Search

CN-121998130-A - Lightweight air detection model based on federal learning and cross-domain environment adaptation method

CN121998130ACN 121998130 ACN121998130 ACN 121998130ACN-121998130-A

Abstract

The invention relates to the technical field of federal learning environment monitoring, and discloses a lightweight air detection model based on federal learning and a cross-domain environment adaptation method. The model comprises an environment perception interface, a cross-domain environment difference measure, a model parameter dynamic adjustment strategy, a federal learning aggregation server and five core modules for parameter evolution. The system collects data of each node through an environment perception interface to construct an environment feature library, a difference measurement module generates an environment difference spectrum representing data distribution heterogeneity, a dynamic adjustment module activates a parameter space division and self-adaptive learning rate mechanism according to the environment difference spectrum, an aggregation server coordinates each node to carry out collaborative optimization based on the mechanism, and a parameter evolution module generates final model parameters through multiple rounds of iteration. According to the method, the detection precision and the robustness of the model under heterogeneous environments are improved by quantifying the environmental differences and implementing parameter differentiation updating.

Inventors

  • LIU YUXIANG
  • SHI YAN
  • YANG WEIGUO

Assignees

  • 上海凌泽信息科技有限公司

Dates

Publication Date
20260508
Application Date
20260127

Claims (10)

  1. 1. Lightweight air detection model based on federal learning, characterized in that the model comprises: The environment perception interface module is configured to acquire real-time environment data streams of the deployment nodes and construct an environment feature library based on the real-time environment data streams; The cross-domain environment difference measurement module is configured to measure the cross-domain environment difference by utilizing the environment feature library and generate an environment difference spectrum; The model parameter dynamic adjustment strategy module is configured to activate parameter space division rules and an adaptive learning rate mechanism according to the environmental difference spectrum; the federal learning aggregation server module is configured to start a collaborative optimization flow based on the parameter space division rule and the self-adaptive learning rate mechanism; the parameter evolution module is configured to conduct multiple parameter evolution on the lightweight air detection model through the collaborative optimization flow, and generates a cross-domain adapted air detection model parameter set.
  2. 2. The federally-learned lightweight air detection model according to claim 1, wherein the cross-domain environmental variability metrics using the environmental profile library generate an environmental variability spectrum comprising: Extracting key environmental characteristic dimensions from the environmental characteristic library, wherein the key environmental characteristic dimensions comprise temperature fluctuation characteristics, humidity distribution characteristics and particulate matter concentration change characteristics; Calculating the distribution distance of key environmental feature dimensions among different deployment nodes to form an environmental feature distance matrix; Performing multi-scale spectrum analysis on the environmental characteristic distance matrix, and extracting principal component components of environmental characteristic differences; An environmental difference spectrum is constructed based on the principal component components, the environmental difference spectrum comprising a short-term environmental fluctuation pattern and a long-term environmental transition trend.
  3. 3. The federally learned lightweight air detection model according to claim 2, wherein the activating parameter space partitioning rules and adaptive learning rate mechanisms according to the environmental difference spectrum comprises: analyzing a short-term environment fluctuation mode and a long-term environment transition trend in the environment difference spectrum; When the short-term environment fluctuation mode is detected to exceed a preset threshold value, triggering a quick response adjustment mechanism, wherein the quick response adjustment mechanism carries out high-frequency fine adjustment on model shallow parameters; when a long-term environment transition trend is identified, starting a depth adaptation adjustment mechanism, wherein the depth adaptation adjustment mechanism carries out structural optimization on deep parameters of the model; and generating a parameter space division rule and an adaptive learning rate mechanism according to output results of the quick response adjustment mechanism and the depth adaptation adjustment mechanism.
  4. 4. The federally learned lightweight air detection model according to claim 3, wherein the initiating a collaborative optimization procedure based on the parameter space partitioning rules and an adaptive learning rate mechanism comprises: The federal learning aggregation server receives the local model parameter updating quantity uploaded by each deployment node; Partitioning the local model parameter updating amount by applying a parameter space partitioning rule to generate a parameter updating partition set; applying an adaptive learning rate mechanism to each parameter updating partition, and calculating a partition specific learning rate; and carrying out weighted aggregation on the parameter updating partition set based on the partition specific learning rate to generate a global model parameter updating instruction.
  5. 5. The federally learned lightweight air test model according to claim 4, wherein the performing multiple rounds of parameter evolution on the lightweight air test model by the collaborative optimization procedure comprises: distributing the global model parameter updating instruction to each deployment node; Each deployment node updates local model parameters based on the global model parameter updating instruction, and acquires a new round of environment data flow; processing a new round of environmental data flow by using the updated local model to generate a model performance evaluation index; Adjusting partition strategy and learning rate setting of next round of parameter updating based on the model performance evaluation index; and repeatedly executing the adjustment process until the model performance evaluation index reaches a stable state.
  6. 6. The federally learned based lightweight air test model according to claim 5, wherein the processing the new environmental data stream with the updated local model to generate the model performance assessment index comprises: inputting a new round of environmental data stream into the updated local model to obtain a model output result; comparing the output result of the model with real monitoring data, and calculating an accuracy index, a recall index and a specificity index; Monitoring the consumption of computing resources in the model reasoning process, and generating an efficiency index; And fusing the precision index, the recall index, the specificity index and the efficiency index to generate a comprehensive model performance evaluation index.
  7. 7. A federally learned lightweight air test model according to claim 3, wherein the fast response tuning mechanism performs high frequency fine tuning for model shallow parameters, comprising: Identifying a shallow parameter subset with highest sensitivity to environmental change in the model, and establishing a real-time mapping relation between the shallow parameter subset and environmental fluctuation characteristics; and dynamically adjusting the learning rate of the shallow parameter subset according to the intensity of the environmental fluctuation characteristic, and setting a shallow parameter updating frequency threshold.
  8. 8. A federally learned lightweight air test model according to claim 3, wherein the depth adaptation mechanism structurally optimizes model deep parameters, comprising: Analyzing the influence path of the long-term environment transition trend on the model functional structure, and identifying deep parameter units needing structural adjustment; And optimizing the deep parameter unit by adopting a progressive parameter updating strategy.
  9. 9. The federally learned lightweight air detection model according to claim 1, wherein the acquiring the real-time environmental data stream of deployment nodes comprises: configuring a multi-mode environment sensor network, and continuously collecting temperature, humidity, air pressure and particulate matter concentration data; Carrying out time sequence alignment and outlier filtering on the collected original environment data; Extracting statistical features and time sequence features of the environmental data, and constructing an environmental feature vector; The environmental feature vectors are organized into a real-time environmental data stream at time stamps.
  10. 10. A federal learning-based lightweight air detection cross-domain environment adaptation method applied to the federal learning-based lightweight air detection model according to any one of claims 1 to 9, comprising the steps of: step 1, collecting real-time environment data streams of deployment nodes, and constructing an environment feature library based on the real-time environment data streams; step 2, using the environment feature library to carry out cross-domain environment difference measurement and generating an environment difference spectrum; step 3, activating a parameter space division rule and an adaptive learning rate mechanism according to the environmental difference spectrum; step 4, starting a collaborative optimization flow based on the parameter space division rule and the self-adaptive learning rate mechanism; And 5, carrying out multiple parameter evolution on the lightweight air detection model through the collaborative optimization flow, and generating a cross-domain adapted air detection model parameter set.

