Deep learning-based gyroscope noise reduction method
By adding random Gaussian noise to the gyroscope data and extracting features using deep learning networks, the problem of low data generalization in gyroscope noise reduction is solved, and more accurate noise suppression and error reduction are achieved.
Patent Information
- Application Number
- CN202410920440.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2024-12-10
AI Technical Summary
The data generalization of the existing gyroscope noise reduction algorithm is low, resulting in a large error between the measurement results and the real value, and it is impossible to effectively deal with multiple noise interferences.
Using the deep learning-based gyroscope noise reduction method, random Gaussian noise is added by preprocessing the gyroscope data set, and the Gyro_se_Net network is built using the RS module and the DC module to extract feature data and expand the receptive field to obtain multi-scale context information. , coupling correction is performed to reduce noise.
Improves the generalization of data and the robustness of feature extraction, provides more accurate data calibration, and reduces measurement errors.
Smart Images
Figure CN119106250A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning gyroscope noise reduction, and in particular to a gyroscope noise reduction method based on deep learning. Background Art
[0002] A gyroscope is a sensor used to measure angular velocity (i.e., rotation speed). It plays an important role in inertial navigation, attitude estimation, and motion control. However, gyroscope signals are often interfered by noise, which may have a negative impact on the application.
[0003] Gyroscope noise reduction aims to reduce or eliminate the noise in the gyroscope signal by adopting various technologies and methods, thereby improving the accuracy and stability of the measurement results. The noise of the gyroscope signal may come from multiple factors such as the sensor itself, environmental interference, and electromagnetic interference. Therefore, gyroscope noise reduction has always been a widely concerned issue in research and application at home and abroad. Foreign research institutions and academia have been committed to developing and improving gyroscope noise reduction algorithms, including common algorithms such as Kalman filtering, unscented Kalman filtering, particle filtering, and adaptive filtering. Domestic scientific research institutions and universities have also conducted a series of studies in the field of gyroscope noise reduction, exploring different filtering algorithms and sensor fusion strategies. The data generalization in the existing methods is low, so that the error between the measurement results and the true value is large. Summary of the invention
[0004] In view of this, the present invention provides a gyroscope noise reduction method based on deep learning to solve the above problems.
[0005] The present invention provides a gyroscope denoising method based on deep learning, comprising: collecting motion data of an object as a gyroscope data set through a gyroscope device; preprocessing the data in the gyroscope data set, wherein the preprocessing comprises adding random Gaussian noise; processing feature data extracted from the preprocessed gyroscope data set by stacking an RS module and a DC module to form a Gyro_se_Net network to obtain output features; wherein the RS module comprises a RES module and a SE module, the feature data respectively pass through the RES module and the SE module, and the result enters the RES module again for feature extraction through element-by-element multiplication, and after passing through a BN batch normalization layer, the result is added to the feature data to obtain the output features; the DC module is used to expand the receptive field of the output features and obtain multi-scale context information; the output features are coupled and corrected with the original data in the gyroscope data set to obtain denoised gyroscope data.
[0006] In another implementation of the present invention, it also includes: acquiring the motion trajectory of the object through a camera to obtain true value data; comparing the original data in the gyroscope data set and the gyroscope data after noise reduction with the true value data to obtain an error analysis result.
[0007] In another implementation of the present invention, the noise amplitude of the added random Gaussian noise is randomly generated according to Gaussian distribution, which is expressed as:
[0008]
[0009] Here, μ is the expected value of the distribution, σ is the standard deviation, and σ squared is the variance.
[0010] In another implementation of the present invention, the output feature is expressed as:
[0011] H(x)=F(x)+x
[0012] Among them, x is the input and F(x) is the feature obtained through the deep network.
[0013] In another implementation of the present invention, the calculation of the receptive field is expressed as:
[0014]
[0015] Among them, RF l+1 is the receptive field size corresponding to the current feature, RF l is the receptive field size corresponding to the previous layer feature, f l+1 is the convolution kernel size of the current convolution layer, and the last s multiplication term represents the stride product of the previous convolution layer.
