An IMU gyroscope noise reduction method and system based on deep learning
By using a lightweight dilated convolutional neural network and a learnable gyroscope intrinsic parameter calibration matrix, the problem of noise and bias separation in low-cost IMU gyroscopes is solved, achieving high-precision attitude estimation, which is applicable to fields such as robotics, drones, and autonomous driving.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING SHENDAOKEXUN SCI TECH DEV CO LTD
- Filing Date
- 2026-06-15
- Publication Date
- 2026-07-24
AI Technical Summary
The angular velocity output by the gyroscope of a low-cost MEMS inertial measurement unit (IMU) contains significant noise and bias, which leads to serious attitude estimation drift. Traditional methods are difficult to separate noise and bias in real time, deep learning methods have a large number of parameters and slow inference speed, and vision-assisted solutions have low reliability in harsh environments.
A lightweight dilated convolutional neural network is used in conjunction with a learnable gyroscope intrinsic parameter calibration matrix. The temporal correlation features between the gyroscope and the accelerometer are extracted through the dilated convolutional neural network, and an attitude increment loss function is constructed to achieve end-to-end optimization. Accelerometer auxiliary features are used to constrain kinematic consistency for gyroscope correction and attitude estimation.
Achieving sub-degree-level pose accuracy without visual dependence, significantly suppressing short-term dynamic deviations and long-term cumulative drift, reducing computational complexity, and suitable for real-time operation on embedded platforms.
Smart Images
Figure CN122448253A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of inertial navigation and deep learning technology; in particular, it relates to a deep learning-based IMU gyroscope noise reduction method and system, which is applicable to high-precision attitude estimation of low-cost MEMS inertial measurement units in fields such as robotics, drones, and autonomous driving. Background Technology
[0002] The essence of three-dimensional attitude estimation is to deduce the spatial orientation of the carrier by measuring the angular velocity obtained by a continuous integrating gyroscope. However, the angular velocity measured by a low-cost MEMS inertial measurement unit (IMU) is mixed with significant noise and bias, and direct integration will cause serious attitude estimation drift in a short period of time.
[0003] The main challenges currently facing the technology can be summarized into five points: First, noise and bias are difficult to separate in real time. The gyroscope output simultaneously contains the true angular velocity, a slowly changing quasi-static bias, and high-frequency random noise. Traditional offline analysis methods cannot distinguish these three components online, leading to a continuous accumulation of errors after integration.
[0004] Secondly, traditional filtering methods have trade-offs. Low-pass or high-pass filtering introduces phase delay while suppressing noise; optimal estimation methods such as Kalman filtering require precise knowledge of the noise statistics and involve large computational loads, making them difficult to run in real time on embedded platforms.
[0005] Third, deep learning models are too bloated. Although denoising methods based on recurrent neural networks can capture temporal dependencies, they have a large number of model parameters, long training cycles, high data requirements, and the serial computation method limits the actual inference speed.
[0006] Fourth, the loss function design is not precise enough. Existing methods often directly use mean square error to optimize gyroscope correction values, but there is no linear correspondence between attitude estimation error and gyroscope correction, so a dedicated optimization target that can directly reflect attitude accuracy is needed.
[0007] Fifth, visual-assisted solutions are scene-dependent. While solutions that fuse camera and inertial data can achieve high accuracy, their performance drops sharply in environments such as low light, rapid motion, or lack of texture, and they are also computationally expensive. Therefore, high-precision pose estimation based solely on inertial data without relying on vision is of great value.
[0008] Existing technological approaches mainly fall into four categories, but each has its own obvious limitations: The wavelet transform denoising approach separates noise and drift through time-frequency analysis, supplemented by neural network compensation. Its drawbacks include the empirical dependence of wavelet parameter selection, processing delays, and limited generalization ability of the neural network.
[0009] Recurrent neural network (RNN) denoising approaches directly learn the noise distribution using time-series models. However, they suffer from drawbacks such as bloated models, computational complexity, slow inference speed, large training data requirements, and gradient decay issues with long sequences.
[0010] Meta-learning adaptive approaches employ multi-stage network structures to achieve small-sample domain transfer. Their drawbacks include complex network structures, the need to jointly optimize multiple loss functions, and high engineering deployment difficulty.
[0011] The vision-inertial fusion approach improves accuracy through multi-sensor fusion. Its drawbacks include heavy reliance on visual input, low reliability in harsh visual environments, and high overall computational cost. Summary of the Invention
[0012] Based on the above analysis, this invention aims to disclose a deep learning-based IMU gyroscope noise reduction method and system, achieving high-precision attitude estimation relying solely on IMU signals. Specifically, the technical problems to be solved by this invention include: difficulty in effectively separating gyroscope noise from bias, phase delay in traditional filtering methods, excessively large parameter size and slow inference speed in existing deep learning methods, lack of a dedicated loss function for attitude increments, and the dependence of vision-inertial fusion methods on visual sensors.
