Method and device for extracting and suppressing magnetic noise of unmanned ground vehicle wheel

Abstract

The application provides a method and device for extracting and suppressing magnetic noise of unmanned ground vehicle wheels, which solves the technical problem that it is difficult to accurately extract and effectively suppress dynamic magnetic noise generated by the rotation of unmanned ground vehicle wheels in real time. It includes collecting data information of the unmanned ground vehicle, preprocessing, and constructing a training data set; based on the training data set, a Refinement U-Mamba model is constructed and trained to obtain the weight data of the trained model, which is loaded into the Refinement U-Mamba model to obtain the trained Refinement U-Mamba model; the real-time collected data information is input into the trained Refinement U-Mamba model synchronously, and the denoising result is output in real time. The application can be widely applied in the field of geomagnetic environment perception technology.

Description

Technical Field

[0001] This application belongs to the field of geomagnetic environment sensing technology, and relates to a method and device for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles. Background Technology

[0002] With the widespread application of unmanned ground vehicles in fields such as geological exploration, target detection, and navigation and positioning, data acquisition technology based on magnetic sensors has become a research hotspot. However, in practical applications, the unmanned ground vehicle itself is a complex source of magnetic interference. During the vehicle's operation, the rotation of the wheels, the operation of the motor, and the vibration of metal components generate complex magnetic field interference signals. These interferences are often large in amplitude and wide in bandwidth, and are highly coupled with the target's magnetic field signal.

[0003] In existing technologies, hardware compensation or traditional digital signal processing methods (such as filtering and principal component analysis) are commonly used to suppress interference. However, hardware compensation methods require extremely high equipment installation precision and are difficult to adapt to dynamic changes in the vehicle's motion state; traditional filtering methods often struggle to distinguish between interference signals with overlapping spectral characteristics and useful signals, easily leading to distortion or loss of effective signals. Furthermore, most existing neural network-based methods rely on conventional convolutional neural networks (CNNs) or recurrent neural networks (RNNs), which, while improving denoising performance to some extent, often fail to effectively capture global dependencies in long sequences of magnetic field data, and their computational efficiency is insufficient for vehicle-mounted detection scenarios with high real-time requirements.

[0004] Therefore, how to construct a magnetic field data processing method that can efficiently learn the complex interference characteristics of unmanned ground vehicles and remove interference such as wheels in real time and with high precision has become a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a method and apparatus for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles, so as to solve the technical problem in the prior art that it is difficult to accurately extract and effectively suppress the dynamic magnetic noise generated by the rotation of the wheels of unmanned ground vehicles in real time.

[0006] To achieve the above objectives, the first embodiment of this application provides a method for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles, comprising the following steps: Collect data from unmanned ground vehicles, preprocess it, and construct a training dataset; Based on the training dataset, a Refinement U-Mamba model is constructed and trained. The weight data after model training is obtained and loaded into the Refinement U-Mamba model to obtain the trained Refinement U-Mamba model. The real-time collected data is synchronously input into the trained Refinement U-Mamba model, and the denoising results are output in real time.

[0007] Preferably, the Refinement U-Mamba model includes a dual-channel input module, an encoder, a bottleneck layer, a decoder, and an output module; The dual-channel input module is used to extract dual-channel features from the training dataset to obtain the main signal features and the guiding signal features; The encoder is used to extract features from the main signal features, refine and gate based on the guide signal features, and pass the jump connection features to the decoder layer by layer through jump connections; Bottleneck layer, used for deep feature transformation between encoder and decoder; The decoder is used to fuse skip connection features and output the fused features. The output module maps the fused features to a noise estimate, subtracts the noise estimate from the training dataset, and outputs the denoising result.

[0008] Preferably, the encoder consists of four cascaded GGRT blocks. In each GGRT block, the guiding signal features are processed and enhanced by a CNN module to obtain enhanced internal guiding features. Based on the enhanced internal guiding path features, the main signal features are input into an improved Mamba-Transformer module. Through Mamba component feature extraction, gated attention incremental guidance, and residual connection, the intermediate features of the main path are obtained.