[0016] The deep learning-based gyroscope denoising method of the present invention can effectively simulate external random interference and improve the generalization of data by preprocessing the collected gyroscope IMU data set and adding random Gaussian noise; through the Gyro_se_Net network, the robustness of IMU feature extraction is enhanced to provide more accurate data for subsequent calibration processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. By reading the detailed description of the following implementations, the advantages and benefits of the solutions become clear to those skilled in the art. The drawings are only used to illustrate the preferred implementations and are not to be considered as limiting the present invention.
[0018] In the attached picture:
[0019] Figure 1The present invention is a flowchart of a gyroscope noise reduction method based on deep learning according to an embodiment of the present invention.
[0020] Figure 2 The figure is a schematic diagram of a gyroscope noise reduction network flow chart according to an embodiment of the present invention.
[0021] Figure 3 A schematic diagram of the Gyro_se_Net network framework of an embodiment of the present invention.
[0022] Figure 4 Schematic diagram of the RS module framework of an embodiment of the present invention.
[0023] Figure 5 Schematic diagram of a DC module framework according to an embodiment of the present invention. DETAILED DESCRIPTION
[0024] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be described clearly and in detail below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in the field based on the embodiments in the embodiments of the present invention should fall within the scope of protection of the embodiments of the present invention.
[0025] Figure 1 A schematic flow chart of a gyroscope noise reduction method based on deep learning provided by an embodiment of the present invention is as follows: Figure 1 As shown, this embodiment mainly includes:
[0026] The motion data of the object is collected through a gyroscope device as a gyroscope data set.
[0027] Exemplarily, a gyroscope device is used to collect data such as acceleration and angular velocity of an object to produce a gyroscope IMU data set.
[0028] The data in the gyroscope data set is preprocessed, and the preprocessing includes adding random Gaussian noise.
[0029] Exemplarily, preprocessing the collected gyroscope IMU data set and adding random Gaussian noise can effectively simulate external random interference and improve the generalization of the data.
[0030] The Gyro_se_Net network is built by stacking RS modules and DC modules to process the feature data extracted from the preprocessed gyroscope data set to obtain output features.
[0031] For example, Figure 2 and Figure 3As shown in the figure, 3 layers of RS modules and 1 layer of DC modules are stacked to build the Gyro_se_Net network to extract features, and the coupling correction is made with the original IMU data to reduce the noise of the gyroscope. Since the RS module improves the representation ability of the network on different channels by learning adaptive channel weights, and integrates the characteristics of quick connections to prevent overfitting, it can effectively extract features and then input them into the DC module with expanded receptive field and multi-scale context information, enhancing the robustness of the neural network for feature extraction.
[0032] Among them, the RS module includes a RES module and a SE module. The feature data passes through the RES module and the SE module respectively, and the result enters the RES module again for feature extraction through element-by-element multiplication. After passing through the BN batch normalization layer, it is added to the feature data to obtain the output feature.
[0033] For example, Figure 4 As shown in the figure, the RS module is constructed by referring to the residual structure and SE (Squeeze-and-Excitation) attention mechanism. The processing process is that the feature data f passes through the RES module and the SE module respectively, and the result enters the RES module again for feature extraction through element-by-element multiplication. After passing through the BN (Batch Normalization) layer, it is added to the feature data f to obtain the output.
[0034] It should be understood that the SE module is a module used to enhance channel attention in deep learning. It aims to improve the representation ability of the network on different channels by learning adaptive channel weights. Its core idea is to establish the correlation between channels through two simple operations, namely compression and excitation.
[0035] Compression: Perform a global average pooling operation on the input feature map to convert the feature map of each channel into a scalar value. This scalar value can be regarded as a global information summary of the channel. Formally, the statistic z is generated by reducing the spatial dimension of u H×W, so that the cth element of z is calculated by the following formula:
[0036]
[0037] Where u is the output result of the upper feature map, H and W are the length and width of the feature map.