[0013] This invention discloses a deep learning-based IMU gyroscope noise reduction method, comprising: S1. Acquire raw IMU measurement data and obtain the true attitude increment. The raw IMU measurement data includes raw gyroscope measurement values and raw accelerometer measurement values. S2. Input the historical measurement data in the IMU raw measurement data into the dilated convolutional neural network, and the dilated convolutional neural network jointly extracts the time-series correlation features between the gyroscope and the accelerometer, and outputs the time-varying correction value of the gyroscope; S3. Establish a gyroscope calibration model; the model uses a learnable gyroscope intrinsic parameter calibration matrix to linearly transform the original gyroscope measurement values, and superimposes the time-varying correction values to obtain the corrected gyroscope data; S4. Perform open-loop integration on the corrected gyroscope data to calculate the estimated attitude increment; S5, based on The group logarithmic mapping calculates the rotation angle error between the estimated attitude increment and the true attitude increment. An attitude increment loss function is constructed based on the rotation angle error. The network parameters of the dilated convolutional neural network and the gyroscope intrinsic parameter calibration matrix are jointly optimized end-to-end based on the attitude increment loss function.
[0014] Furthermore, the attitude increment loss function is constructed based on the Huber loss function; the time scale is... Loss items at time Represented as: ; in, From time At the time The actual attitude increment, From time At the time The transpose of the estimated attitude increment; for Logarithmic mapping; For vector norm, Here is the Huber loss function. Indicates the first Summing at each initial time.
[0015] Furthermore, the attitude increment loss function includes at least two loss terms with different time scales. The size of the time scale of each loss term is determined according to the ratio between the output frequency of the true attitude increment and the IMU sampling frequency, so that the calculation frequency of the attitude increment error at each time scale is on the same order of magnitude as the output frequency of the true attitude increment. Moreover, each loss term covers historical IMU measurement data of different lengths to simultaneously constrain short-term dynamic error and long-term cumulative drift.
[0016] Furthermore, the dilated convolutional neural network adopts an M-layer dilated convolutional structure. The first M-1 layers are configured with an exponentially increasing dilation rate, and the kernel size is matched with the local temporal feature scale determined according to the IMU raw measurement data, so that the temporal receptive field coverage matches the historical IMU measurement data for a period of time that meets the attitude estimation accuracy requirements. The Mth layer adopts a convolutional structure with a kernel size of 1, which is used to map the feature channels into three-axis gyroscope correction values for output.
[0017] Furthermore, the first layer of the dilated convolutional neural network uses six-dimensional IMU measurements as joint input, including three-axis gyroscope measurements and three-axis accelerometer measurements; the first M-1 layers of the dilated convolutional neural network fuse the auxiliary features and gyroscope temporal features layer by layer by sharing dilated convolutional kernels, and the Mth layer maps the fused features to three-axis gyroscope time-varying correction values.
[0018] Furthermore, before the end-to-end joint optimization, the gyroscope intrinsic parameter calibration matrix is initialized to an identity matrix, and the gyroscope time-varying correction value is initialized to a zero vector. This makes the corrected gyroscope data in the early stage of training close to the original gyroscope measurement value, so as to constrain the attitude estimation to evolve along the open-loop integral trajectory of the original measurement value in the early stage of training, and avoid the attitude estimation divergence caused by over-correction in the early stage of the network.
[0019] Furthermore, during the training phase of the end-to-end joint optimization, the open-loop integral employs a tree-structured batch attitude integral algorithm. The matrix multiplication by rotation is organized into a binary tree structure, and the time complexity is reduced from [previous value] to [new value] through parallel computation. Reduce to .
[0020] Furthermore, in the gyroscope calibration model, the linear transformation is achieved by performing matrix multiplication between the gyroscope intrinsic parameter calibration matrix and the original gyroscope measurement value, which is used to compensate for axis deviation and scaling factor error; the time-varying gyroscope correction value is output by the dilated convolutional neural network and added to the linear transformation result to obtain the corrected gyroscope data.
[0021] Furthermore, the gyroscope intrinsic parameter calibration matrix is decomposed into the product of the scaling factor matrix and the axis deviation matrix; The scaling factor matrix and the axis deviation matrix are constrained by diagonal dominance or orthogonality constraints during training, which ensures that the optimization results of the gyroscope intrinsic parameter calibration matrix remain physically interpretable and avoids the intrinsic parameter matrix deviating from the physical meaning of the inertial sensor calibration due to end-to-end joint training.
