Lightweight magnetic compensation neural network method for aerial computing platform of multi-rotor unmanned aerial vehicle

Abstract

The invention is suitable for the field of aeromagnetic measurement, and provides a lightweight magnetic compensation neural network method for a multi-rotor unmanned aerial vehicle aerial computing platform, and the method comprises the steps: constructing a teacher model: pre-training a large neural network as the teacher model, the teacher model being a multi-hidden-layer deep neural network or a back-propagation neural network, the learning module is used for learning magnetic interference characteristics in aeromagnetic data; building a student model and executing knowledge distillation; building a lightweight three-layer feed-forward full-connection neural network as the student model, and migrating knowledge of the teacher model to the student model through knowledge distillation; and deploying a lightweight model to an unmanned aerial vehicle platform: migrating the weight and bias of the trained student model to a computing platform carried by the multi-rotor unmanned aerial vehicle, and realizing real-time magnetic interference compensation of aeromagnetic data. According to the invention, the structure of the whole neural network can be lightened, and the compensation calculation of the aeromagnetic survey data can be completed with low cost and high efficiency under the condition of limited calculation resources.

Description

Technical Field

[0001] The present invention relates to the field of aeromagnetic measurement, and in particular to a lightweight magnetic compensation neural network method for a multi-rotor unmanned aerial vehicle (UAV) aviation computing platform. Background Art

[0002] Aeromagnetic 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 drones (UAVs) as exploration platforms has made aeromagnetic surveying a more viable data measurement technique in terms of data collection efficiency and quality. However, aeromagnetic data contains both geomagnetic information and magnetic interference generated by the aircraft. Eliminating magnetic interference is one method for improving the accuracy of aeromagnetic data. An effective aeromagnetic compensation method will further enhance the accuracy of measurement data. Existing high-precision methods for compensating for magnetic interference utilize neural networks to learn the interference and apply the learned biases and weights to compensate for it. Commonly used neural networks for aeromagnetic interference compensation include back-propagation neural networks (BPNNs) and deep neural networks (DNNs). However, while these neural networks offer good compensation for aeromagnetic interference, they require significant computational resources due to their complex structures, including numerous hidden layers and neurons. The computing resources of multi-rotor UAV platforms are limited. For real-time compensation calculations, large-scale and computationally complex neural networks are expensive to run and slow to execute. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of the present invention is to propose a lightweight magnetic compensation neural network method for a multi-rotor UAV aviation computing platform to solve the technical problems mentioned in the background technology.

[0004] A lightweight magnetic compensation neural network method for a multi-rotor UAV aviation computing platform, the method comprising: S1. Build a teacher model: pre-train a large neural network as a teacher model. The teacher model is a multi-hidden layer deep neural network (DNN) or back propagation neural network (BPNN) for learning magnetic interference characteristics in aeromagnetic data. S2. Build a student model (a small neural network with only input, hidden, and output layers) 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 the lightweight model to the drone platform: Migrate the weights and biases of the trained student model to the embedded computing platform (such as Raspberry Pi) on the multi-rotor drone to achieve real-time magnetic interference compensation of aeromagnetic data.

[0005] Preferably, step S2 includes: S2.1. Design of total loss function ,in: (Cross-Entropy Loss) is the standard cross-entropy loss of the student model for the hard labels of the training data; ; (Knowledge Distillation Loss) is the distillation loss of the student model imitating the soft label distribution of the teacher model: ; Among them, L is the total loss function, Tune hyperparameters for weights, is the true label of the i-th sample, T is the distillation temperature parameter, is the output probability distribution of the teacher model at temperature T, Predictions for the student model at the same temperature.

[0006] Preferably, and The calculation formula is: ; ; in, is the original output value of the i-th data before the teacher model is processed by the softmax activation function, is the original output value of the j-th data before the teacher model is processed by the softmax activation function, is the original output value of the i-th data before the student model is processed by the softmax activation function. It is the original output value of the j-th data before the student model is processed by the softmax activation function.

[0007] Preferably, the real-time magnetic interference compensation in step S3 is performed by the computing platform calling the student model to perform online compensation calculation on the data while the drone collects the aeromagnetic data.

[0008] Beneficial effects achieved by the present invention: Compared with large neural networks that perform compensation calculations directly on computing devices on a multi-rotor UAV aeromagnetic survey platform, the present invention can lightweight the structure of the entire neural network and achieve low-cost and high-efficiency compensation calculations for aeromagnetic survey data under limited computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Schematic diagram of the knowledge distillation principle in an embodiment of the present invention.