[0009] Preferably, the improved Mamba-Transformer module includes a Mamba component and a gated attention increment submodule; The gated attention increment submodule is used to receive enhanced internal guiding features as query and gate signal sources, and to receive intermediate features of the main path as keys and values. After performing layer normalization and dimension rearrangement on the query, key, and value, the rearranged internal guidance path features are obtained. A gating signal is generated through linear transformation, and the gating value is obtained after passing through the Sigmoid activation function. The gating value is then applied element by element to the main signal features to obtain the gated main features. After normalization, the gated main features are obtained after layer normalization. The rearranged internal guidance path features and the gated main features after layer normalization are input into a multi-head attention mechanism to obtain the attention output. After regularization, the attention contribution is obtained. After adding the attention contribution to the intermediate features of the main path, the result is normalized and then input into the multilayer perceptron to obtain the MLP contribution. This contribution is then added to the attention contribution to obtain the total increment.

[0010] Preferably, the data information of the unmanned ground vehicle includes raw magnetic field data, raw angle data, clean wheel noise data, reference signal, background magnetic field signal, and the relationship between clean wheel noise data and raw angle data.

[0011] Preferably, the preprocessing process includes: unwinding the original angle data to obtain the total rotation angle of the wheel, performing a cosine transform to obtain a smooth periodic waveform whose frequency is proportional to the wheel speed, and obtaining a guide signal; The noisy signal is obtained by randomly selecting equal-length segments from the reference signal and the clean round-noise data and linearly adding them point by point.

[0012] Preferably, the process of constructing the training dataset includes: combining the noisy signal, the corresponding guiding signal, and the reference signal to form training samples; The training dataset is obtained by combining the original magnetic field data, the clean wheel noise data, the background magnetic field signal, and the relationship between the clean wheel noise data and the original angle data.

[0013] Preferably, the process of obtaining weight data includes: inputting the training dataset into the Refinement U-Mamba model for training, using the mean square error between the estimated noise value and the clean round-robin noise data as the loss function, optimizing the network parameters until convergence, and obtaining the weight data after model training.

[0014] Preferably, the formula for the loss function is: ; In the formula, For loss function, For the first An estimated value of the noise. For the first A clean, noise-generating data set. This is the signal length.

[0015] The second embodiment of this application provides a device for extracting and suppressing magnetic noise from the wheels of an unmanned ground vehicle, comprising: an unmanned ground vehicle, a first magnetometer inside the unmanned ground vehicle, an absolute rotary encoder on the wheel axle of the unmanned ground vehicle, a mast on one side of the unmanned ground vehicle, and a second main magnetometer at the top of the mast.

[0016] The method and apparatus for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles provided in this application have the following advantages compared with the prior art: This application provides a method and apparatus for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles. By collecting data information from unmanned ground vehicles and preprocessing it to construct a training dataset, reliable data support is provided for introducing guidance signals that are highly correlated with the noise source. It can achieve accurate noise extraction based on the kinematic state information of the noise source, and significantly improve the extraction reliability under conditions where the signal and interference characteristics are similar and the spectrum is mixed.

[0017] Next, a Refinement U-Mamba model was constructed and trained based on the training dataset. This model integrates the multi-scale feature extraction capabilities of the U-Net model with the long-term temporal dependency modeling capabilities of Mamba, forming an efficient processing architecture suitable for complex geophysical signals and improving the depth and computational efficiency of feature learning. Simultaneously, the Guided-Gated Refined Mamba-Transformer (GGRT) block and the Gated Attention Increment (GAD) module were used to dynamically regulate the main signal processing process using guiding information, further enhancing the extraction accuracy of wheel magnetic noise components and enabling efficient adaptation to nonlinear and time-varying interference scenarios caused by wheel rotation.

