Vehicle-mounted MEMS-IMU error compensation method based on MH-CNN
By constructing a parallel-serial hybrid architecture and feature fusion technology using MH-CNN, and designing a loss function, the problem of insufficient error compensation caused by the differences in error characteristics of each axis of the vehicle-mounted MEMS-IMU was solved, achieving high-precision error compensation and improving the accuracy of the navigation system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-24
- Publication Date
- 2026-04-10
AI Technical Summary
Existing deep learning methods fail to fully consider the differences in error characteristics of each axis of automotive MEMS-IMU, resulting in limited error compensation effects. In particular, the problem of error correction for different axes during vehicle operation has not been effectively solved.
A parallel-serial hybrid architecture is constructed using a multi-head convolutional neural network (MH-CNN). Through a multi-head feature extraction module, a feature fusion module, and an error compensation output module, the loss function is designed as the Huber loss function based on attitude increment, which achieves high-precision compensation for errors on each axis of the MEMS-IMU.
It effectively compensates for various errors of MEMS-IMU, such as zero bias, scale factor error, cross-axis coupling error, and random walk, thereby improving the navigation accuracy of the vehicle inertial navigation system.
Smart Images
Figure SMS_8 
Figure SMS_10 
Figure SMS_20
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of inertial navigation, and particularly relates to a vehicle-mounted micro-electromechanical inertial measurement unit (MEMS-IMU) error compensation method based on a multi-head convolutional neural network (MH-CNN). The method is based on the MH-CNN to compensate for three-axis scale factor errors, zero bias errors, and nonlinear noise of the MEMS-IMU, and belongs to the technical method. BACKGROUND
[0002] An inertial navigation system (INS) has the advantage of providing navigation information without interruption, but can only maintain high precision in the short term. Its error will accumulate over time and eventually diverge exponentially. The navigation accuracy of the INS mainly depends on the type and quality of the inertial measurement unit (IMU). High-precision IMUs represented by laser gyroscopes, fiber optic gyroscopes, and fiber optic accelerometers have small positioning errors but high costs, ranging from tens of thousands to hundreds of thousands of yuan, and rely on time-consuming initialization processes. With the development of micro-electromechanical system (MEMS) manufacturing processes, MEMS gyroscopes and accelerometers have gradually become small, low-power, and low-cost, and have been increasingly widely used in intelligent vehicles, drones, and robot navigation and positioning fields. However, due to manufacturing process reasons, the error characteristics of MEMS-IMU (hereinafter referred to as MIMU) are more complex, with poor regularity of zero bias, scale factor error, cross-axis coupling error, and white noise, making modeling difficult. Without external assistance, this can cause unbounded drift in the double integration of dead reckoning in the INS, and the positioning error will diverge exponentially with time, making the navigation information unusable within a few seconds. Therefore, modeling and compensating for MIMU errors is an effective method to improve the navigation accuracy of the INS. For MIMU, due to the complex and highly nonlinear error sources, traditional Kalman filter (KF)-based methods cannot effectively model all error sources. With the rapid development of data processing methods and machine learning technologies, deep learning methods based on neural networks have shown strong nonlinear mapping and fitting capabilities. The high-frequency available data in the INS also provides a basis for the application of deep learning in it. Therefore, it is necessary to use deep learning technology to supplement traditional error compensation methods to effectively reduce MIMU errors.
[0003] During vehicle driving, large angle changes in heading direction and small changes in horizontal attitude usually occur, which results in richer data information of the sky direction gyro and easier error correction, while the errors of the front and right direction gyros are more difficult to be perceived; the accelerometers also face the situation that the output ranges of the east and north directions are larger and the sky direction velocity changes little, which poses a challenge to the simultaneous correction of errors of different axis accelerometers; the existing deep learning methods such as OriNet use recurrent neural networks or ordinary convolutional neural networks for IMU noise reduction, but do not fully consider the differences in error characteristics of each axis of the MIMU, resulting in limited compensation effect. SUMMARY
[0004] The present application aims to at least solve one of the technical problems existing in the prior art.
