Lightweight Magnetic Compensation Neural Network Method for Multi-rotor UAV Aerial Computing Platform

Abstract

This invention relates to the field of aeromagnetic measurement and provides a lightweight magnetic compensation neural network method for a multi-rotor UAV aeronautical computing platform. The method includes: constructing a teacher model: pre-training a large neural network as the teacher model, wherein the teacher model is a deep neural network with multiple hidden layers or a backpropagation neural network, used to learn the magnetic interference characteristics in aeromagnetic data; constructing a student model and performing knowledge distillation: establishing a lightweight three-layer feedforward fully connected neural network as the student model, and transferring knowledge from the teacher model to the student model through knowledge distillation; deploying the lightweight model to the UAV platform: transferring the weights and biases of the trained student model to the computing platform carried by the multi-rotor UAV to achieve real-time magnetic interference compensation of the aeromagnetic data. This invention can lightweight the entire neural network structure, enabling low-cost and high-efficiency completion of aeromagnetic measurement data compensation calculations under limited computing resources.

Description

Technical Field

[0001] This invention relates to the field of aeromagnetic measurement, and more particularly to a lightweight magnetic compensation neural network method for a multi-rotor unmanned aerial vehicle (UAV) aeronautical computing platform. Background Technology

[0002] Aerial magnetic surveying, as an important geophysical exploration method, provides crucial information about a region's geological structure and mineral resources. In recent years, the development of multi-rotor UAVs as exploration platforms has made aerial magnetic surveying a more feasible data measurement technology in terms of data collection efficiency and data quality. However, aerial magnetic data contains geomagnetic information and magnetic interference information generated by the aircraft. Eliminating magnetic interference is one way to improve the accuracy of aerial magnetic data. An effective aerial magnetic compensation method will further improve the accuracy of the measurement data. Existing high-precision magnetic interference compensation methods utilize neural networks to learn about magnetic interference and use the learned biases and weights to perform compensation operations. Commonly used neural networks for aerial magnetic interference compensation include backpropagation neural networks (BPNN) and deep neural networks (DNN). However, while these neural networks have good compensation effects for aerial magnetic interference, their complex structures with numerous hidden layers and neurons require significant computational resources. The computing resources of the computing devices on multi-rotor UAV platforms are limited. For situations requiring real-time compensation calculations on computing devices, large-scale and computationally complex neural networks are costly to operate and slow to compute. Summary of the Invention

[0003] In view of the shortcomings of the prior art, the purpose of this invention is to propose a lightweight magnetic compensation neural network method for a multi-rotor unmanned aerial vehicle (UAV) aeronautical computing platform to solve the technical problems mentioned in the background art.

[0004] A lightweight magnetic compensation neural network method for a multi-rotor unmanned aerial vehicle (UAV) avionics computing platform, the method comprising:

[0005] S1. Constructing a teacher model: A pre-trained large neural network is used as a teacher model, which is a deep neural network (DNN) with multiple hidden layers or a backpropagation neural network (BPNN) to learn the magnetic interference features in aeromagnetic data;

[0006] S2. Construct a student model (a small neural network with only an input layer, hidden layer, and output layer) and perform knowledge distillation: Build a lightweight three-layer feedforward fully connected neural network as the student model, and transfer the knowledge from the teacher model to the student model through knowledge distillation;

[0007] S3. Deploy lightweight models to UAV platforms: Transfer the weights and biases of the trained student models to the embedded computing platform (such as Raspberry Pi) carried by multi-rotor UAVs to achieve real-time magnetic interference compensation for aeromagnetic data.

[0008] Preferably, step S2 includes:

[0009] S2.1 Design the total loss function ,in:

[0010] (Cross-Entropy Loss) is the standard cross-entropy loss of the student model on the hard labels of the training data;

[0011] ;

[0012] (Knowledge Distillation Loss) is the distillation loss used by the student model to mimic the soft label distribution of the teacher model.

[0013] ;

[0014] Where L is the total loss function, For weighted adjustment of hyperparameters, This is the true label of the i-th sample, and T is the distillation temperature parameter. This represents the output probability distribution of the teacher model at temperature T. The predictions of the student model at the same temperature.

[0015] Preferably, and The calculation formula is:

[0016] ;

[0017] ;

[0018] in, This represents the original output value of the i-th data point in the teacher model before processing by the softmax activation function. This represents the raw output value of the j-th data point in the teacher model before processing by the softmax activation function. This represents the original output value of the i-th data point in the student model before processing by the softmax activation function. This represents the raw output value of the j-th data point in the student model before it is processed by the softmax activation function.

[0019] Preferably, the real-time magnetic interference compensation described in step S3 is performed by the computing platform calling the student model to perform online compensation calculations on the data while the UAV is collecting aeromagnetic data.

[0020] The beneficial effects achieved by this invention are as follows:

[0021] Compared with large neural networks performing direct compensation calculations on the computing equipment of a multi-rotor UAV aero-magnetic measurement platform, this invention can reduce the structure of the entire neural network and achieve low-cost and high-efficiency compensation calculations for aero-magnetic measurement data under limited computing resources. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the principle of knowledge distillation in an embodiment of the present invention.