[0018] Finally, the real-time collected data is synchronously input into the trained Refinement U-Mamba model, and the denoising results are output. This enables real-time extraction and suppression of wheel magnetic noise on unmanned ground vehicle platforms without the need for physical avoidance methods such as sensor elevation. It is more in line with the trend of miniaturization of unmanned equipment, and while simplifying hardware deployment, it significantly improves the purity and measurement accuracy of magnetic field data. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Figure 1 A flowchart illustrating the overall process for extracting and suppressing magnetic noise from the wheels of an unmanned ground vehicle, as provided in an embodiment of this application. Figure 2 A schematic diagram of the structure of a device for extracting and suppressing magnetic noise from the wheels of an unmanned ground vehicle provided in an embodiment of this application; Figure 3 This is a schematic diagram of the Refinement U-Mamba model provided in one embodiment of this application; Figure 4 A schematic diagram of the structure of a boot-gated refining Mamba-TransformerGGRT block provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a gated attention increment (GAD) provided in an embodiment of this application; Figure 6A flowchart of a noise reduction method for unmanned ground vehicles provided in an embodiment of this application.

[0020] In the diagram: 1. First magnetometer; 2. Absolute rotary encoder; 3. Mast; 4. Second main magnetometer. Detailed Implementation

[0021] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0022] Example 1, please refer to Figure 1 This is the first embodiment of the present application, which provides a method for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles, including: It should be noted that the embodiments of this application accurately identify and separate the fundamental frequency interference caused by wheel rotation from the measurement data of the UGV (Unmanned Ground Vehicle) on-board magnetometer through a guided deep learning noise extraction model, thereby obtaining a high-purity background magnetic field signal, providing high-quality input data for subsequent geomagnetic anomaly interpretation or TL compensation (carrier magnetic interference compensation) and other processing procedures.

[0023] The overall work process is as follows Figure 1 As shown, the specific implementation method is as follows: S1: Collect data information from unmanned ground vehicles, including raw magnetic field data, raw angle data, clean wheel noise data, reference signal, background magnetic field signal, and the relationship between clean wheel noise data and raw angle data.

[0024] It should be noted that you can refer to [link / reference]. Figure 2 First, main sensors are installed on the UGA (Unmanned Ground Vehicle), including a set of first magnetometers (one scalar magnetometer and one vector magnetometer) and an absolute rotary encoder. Specifically, a first main magnetometer is installed at a target location inside the UGA to collect raw magnetic field data containing various interferences. An absolute rotary encoder is installed on the wheel closest to the first magnetometer inside the cabin to obtain the wheel's precise rotation angle in real time. The scalar magnetometer and vector magnetometer in the first magnetometer are positioned opposite each other and firmly installed at the target location inside the UGA to ensure no relative displacement occurs during UGA operation. Simultaneously, the rotary encoder is installed on the wheel axle or hub of the wheel closest to the first magnetometer. This encoder is an absolute rotary encoder with power-off retention to ensure a clear correspondence between its output raw angle data and the wheel's physical rotation state.

[0025] Next, reference sensors were installed on the UGA (Unmanned Ground Vehicle) to collect relatively clean background magnetic field signals for creating the model's training dataset. The reference sensors consisted of a pair of second magnetometers of the same model as the first magnetometer, mounted on a non-magnetic mast. The reference sensors were kept at a distance of more than 1.5 meters from the main ferromagnetic structure of the UGA to ensure that the data collected by the reference sensors could represent the background magnetic field to the greatest extent possible, with minimal interference from the vehicle's own magnetic field.

[0026] Then, the UGA (Unmanned Ground Vehicle) was placed in an area with a stable magnetic field environment, and the relationship between the magnetic field strength data of the second magnetometer (i.e., the pure wheel noise data) and the raw angle data of the rotary encoder was recorded to create a training dataset for the model.

[0027] Specifically, clean wheel noise data was collected in relatively stable indoor or outdoor environments. To simulate the synchronous rotation of multiple wheels during real-world driving, a non-magnetic synchronous belt was used to connect the front and rear axles of the UGA (Unmanned Ground Vehicle). The vehicle's power system was then activated, causing the wheels to spin in place. During this time, a unified data acquisition module synchronously recorded readings from a second magnetometer (clean wheel noise data) and raw angle data from an absolute rotary encoder at a sampling rate of at least 125Hz. Data synchronization employed a high-precision hardware timestamp mechanism to ensure a strict time correspondence between each set of clean wheel noise data and raw angle data.