[0022] Another aspect of this invention discloses a deep learning-based IMU gyroscope noise reduction system, comprising: an IMU chip and a processor; wherein, An IMU chip is used to acquire raw IMU measurement data; the raw IMU measurement data includes raw gyroscope measurement values and raw accelerometer measurement values. The processor, connected to the IMU chip, is used to process the raw measurement data of the IMU using the gyroscope correction model trained by the deep learning-based IMU gyroscope noise reduction method described above, and output the corrected gyroscope data and estimated attitude increment.
[0023] Compared with traditional methods, the present invention has the following technical advantages: This invention replaces traditional recurrent neural networks with lightweight dilated convolutional neural networks, employing fully parallel high-speed inference to reduce network parameters, significantly shorten training time for single datasets, and substantially lower the deployment threshold. A large receptive field of up to 2.24 seconds is constructed through a progressive dilation rate, effectively capturing long-distance temporal dependencies with a lightweight structure. Learnable intrinsic parameter calibration is combined with time-varying correction values in the network output to achieve synchronous adaptive compensation for IMU inherent errors and dynamic drift, with the initial network output approaching zero to ensure training stability. A rotation-group-based attitude increment loss function directly constrains attitude estimation accuracy, and a multi-timescale design simultaneously suppresses short-term dynamic bias and long-term cumulative drift. Accelerometer auxiliary features are integrated, utilizing the physical constraint relationship between acceleration change and angular velocity to ensure the correction output conforms to real kinematic laws, thereby achieving sub-degree attitude accuracy under purely inertial conditions without visual dependence. Attached Figure Description
[0024] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of the deep learning-based IMU gyroscope noise reduction method in an embodiment of the present invention; Figure 2 This is a schematic diagram of a dilated convolutional neural network with a five-layer dilated convolutional structure in an embodiment of the present invention; Figure 3 This is a comparison diagram of relative attitude error on the test sequence in an embodiment of the present invention; Figure 4 This is a comparison diagram of the roll, pitch, and yaw angle estimation results for two sequences in an embodiment of the present invention. The two sequences are EuRoC MH 04 difficult and TUM-VI room 4; Figure 5 This is a schematic diagram of the gyroscope correction signal transformation on the TUM-VI room 4 sequence in an embodiment of the present invention. Detailed Implementation
[0025] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.
[0026] Example 1 One embodiment of the present invention discloses a deep learning-based IMU gyroscope noise reduction method, such as... Figure 1 As shown, it includes the following steps: S1. Acquire raw IMU measurement data and obtain the true attitude increment. The raw IMU measurement data includes raw gyroscope measurement values and raw accelerometer measurement values. S2. Input the historical measurement data in the IMU raw measurement data into the dilated convolutional neural network, and the dilated convolutional neural network jointly extracts the time-series correlation features between the gyroscope and the accelerometer, and outputs the time-varying correction value of the gyroscope; S3. Establish a gyroscope calibration model; the model uses a learnable gyroscope intrinsic parameter calibration matrix to linearly transform the original gyroscope measurement values, and superimposes the time-varying correction values to obtain the corrected gyroscope data; S4. Perform open-loop integration on the corrected gyroscope data to calculate the estimated attitude increment; S5. Calculate the rotation angle error between the estimated attitude increment and the true attitude increment based on the SO(3) group logarithmic mapping, construct the attitude increment loss function based on the rotation angle error, and perform end-to-end joint optimization of the network parameters of the dilated convolutional neural network and the gyroscope intrinsic parameter calibration matrix based on the attitude increment loss function.
[0027] Specifically, in S1, raw IMU data (three-axis gyroscope, three-axis accelerometer) and visual inertial odometry ground truth (VIO ground truth) are collected; the true attitude increment is obtained using the VIO ground truth. The data is split into a training set (the first 50 seconds of each sequence), a validation set (the remaining parts), and a test set (independent sequences). The data is then normalized to make it suitable for neural network input.
[0028] The IMU used in this embodiment is a low-cost MEMS IMU, such as ADIS16448 (EuRoC dataset) or BMI160 (TUM-VI dataset). Specifically, in S2, the dilated convolutional neural network receives six-dimensional historical measurement data from the IMU's raw measurement data, including raw gyroscope measurements and raw accelerometer measurements. It extracts auxiliary features from the raw accelerometer measurements using the physical constraint relationship between acceleration changes and angular velocities at adjacent time points. These auxiliary features reflect the variation pattern of accelerometer measurements caused by changes in angular velocity. The dilated convolutional neural network then fuses these auxiliary features with gyroscope temporal features to output the gyroscope time-varying correction value, thereby constraining the kinematic consistency between the gyroscope time-varying correction value and the raw accelerometer measurements.
