IMU (inertial measurement unit) noise reduction method based on self-supervised training for inertial aided navigation

Abstract

The invention discloses an IMU (inertial measurement unit) noise reduction method based on self-supervised training for inertial aided navigation. The method comprises the following steps: designing a deep neural network model according to the characteristic of high release frequency of an inertial sensor; the read IMU data is subjected to noise reduction processing of sequence data through a structured state space model in combination with a self-supervised training mode; the method comprises the following steps: establishing a mathematical model of an IMU (Inertial Measurement Unit), carrying out self-supervised training through the design of a deep neural network, carrying out model training by utilizing the idea of reinforcement learning, and reducing the noise of IMU data into real data more fitting the physical world by the model through a minimum mean square error loss function so as to meet various inertial aided navigation algorithms. Whether the requirements are met or not is tested, and performance evaluation is conducted on a result. According to the method, the confidence coefficient of IMU data can be improved, meanwhile, other sensors are assisted in cooperative positioning navigation, the requirements of various navigation tasks are met, short-time and effective accurate navigation can be supported, and the task safety is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of high-end equipment manufacturing, especially the technical field of unmanned aerial vehicles. Specifically, it discloses a self-supervised training-based IMU (Inertial Measurement Unit) noise reduction method for inertial aided navigation. Background Art

[0002] Inertial Aided Navigation (IAN) technology provides more stable and accurate navigation capabilities in complex environments by combining an Inertial Navigation System (INS) with other navigation technologies, such as vision, lidar, etc. In recent years, with the rapid development of intelligent driving technology and the robotics field, inertial aided navigation technology has been increasingly widely used in the intelligent driving technology and the robotics field, especially performing well in GPS (Global Positioning System) denied environments such as indoor and outdoor environments.

[0003] In visual inertial navigation, the inertial aided navigation system realizes the complementary advantages of both by utilizing the autonomy and continuity of the INS and the environmental perception and high-precision positioning of visual navigation. The INS system consists of an accelerometer and a gyroscope, which calculate speed and position by measuring linear acceleration and angular velocity, but there is a problem that excessive noise will cause error accumulation and it cannot directly rely on IMU data for navigation and positioning. Visual navigation uses the image information obtained by a camera for environmental perception and positioning, but it will fail in cases of poor lighting conditions or insufficient visual features. Transient environmental disturbances, whether in the field of intelligent driving or the robotics field, will pose a threat to mission safety, resulting in economic and personnel losses. Summary of the Invention

[0004] To solve the above problems existing in the prior art, the present invention provides a self-supervised training-based IMU noise reduction method for inertial aided navigation, including the following steps:

[0005] S1. Discretize the original IMU data x(t) of length L through zero-order hold technology and perform sampling to obtain IMU time series data;

[0006] S2. Input the IMU time series data into a deep neural network to obtain the output feature f(x(t));

[0007] Input the output feature f(x(t)) of the deep neural network into the SSM neural network and process it using the following formulas (1) and (2):

[0008] h(t) = A·h(t - 1)+B·f(x(t)) (1)

[0009] y(t) = C·h(t)+D·f(x(t)) (2)

[0010] Where h(t) is the latent state, A, B, C, and D are relevant parameter matrices with shapes matching f(x(t)) and h(t), and y(t) is the output state, which is consistent with the dimension of the IMU time series data;

[0011] S3. Use the self-supervised training method to perform regression analysis on the current IMU data through historical data;

[0012] S4. Construct a training data set and a test data set based on the collected real IMU data, where the training data accounts for 80% and the test data accounts for 20%;

[0013] After obtaining the model parameters through training on the training data set, fix the model parameters;

[0014] Perform testing on the test data set until the MSE error is lower than 0.001;

[0015] S5. Use the trained model to perform noise reduction processing on the input raw IMU data.

[0016] Furthermore, the discretization process of the raw IMU data x(t) of length L through the zero-order hold technique in S1 includes:

[0017] S1.1. When receiving the signal transmitted by the IMU, if the length of the raw IMU data x(t) is greater than L, delete the earliest incoming data in the signal queue data;

[0018] Otherwise, add the data of the new signal into the data queue;

[0019] S1.2. Sample the continuous input signals according to the input time step to generate IMU time series data.