[0028] Finally, the data collected by the reference sensor in the actual experiment and after rigorous filtering and TL compensation was used as the reference signal.

[0029] Specifically, the reference signal originates from data collected by a reference sensor mounted on the mast in the actual experiment. First, a low-pass Butterworth filter with a cutoff frequency of 1Hz is applied to filter out high-frequency electronic noise from the reference sensor itself and other high-frequency electromagnetic interference from the environment. Then, standard TL compensation (carrier magnetic interference compensation) is performed on the filtered data to eliminate the weak but still present platform interference caused by attitude changes (maneuvers) during UGA (Unmanned Ground Vehicle) operation. The signal data after these two steps is considered representative of the real geomagnetic environment, free from wheel interference and platform maneuvering interference, and can be used as a true reference signal.

[0030] S2: Preprocess the data information of unmanned ground vehicles to build a training dataset.

[0031] The preprocessing of data from unmanned ground vehicles includes: first, unwinding the collected raw wheel rotation angle data to obtain the total wheel rotation angle. Because the angle of an absolute rotary encoder jumps back to 0 degrees after reaching 360 degrees, the unwinding algorithm detects this discontinuous jump and accumulates a 2. The integer multiples of the angle are converted into a continuously monotonically increasing sequence of angles to truly reflect the total rotation of the wheel.

[0032] Then, the total rotation angle of the wheel after unwinding. Perform cosine transform and calculate This transformation converts the linearly increasing angle signal into a smooth, periodic waveform within the range of [-1, 1], with a frequency proportional to the wheel speed. This form is more easily recognized by neural networks as a stable pattern feature and can be used as a guiding signal.

[0033] Finally, the dataset is synthesized. A segment of 1440 data points is randomly extracted from the acquired reference signal, and a segment of the same length is also randomly extracted from the acquired clean, noisy data. These two segments are then linearly added point by point to generate a synthesized noisy signal.

[0034] The noisy signal, its corresponding absolute value rotary encoder guide signal, and the reference signal serving as the ground truth together form a training sample. This process is repeated, and together with the original magnetic field data, the clean wheel noise data, the background magnetic field signal, and the relationship between the clean wheel noise data and the original angle data, it forms the training dataset used as the model.

[0035] S3: Construct a Refinement U-Mamba model. Input the training dataset into the Refinement U-Mamba model for training. Obtain the weight data after model training and load it into the Refinement U-Mamba model to obtain the trained Refinement U-Mamba model.

[0036] It should be noted that you can refer to [link / reference]. Figure 3 This application provides a wheel magnetic noise extraction model (i.e., the Refinement U-Mamba model), which is a guided dual-channel deep learning architecture whose core is to accurately extract the wheel noise component of the main signal. The model adopts the classic U-Net structure and mainly consists of five parts: a dual-channel input module, an encoder, a bottleneck layer, a decoder, and an output module.

[0037] Specifically, the dual-channel input module includes a two-channel one-dimensional convolutional layer (Conv1D), a layer normalization layer (LN), and a GELU activation function. (Noisy signal (main input)) With the guide signal (guide input) Initial feature embedding is performed in parallel through independent 1D convolutional layers. This is the signal length. The kernel size of this convolutional layer is 7, which will convert the noisy signal... and guidance signal Embedded into a 32-dimensional feature space, the data is sequentially input into a one-dimensional convolutional layer (Conv1D), a layer normalization layer (LN), and a GELU activation function to obtain the main signal features. and guiding signal characteristics .

[0038] The encoder consists of four cascaded Guided-Gated Refining Mamba-Transformer (GGRT) blocks, responsible for extracting features from the main signal features and refining and gating them using information from the guide signal features. The outputs of each layer of the encoder are passed to the decoder via skip connections, passing the skip connection features to the skip connection after the skip connection.

[0039] Specifically, in each GGRT block, the main signal characteristics and guiding signal characteristics It is processed collaboratively. The guiding feature modulates the extraction process of the main feature through a gated attention increment (GAD) block, whose output increment... It was added back to the main path, enabling guided refinement: ; In the formula, This represents the output characteristics of the main path at the current stage.