[0023] Figure 2 This is a structural diagram of knowledge distillation according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1 to 2 This invention provides a lightweight magnetic compensation neural network method for a multi-rotor unmanned aerial vehicle (UAV) aeronautical computing platform, the method comprising:

[0026] S1. Constructing a teacher model: A pre-trained large neural network is used as a teacher model, which is a deep neural network (DNN) with multiple hidden layers or a backpropagation neural network (BPNN) to learn the magnetic interference features in aeromagnetic data;

[0027] S2. Construct a student model (a small neural network with only an input layer, hidden layer, and output layer) and perform knowledge distillation: Build a lightweight three-layer feedforward fully connected neural network as the student model, and transfer the knowledge from the teacher model to the student model through knowledge distillation;

[0028] S3. Deploy lightweight models to UAV platforms: Transfer the weights and biases of the trained student models to the embedded computing platform (such as Raspberry Pi) carried by multi-rotor UAVs to achieve real-time magnetic interference compensation for aeromagnetic data.

[0029] In this invention, the neural network serving as the teacher model is much larger than the student model. The main idea is to train an efficient student model under the guidance of the teacher model to achieve comparable accuracy. Supervisory signals from the teacher model (often referred to as the "knowledge" learned by the teacher model) help the student model mimic the teacher model's behavior. The predictions obtained through neural network learning contain weight information and can be used as a supervisor to transfer knowledge information from the teacher model to the student model. Finally, the weight and bias information of the entire student model are transferred to the computing platform carried by the multi-rotor UAV airborne magnetic surveying system.

[0030] In this embodiment, step S2 includes:

[0031] S2.1 Design the total loss function ,in:

[0032] (Cross-Entropy Loss) is the standard cross-entropy loss of the student model on the hard labels of the training data;

[0033] ;

[0034] (Knowledge Distillation Loss) is the distillation loss used by the student model to mimic the soft label distribution of the teacher model.

[0035] ;

[0036] Where L is the total loss function, For weighted adjustment of hyperparameters, by To balance the two types of loss in knowledge distillation, so as to make the compensation effect of the student model closer to that of the teacher model. This is the true label of the i-th sample, and T is the distillation temperature parameter. This represents the output probability distribution of the teacher model at temperature T. The predictions of the student model at the same temperature.

[0037] and The calculation formula is:

[0038] ;

[0039] ;

[0040] in, This represents the original output value of the i-th data point in the teacher model before processing by the softmax activation function. This represents the raw output value of the j-th data point in the teacher model before processing by the softmax activation function. This represents the original output value of the i-th data point in the student model before processing by the softmax activation function. This represents the raw output value of the j-th data point in the student model before it is processed by the softmax activation function.

[0041] In this embodiment, the real-time magnetic interference compensation mentioned in step S3 is performed by the computing platform calling the student model to perform online compensation calculations on the data while the UAV collects aeromagnetic data.

[0042] To enable rapid processing on the computing platform mounted on the multi-rotor UAV, the weights and biases of the student model, obtained through knowledge distillation of the teacher model mentioned above, need to be transferred to the Raspberry Pi. A newly generated neural network on the Raspberry Pi will then be used to perform compensation operations. This allows the computing platform to perform high-precision, high-efficiency, real-time aeromagnetic data compensation after the multi-rotor UAV's aeromagnetic measurement platform collects data.

[0043] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made using the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A lightweight magnetic compensation neural network method for a multi-rotor unmanned aerial vehicle (UAV) aeronautical computing platform, characterized in that, The method includes: S1. Constructing a teacher model: A pre-trained large neural network is used as a teacher model. The teacher model is a deep neural network with multiple hidden layers or a backpropagation neural network, which is used to learn the magnetic interference features in aeromagnetic data. S2. Construct a student model and perform knowledge distillation: Build a lightweight three-layer feedforward fully connected neural network as the student model, and transfer the knowledge of the teacher model to the student model through knowledge distillation. S3. Deploy lightweight models to UAV platforms: Transfer the weights and biases of the trained student models to the computing platform carried by multi-rotor UAVs to achieve real-time magnetic interference compensation for aeromagnetic data. Step S2 includes: S2.1 Design the total loss function ,in: The standard cross-entropy loss for the student model on the hard labels of the training data: ； Distillation loss for student models to mimic the soft label distribution of teacher models: ； Where L is the total loss function, For weighted adjustment of hyperparameters, This is the true label of the i-th sample, and T is the distillation temperature parameter. This represents the output probability distribution of the teacher model at temperature T. For the student model's predictions at the same temperature; and The calculation formula is: ； ； in, This represents the original output value of the i-th data point in the teacher model before processing by the softmax activation function. This represents the raw output value of the j-th data point in the teacher model before processing by the softmax activation function. This represents the original output value of the i-th data point in the student model before processing by the softmax activation function. This represents the original output value of the j-th data point in the student model before it is processed by the softmax activation function. The real-time magnetic interference compensation mentioned in step S3 is performed by the computing platform calling the student model to perform online compensation calculations on the data while the UAV is collecting aeromagnetic data.

Images