Search

CN-122019173-A - Edge computing implementation method and system based on cloud edge cooperation

CN122019173ACN 122019173 ACN122019173 ACN 122019173ACN-122019173-A

Abstract

The invention belongs to the technical field of edge calculation, and discloses a cloud edge cooperation-based edge calculation implementation method and a cloud edge cooperation-based edge calculation implementation system. The method comprises the steps of extracting a local distortion signature which does not contain original privacy data by an edge node based on an reasoning process of a local model, uploading the local distortion signature to a cloud, constructing and training a variation diffusion residual error model by the cloud, generating a lightweight distortion compensation table for each edge node based on a learned reversible mapping, transmitting the distortion compensation table and a basic model to the edge node together, replacing the local model by the edge node by the transmitted basic model, and executing deformable compensation on the basic model by loading the distortion compensation table so as to correct expected distortion of the basic model relative to data distribution of the edge node, thereby improving consistency of cross-node reasoning. The method can effectively correct the model distribution distortion of each edge node caused by long-term independent training, and remarkably improves the consistency of the cross-node reasoning result, the feasibility of cloud model aggregation and the stability of edge model updating.

Inventors

  • KANG JUNYAN

Assignees

  • 云边云科技(上海)有限公司

Dates

Publication Date
20260512
Application Date
20260202

Claims (10)

  1. 1. The edge computing implementation method based on cloud edge cooperation is characterized by comprising the following steps of: Step S10, extracting a local distortion signature which does not contain original privacy data on the side of an edge node based on an inference process of a local model, wherein the local distortion signature comprises an input distribution offset characteristic, an inference path activation mode characteristic and a gradient direction statistical characteristic, and uploading the local distortion signature to a cloud; Step S20, constructing and training a variation diffusion residual model at the cloud side, wherein the variation diffusion residual model is used for learning reversible mapping from local distortion signatures of edge nodes to cloud reference distribution, and generating a lightweight distortion compensation table for each edge node based on the reversible mapping, and the lightweight distortion compensation table comprises compensation parameters for carrying out weight micro-deformation, activation domain weight alignment and reasoning boundary adjustment on the node model; And step S30, the distortion compensation table and the basic model are issued to the edge node, the edge node replaces the local model by using the issued basic model, and the distortion compensation table is loaded to execute deformable compensation on the basic model so as to correct the expected distortion of the basic model relative to the data distribution of the edge node, thereby improving the consistency of the cross-node reasoning.
  2. 2. The edge computing implementation method based on cloud edge collaboration according to claim 1, wherein the extracting of the local distortion signature not including the original privacy data based on the reasoning process of the local model comprises: s11, collecting reasoning data of a local model of a preset time window, and extracting input distribution offset characteristics, reasoning path activation mode characteristics and gradient direction statistical characteristics based on the reasoning data, wherein: the input distribution offset feature is used for representing the dynamic offset degree of the current input data distribution of the edge node relative to cloud reference distribution, and is obtained at least by comparing the statistics of the current reasoning input with the historical cloud reference input distribution; The inference path activation mode features are used for representing distortion and deformation of an internal feature space of the local model, and are obtained at least by analyzing sparsity of activation tensors of key layers of the model, frequency domain energy distribution and statistical differences between the frequency domain energy distribution and cloud reference activation; The gradient direction statistical features are used for representing the long-term offset trend of the local model local training updating direction relative to a reference coordinate system, and at least comprise direction distribution features in a low-dimensional projection space, and deviation degree indexes of the current updating direction and the historical updating direction; and step S12, compressing and encoding the input distribution offset characteristic, the inference path activation mode characteristic and the gradient direction statistical characteristic to generate a low-dimensional signature vector as a local distortion signature.
  3. 3. The edge computing implementation method based on cloud edge collaboration, as set forth in claim 1, is characterized by constructing and training a variational diffusion residual model for learning a reversible mapping from local distortion signatures of edge nodes to cloud reference distribution, and comprising: S21, constructing a variation diffusion residual model comprising an encoder, a multi-step diffusion processor and a residual mapper; The multi-step diffusion processor is used for performing a forward diffusion process of multi-step noise addition on the characterization in the latent space, and the residual mapper is used for gradually denoising the noisy latent space characterization in a reverse process and learning the residual between the noisy latent space characterization and cloud reference distribution so as to approximate the reversible mapping; step S22, training the variational diffusion residual model by minimizing the difference between the output of the residual mapper and cloud reference distribution.
  4. 4. The edge computing implementation method based on cloud edge collaboration according to claim 1, wherein the multi-step diffusion processor is configured to perform a multi-step noise addition forward diffusion process on the characterization in the latent space, and the method comprises: Receiving the latent space representation from the encoder and setting a total diffusion step number K; setting the noise adding intensity of each diffusion step by adopting a cosine scheduling strategy, and sequentially executing noise adding operation according to the diffusion step sequence, wherein the cosine scheduling strategy ensures that the noise adding intensity is relatively low in the earlier diffusion step so as to keep structural information, and the noise adding intensity is gradually enhanced in the later diffusion step so as to accelerate the noising; after the diffusion treatment of the K step, the noise adding latent space representation output by the K step is ensured to be pure noise vector conforming to standard Gaussian distribution.
  5. 5. The method for realizing edge calculation based on cloud edge cooperation as set forth in claim 1, wherein generating a lightweight distortion compensation table for each edge node based on the reversible mapping comprises: Step S23, inputting the local distortion signature of the edge node into a reversible mapping function of the variation diffusion residual error model to obtain a corresponding reverse distortion form as a reverse distortion vector; and S24, analyzing and distributing the reverse distortion vector to a weight deformation branch, an activation offset branch and a decision boundary branch to generate compensation parameters for weight micro deformation, activation domain compensation and reasoning boundary compensation, and generating the distortion compensation table based on the compensation parameters.
  6. 6. The method for realizing edge calculation based on cloud edge cooperation according to claim 1, wherein the deformable compensation specifically comprises weight micro deformation compensation, activation domain compensation and inference boundary compensation, and the method is characterized in that: The weight micro deformation compensation is specifically to locally stretch, compress or rotate model weights according to deformation factors in the distortion compensation table; the activation domain compensation specifically includes introducing a leachable offset to the activation output of a model specific layer to stabilize the activation variance; the reasoning boundary compensation is specifically to adjust the boundary softening coefficient or the normalization scale of the model output layer.
  7. 7. An edge computing implementation system based on cloud edge cooperation is characterized by comprising a cloud end, a plurality of edge nodes and communication equipment; The edge node is configured to extract a local distortion signature which does not contain original privacy data based on an reasoning process of a local model, wherein the local distortion signature comprises an input distribution offset characteristic, a reasoning path activation mode characteristic and a gradient direction statistical characteristic, and upload the local distortion signature to a cloud; the cloud end is configured to construct and train a variation diffusion residual model, which is used for learning reversible mapping from local distortion signatures of edge nodes to cloud end reference distribution, generating a lightweight distortion compensation table for each edge node based on the reversible mapping, and comprises compensation parameters for carrying out weight micro-deformation, activation domain weight alignment and reasoning boundary adjustment on the node model; The edge node is further configured to replace a local model thereof with the issued base model and to improve consistency of cross-node reasoning by loading the distortion compensation table to perform deformable compensation on the base model to correct expected distortion thereof relative to the edge node data distribution; The communication device is configured to realize data transmission between the edge node and the cloud.
  8. 8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method according to any of the claims 1-6 when executing the computer program.
  9. 9. A storage medium having stored thereon a computer program, which when executed by a processor performs the steps of the method according to any of claims 1-6.
  10. 10. Computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, realizes the steps of the method according to any of claims 1-6.