[0040] The bottleneck layer performs deep feature transformation between the encoder and decoder, and it consists of a Mamba module.

[0041] The decoder also consists of four cascaded Mamba-Transformer blocks, which gradually reconstruct the target noise signal by fusing the skip connection features from the encoder.

[0042] Finally, the output module maps the fused features from the decoder output to a noise estimate—that is, an estimate of the noise component in the network's predicted noisy signal—using a one-dimensional convolutional layer (Conv1D) with a kernel size of 7, a layer normalization layer (LN), and a GELU activation function. The noise estimate is then subtracted from the original noisy signal to obtain the final denoising result.

[0043] Among them, such as Figure 4As shown, the Guided-Gated Refinement Mamba-Transformer (GGRT) block is the core processing unit for refining the U-Mamba network encoder. The design goal of this module is to integrate the guiding signal into Mamba's powerful sequence modeling capabilities, thereby achieving precise control and continuous refinement of the main signal feature extraction. Each GGRT block receives main signal features from the previous stage. and guiding signal characteristics As input, B is the batch size, C is the feature dimension (configured to 32), and L is the sequence length. The feature dimension is not limited here and can be chosen according to the actual situation.

[0044] The processing flow within the GGRT block begins with the boot path. Characteristics of the input boot signals... First, the internal guided features are processed and enhanced using a CNN module. This module consists of a 1D convolutional layer with a kernel size of 3, layer normalization (LayerNorm), and a GELU activation function. After processing, the enhanced internal guided features are obtained. : ; Meanwhile, main signal characteristics The data is fed into an improved Mamba-Transformer module for processing. This module contains two Mamba components and a Gated Attention Delta (GAD) submodule, which enhances the internal guided features. The query (Q) signal serves as the attention mechanism. All improved Mamba-Transformer modules in the network are configured with a state dimension of 8, a convolutional kernel dimension of 2, and a spread factor of 2. The processing flow is as follows, with residual connections applied at each step: ; In the formula, The intermediate features of the main path To output the increment, It is the sum of intermediate features and output increments. and These are two standard Mamba components. This refers to the output characteristics of the main path at the current stage.

[0045] Furthermore, such as Figure 5 As shown, the Gated Attention Delta (GAD) submodule receives enhanced internal guiding features. As a source of both query (Q) and gated signals, and to receive intermediate features from the main path (i.e., in GGRT) The GAD module serves as the source of both the key (K) and value (V). Its function is to calculate an incremental (Delta) signal. The signal is then added back to the main path, thereby adjusting its feature representation based on the guidance information.

[0046] Furthermore, the computation process of the GAD module begins with layer normalization (LayerNorm) of the input query stream and key / value stream. First, the rearranged internal guiding features are used. (Enhanced internal guidance features) The result after layer normalization and dimension rearrangement is used to generate a gated signal through a linear transformation. The gate value is then obtained by passing it through the Sigmoid activation function σ. The formula is: ; These gate values The main signal features are applied element-wise to the unnormalized main signal features to obtain the gated main features. Then, the gated main features Normalization is performed using the following formula: ; In the formula, For element-wise multiplication, The input GAD module contains the original main features, which are the intermediate features of the main path within the GGRT module. These are the gated main features after layer normalization.

[0047] Next, the rearranged internal guiding features and gated main features after layer normalization The input is fed into a multi-head attention mechanism with four attention heads to obtain the attention output. After Dropout (regularization) processing, the attention contribution is formed. : ; Subsequently, attention contribution Intermediate features added to the main path The summation result, after layer normalization, is fed into a multilayer perceptron (MLP). This MLP consists of two linear layers: the first linear layer expands the feature dimension by a factor of 2, followed by GELU activation and Dropout (regularization); the second linear layer projects it back to the original dimension, followed by another Dropout (regularization) process. This process generates the MLP contribution. : ; In the formula, Input signal for the multilayer sensor The input signal to the multilayer perceptron is after layer normalization. This is the output of the multilayer perceptron.

[0048] Finally, contribute attention and MLP contributions Add them together to get the total increment. : ; The total increment This is the final output of the GAD module.