[0029] The dilated convolutional neural network adopts an M-layer dilated convolutional structure. The first M-1 layers are configured with an exponentially increasing dilation rate and are matched with the kernel size determined according to the local temporal feature scale of the IMU raw measurement data, so that the temporal receptive field coverage matches the historical IMU measurement data for a period of time that meets the requirements of attitude estimation accuracy. The Mth layer adopts a convolutional structure with a kernel size of 1, which is used to map the feature channels into three-axis gyroscope correction values for output.
[0030] The first layer of the dilated convolutional neural network uses six-dimensional IMU measurements as joint inputs, including three-axis gyroscope measurements and three-axis accelerometer measurements. The first M-1 layers of the dilated convolutional neural network fuse the auxiliary features and gyroscope temporal features layer by layer by sharing dilated convolutional kernels. The Mth layer maps the fused features to three-axis gyroscope time-varying correction values.
[0031] like Figure 2 As shown, M is 5, that is, a five-layer dilated convolutional structure is adopted, with only 77,052 parameters, to achieve lightweight and efficient inference.
[0032] The specific configuration is shown in Table 1:
[0033] Each convolutional layer is followed by batch normalization and the GELU activation function. The receptive field calculation formula is: ; in, The kernel size is the convolution kernel size. For the first Layer expansion rate. At a sampling rate of 200Hz, 448 sampling points correspond to 2.24 seconds of historical information.
[0034] The first four layers of this network structure are configured with exponentially increasing dilation rates (1, 4, 16, 64), and are paired with a convolutional kernel size of 7, ensuring that the temporal receptive field covers 2.24 seconds of historical IMU measurement data, matching the accuracy requirements for attitude estimation. The fifth layer uses a convolutional structure with a kernel size of 1 to map 128 feature channels to three-axis gyroscope correction values as output. The entire dilated convolutional neural network has only 77,052 parameters, less than 100,000, achieving lightweight and efficient inference.
[0035] Network through formula Calculate the gyroscope correction value; in, Functions defined for dilated convolutional neural networks; Indicates from time At the time IMU measurement sequences; each Includes data from a three-axis gyroscope and a three-axis accelerometer; For the length of the receptive field; The sampling interval is... The sampling point number; This is the output time-varying correction value for the three-axis gyroscope.
[0036] The advantage of this model is that: The network output is initialized to be close to zero to ensure that the estimated value is close to the original measurement value in the early stage of training. The intrinsic parameter calibration matrix is jointly optimized with the neural network to adapt to specific IMU characteristics.
[0037] Specifically, in S3, in the established gyroscope calibration model, the linear transformation is achieved by performing matrix multiplication between the gyroscope intrinsic parameter calibration matrix and the original gyroscope measurement value, which is used to compensate for axis deviation and scaling factor error; the time-varying gyroscope correction value is output by the dilated convolutional neural network and added to the linear transformation result to obtain the corrected gyroscope data.
[0038] The formula is expressed as follows: ; in, For the first The gyroscope data after secondary correction; This is a learnable gyro intrinsic parameter calibration matrix; For the first The original gyroscope measurement value; For the first The time-varying correction amount; The learnable gyro intrinsic parameter calibration matrix is decomposed into the product of the scale factor matrix and the axis deviation matrix, expressed as: ; in, This is a scaling factor matrix; This is the axis deviation matrix.
[0039] During training, the scaling factor matrix and axis deviation matrix are constrained by diagonal dominance or orthogonality constraints, which ensures that the optimization results of the gyroscope intrinsic parameter calibration matrix remain physically interpretable and avoids the intrinsic parameter matrix deviating from the physical meaning of the inertial sensor calibration due to end-to-end joint training.
[0040] No. The time-varying correction amount Represented as: ; in, Output by neural network This is a static bias.
[0041] Before end-to-end joint optimization, the gyroscope intrinsic parameter calibration matrix is... Initialize to identity matrix , gyroscope time-varying correction value Initializing the vector to zero ensures that the corrected gyroscope data in the early stages of training is close to the original gyroscope measurements. This constrains the attitude estimation to evolve along the open-loop integral trajectory of the original measurements in the early stages of training, thus avoiding over-correction in the early stages of the network that could lead to attitude estimation divergence.