[0020] Furthermore, in S2, the deep neural network adopts a fully connected neural network, the input data dimension is 6, the deep neural network is divided into three layers. The input of the first layer is 6-dimensional, the output is 256-dimensional, and the kernel function adopts the relu function; the input of the second layer is 256-dimensional, the output is 128-dimensional, and the kernel function adopts the relu function; the input of the third layer is 128-dimensional, the output is 256-dimensional, and the kernel function adopts the relu function.

[0021] Furthermore, S3 also includes:

[0022] Define the MSE loss function between the IMU data y(t) output in S2 and the original IMU data x(t) as the loss function of deep learning;

[0023] Using the fully connected neural network and SSM neural network model constructed in S2, taking the IMU data at the current n moments as input, predicting the IMU value at the next moment, and calculating the MSE error between it and the true IMU data;

[0024] Adopt the Adam optimization algorithm to iteratively update the model parameters until the error is less than 0.0001 and then stop training.

[0025] The present invention can provide a more accurate navigation and positioning effect in inertial-aided navigation. When other sensors fail, the IMU can support short-term and effective accurate navigation, support the task implementation, and maintain the safe progress of the task. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the structure diagram of the deep neural network in the present invention;

[0027] Figure 2 It is the evaluation effect diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0028] The technical solution of the present invention will be further described below through the drawings and embodiments. Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those of ordinary skill in the field to which the present invention belongs.

[0029] In a specific embodiment of the present invention, it includes the following steps:

[0030] S1. For the original IMU data x(t), perform discretization processing and sampling of the data through the zero-order hold technology, specifically including:

[0031] S1.1. Optimize the processing flow of the IMU signal by using the sliding window method.

[0032] When receiving the signal transmitted by the IMU, if the length of the original IMU data x(t) is greater than L, delete the earliest transmitted data in the signal queue data, otherwise add the data of the new signal into the data queue. Thus, the signal queue maintains a fixed length.

[0033] Specifically, when receiving the signal transmitted by the IMU, the system stores and retains the current signal, and continuously monitors the reception status of subsequent signals. When a new signal arrives, the new and old signals are fused to establish a continuous and stable signal sequence. This processing method can effectively reduce the situation of data loss or breakpoints, and ensure the integrity and consistency of the signal during the experiment.

[0034] S1.2. Sample the continuous input signal according to the input time step Δ to generate discrete output and obtain sampling data. The holding time step Δ can be set to 0.01 s.

[0035] S2. Use the IMU time series data obtained by sampling in S1 and input it into a deep neural network to extract the true features of the IMU data from complex noise.

[0036] The designed deep neural network is as Figure 1 shown. A fully connected network is adopted. The input data dimension is 6. The deep neural network is divided into three layers. The input of the first layer is 6 - dimensional and the output is 256 - dimensional. The kernel function uses the relu function. The input of the second layer is 256 - dimensional and the output is 128 - dimensional. The kernel function uses the relu function. The input of the third layer is 128 - dimensional and the output is 256 - dimensional. The kernel function still uses the relu function. The output is defined as f(x(t)).

[0037] Then establish an output denoising model and input the features output by the fully connected network into a deep structured state space model (Deep Structured State Space Models, SSM) neural network.

[0038] The output features of the fully connected network are f((x(t)), which are used as the input of the SSM (State Space Model) model. They are sequence data related to the time series and are processed using the following formula:

[0039] h(t) = A·h(t - 1)+B·f(x(t)) (1)

[0040] y(t) = C·h(t)+D·f(x(t)) (2)

[0041] where h(t) is the latent state, which is self - saved after being initialized by the model. A, B, C, and D are relevant parameter matrices whose shapes match f(x(t)) and h(t). y(t) is the output state, with a dimension of 6, which is consistent with the IMU dimension.

[0042] S3. Use a self - supervised training method to perform regression analysis on the current IMU data through historical data to improve the model's prediction ability for IMU signals.

[0043] Define the mean squared error (MSE) between the output IMU data y(t) and the original IMU data x(t) in S2 as the loss function for deep learning. During training, use the fully connected neural network and the state space model (SSM) neural network model constructed in S2. Take the IMU data at the current n time instants as input, predict the IMU value at the next time instant, and calculate the MSE error between it and the true IMU data. Use the Adam optimization algorithm to iteratively update the model parameters until the error is less than 0.0001 to stop training, ensuring the convergence and prediction accuracy of the model.

[0044] S4. Construct a training dataset and a test dataset based on the collected real IMU data, where the training data accounts for 80% and the test data accounts for 20%.

[0045] Train on the training dataset to obtain the optimal model parameters, and fix the model parameters after training is completed.