[0049] After constructing the Refinement U-Mamba model, the training dataset is input into the Refinement U-Mamba model to train the model. The mean square error between the noise estimate and the clean wheel noise data is used as the loss function to optimize the network parameters until convergence. The weight data of the Refinement U-Mamba model after training is saved and loaded into the Refinement U-Mamba model of the vehicle data processing unit.

[0050] Specifically, during the model training phase, the model receives synthesized noisy signals. and guidance signal Its output is an estimate of the noise. This estimate is compared with the clean round-noise data. The mean squared error (MSE) between the two sides was used as the loss function. To optimize network parameters. The formula for calculating the loss function is: ; By minimizing this loss function, the model is trained to accurately separate the noise component that matches the guiding signal pattern from the noisy signal.

[0051] Furthermore, after model training is complete, the magnetometer mounted on the external mast of the UGV (Unmanned Ground Vehicle) is removed from the UGV. This operation aims to restore the UGV to its most simplified, operationally suitable hardware configuration, retaining only the main sensors within the cabin. On the software side, the generated model weight data is loaded onto the Refinement U-Mamba model deployed in the onboard data processing unit.

[0052] S4: The raw magnetic field data collected in real time and the preprocessed guidance signal are synchronously input into the trained Refinement U-Mamba model, and the magnetic field data after removing wheel interference is output in real time, that is, the denoising result.

[0053] It should be noted that during the actual surveying mission performed by the UGV (Unmanned Ground Vehicle), the data acquisition system will collect raw magnetic field data from the first magnetometer inside the cabin and raw angle data from the absolute value rotary encoder in the wheels in real time and synchronously. The collected raw angle data is first unwound and cosine transformed to generate a real-time guidance signal. Subsequently, a fixed length of real-time raw magnetic field data (e.g., 1440 data points) is used as the main input, and the guidance signal of the corresponding time period is used as the guidance input, both fed into a weighted Refinement U-Mamba network. The network performs one forward propagation calculation, and its output is the wheel noise estimate separated from the raw magnetic field data of that segment. Finally, the noise estimate output by the network is subtracted from the input raw magnetic field data to obtain the high-purity magnetic field data with wheel interference removed for that time period, i.e., the denoising result. This process is continuously executed throughout the mission in a sliding window manner.

[0054] It should be noted that the original magnetic field data is the magnetometer sampling data after being processed by LPF (Low-Pass Filter) and DC removal; the encoder guide signal is the original angle data of the absolute value rotary encoder in the wheel that has been synchronized, corresponds to the magnetometer sampling point, and has been unwound.

[0055] Example 2, please refer to Figure 2 This is the second embodiment of the present application. This embodiment provides a device for extracting and suppressing magnetic noise from the wheels of an unmanned ground vehicle, including: an unmanned ground vehicle (UGV), a first magnetometer 1 installed inside the UGV, an absolute value rotary encoder 2 installed on the wheel axle of the UGV, a mast 3 installed on one side of the UGV, and a second main magnetometer 4 installed at the top of the mast 3.

[0056] Example 3: One-time noise removal instance.

[0057] Please see Figure 6 Step 1: Install the main sensors, including a first magnetometer 1 and an absolute rotary encoder 2. The first magnetometer is installed at the target location inside the cabin of the unmanned ground vehicle (UGV) to collect raw magnetic field data containing various interferences. An absolute rotary encoder is installed on the axle of the wheel closest to the main magnetometer inside the cabin to obtain the precise rotation angle of that wheel in real time.

[0058] Step 2: Install a reference sensor to acquire a relatively clean background magnetic field signal.

[0059] Step 3: Collect pure wheel noise. Place the unmanned ground vehicle (UGV) in an area with a stable magnetic field environment and record the relationship between the reading of the main magnetometer inside the cabin (which can be regarded as pure vehicle noise data) and the angle data of the absolute value rotary encoder.

[0060] Step 4: Obtain a clean reference magnetometer data that has been rigorously filtered and compensated and acquired through a reference sensor in actual experiments, as the reference signal.

