Furthermore, the structured dataset is subjected to integrity checks, field standardization, and sequential arrangement of records. Sample inputs are extracted, and the global parameter set is traversed parameter by parameter. Combined with the backpropagation calculation process of the distributed model, iterative update weight information is extracted. Each sample input is paired with the updated distributed model parameter vector and iterative update weight information to form an input-parameter-weight triplet. The triplet is then sequentially input into the distributed model for forward inference to generate the predicted output for each sample. The error value is calculated by comparing the predicted output with the actual output. The error value is used for backpropagation within the distributed model to update the internal parameters of the distributed model according to the gradient direction. The forward inference, error calculation, and parameter update steps are repeated in each iteration to gradually optimize the internal parameters of the distributed model. After completing a predetermined number of iterations, the trained distributed model is obtained.

[0112] This embodiment also provides a computer device applicable to the distributed model training method based on edge network devices, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the distributed model training method based on edge network devices as proposed in the above embodiment.

[0113] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0114] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the distributed model training method based on edge network devices as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0115] In summary, this invention achieves global perception and encoding of network topology and hierarchical relationships between nodes by cleaning, formatting, and transforming local training data in edge networks, combining nonlinear mapping and fractional aggregation of local training results to form a multi-level aggregation structure and generate a global topology parameter set; and by generating a global parameter set through sample vector extraction and cross-sample collaborative feature aggregation, it achieves deep fusion of information between different nodes and samples, thereby enhancing the training accuracy, robustness, and scalability of the distributed model in heterogeneous edge network environments. At the same time, it optimizes the model parameters overall, enabling the final distributed model to have high performance and adaptability.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for distributed model training based on edge network devices, characterized in that: Comprising, Collect local training data in a distributed edge network and perform cleaning, formatting and feature conversion to generate a structured data set; Perform local training operations on the structured data set to obtain a local update result; Extract and format the local update result to generate an intermediate parameter set; Perform logical hierarchical aggregation operations on the intermediate parameter set to form a multi-level aggregation structure, and generate a global topology parameter set through topology-aware mapping; Perform unified fusion calculation on the global topology parameter set to obtain a fusion result set, and perform sample vector extraction and collaborative feature aggregation to obtain a global parameter set; Based on the global parameter set and the structured data set, perform parameter adaptation optimization on the distributed model to obtain a trained distributed model.

2. The edge-network-device-based distributed model training method of claim 1, wherein: The specific steps of generating a structured data set are as follows, Collect local training data in a distributed edge network, perform time series synchronization processing, generate an original training data set, and perform anomaly value detection and cleaning to obtain a standardized data set; Map the standardized data set to a unified feature space and perform non-linear transformation and standardization to generate a formatted data set; Perform feature interaction and mapping processing on the formatted data set to generate a structured data set. 3.The edge network device based distributed model training method of claim 1, wherein: The specific steps of obtaining a local update result are as follows, Perform non-linear mapping and feature interaction calculation on the structured data set to obtain a local model parameter vector, and perform aggregated gradient update and non-linear fractional calculation to form a sample local tuning value; Perform aggregated training operations on the sample local tuning value to generate a local update result. 4.The edge network device based distributed model training method of claim 1, wherein: The specific steps of generating an intermediate parameter set are as follows, Extract sample feature values from the local update result, perform non-linear combination mapping to obtain a mapping feature set, and perform fractional aggregation operations to generate an intermediate parameter vector; Perform structured organization and unified formatting processing on the intermediate parameter vector to form an intermediate parameter set. 5.The edge network device based distributed model training method of claim 1, wherein: The specific steps of forming a multi-level aggregation structure are as follows, Perform hierarchical mapping and inter-layer interaction processing on the intermediate parameter set to obtain a hierarchical mapping feature set; Perform intra-layer normalization and group aggregation on the hierarchical mapping feature set to obtain a hierarchical aggregation value set, and perform cross-layer combination and non-linear mapping to generate a multi-level aggregation structure vector; Perform full-sample summarization and integration operations on the multi-level aggregation structure vector to obtain a multi-level aggregation structure. 6.The edge network device based distributed model training method of claim 1, wherein: The specific steps of generating a global topology parameter set are as follows, Combine the multi-level aggregation structure and the hierarchical mapping feature set, and perform topology-aware mapping to generate a topology-aware vector; Perform non-linear mapping and inter-layer feature interaction fusion operations on the topology-aware vector to obtain a sample topology fusion value set; Perform aggregation operations and standardization processing on the sample topology fusion value set to generate a global topology parameter set.

7. The edge-network-device-based distributed model training method of claim 6, wherein: The specific steps of performing aggregation operations and standardization processing on the sample topology fusion value set to generate a global topology parameter set are as follows, Extract the running availability, bandwidth normalization score and synchronization success rate from the sample topology fusion value set, perform geometric mean operation to obtain a sample reliability score; The sample reliability score is combined with the sample topology fusion value set to perform correlation mapping and weight normalization operation, to obtain a topology importance coefficient; The topology importance coefficient is subjected to multi-dimensional parameter mapping and tensorization aggregation operation, to generate a dimension aggregation value, and subjected to multi-dimensional orthogonal projection and standardization synthesis operation, to obtain a global topology parameter set. 8.The edge network device based distributed model training method of claim 1, wherein: The fusion result set is obtained through the following specific steps, A global topology parameter vector is extracted from the global topology parameter set, subjected to nonlinear mapping and weight interaction fusion, to obtain a comprehensive topology feature vector; The comprehensive topology feature vector is subjected to fractional fusion calculation, to generate a fusion result vector, and subjected to summary aggregation, to form a fusion result set. 9.The edge network device based distributed model training method of claim 1, wherein: The global parameter set is obtained through the following specific steps, A fusion sample vector is extracted from the fusion result set, and subjected to cross-sample similarity mapping and asymmetric interaction combination operation, to generate a collaborative feature set; The collaborative feature set is subjected to fractional nonlinear aggregation operation, to generate a global parameter vector, and subjected to set aggregation processing, to generate a global parameter set. 10.The edge network device based distributed model training method of claim 1, wherein: The trained distributed model is obtained through the following specific steps, The global parameter set and the structured data set are uniformly fused and parameter integrated according to sample corresponding relationship, to generate an updated distributed model parameter vector; The updated distributed model parameter vector is combined with the structured data set and the global parameter set, and input into the distributed model to perform forward reasoning and error calculation, and the distributed model internal parameters are iteratively optimized through back propagation, to obtain a trained distributed model.

Images