[0042] In this embodiment, accelerometer information is used to assist gyroscope correction. When the velocity changes slowly, the acceleration difference between adjacent moments satisfies: ; This relationship indicates that acceleration changes also contain angular velocity information. By jointly processing gyroscope and acceleration data, the neural network can more accurately estimate the gyroscope correction value. The dilated convolutional neural network utilizes the physical constraint relationship between acceleration changes and angular velocities at adjacent time points to extract auxiliary features from the original accelerometer measurements. These auxiliary features reflect the variation pattern of accelerometer measurements caused by changes in angular velocity. The dilated convolutional neural network fuses these auxiliary features with gyroscope temporal features to output a time-varying gyroscope correction value, thereby constraining the kinematic consistency between the time-varying gyroscope correction value and the original accelerometer measurements.
[0043] Specifically, feature fusion is performed layer by layer from layer 1 to layer 4 of the dilated convolutional neural network. The first layer uses 6-dimensional IMU measurements (3-axis gyroscope + 3-axis acceleration) as joint input. The second to fourth layers further fuse features by sharing dilated convolutional kernels. Through weight learning of dilated convolutions layer by layer, the temporal correlation features of gyroscope and acceleration are adaptively fused.
[0044] Specifically, in S4, open-loop integration is performed on the corrected gyroscope data to calculate the estimated attitude increment; during the training phase of end-to-end joint optimization, the open-loop integration adopts a tree-structured batch attitude integration algorithm. The matrix multiplication by rotation is organized into a binary tree structure, and the time complexity is reduced from [previous value] to [new value] through parallel computation. Reduce to .
[0045] For example, It only requires 5 batch GPU operations, which is about 6 times faster than serial computing.
[0046] This algorithm is used in the training phase of end-to-end joint optimization, which can significantly improve the computational efficiency of attitude integrals, shorten training time, and improve training efficiency.
[0047] Since it is difficult to directly obtain the true angular velocity at high frequencies (200Hz), a loss function based on attitude increments is established in S5.
[0048] From time arrive True attitude increment The true stance provided by VIO ground truth and Calculated; ; in, From time At the time The actual attitude increment, For incremental time scales, For the time of measurement The transpose of the true posture; For the time of measurement The true attitude.
[0049] From time arrive Estimated attitude increment gyroscope data after integral correction through The result is obtained by multiplying the exponents together; ; The sampling period.
[0050] The pose increment loss function is constructed based on the Huber loss function, where the Huber loss function is: ; in, For time scale Loss items at that time From time At the time The actual attitude increment; From time At the time The transpose of the estimated attitude increment; For SO(3) logarithmic mapping; map the rotation matrix to Lie algebra; It is the vector norm; The Huber loss function is used, with a parameter set to 0.005 to provide robustness against outliers. The Huber loss function uses squared loss for small errors and linear loss for large errors, as shown in the formula: or ; Indicates the first Summing at each initial time.
[0051] The attitude increment loss function established in this embodiment includes at least two loss terms with different time scales. The time window length of each loss term is determined according to the ratio between the output frequency of the true attitude increment and the IMU sampling frequency, so that the calculation frequency of the attitude increment error at each time scale is on the same order of magnitude as the output frequency of the true attitude increment. Furthermore, each loss term covers historical IMU measurement data of different lengths to simultaneously constrain short-term dynamic error and long-term cumulative drift.
[0052] Preferably, loss terms with time scales of 16 and 32, matching the VIO ground truth frequency, are used, corresponding to attitude increment errors of approximately 12Hz and 6Hz, respectively. The total loss function is: .
[0053] This loss function has the following advantages: Optimize directly at the attitude level, which aligns with the final goal; Logarithmic mappings guarantee rotation invariance; Huber loss provides robustness.
[0054] Specifically, the training strategy of this embodiment includes the following aspects: Optimizer: The ADAM optimizer is used with an initial learning rate of 0.01. Combined with the cosine annealing restart learning rate scheduling strategy, the learning rate is dynamically adjusted during training, which helps to escape local optima.
[0055] Regularization: Weight decay (0.1) and Dropout (probability 0.1) are used to prevent overfitting. Batch normalization after each convolutional layer also helps stabilize the training process.
[0056] Data augmentation: A standard deviation of 0.01 is randomly added to each training round. Gaussian noise is used to improve the model's generalization ability. This data augmentation strategy makes the model more adaptable to unseen noise patterns.
[0057] Training time: 1800 training epochs, approximately 5 minutes per dataset on a GTX 1080 GPU. Due to the small number of network parameters and the use of parallel computing, the training efficiency is high, making it suitable for rapid iteration and transfer learning.
[0058] Initialization strategy: Gyroscope intrinsic parameter calibration matrix Initialize to identity matrix gyroscope time-varying correction value Initialize to zero vector to make the corrected gyroscope data in the early stage of training close to the original gyroscope measurement value, and avoid over-correction in the early stage of the network, which would cause the attitude estimation to diverge.
