Search

CN-121981209-A - Robot personalized federal learning method based on guide filtering feature fusion

CN121981209ACN 121981209 ACN121981209 ACN 121981209ACN-121981209-A

Abstract

The invention provides a robot personalized federation learning method based on guide filtering feature fusion, which comprises the following steps of 1, system initialization and guide model deployment, 2, two-channel heterogeneous manifold coding and heterogeneous space manifold alignment, wherein in each round of federation training, a server broadcasts global guide model parameters to all online robots, the robots input sample data locally acquired by the robots into a local model and a received global guide model at the same time, 3, guide filtering feature fusion based on enhanced manifold alignment, 4, orthogonal decoupling reasoning mechanism and construction of a joint loss function with feature decoupling constraint, and 5, parameter decoupling optimization and co-evolution. The method solves the problem of characteristic alignment of the heterogeneous robot system and improves the compatibility of the system. The method effectively solves the problems of inconsistent characteristic dimensions and topology difference caused by different hardware architectures in the multi-robot system.

Inventors

  • LIU MIAO
  • HUANG KEMING
  • CHEN YIYANG
  • SUN ZHENXING

Assignees

  • 无锡学院

Dates

Publication Date
20260505
Application Date
20260203

Claims (10)

  1. 1. The personalized federal learning method for the robot based on the guide filtering feature fusion is characterized by comprising the following steps of: Step 1, system initialization and guide model deployment; Step 2, aligning the dual-channel heterogeneous manifold code with the heterogeneous space manifold, wherein in each round of federal training, the server guides the global guiding model parameters Broadcast to all online robots, robot k locally collects sample data Simultaneously inputting the local model and the received global guiding model; step 3, based on the guide filtering feature fusion of the enhanced manifold alignment; Step 4, an orthogonal decoupling reasoning mechanism and constructing a joint loss function with characteristic decoupling constraint; and 5, parameter decoupling optimization and co-evolution.
  2. 2. The method of claim 1, wherein step 1 comprises: Step 1-1, initializing a server side, namely constructing and initializing a lightweight global guiding model GM by the server, wherein parameters of the global guiding model GM are recorded as follows The feature extraction function is noted as ; Step 1-2, client local initialization heterogeneous model components, namely, each of K robot clients in a system initializes a local private model, different robots deploy local feature extraction networks of different architectures Initializing a matrix containing dimension reduction for each client And dimension-increasing matrix Nonlinear bottleneck structure, guided filter layer, guided parameter generation network of (c) And a dual-head prediction module; The nonlinear bottleneck structure is used for filtering out hardware noise and extracting manifold structures; The guide filter layer is a characteristic distribution alignment layer based on second-order statistics; the dual-head prediction module includes a global consistency head And a personalized decoupling head 。
  3. 3. The method according to claim 2, wherein step 2 comprises: step 2-1, performing the following two-channel isomerous manifold coding: channel one, general knowledge manifold coding, utilizing global steering model Extracting general semantic features with cross-scene consistency and outputting features Represented as ; A reference manifold space for group sharing is formed; Representing the ith sample data locally collected by the kth robot client; channel two, personalized heterogeneous manifold coding, namely utilizing local heterogeneous network specific to each client Extracting characteristic features adapting to local physical environment and generating channel features ; Step 2-2, manifold pair salvo shadow of isomerism space: The method comprises the steps of designing a nonlinear bottleneck structure comprising dimension reduction compression, nonlinear activation and dimension rising reconstruction on a projection layer, forcibly filtering non-semantic noise introduced by heterogeneous hardware by utilizing an information bottleneck principle, extracting a manifold structure of data essence, and firstly splicing two paths of features to obtain new features Spliced dimension Is that Then, the dimension is reduced and then increased, and the calculation formula is as follows: , Wherein, the To reduce the dimension of the compressed matrix, the intermediate dimension ; As a dimension of the global steering feature, For the dimension of the local personalization feature, For the projection layer of the original features, Activating a function for a Gaussian error linear unit; for the upgoing dimension reconstruction matrix, the method is used for remapping the purified low-dimension manifold features back to the unified feature space preset by the system ; Representing real space.
  4. 4. A method according to claim 3, wherein step 3 comprises: step 3-1, nonlinear feature enhancement path: Non-linear mapping enhancement of projection features: , Wherein, the For the layer normalization operation, The learnable parameters of the path are enhanced for non-linear features, Representing the enhanced feature vector; step 3-2, constructing a guiding parameter generator based on second order statistics by using the global guiding feature Correcting distribution drift of local features using global steering features To dynamically generate affine transformation parameters for correcting the local feature distribution; First, a parameter generation network is constructed The parameter generating network The system comprises a dimension reduction full-connection layer, a ReLU activation function and a dimension increase full-connection layer; Directing global features Input parameter generation network Two sets of statistical correction parameters, scaling factor vectors, are obtained by nonlinear mapping And a translation factor vector The formula is: , Wherein, the The dimension is consistent with the number C of the characteristic channels; step 3-3, directing the filter layer to perform statistic-based feature modulation by first projecting features locally Carrying out standardization processing, then carrying out affine transformation by utilizing global guiding parameters, and mapping the affine transformation to a global unified feature statistical space, wherein a specific calculation formula is as follows: , , Wherein, the Representing the mean of the local features in the spatial dimension; representing the variance of the local feature in the spatial dimension; a small constant to prevent zero denominator; representing channel-by-channel element multiplication, last term Is residual connection; representing the local characteristics obtained after normalization; representing the characteristics of the final fusion.
  5. 5. The method of claim 4, wherein step 4 comprises: Step 4-1, designing an orthogonal decoupling reasoning mechanism based on a double-head architecture, wherein the orthogonal decoupling reasoning mechanism forces the fused characteristics through structural independence Projection in a high-dimensional space along different semantic directions specifically includes: personalized decoupling head, constructing local classifier ; Global consistency header building Global classifier ; And 4-2, constructing a joint loss function with characteristic decoupling constraint.
  6. 6. The method of claim 5, wherein step 4-2 comprises, for a sample of client k The loss function is defined as: , , , Wherein, the In order for the cross-entropy loss to occur, And For the balance coefficient, satisfy , Represents the joint loss value of the kth robot, A true label representing the sample is presented, Representing a classification task loss; representing the loss of feature orthogonality constraint, and T represents the transpose.
  7. 7. The method of claim 6, wherein step 5 comprises: Step 5-1, independent optimization of local manifold space: Based on local data, the robot client minimizes the total loss value by using a random gradient descent algorithm Updating private components of heterogeneous manifold encoder parameters Guiding filtering fusion layer Orthogonal decoupling inference head ; Step 5-2, the aggregation evolution of the global oriented knowledge; and 5-3, cooperatively iterating, namely repeatedly executing the steps 2 to 5-2 until the global guiding model converges or reaches a preset training round.
  8. 8. The method of claim 7, wherein step 5-2 comprises the client computing and uploading information about the global steering model only The server gathers feedback from all online clients, performs a weighted average aggregation to evolve general knowledge: , Wherein, the Representing the updated global steering model parameters of the t+1st round, Representing the set of robot clients selected to participate in the training in round t, Represents the total number of robots involved in the present round of training, And the updated global guiding model parameters uploaded by the kth robot are represented.
  9. 9. An electronic device comprising a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method of any of claims 1 to 8.
  10. 10. A storage medium storing a computer program or instructions which, when run on a computer, performs the steps of the method of any one of claims 1 to 8.

