Reservoir area water bloom rapid monitoring method and device based on unmanned aerial vehicle swarm coordination

A reservoir area water bloom rapid monitoring device based on unmanned aerial vehicle swarm coordination, comprising a central server and unmanned aerial vehicles as clients;

• wherein reservoir area water bloom image data is collected by using unmanned aerial vehicle swarm;
• each of the unmanned aerial vehicles collects and saves the reservoir area water bloom image data, and a reservoir area water bloom rapid monitoring model is trained through federal learning algorithm;
• during r-th round of training, the central server sends r−1th round of global model update Δ{tilde over (w)}r−1 to N unmanned aerial vehicles;
• the N unmanned aerial vehicles receiving global model update Δ{tilde over (w)}r−1 update a local model to
• w
• i
• r
• =
• w
• i
• r
• -
• 1
• +
• △
• ⁢
• w
• ~
• r
• -
• 1
• ,
• wherein
• w
• i
• r
• is a local model after updating Δ{tilde over (w)}r−1 and
• w
• i
• r
• -
• 1
• is a local model before updating Δ{tilde over (w)}r−1;
• after updating the local model, unmanned aerial vehicles participating in federal training defines loss function queue according to water bloom image data collected by itself, and calculates a historical average loss value and a current average loss value;
• the unmanned aerial vehicles use a ratio of historical loss function to current loss function to determine a communication quantization coefficient for the local model, and further introduces a fading learning rate to improve a problem of quantization effect gradually weakening with communication rounds iteration;
• the unmanned aerial vehicles adopts random and uniform quantization operation to compress local model update, is quantized local model update of unmanned aerial vehicle ci, and Δ{tilde over (w)}i is transmitted and communicated to the central server, and the central server prepares for next round communication after receiving model update;
• communication between the central server and the unmanned aerial vehicle swarm continues, so as to iteratively update the local model of the unmanned aerial vehicles and a global model of the central server until the global model converges or reaches a specified communication round;
• on a basis of ensuring accuracy of the global model, federal learning communication cost is greatly reduced, convergence speed of the reservoir area water bloom monitoring model is accelerated, and communication volume in a training process of the reservoir area water bloom monitoring model is greatly reduced, a final obtained global model is the reservoir area water bloom rapid monitoring model;
• in actual monitoring, a global model obtained by deployment is capable of being applied to newly collected image data to realize real-time water bloom monitoring;
• when new image data are collected, the new image data are sent to a deployed model for processing, and model outputs prediction results of whether water bloom exists in water body;
• the results are capable of being presented to users through visual interface or alarm system, thus monitoring personnel are helped to take timely measures to deal with water quality problems.

A computer device, comprising a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein when the processor executes the computer program, a reservoir area water bloom rapid monitoring method based on unmanned aerial vehicle swarm coordination is realized, in the reservoir area water bloom rapid monitoring method based on unmanned aerial vehicle swarm coordination, compressing model update between unmanned aerial vehicles and a central server by local update quantification technology, collecting reservoir area water bloom image data by unmanned aerial vehicle swarm, and obtaining a global model by aggregating a local model trained on unmanned aerial vehicles, the global model being a reservoir area water bloom rapid monitoring model, wherein the method comprises following steps:

• S1:
• according to water bloom image data taken by the unmanned aerial vehicles, training the local model by the unmanned aerial vehicles through federal learning algorithm, the local model is the reservoir area water bloom rapid monitoring model;
• in a training process of r-th round, updating r−1th round trained global model Δ{tilde over (w)}r−1 and sending to N unmanned aerial vehicles by the central server;
• S2:
• updating the local model to
• w
• i
• r
• =
• w
• i
• r
• -
• 1
• +
• △
• ⁢
• w
• ~
• r
• -
• 1
• by the N unmanned aerial vehicles receiving global model update Δ{tilde over (w)}r−1, wherein
• w
• 1
• r
• is a local model after updating Δ{tilde over (w)}r−1 and
• w
• i
• r
• -
• 1
• is a local model before updating Δ{tilde over (w)}r−1;
• S3:
• after updating the local model, defining a loss function queue according to the water bloom image data collected by unmanned aerial vehicles participating in federal training, and calculating a historical average loss value and a current average loss value;
• S4:
• determining a communication quantization coefficient for the local model by using a ratio of a historical loss function to a current loss function, and further introducing a fading learning rate to improve a problem of quantization effect gradually weakening with communication round iteration;
• S5:
• compressing local model update Δ{tilde over (w)}i by random and uniform quantization operation, wherein Δ{tilde over (w)}i is quantized local model update of unmanned aerial vehicle ci;
• transmitting and communicating Δ{tilde over (w)}i to the central server;
• after receiving model update, preparing next round communication by the central server, and repeating S1-S5 until the global model converges or reaches a specified communication round, greatly reducing federal learning communication cost on a basis of ensuring accuracy of the global model, accelerating convergence speed of the reservoir area water bloom monitoring model, and greatly reducing communication volume in a training process of the reservoir area water bloom monitoring model, wherein a final obtained global model is the reservoir area water bloom rapid monitoring model.