[0059] Optionally, the hardware implementation in this embodiment adopts the following configuration: IMU sensor: ADIS16448 (EuRoC dataset) or BMI160 (TUM-VI dataset), is a low-cost MEMS inertial measurement unit with a sampling frequency of 200Hz.
[0060] Training platform: A workstation equipped with a GTX 1080 GPU for end-to-end joint optimization training of the model.
[0061] Inference platform: Embedded devices that support PyTorch (such as the NVIDIA Jetson series) for real-time inference and pose estimation. With only 77,052 model parameters, the computational complexity is low, making it suitable for real-time operation on embedded platforms.
[0062] A specific software implementation process in this embodiment includes: Step 1: Data Acquisition and Preprocessing; Acquire raw IMU data (three-axis gyroscope, three-axis accelerometer) and VIO ground truth pose data. Split the data into a training set (first 50 seconds of each sequence), a validation set (the remaining portion), and a test set (independent sequences). Normalize the data to make it suitable for neural network input.
[0063] Step 2: Network Initialization; Initialize the parameters of the dilated convolutional neural network and the intrinsic parameter calibration matrix. Initialize to identity matrix Gyroscope correction value Initialize to a zero vector; Network layer parameters are set as follows: five dilated convolutional layers with a kernel size of 7, and the number of channels are 16, 32, 64, 128, and 3 respectively, with dilation rates of 1, 4, 16, 64, and 1 respectively.
[0064] Step 3, Forward Inference: Input N=448 historical IMU measurements into the network and calculate the gyroscope correction values. According to the formula Obtain the corrected gyroscope data The network simultaneously processes gyroscope and acceleration data to extract temporal correlation features.
[0065] Step 4, Attitude Integration; using the formula The attitude increment is calculated by performing an open-loop integral on the corrected gyroscope data. During the training phase, a tree-structured batch pose integration algorithm is used to... The matrix multiplication by rotation is organized into a binary tree structure, and the time complexity is reduced from [previous value] to [new value] through parallel computation. Reduce to .
[0066] Step 5, Loss Calculation; using the Huber loss function (parameters) Provides robustness; the total loss function is This corresponds to the attitude increment error at two different time scales.
[0067] Step 6, Backpropagation and Optimization: Calculate the gradient through backpropagation, and update the network parameters and intrinsic parameter calibration matrix using the ADAM optimizer (initial learning rate 0.01, cosine annealing restarts the learning rate scheduling). Apply weight decay (weightdecay=0.1) and Dropout (probability 0.1) to prevent overfitting.
[0068] Step 7: Iterative training; repeat steps 3-6 for a total of 1800 rounds, adding Gaussian noise with a standard deviation of 0.01 deg / s for data augmentation in each round. Evaluate the validation set performance after each round of training and save the optimal model.
[0069] Step 8: Model Deployment; Export the trained model in ONNX format and deploy it to the target embedded platform for real-time inference. Due to the small number of model parameters and high computational efficiency, the inference latency is less than 5ms, meeting the real-time requirements.
[0070] Simulation verification results To verify the effectiveness of this invention, experiments were conducted on the publicly available EuRoC and TUM-VI datasets. Comparison methods included: raw IMU (pure inertial), calibrated IMU (pure inertial), OriNet (LSTM-based method, pure inertial), VINS-Mono (monocular vision-inertial odometry framework, vision-inertial fusion), and Open-VINS (visual-inertial estimator based on extended Kalman filter, vision-inertial fusion).
[0071] (1) Comparison of absolute attitude error (AOE); Absolute attitude error (AOE) is calculated as the difference between the estimated attitude value and the true attitude value. For data with a sequence length of M, AOE is defined as: ; in for Logarithmic mapping, The vector norm is given. Table 1 shows the absolute attitude error (3D attitude / yaw angle, in degrees) of each method on the test sequence. In the table, the OriNet column is marked with "—" on the TUM-VI dataset, indicating that there are no experimental results for OriNet on this dataset.
[0072]
[0073] Table 2: Comparison of Absolute Attitude Errors (3D Attitude / Yaw Angle, Unit: Degrees) As shown in Table 2, the proposed method outperforms or approaches Open-VINS on most test sequences, and significantly outperforms the calibrated IMU, VINS-Mono, and OriNet methods. On the EuRoC dataset, the proposed method has a mean absolute attitude error of 0.89 degrees and a yaw angle error of 0.78 degrees, representing an improvement of approximately 40% compared to the calibrated IMU. On the TUM-VI dataset, the mean absolute attitude error is 0.57 degrees and the yaw angle error is 0.57 degrees, representing an improvement of approximately 50% compared to the calibrated IMU.