Description

Lightweight air detection model based on federal learning and cross-domain environment adaptation method Technical Field The invention relates to the technical field of federal learning environment monitoring, in particular to a lightweight air detection model based on federal learning and a cross-domain environment adaptation method. Background Currently, an air quality monitoring system based on the internet of things generally adopts a centralized model training method, and all data acquired by sensor nodes distributed at different geographic positions need to be transmitted to a central server. The method not only generates huge data transmission cost, but also brings security and privacy risks of sensitive environment data leakage. Federal learning techniques allow models to be trained at local nodes, only uploading model parameters for aggregation, providing a potential solution to these challenges. The direct application of standard federal learning frameworks to air quality detection scenarios faces obstacles. The mainstream algorithm builds on the ideal assumption that the data is independent and distributed. In reality, the data distribution of different nodes has high heterogeneity due to the remarkable difference of geographic positions, pollution source constitution and meteorological conditions. Such environmental differences result in the aggregated global model performing poorly on specific local nodes and being poorly generalized. The disadvantage of the prior art is that its optimization strategy is disjoint from the physical environment. The aggregation process only mechanically processes model parameters, completely ignoring environmental sources that lead to parameter differences. The system cannot quantitatively evaluate the similarity and the difference of the environmental characteristics of different nodes, and lacks the perception capability of the environmental heterogeneity. Traditional parameter updating strategies are consistent across all model components, failing to distinguish between sensitivity of different parameters to environmental changes. The extensive aggregation method makes the model difficult to realize accurate and robust detection effect in complex and changeable real environments. There is a need for a new federal learning approach that can sense and accommodate environmental differences. Disclosure of Invention The invention aims to provide a lightweight air detection model based on federal learning and a cross-domain environment adaptation method, so as to solve the problems in the background technology. To achieve the above object, the present invention provides a lightweight air detection model based on federal learning, the model comprising: The environment perception interface module is configured to acquire real-time environment data streams of the deployment nodes and construct an environment feature library based on the real-time environment data streams; The cross-domain environment difference measurement module is configured to measure the cross-domain environment difference by utilizing the environment feature library and generate an environment difference spectrum; The model parameter dynamic adjustment strategy module is configured to activate parameter space division rules and an adaptive learning rate mechanism according to the environmental difference spectrum; the federal learning aggregation server module is configured to start a collaborative optimization flow based on the parameter space division rule and the self-adaptive learning rate mechanism; the parameter evolution module is configured to conduct multiple parameter evolution on the lightweight air detection model through the collaborative optimization flow, and generates a cross-domain adapted air detection model parameter set. Preferably, the performing cross-domain environmental difference measurement by using the environmental feature library, generating an environmental difference spectrum includes: Extracting key environmental characteristic dimensions from the environmental characteristic library, wherein the key environmental characteristic dimensions comprise temperature fluctuation characteristics, humidity distribution characteristics and particulate matter concentration change characteristics; Calculating the distribution distance of key environmental feature dimensions among different deployment nodes to form an environmental feature distance matrix; Performing multi-scale spectrum analysis on the environmental characteristic distance matrix, and extracting principal component components of environmental characteristic differences; An environmental difference spectrum is constructed based on the principal component components, the environmental difference spectrum comprising a short-term environmental fluctuation pattern and a long-term environmental transition trend. Preferably, the activating a parameter space division rule and an adaptive learning rate mechanism accordi