[0038] Excitation: The correlation between channels is learned by introducing two fully connected (or convolutional) layers. First, the compressed feature vector is input into a fully connected layer (or convolutional layer) and activated using a nonlinear activation function (such as ReLU). Then, it is input into another fully connected layer (or convolutional layer) to reduce its dimension to a small fraction of the original number of channels. Finally, the sigmoid function is used to limit the output between 0 and 1 to obtain the channel attention weight. The formula of the excitation module is as follows:
[0039] s=F ex (z,V)=σ(g(z,V))=σ(V2δ(V1z))
[0040] Where δ(·) represents the ReLU activation function, σ(·) represents the sigmoid activation function, and V is the weight parameter, which is a learnable parameter matrix used to adjust the expressiveness of the feature.
[0041] Reweighting: Multiply the learned channel attention weights by the original feature maps to enhance useful feature channels and suppress redundant and noisy channels.
[0042]
[0043] In the formula, F scale () represents a scaling function used to scale the feature map according to the channel attention weight, s represents the channel attention weight vector, and u represents the input feature map.
[0044] It should be understood that the RS module uses the idea of "quick connection" to solve the problems of gradient vanishing and gradient exploding in deep neural network training by adding additional connections. By introducing shortcut connections, the RS module allows the output of a certain layer to directly skip one or more layers and connect to the input of the subsequent layer. Even if some layers do not make any meaningful transformations, the information of the previous layers can still be passed without causing too much loss in the gradient.
[0045] The DC module is used to expand the receptive field of the output feature and obtain multi-scale context information.
[0046] For example, Figure 5 As shown in the figure, a DC (Dilated convolution) module is constructed based on the dilated convolution to expand the receptive field and obtain multi-scale context information. Compared with the traditional convolution operation, the dilated convolution introduces a dilation rate parameter to control the interval between each element in the convolution kernel. In the traditional convolution operation, each element of the convolution kernel is multiplied and summed with the element at the corresponding position in the input feature map to generate an element in the output feature map. In the dilated convolution, the elements of the convolution kernel are no longer adjacent to each other, but are expanded by increasing the interval.
[0047] The dilation rate is represented by a positive integer, and common dilation rates are 1, 2, 3, etc. When the dilation rate is 1, it is a normal convolution operation. When the dilation rate is greater than 1, there will be gaps between the elements of the convolution kernel, which can increase the size of the receptive field and capture a wider range of context information. By increasing the dilation rate, the receptive field of the network can be increased without increasing parameters and computation, thereby improving the model's ability to model large-size objects or context information.
[0048] The output feature is coupled with the original data in the gyroscope data set to obtain gyroscope data after noise reduction.
[0049] Exemplarily, the IMU data tested by the gyroscope denoising network whose backbone network is Gyro_se_Net is closer to the real value data.
[0050] The deep learning-based gyroscope denoising method of the present invention can effectively simulate external random interference and improve the generalization of data by preprocessing the collected gyroscope IMU data set and adding random Gaussian noise; through the Gyro_se_Net network, the robustness of IMU feature extraction is enhanced to provide more accurate data for subsequent calibration processing.
[0051] In another implementation of the present invention, it also includes: acquiring the motion trajectory of the object through a camera to obtain true value data; comparing the original data in the gyroscope data set and the gyroscope data after noise reduction with the true value data to obtain an error analysis result.
[0052] In another implementation of the present invention, the noise amplitude of the added random Gaussian noise is randomly generated according to Gaussian distribution, which is expressed as:
[0053]
[0054] Here, μ is the expected value of the distribution, σ is the standard deviation, and σ squared is the variance.
[0055] For example, by adding noise to the gyroscope IMU data to simulate random interference, the effect of data enhancement is achieved. When adding Gaussian white noise, the amplitude distribution obeys the Gaussian distribution, and the probability spectrum distribution obeys the uniform distribution. Its characteristic is that at each sampling point in the noise signal, the amplitude of the noise is randomly generated according to the Gaussian distribution.