[0074] (2) Comparison of relative attitude error (ROE); Relative attitude error, calculated as the attitude difference between sub-trajectories, is a standard metric for evaluating odometry methods. For a displacement distance of... sub-trajectories , Defined as: ; in Indicates from time arrive The actual attitude increment, To estimate the attitude increment. Figure 3 The relative pose error of each method on the test sequence is shown; from Figure 3 As can be seen, the ROE curve of the proposed method is close to that of Open-VINS. The ROE is less than 1 degree for most displacement distances, which is significantly better than the original IMU, the calibrated IMU and the VINS-Mono method. This verifies that the proposed method can effectively suppress attitude cumulative drift under pure inertial conditions.
[0075] Combined Table 1 and Figure 3The comparison results show that the proposed method, using only an IMU and without relying on a visual sensor, achieves absolute and relative pose error levels comparable to Open-VINS, which uses IMU and camera fusion. Compared to comparison methods such as VINS-Mono and OriNet, as well as baseline methods such as the original IMU and calibrated IMU, the proposed method has significant advantages in pose estimation accuracy and significant application advantages in visually constrained scenarios such as low light, fast motion, and texture loss.
[0076] (3) Comparison of attitude estimation results; Figure 4 The results show the roll, pitch, and yaw angle estimations on the EuRoC MH 04 difficult and TUM-VI room 4 sequences.
[0077] comprehensive Figure 4 The comparison of roll, pitch, and yaw angles between the two sequences shows that the original IMU data exhibits significant drift within seconds. Although the calibrated IMU improves short-term accuracy through intrinsic parameter calibration, the long-term drift problem persists. The method of this invention, through joint optimization of deep learning denoising and intrinsic parameter calibration, can closely track the true attitude and achieve high-precision attitude estimation in all three degrees of freedom: roll, pitch, and yaw. It particularly demonstrates excellent performance in yaw angle estimation, effectively solving the technical challenge of severe yaw drift in purely inertial methods.
[0078] (4) Gyroscope correction signal analysis; Figure 5 The gyroscope correction signal on the TUM-VI room 4 sequence is shown. The correction amplitude is approximately 1 deg / s, but it can significantly improve the attitude estimation accuracy, verifying that the deep learning model can learn an effective error compensation pattern.
[0079] from Figure 5 It can be seen that the gyroscope correction signal has the following characteristics: ① The correction signal exhibits time-varying characteristics, indicating that the neural network can adaptively estimate the time-varying bias; ② The correlation between the correction signal and the changing trend of the original gyroscope data indicates that the network has learned the temporal patterns of noise and bias; ③ The amplitude of the correction signal is within a reasonable range, avoiding attitude estimation divergence caused by overcorrection.
[0080] Analysis of the gyroscope correction signal shows that the extended convolutional neural network of the present invention can effectively extract the IMU temporal features and learn accurate error compensation patterns, thereby achieving high-precision attitude estimation.
[0081] Example 2 This embodiment discloses a deep learning-based IMU gyroscope noise reduction system, including: an IMU chip and a processor; wherein, An IMU chip is used to acquire raw IMU measurement data; the raw IMU measurement data includes raw gyroscope measurement values and raw accelerometer measurement values. The processor, connected to the IMU chip, is used to process the raw IMU measurement data using a gyroscope correction model trained by the deep learning-based IMU gyroscope noise reduction method as described in Embodiment 1, and output the corrected gyroscope data and estimated attitude increment.
[0082] In this embodiment, the more specific technical details and beneficial effects are the same as those disclosed in Embodiment 1. Please refer to them for details, and they will not be repeated here.
[0083] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A deep learning-based IMU gyroscope noise reduction method, characterized in that, include: S1. Acquire raw IMU measurement data and obtain the true attitude increment. The raw IMU measurement data includes raw gyroscope measurement values and raw accelerometer measurement values. S2. Input the historical measurement data in the IMU raw measurement data into the dilated convolutional neural network, and the dilated convolutional neural network jointly extracts the time-series correlation features between the gyroscope and the accelerometer, and outputs the time-varying correction value of the gyroscope; S3. Establish a gyroscope calibration model; the model uses a learnable gyroscope intrinsic parameter calibration matrix to linearly transform the original gyroscope measurement values, and superimposes the time-varying correction values to obtain the corrected gyroscope data; S4. Perform open-loop integration on the corrected gyroscope data to calculate the estimated attitude increment; S5, based on The group logarithmic mapping calculates the rotation angle error between the estimated attitude increment and the true attitude increment. An attitude increment loss function is constructed based on the rotation angle error. The network parameters of the dilated convolutional neural network and the gyroscope intrinsic parameter calibration matrix are jointly optimized end-to-end based on the attitude increment loss function.