[0005] To this end, the present application provides a vehicle-mounted MEMS-IMU error compensation method based on MH-CNN. The method of the present application models the error characteristics of each axis of the MIMU based on MH-CNN, and realizes high-precision compensation of multi-axis errors of the MIMU.
[0006] The technical solution of the present application is as follows:
[0007] According to one aspect, a vehicle-mounted MEMS-IMU error compensation method based on MH-CNN is provided, which comprises:
[0008] Step one, constructing a deep learning multi-head neural network model, wherein:
[0009] The multi-head neural network model adopts a parallel-serial hybrid architecture, including a multi-head feature extraction module, a feature fusion module and an error compensation output module, the gyro and accelerometer of each axis in the feature extraction module correspond to an independent convolutional subnetwork (head), which is used to extract error characteristics specific to the axis, each subnetwork adopts a dilated convolution structure to expand the receptive field and capture long-time dependent features; the feature fusion module is used to integrate multi-axis features; the error compensation module generates the final error compensation amount;
[0010] Step two, obtaining input data and pre-processing the input data, wherein the input data is the original angular velocity and acceleration sequence of the IMU;
[0011] Step three, designing a loss function, wherein the loss function is a Huber loss function based on attitude increment;
[0012] Step four, using the data obtained in step two as training data, training based on step one and step three, learning the nonlinear error of the MIMU based on network mixing, to realize error compensation of the gyro and accelerometer.
[0013] Further, the structure of each "head" in the feature extraction module is as follows:
[0014] layer number layer type kernel size dilation rate input channels output channels 1 dilated convolution 7 1 2 16 2 dilated convolution 7 4 16 32 3 dilated convolution 7 16 32 64 4 dilated convolution 7 64 64 128
[0015] Further, in the multi-head feature extraction module, after each layer of dilated convolution, there is a BatchNorm layer and a GELU activation function.
[0016] Further, the step two specifically includes:
[0017] Suppose the length of the original angular velocity and acceleration sequence of the IMU is N, the data of each axis is respectively processed by Z-Score normalization, including:
[0018] Calculate the mean value of each channel in the training set and the standard deviation :
[0019] For the original value of any sample on the jth channel , its normalized value is:
[0020] .
[0021] Further, the mean value of each channel in the training set and the standard deviation are calculated by the following formula:
[0022]
[0023] where j=1,…,6 represents the number of channels of the sensor, N is the total number of samples in the training set for this channel, is the original measurement value of the ith sample in the training set on the jth channel.
[0024] Further, in step three, the loss function is as follows:
[0025]
[0026] where is the true attitude increment from time i to time i+j, and are the rotation matrices at time i and time i+j respectively, is the angular velocity between time i and i+j, is the attitude increment estimation value (calculated by network output) from time i to time i+j, is the logarithmic mapping (Lie algebra) on SO(3), is the Huber loss function.
[0027] According to another aspect, there is provided a computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the method described above when executing the computer program.
[0028] The technical solution described above is based on the existing research on intelligent inertial navigation based on deep learning, and a model optimization is performed for a vehicle-mounted system. An error compensation method based on MH-CNN is proposed. A multi-head subnetwork and a feature fusion technology are used to design an optimized error compensation model for a vehicle-mounted MIMU. The three-axis error is effectively compensated. Through the design of the loss function, the CNN model constructed can effectively learn and compensate for multiple errors such as zero bias, scale factor error, cross-axis coupling error, and random walk, thereby improving the navigation accuracy of the low-cost IMU. DETAILED DESCRIPTION
[0029] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. The description of the at least one exemplary embodiment is actually only illustrative, but not as any limitation on the application and use of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0030] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof.
[0031] Unless specifically stated otherwise, the relative arrangement of components and steps, numerical expressions, and numerical values set forth in the examples herein are not meant to limit the scope of the present application. Techniques, methods, and devices known to those of ordinary skill in the relevant art can not be discussed in detail, but should be considered as part of the enabling disclosure where appropriate. In all examples shown and discussed herein, any specific value should be interpreted as merely exemplary, and not as a limitation. Therefore, other examples of the exemplary embodiments can have different values.