[0061] Step 5: Data preprocessing and synthesis. The collected angle data is unwound and subjected to cosine transform to generate a periodic guiding signal. A segment of the reference signal is randomly extracted, and simultaneously, an equal-length segment is randomly obtained from the clean, noisy data. The two segments are linearly added point-by-point to obtain the noisy signal. This process is repeated to form the training dataset.

[0062] Step 6: Model training and weight generation. Using the designed Refinement U-Mamba model, load the generated training dataset for training to obtain model weight data.

[0063] Step 7: Remove the reference sensor, keeping only the main sensor. Load the trained model weight data into the model.

[0064] Step 8: Use the real-time acquired raw magnetic field data inside the cabin as the main input to the model, and use the synchronously acquired and preprocessed guidance signal as the guidance input to the model. The model will output the magnetic field data after removing wheel interference in real time to obtain the denoising result.

[0065] Specifically, the absolute rotary encoder in step 1 of this implementation scheme should have the functions of manually calibrating the zero point and saving the position after power failure.

[0066] Specifically, in step 2 of this implementation plan, the mounting mast of the reference sensor should be made of non-ferromagnetic material and be at least 2 meters away from the main magnetic noise source of the vehicle body.

[0067] Specifically, in step 3 of this implementation plan, during the process of collecting pure wheel noise, the wheels should be controlled to rotate at different speeds, both uniformly and variablely, to obtain corresponding data under different rotation modes.

[0068] Specifically, the clean data in step 4 of this implementation plan is geomagnetic data with certain fluctuations after preprocessing filtering, TL compensation, and DC removal.

[0069] Specifically, the process of generating the periodic guidance signal and training dataset in step 5 of this implementation scheme is as follows: The original data is unwound, and the unwound unidirectional angle signal is low-pass filtered with a cutoff frequency of 0.5Hz. The angle signal after low-pass filtering is then cosine transformed to form the final cosine signal.

[0070] From the magnetometer data obtained in step 4, 1440 consecutive data points from a random starting point are selected as the reference signal.

[0071] In the noise data obtained in step 3, 1440 consecutive data points from a random starting point are selected as noise signals; then, 1440 consecutive data points from the same starting point are selected from the periodic guidance signals as guidance signals.

[0072] The reference signal and the noise signal are added together to obtain the noisy signal. The noisy signal, the reference signal, the noise signal, and the guiding signal form a training dataset. The above process is repeated to generate the training dataset.

[0073] Specifically, in this implementation scheme, the training parameters in step 6 are as follows: the ratio of training set to test set is 8:2; the batch size is 16; the learning rate is 1e-3; and the epoch is 500.

[0074] Specifically, in this implementation scheme, the weights in step 7 are the weight data obtained in step 6; all hyperparameters of the model are completely consistent with the model in step 6. At the same time, it is necessary to ensure that after removing the reference sensor, all other hardware and structural configurations remain completely unchanged.

[0075] Specifically, in step 8 of this implementation scheme, the original magnetic field data is the magnetometer sampling data after being processed by LPF (Low-Pass Filter) and DC removal; the encoder guide signal is the synchronized wheel encoder data stream (angle data) corresponding to the magnetometer sampling point and after unwinding processing.

[0076] It should be noted that the above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles, characterized in that, Includes the following steps: Collect data from unmanned ground vehicles, preprocess it, and construct a training dataset; Based on the training dataset, a Refinement U-Mamba model is constructed and trained. The weight data after model training is obtained and loaded into the Refinement U-Mamba model to obtain the trained Refinement U-Mamba model. The real-time collected data is synchronously input into the trained Refinement U-Mamba model, and the denoising results are output in real time.

2. The method for extracting and suppressing magnetic noise from unmanned ground vehicle wheels as described in claim 1, characterized in that, The Refinement U-Mamba model includes a dual-channel input module, an encoder, a bottleneck layer, a decoder, and an output module; The dual-channel input module is used to perform dual-channel feature extraction on the training dataset to obtain main signal features and guiding signal features; The encoder is used to extract features from the main signal features, refine and gate based on the guiding signal features, and pass the jump connection features to the decoder layer by layer through jump connections; The bottleneck layer is used to perform deep feature transformation between the encoder and the decoder; The decoder is used to fuse the skip connection features and output the fused features; The output module is used to map the fused features to a noise estimate, subtract the noise estimate from the training dataset, and output the denoising result.