Description

Edge computing implementation method and system based on cloud edge cooperation Technical Field The invention relates to the technical field of edge calculation, in particular to a cloud edge cooperation-based edge calculation implementation method and system. Background With the wide application of edge computing, artificial intelligence models are increasingly deployed on various types of edge nodes (such as factories, markets, traffic intersections, and the like) for long-term autonomous training and local adaptation. However, due to the independence of the environments of the edge nodes and the heterogeneity of the data of the edge nodes, the model can generate model distribution Distortion (disturbance Drift) phenomenon in the long-term operation process, and the model is characterized in that the reasoning results of the same task on different edge nodes are not comparable, the cloud is difficult to effectively aggregate the models or the reasoning results uploaded by different nodes, the model performance fluctuation is obvious, the stability is reduced when the nodes are switched or updated, and the new model issued by the cloud is inconsistent in performance on different nodes and is uncontrollable in updating. Meanwhile, federal learning, model alignment, light-weight distillation, and other methods in the prior art generally require similar data distribution, consistent model structure, or uniform training environment and update period, which are difficult to meet in actual industrial scenes. Therefore, a technical scheme capable of realizing real-time alignment and distortion compensation of the cross-node on the premise of not depending on unified data and unified training environment is needed. Disclosure of Invention The invention aims to provide a cloud edge cooperation-based edge computing implementation method, a cloud edge cooperation-based edge computing implementation system, computer equipment, a storage medium and a computer program product, so as to solve the problems in the background art. The first aspect of the invention provides an edge computing implementation method based on cloud edge cooperation, which comprises the following steps: Step S10, extracting a local distortion signature which does not contain original privacy data on the side of an edge node based on an inference process of a local model, wherein the local distortion signature comprises an input distribution offset characteristic, an inference path activation mode characteristic and a gradient direction statistical characteristic, and uploading the local distortion signature to a cloud; Step S20, constructing and training a variation diffusion residual model at the cloud side, wherein the variation diffusion residual model is used for learning reversible mapping from local distortion signatures of edge nodes to cloud reference distribution, and generating a lightweight distortion compensation table for each edge node based on the reversible mapping, and the lightweight distortion compensation table comprises compensation parameters for carrying out weight micro-deformation, activation domain weight alignment and reasoning boundary adjustment on the node model; And step S30, the distortion compensation table and the basic model are issued to the edge node, the edge node replaces the local model by using the issued basic model, and the distortion compensation table is loaded to execute deformable compensation on the basic model so as to correct the expected distortion of the basic model relative to the data distribution of the edge node, thereby improving the consistency of the cross-node reasoning. The second aspect of the invention provides an edge computing implementation system based on cloud edge cooperation, which comprises a cloud end, a plurality of edge nodes and communication equipment; The edge node is configured to extract a local distortion signature which does not contain original privacy data based on an reasoning process of a local model, wherein the local distortion signature comprises an input distribution offset characteristic, a reasoning path activation mode characteristic and a gradient direction statistical characteristic, and upload the local distortion signature to a cloud; the cloud end is configured to construct and train a variation diffusion residual model, which is used for learning reversible mapping from local distortion signatures of edge nodes to cloud end reference distribution, generating a lightweight distortion compensation table for each edge node based on the reversible mapping, and comprises compensation parameters for carrying out weight micro-deformation, activation domain weight alignment and reasoning boundary adjustment on the node model; The edge node is further configured to replace a local model thereof with the issued base model and to improve consistency of cross-node reasoning by loading the distortion compensation table to perform deformable compensation on the ba