[0032] In one embodiment of the present application, a MH-CNN-based vehicle-mounted MEMS-IMU error compensation method is provided, which comprises:
[0033] Step one, model construction, including:
[0034] The multi-head neural network of this step adopts a parallel-serial hybrid architecture, including a multi-head feature extraction module, a feature fusion module, and an error compensation output module. The gyroscope and accelerometer of each axis in the feature extraction module correspond to an independent convolutional subnetwork (head) for extracting error features specific to the axis. Each subnetwork adopts a dilated convolution structure to expand the receptive field and capture long-term dependency features. The feature fusion module is used to integrate multi-axis features. The error compensation module generates the final error compensation amount.
[0035] The structure of each "head" in the feature extraction module is shown in Table 1.
[0036] Table 1 Subnetwork structure
[0037] layer number layer type kernel size dilation rate input channels output channels 1 dilated convolution 7 1 2 16 2 dilated convolution 7 4 16 32 3 dilated convolution 7 16 32 64 4 dilated convolution 7 64 64 128
[0038] After each layer of dilated convolution, there is a BatchNorm layer and a GELU activation function.
[0039] Step two, input data processing, including:
[0040] The input is the original angular velocity and acceleration sequence of the IMU, with a length of N. The data for each axis is normalized by Z-Score.
[0041] First, calculate the mean and standard deviation of each channel in the training set:
[0042]
[0043] where j = 1, …, 6 represents the number of channels of the sensor, N is the total number of samples in the training set for the jth channel, is the original measurement value of the ith sample in the training set on the jth channel.
[0044] Then, for any sample on the jth channel, the normalized value of the original value is:
[0045]
[0046] Step three, loss function design, including:
[0047] Use the Huber loss function based on pose increments:
[0048]
[0049] where is the true pose increment from time i to time i+j, and are rotation matrices at time i and time i+j respectively, is the angular velocity between time i and i+j, is the pose increment estimate from time i to time i+j (computed by the network output), is the logarithmic map (Lie algebra) on SO(3), is the Huber loss function.
[0050] Step four, training.
[0051] In the embodiment of the application, the ADAM optimizer and the cosine annealing learning rate are used, and the dropout layer and the weight decay strategy are used to prevent overfitting.
[0052] That is, the IMU data is high-frequency and has obvious local features, the error correction task has low demand for global information of sequence data, the error can be efficiently extracted through convolution operation, and the requirement for computing resources is lower, therefore, in the embodiment of the application, high-precision navigation information is used as training data, the nonlinear error of the MIMU is learned based on the CNN, the error compensation of the gyroscope and the accelerometer is realized, and thus the autonomous navigation precision is improved.
[0053] In addition, under the vehicle-mounted environment, the data diversity of the IMU is limited, and the error of part of the axes is difficult to perceive, for example, the change levels of the heading angle, the roll angle and the pitch angle during the driving of the vehicle are not in the same order of magnitude, and the error levels also present the multi-scale characteristics, and the loss function and the network structure based on the single CNN network model cannot simultaneously learn the multi-scale errors of different axes, therefore, in the embodiment of the application, the CNN model is optimized and improved, and the errors of different axes are simultaneously compensated based on the MH-CNN method.
[0054] It can be seen that, on the basis of the existing intelligent inertial navigation research based on deep learning, the model optimization is carried out for the vehicle-mounted system, an error compensation method based on the MH-CNN is proposed, the multi-head sub-network and the feature fusion technology are adopted, the optimized error compensation model for the vehicle-mounted MIMU is designed, the three-axis errors are effectively compensated, through the design of the loss function, the CNN model constructed can effectively learn and compensate various errors such as the zero bias, the scale factor error, the cross-axis coupling error and the random walk, and thus the navigation precision of the low-cost IMU is improved.
[0055] According to another embodiment, a computer device is provided, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of the above-mentioned embodiment when executing the computer program.