[0046] Subsequently, test on the test dataset. If the MSE error is lower than 0.001, it can be determined that the model converges, indicating that the model has strong generalization ability on unseen data and can stably predict IMU signals.

[0047] S5. Use the trained model to perform noise reduction processing on the input original IMU data.

[0048] After verifying the model convergence on the test dataset, to further evaluate the actual effect of IMU data noise reduction, input the output IMU data y(t) and the original IMU data x(t) into VINS-Mono for comparative testing.

[0049] VINS-Mono, as a visual-inertial odometry-based SLAM system, can intuitively evaluate the accuracy and stability of IMU data.

[0050] Specifically, use VINS-Mono to input the original IMU data and the noise-reduced IMU data respectively, and compare their trajectory estimation results. If the noise-reduced IMU data is superior to the original data in terms of trajectory smoothness, estimation accuracy, and attitude stability, it indicates that the noise reduction strategy is effective. In addition, the noise reduction effect can be further quantified by calculating the absolute trajectory error (ATE) and the relative trajectory error (RTE), and the results are as Figure 2 shown, ensuring that the optimized IMU data can provide higher reliability and accuracy in practical applications.

[0051] The following introduces a specific embodiment for verifying the solution of the present invention:

[0052] The IMU data in the dataset is divided into two groups. One group uses the original data without processing and inputs it into the inertial-aided navigation algorithm; the other group of IMU data generates denoised data through the neural network proposed by the present invention, and then inputs the denoised data into inertial-aided navigation. Compare the navigation and positioning effects between the two, and the verification results are as Figure 2 shown. In the three highlighted bounding box areas, through IMU denoising, the IMU time series data is closer to the true motion trajectory.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solution of the present invention, and these modifications or equivalent replacements cannot make the modified technical solution deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A self-supervised training-based IMU noise reduction method for inertial-aided navigation, characterized in that, It includes the following steps: S1. Discretize the original IMU data x(t) of length L by zero-order hold technology and sample it to obtain IMU time series data; S2. Input the IMU time series data into a deep neural network to obtain the output feature f(x(t)); Input the output feature f(x(t)) of the deep neural network into the SSM neural network and process it using the following formulas (1) and (2): h(t) = A·h(t - 1) + B·f(x(t)) (1) y(t) = C·h(t) + D·f(x(t)) (2) where h(t) is the latent state, A, B, C, and D are relevant parameter matrices with shapes matching f(x(t)) and h(t), and y(t) is the output state, which is consistent with the dimension of the IMU time series data; S3. Use a self-supervised training method to perform regression analysis on the current IMU data through historical data; S4. Construct a training data set and a test data set based on the collected real IMU data, where the training data accounts for 80% and the test data accounts for 20%; After obtaining the model parameters through training on the training data set, fix the model parameters; Perform testing on the test data set until the MSE error is lower than 0.001; S5. Use the trained model to perform noise reduction processing on the input original IMU data.

2. A self-supervised training-based IMU noise reduction method for inertial aided navigation according to claim 1, characterized in that, The discretization process of the original IMU data x(t) of length L by the zero-order hold technology described in S1 includes: S1.

1. When receiving the signal transmitted by the IMU, if the length of the original IMU data x(t) is greater than L, delete the earliest transmitted data in the signal queue; Otherwise, add the data of the new signal into the data queue; S1.

2. Sample the continuous input signals according to the input time step to generate IMU time series data.

3. A self-supervised training-based IMU noise reduction method for inertial-aided navigation according to claim 2, characterized in that In S2, the deep neural network uses a fully connected neural network, the input data dimension is 6, the deep neural network is divided into three layers. The input of the first layer is 6-dimensional, the output is 256-dimensional, and the kernel function uses the relu function; the input of the second layer is 256-dimensional, the output is 128-dimensional, and the kernel function uses the relu function; the input of the third layer is 128-dimensional, the output is 256-dimensional, and the kernel function uses the relu function.

4. A self-supervised training-based IMU noise reduction method for inertial-aided navigation according to claim 3, characterized in that, S3 also includes: Define the MSE loss function of the IMU data y(t) and the original IMU data x(t) output in S2 as the loss function of deep learning; Use the fully connected neural network and the SSM neural network model constructed in S2, take the IMU data at the current n time instants as the input, predict the IMU value at the next time instant, and calculate the MSE error between it and the real IMU data; Adopt the Adam optimization algorithm to iteratively update the model parameters until the error is less than 0.0001 and stop training.

Images