Description

Robot personalized federal learning method based on guide filtering feature fusion Technical Field The invention belongs to the field of cooperative control of artificial intelligence and robots, and particularly relates to a robot personalized federal learning method based on guided filter feature fusion. Background In the age background of rapid development of digital economy and intelligent internet of things (AIoT), multi-Robot collaboration Systems (Multi-Robot Systems) are gradually evolving from single tasks to group intelligence in complex scenarios. Through collaborative learning, robot groups can share experience and complementary knowledge, so that limitations of individual data and computing power are overcome. However, in practical deployment and application, collaborative learning of multi-robot systems faces severe "dual heterogeneous" and "data privacy" challenges, severely limiting the application performance of the prior art: 1. data privacy security challenges Visual images, laser radar point clouds and user interaction data collected by robots during tasks (e.g., home services, inspection) often contain sensitive privacy information (e.g., home layout, user habits). Limited by data privacy regulations such as GDPR, these data cannot be nor should they be uploaded to a cloud central server for centralized training. 2. Data dependent co-distribution (Non-IID) challenges Since different robots are in different physical environments (e.g., light, terrain differences) or perform different tasks (e.g., grabbing, navigating), there are significant differences in the distribution of their local data. This data heterogeneity results in divergent update directions for each robot's local model, the global model of a traditional federal learning algorithm (e.g., fedAvg) is difficult to converge, and the individuality on a particular robot is poor. 3. System and model heterogeneous challenges In reality, a robot cluster often consists of devices with different batches and different models, and the computing power (CPU/GPU/NPU), storage resources and sensor configurations are different. This means that different robots can only run neural network models of different architecture or scale (e.g., resNet and MobileNet coexist). The model isomerism causes the model parameter space of each client to be inconsistent, so that the server cannot directly perform parameter aggregation. In order to solve the above problems, the prior art mainly adopts strategies such as parameter decoupling (e.g. FedPer, fedRep) or knowledge distillation (e.g. FedMD, fedProto). Although the parameter decoupling method can keep the personalized layer, the static model division is difficult to adapt to the dynamic changing physical environment of the robot; Knowledge distillation methods, while supporting heterogeneous models, typically rely on a common data set to assist in knowledge migration. In the field of robots, it is extremely difficult and costly to construct a high quality public data set that can cover all real physical environment changes; existing feature fusion methods (such as simple stitching or weighting) have difficulty in handling semantic conflicts between heterogeneous features (e.g., the same feature represents different meanings in different scenarios), and lack dynamic adaptive mechanisms. In summary, the existing methods have limitations in terms of: (1) Lack of a mechanism for realizing efficient knowledge sharing among heterogeneous models without common data dependence; (2) The conflict and redundancy problems between the global general knowledge and the local personalized features cannot be effectively solved; (3) The communication overhead is large, and the real-time requirement of the robot network bandwidth limitation is difficult to meet. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a personalized federal learning method of a robot based on guided filter feature fusion, which aims to solve the following problems on the premise of not uploading local original data of the robot: 1. collaborative training difficulty caused by isomerism (different model architectures and calculation force differences) of a robot system; 2. The global model performance is reduced due to Non-independent co-distribution (Non-IID) of robot data; 3. high communication overhead caused by traditional large-model transmission; 4. Conflict between global general knowledge and local personalized requirements of the robot. The method specifically comprises the following steps: Step 1, system initialization and guide model deployment; Step 2, aligning the dual-channel heterogeneous manifold code with the heterogeneous space manifold, wherein in each round of federal training, the server guides the global guiding model parameters Broadcast to all online robots, robot k locally collects sample dataSimultaneously inputting the local model and the received global guiding model; step 3, based o