[0056] In another implementation of the present invention, the output feature is expressed as:
[0057] H(x)=F(x)+x
[0058] Among them, x is the input and F(x) is the feature obtained through the deep network.
[0059] In another implementation of the present invention, the calculation of the receptive field is expressed as:
[0060]
[0061] Among them, RF l+1 is the receptive field size corresponding to the current feature, RF l is the receptive field size corresponding to the previous layer feature, f l+1 is the convolution kernel size of the current convolution layer, and the last s multiplication term represents the stride product of the previous convolution layer.
[0062] Exemplarily, a DC module is constructed based on dilated convolution, which mainly includes 5 layers of dilated convolution with dilation rates of 1, 4, 8, 32, and 1, respectively. It can effectively expand the receptive field and is conducive to obtaining multi-scale contextual information.
[0063] Deep learning uses neural networks with multiple hidden layers to learn and represent data. Compared with traditional shallow neural networks, deep learning models perform better in handling complex signal processing tasks. Deep learning models can better capture the abstract features and structure of signals by extracting high-level features of input signals layer by layer. Compared with traditional manual feature extraction methods, deep learning can automatically learn more discriminative and representative feature representations. This enables deep learning to better understand the statistical laws and patterns of signals in signal processing, and improve the accuracy and robustness of signal processing. Deep learning models usually require a large amount of training data, but they also have the ability to process large-scale data. In the field of signal processing, by using a large amount of training data, deep learning models can more accurately learn the statistical laws and patterns of signals.
[0064] Thus far, specific embodiments of the present invention have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired results. Additionally, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing may be advantageous.
[0065] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.
[0066] In the description of the present invention, the terms "first" and "second" are only used to facilitate the description of different components or names, and cannot be understood as indicating or implying a sequential relationship, relative importance, or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.
[0067] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.
[0068] It should be noted that although the specific embodiments of the present invention are described in detail in conjunction with the accompanying drawings, it should not be understood as limiting the scope of protection of the present invention. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still belong to the scope of protection of the present invention.
[0069] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be improper limitations of the embodiments of the present invention.
[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A gyroscope noise reduction method based on deep learning, characterized in that: include: Collecting the motion data of the object as a gyroscope data set through a gyroscope device; Preprocessing the data in the gyroscope data set, the preprocessing comprising adding random Gaussian noise; The Gyro_se_Net network is built by stacking RS modules and DC modules to process the feature data extracted from the preprocessed gyroscope data set to obtain output features; The RS module includes a RES module and a SE module. The feature data passes through the RES module and the SE module respectively. The result enters the RES module again for feature extraction through element-by-element multiplication. After passing through the BN batch normalization layer, it is added to the feature data to obtain the output feature. The DC module is used to expand the receptive field of the output feature and obtain multi-scale context information; The output feature is coupled with the original data in the gyroscope data set to obtain gyroscope data after noise reduction.
2. The method according to claim 1, characterized in that Also includes: Acquire the motion trajectory of the object through a camera to obtain true value data; The original data in the gyroscope data set and the denoised gyroscope data are respectively compared with the true value data to obtain an error analysis result.
3. The method according to claim 1, characterized in that The noise amplitude of the added random Gaussian noise is randomly generated according to Gaussian distribution, which is expressed as: Here, μ is the expected value of the distribution, σ is the standard deviation, and σ squared is the variance.
4. The method according to claim 1, characterized in that: The output features are expressed as: H(x)=F(x)+x Among them, x is the input and F(x) is the feature obtained through the deep network.
5. The method according to claim 1, characterized in that The calculation of the receptive field is expressed as: Among them, RF l+1 is the receptive field size corresponding to the current feature, RF l is the receptive field size corresponding to the previous layer feature, f l+1 is the convolution kernel size of the current convolution layer, and the last s multiplication term represents the stride product of the previous convolution layer.