[0010] Figure 2 This is a knowledge distillation structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0012] See also Figures 1 to 2 , an embodiment of the present invention provides a lightweight magnetic compensation neural network method for a multi-rotor UAV aviation computing platform, the method comprising: S1. Build a teacher model: pre-train a large neural network as a teacher model. The teacher model is a multi-hidden layer deep neural network (DNN) or back propagation neural network (BPNN) for learning magnetic interference characteristics in aeromagnetic data. S2. Build a student model (a small neural network with only input, hidden, and output layers) 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 the lightweight model to the drone platform: Migrate the weights and biases of the trained student model to the embedded computing platform (such as Raspberry Pi) on the multi-rotor drone to achieve real-time magnetic interference compensation of aeromagnetic data.

[0013] The neural network used as the teacher model in this invention is much larger than the student model. The key idea is to train an efficient student model under the guidance of the teacher model to achieve comparable accuracy. The supervisory signal from the teacher model (often referred to as the "knowledge" learned by the teacher model) helps the student model imitate the teacher's behavior. The predictions of the model learned through neural network learning contain weight information and serve as a supervisor to transfer knowledge from the teacher model to the student model. Finally, the weights and bias information of the entire student model are migrated to the computing platform onboard the multirotor drone aeromagnetic survey.

[0014] In this embodiment, step S2 includes: S2.1. Design of total loss function ,in: (Cross-Entropy Loss) is the standard cross-entropy loss of the student model for the hard labels of the training data; ; (Knowledge Distillation Loss) is the distillation loss of the student model imitating the soft label distribution of the teacher model: ; Among them, L is the total loss function, Tune hyperparameters for weights by To balance the two losses in knowledge distillation, so that the compensation effect of the student model is closer to the compensation effect of the teacher model, is the true label of the i-th sample, T is the distillation temperature parameter, is the output probability distribution of the teacher model at temperature T, Predictions for the student model at the same temperature.

[0015] and The calculation formula is: ; ; in, is the original output value of the i-th data before the teacher model is processed by the softmax activation function, is the original output value of the j-th data before the teacher model is processed by the softmax activation function, is the original output value of the i-th data before the student model is processed by the softmax activation function. It is the original output value of the j-th data before the student model is processed by the softmax activation function.

[0016] In this embodiment, the real-time magnetic interference compensation in step S3 is performed by the computing platform calling the student model to perform online compensation calculation on the data while the drone collects the aeromagnetic data.

[0017] To enable rapid processing on the multirotor drone's onboard computing platform, the weights and biases of the student model, learned from the teacher model through knowledge distillation, are transferred to the Raspberry Pi and compensated using the newly generated neural network on the Pi. This allows the computing platform to perform high-precision, efficient, and real-time aeromagnetic data compensation after the multirotor drone's aeromagnetic platform collects data.

[0018] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the description of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A lightweight magnetic compensation neural network method for a multi-rotor UAV aviation computing platform, characterized in that: The method comprises: S1. Build a teacher model: pre-train a large neural network as a teacher model. The teacher model is a deep neural network with multiple hidden layers or a back-propagation neural network, which is used to learn the magnetic interference characteristics in the aeromagnetic data. S2. Build 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 the lightweight model to the UAV platform: Migrate the weights and biases of the trained student model to the computing platform on the multi-rotor UAV to achieve real-time magnetic interference compensation for aeromagnetic data.

2. The lightweight magnetic compensation neural network method for the multi-rotor UAV aviation computing platform according to claim 1 is characterized in that: Step S2 includes: S2.

1. Design of total loss function ,in: The standard cross entropy loss for the student model on the hard labels of the training data is: ； Distillation loss for the student model to imitate the soft label distribution of the teacher model: ； Among them, L is the total loss function, Tune hyperparameters for weights, is the true label of the i-th sample, T is the distillation temperature parameter, is the output probability distribution of the teacher model at temperature T, Predictions for the student model at the same temperature.

3. The lightweight magnetic compensation neural network method for the multi-rotor UAV aviation computing platform according to claim 2 is characterized in that: and The calculation formula is: ； ； in, is the original output value of the i-th data before the teacher model is processed by the softmax activation function, is the original output value of the j-th data before the teacher model is processed by the softmax activation function, is the original output value of the i-th data before the student model is processed by the softmax activation function. It is the original output value of the j-th data before the student model is processed by the softmax activation function.

4. The lightweight magnetic compensation neural network method for a multi-rotor UAV aviation computing platform according to claim 1 is characterized in that: The real-time magnetic interference compensation in step S3 is that while the UAV is collecting aeromagnetic data, the computing platform calls the student model to perform online compensation calculations on the data.

Images