[0056] Features as described and / or illustrated above in relation to one embodiment can be used in the same or analogous manner in one or more other embodiments, and / or combined with or to replace features in other embodiments.
[0057] It should be emphasized that the term comprises / comprising, when used in this text, refers to the presence of the feature, whole, step or component, but does not exclude the presence or addition of one or more other features, wholes, steps, components or combinations thereof.
[0058] The above method of the present application can be realized by hardware, or by hardware in combination with software. The present application relates to a computer readable program which, when executed by a logic component, causes the logic component to realize the above-described apparatus or constituent components, or to realize the above-described various methods or steps. The present application also relates to a storage medium for storing the above program, such as a hard disk, a magnetic disk, an optical disk, a DVD, a flash memory, etc.
[0059] The many features and advantages of these embodiments are apparent from the detailed specification, and thus, it is intended by the appended claims to cover all such features and advantages of the embodiments within their true spirit and scope. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the embodiments of the present application to the exact construction and operation described and illustrated, and accordingly, all suitable modifications and equivalents can be resorted to, falling within the scope of the embodiments.
[0060] Portions of the present application that are not specifically described herein are well known to those skilled in the art.
Claims
1. A method for vehicle MEMS-IMU error compensation based on MH-CNN, characterized in that, The error compensation method comprises: Step one, constructing a deep learning multi-head neural network model, wherein: The multi-head neural network model adopts a parallel-serial hybrid architecture, comprising a multi-head feature extraction module, a feature fusion module and an error compensation output module, the feature extraction module corresponds to an independent convolutional subnetwork (head) for each axis of the gyroscope and accelerometer, which is used to extract error characteristics specific to the axis, each subnetwork adopts a dilated convolution structure to expand the receptive field and capture long-time dependent features; the feature fusion module is used to integrate multi-axis features; the error compensation module generates the final error compensation amount; Step two, obtaining input data and preprocessing the input data, wherein the input data is the original angular velocity and acceleration sequence of the IMU; Step three, designing a loss function, wherein the loss function is a Huber loss function based on the attitude increment; Step four, using the data obtained in step two as training data, training based on steps one and three, and learning the nonlinear error of the MIMU based on the network mixing to achieve error compensation for the gyroscope and accelerometer.
2. The MH-CNN-based vehicle-mounted MEMS-IMU error compensation method according to claim 1, characterized in that, The structure of each "head" in the feature extraction module is as follows:
3. The MH-CNN-based vehicle-mounted MEMS-IMU error compensation method according to claim 2, characterized in that, In the multi-head feature extraction module, after each layer of dilated convolution, there is a BatchNorm layer and a GELU activation function.
4. The MH-CNN-based vehicle-mounted MEMS-IMU error compensation method according to claim 2 or 3, characterized in that, The step two specifically comprises: Let the length of the original angular velocity and acceleration sequence of the IMU be N, and perform Z-Score normalization processing on the data of each axis respectively, including: Compute the mean μ and standard deviation σ for each channel in the training set j and standard deviation σ j : For any one sample, the original value x on the jth channel j The normalized value is:
5. The MH-CNN-based vehicle-mounted MEMS-IMU error compensation method according to claim 4, characterized in that, The mean μ and standard deviation σ of each channel in the training set are calculated by the following equations j j : where j = 1,..., 6 represents the channel number of the sensor, N is the total number of samples in the training set for this channel, x i,j is the original measurement of the i-th sample in the training set on the j-th channel.
6. The MH-CNN-based vehicle-mounted MEMS-IMU error compensation method according to claim 4 or 5, characterized in that, In step three, the loss function is as follows: L = L 16 + L 32 where δR i,i+j is the real pose increment from time i to time i + j, R i and R i+j are the rotation matrices at time i and time i + j, respectively, ω k is the angular velocity between time i and i + j, is the pose increment estimate from time i to time i + j (computed by the network output), log(·) is the logarithmic map on SO(3) (Lie algebra), and p(·) is the Huber loss function.
7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the method of claims 1-6 when executing the computer program.