2. The deep learning-based IMU gyroscope noise reduction method according to claim 1, characterized in that, The attitude increment loss function is constructed based on the Huber loss function; the time scale is... Loss items at time Represented as: ； in, From time At the time The actual attitude increment, From time At the time The transpose of the estimated attitude increment; for Logarithmic mapping; For vector norm, Here is the Huber loss function. Indicates the first Summing at each initial time.
3. The deep learning-based IMU gyroscope noise reduction method according to claim 2, characterized in that, The attitude increment loss function includes at least two loss terms with different time scales. The size of the time scale of each loss term is determined according to the ratio between the output frequency of the true attitude increment and the IMU sampling frequency, so that the calculation frequency of the attitude increment error at each time scale is on the same order of magnitude as the output frequency of the true attitude increment. Furthermore, each loss term covers historical IMU measurement data of different lengths to simultaneously constrain short-term dynamic errors and long-term cumulative drift.
4. The deep learning-based IMU gyroscope noise reduction method according to claim 2, characterized in that, The dilated convolutional neural network adopts an M-layer dilated convolutional structure. The first M-1 layers are configured with an exponentially increasing dilation rate and are matched with the kernel size determined according to the local temporal feature scale of the IMU raw measurement data, so that the temporal receptive field coverage matches the historical IMU measurement data for a period of time that meets the requirements of attitude estimation accuracy. The Mth layer adopts a convolutional structure with a kernel size of 1, which is used to map the feature channels into three-axis gyroscope correction values for output.
5. The deep learning-based IMU gyroscope noise reduction method according to claim 4, characterized in that, The first layer of the dilated convolutional neural network uses six-dimensional IMU measurements as joint inputs, including three-axis gyroscope measurements and three-axis accelerometer measurements. The first M-1 layers of the dilated convolutional neural network fuse the auxiliary features and gyroscope temporal features layer by layer by sharing dilated convolutional kernels. The Mth layer maps the fused features to three-axis gyroscope time-varying correction values.
6. The deep learning-based IMU gyroscope noise reduction method according to claim 4, characterized in that, Before the end-to-end joint optimization, the gyroscope intrinsic parameter calibration matrix is initialized to an identity matrix, and the gyroscope time-varying correction value is initialized to a zero vector. This makes the corrected gyroscope data in the early stage of training close to the original gyroscope measurement value, so as to constrain the attitude estimation to evolve along the open-loop integral trajectory of the original measurement value in the early stage of training, and avoid the attitude estimation divergence caused by over-correction in the early stage of the network.
7. The deep learning-based IMU gyroscope noise reduction method according to claim 1, characterized in that, During the training phase of the end-to-end joint optimization, the open-loop integration employs a tree-structured batch attitude integration algorithm. The matrix multiplication by rotation is organized into a binary tree structure, and the time complexity is reduced from [previous value] to [new value] through parallel computation. Reduce to .
8. The deep learning-based IMU gyroscope noise reduction method according to claim 1, characterized in that, In the gyroscope calibration model, the linear transformation is achieved by performing matrix multiplication between the gyroscope intrinsic parameter calibration matrix and the original gyroscope measurement value, which is used to compensate for axis deviation and scaling factor errors; the time-varying gyroscope correction value is output by the dilated convolutional neural network and added to the linear transformation result to obtain the corrected gyroscope data.
9. The deep learning-based IMU gyroscope noise reduction method according to claim 8, characterized in that, The gyroscope intrinsic parameter calibration matrix is decomposed into the product of the scaling factor matrix and the axis deviation matrix; The scaling factor matrix and the axis deviation matrix are constrained by diagonal dominance or orthogonality constraints during training, which ensures that the optimization results of the gyroscope intrinsic parameter calibration matrix remain physically interpretable and avoids the intrinsic parameter matrix deviating from the physical meaning of the inertial sensor calibration due to end-to-end joint training.
10. A deep learning-based IMU gyroscope noise reduction system, characterized in that, include: IMU chip and processor; among which... An IMU chip is used to acquire raw IMU measurement data; the raw IMU measurement data includes raw gyroscope measurement values and raw accelerometer measurement values. The processor, connected to the IMU chip, is used to process the raw measurement data of the IMU using the gyroscope correction model trained by the deep learning-based IMU gyroscope noise reduction method as described in any one of claims 1 to 9, and output the corrected gyroscope data and estimated attitude increment.