3. The method for extracting and suppressing magnetic noise from unmanned ground vehicle wheels as described in claim 2, characterized in that, The encoder consists of four cascaded GGRT blocks. In each GGRT block, the guiding signal features are processed and enhanced by a CNN module to obtain enhanced internal guiding features. Based on the enhanced internal guidance path features, the main signal features are input into the improved Mamba-Transformer module. Through Mamba component feature extraction, gated attention incremental guidance, and residual connection, the intermediate features of the main path are obtained.

4. The method for extracting and suppressing magnetic noise from unmanned ground vehicle wheels as described in claim 3, characterized in that, The improved Mamba-Transformer module includes Mamba components and a gated attention increment submodule; The gated attention increment submodule is used to receive the enhanced internal guidance features as a query and gated signal source, and to receive the intermediate features of the main path as keys and values. After performing layer normalization and dimension rearrangement on the query, key, and value, the rearranged internal guidance path features are obtained. A gated signal is generated through linear transformation, and a gated value is obtained after passing through the Sigmoid activation function. This value is then applied element-wise to the main signal features to obtain the gated main features. After normalization, the gated main features are obtained after layer normalization. The rearranged internal guidance path features and the gated main features after layer normalization are input into a multi-head attention mechanism to obtain the attention output. After regularization, the attention contribution is obtained. The attention contribution is added to the intermediate features of the main path, and after layer normalization, it is input into a multilayer perceptron to obtain the MLP contribution. The total increment is obtained by adding the attention contribution to the MLP contribution.

5. The method for extracting and suppressing magnetic noise from unmanned ground vehicle wheels as described in claim 1, characterized in that, The data information of the unmanned ground vehicle includes raw magnetic field data, raw angle data, clean wheel noise data, reference signal, background magnetic field signal, and the relationship between the clean wheel noise data and the raw angle data.

6. The method for extracting and suppressing magnetic noise from unmanned ground vehicle wheels as described in claim 5, characterized in that, The preprocessing process includes: unwinding the original angle data to obtain the total rotation angle of the wheel, performing a cosine transform to obtain a smooth periodic waveform whose frequency is proportional to the wheel speed, and obtaining a guide signal; The noisy signal is obtained by randomly selecting equal-length segments from the reference signal and the clean round-noise data and linearly adding them point by point.

7. The method for extracting and suppressing magnetic noise from unmanned ground vehicle wheels as described in claim 6, characterized in that, The process of constructing the training dataset includes: combining the noisy signal, the corresponding guiding signal, and the reference signal to form training samples; The training dataset is obtained by combining the original magnetic field data, the clean wheel noise data, the background magnetic field signal, and the relationship between the clean wheel noise data and the original angle data.

8. The method for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles as described in claim 5, characterized in that, The process of obtaining the weight data includes: inputting the training dataset into the Refinement U-Mamba model for training, using the mean square error between the noise estimate and the clean round-robin noise data as the loss function, optimizing the network parameters until convergence, and obtaining the weight data after model training.

9. The method for extracting and suppressing magnetic noise from the wheels of unmanned ground vehicles as described in claim 8, characterized in that, The formula for the loss function is: ； In the formula, For loss function, For the first An estimated value of the noise. For the first A clean, noise-generating data set. This is the signal length.

10. A method for extracting and suppressing magnetic noise from the wheels of an unmanned ground vehicle, applied to the system for extracting and suppressing magnetic noise from the wheels of an unmanned ground vehicle as described in any one of claims 1-9, comprising an unmanned ground vehicle, characterized in that, The unmanned ground vehicle is equipped with a first magnetometer inside, and an absolute rotary encoder is installed on the wheel axle of the unmanned ground vehicle. A mast is installed on one side of the unmanned ground vehicle, and a second main magnetometer is installed on the top of the mast.

Images