Angular acceleration sensor denoising method based on expanded convolutional neural network

Abstract

The invention provides an angular acceleration sensor denoising method based on an expanded convolutional neural network, which introduces the expanded convolutional neural network to correct angular acceleration data on the basis of the angular acceleration data measured by an angular acceleration sensor, and designs an angular acceleration sensor and deep learning combined denoising method. According to the method, the data processing amount and the training effect of deep learning are considered, the training speed is greatly improved by adopting operation applied to multiple instances in parallel, meanwhile, the learning efficiency is improved by adopting a cosine annealing learning rate adjustment algorithm, the training time is greatly shortened, and the training efficiency is improved. Therefore, we can obtain accurate attitude estimation through a low-cost angular acceleration sensor. Compared with a common neural network training mode, the method of the invention has the advantages of robustness for overfitting, high training efficiency and fast training speed. The angular acceleration sensor introduced in the product is low in cost, and large-scale application can be conveniently carried out.

Description

Technical Field

[0001] The invention belongs to the technical field of automatic control, and in particular relates to an angular acceleration sensor denoising method based on an extended convolutional neural network. Background Art

[0002] Attitude estimation technology is the key and foundation for path planning and motion control of carriers such as drones and handheld devices. Therefore, researchers have conducted a lot of research on how to obtain accurate attitude information of carriers. In the field of attitude estimation, there is an angular acceleration sensor as an inertial device that can directly obtain the angular acceleration data of the carrier's motion. The collected angular acceleration data can be integrated twice to obtain the attitude information of the carrier's motion, which improves the availability of low-cost inertial navigation systems.

[0003] As a sensor that directly measures angular acceleration, the angular acceleration sensor itself has noise interference when collecting data. The attitude information obtained by integrating the angular acceleration twice is greatly affected by noise. Traditional methods should select modeling to eliminate noise, but the solution process is relatively complicated. The dilated convolutional neural network can extract features from a large amount of data and train specific parameters by establishing an appropriate loss function to achieve the target requirements. Therefore, the dilated convolutional neural network can be used to denoise the angular acceleration sensor and improve the accuracy of attitude estimation. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide an angular acceleration sensor denoising method based on a dilated convolutional neural network, which can use the dilated convolutional neural network to denoise the angular acceleration sensor and improve the accuracy of posture estimation.

[0005] A method for denoising an angular acceleration sensor based on a dilated convolutional neural network, comprising:

[0006] The angular acceleration sensor is used to collect the three-axis angular acceleration β at the current time n n , and the three-axis angular accelerations at multiple moments before, are input into the dilated convolutional neural network to obtain the output value of the network, i.e., the angular acceleration correction

[0007] The corrected angular acceleration value is obtained according to the defined angular acceleration sensor error compensation model

[0008]

[0009] Among them, β n is the three-axis angular acceleration measured by the n sensor at the current moment, is the angular acceleration correction, is the corrected angular acceleration value, is the learnable error coefficient matrix;

[0010] Using the corrected angular acceleration Perform posture calculation to obtain the posture change R between two frames of data n ;

[0011] The true value of the attitude change is obtained, and the attitude change R calculated by the corrected angular acceleration value is obtained. n Compare and pass the difference into the loss function to calculate the loss value. The specific expression is as follows:

[0012]

[0013] Among them, y represents the reference value, Δ is the difference between the true value of the posture change and the estimated value of the posture change, and δ is the set threshold hyperparameter;

[0014] According to the loss value calculated by the loss function, the gradient is calculated and back-propagated to update the parameters of the dilated convolutional neural network and the learnable parameters of the error compensation model;

[0015] When the training stop condition is reached, the matrix based on the training completion The dilated convolutional neural network uses the input angular acceleration data to obtain the angular acceleration correction The angular acceleration sensor is denoised according to formula (1).

[0016] Preferably, the structure of the dilated convolutional neural network includes the following structures in sequence: a padding layer that performs padding operations on input data; a convolution layer that convolves the data with a set convolution kernel; a normalization layer; a nonlinear layer; a random inactivation layer; and a fully connected layer that outputs data of a predetermined dimension.

[0017] Better, learnable error coefficient matrix Initialize at the beginning of training:

[0018]

[0019] Among them, R is the matrix of corresponding dimension of random assignment, and I is the unit matrix.

[0020] Preferably, the Adam optimizer is used to update the parameters of the dilated convolutional neural network.

[0021] The present invention has the following beneficial effects:

[0022] Based on the angular acceleration data measured by the angular acceleration sensor, the present invention introduces an expanded convolutional neural network to correct the angular acceleration data, and designs a denoising method combining the angular acceleration sensor with deep learning; the method takes into account the data processing volume and training effect of deep learning, and adopts operations applied in parallel to multiple instances, which greatly improves the training speed, and adopts a cosine annealing learning rate adjustment algorithm to improve the learning efficiency and greatly reduce the training time, so we can obtain accurate posture estimation through a low-cost angular acceleration sensor; compared with the general neural network training method, the method of the present invention is robust to overfitting, has high training efficiency and fast training speed. The angular acceleration sensor introduced by this product has a low cost and is convenient for large-scale application. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The present invention is a flow chart of the method.

[0024] Figure 2 This is a network layer relationship diagram of the expanded convolutional neural network in the present invention. DETAILED DESCRIPTION

[0025] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0026] In view of the noise problem in the existing angular acceleration sensor measurement, the powerful learning and processing capabilities of the neural network are used to correct the original angular acceleration data. The present invention provides an angular acceleration sensor denoising method based on a dilated convolutional neural network, and the overall steps are as follows:

[0027] Step 1: Parameter initialization. Initialize network construction parameters such as input signal channels, convolution kernel size, padding layer parameters, etc.; initialize training parameters such as learning rate, regularization coefficient, etc.; initialize optimization parameters such as T_0, T_mult, minimum learning rate, etc.; initialize compensation model parameters such as error coefficient matrix Initialize and set them as optimizable parameters.

[0028] Step 2: Data collection. Collect the three-axis angular acceleration β through the angular acceleration sensor n .

[0029] Step 3: Data preprocessing: De-mean and normalize the angular acceleration data.

[0030] Step 4: Network extracts information. The angular acceleration data γ nThe input is in the dilated convolutional neural network, and passes through multiple layers of the network, including: 1. A padding layer that pads the input data to ensure that key information at the data boundary is retained and to maintain the dimensional consistency of the data when it is transmitted between neural network layers; 2. A convolution layer that convolves the incoming data with the set convolution kernel and expands the data by setting parameters to increase the receptive field of the model; 3. A normalization layer that improves the convergence speed and prevents overfitting; 4. A nonlinear layer that gives the model nonlinear characteristics and enhances its ability to approximate any nonlinear function, thereby improving the expressiveness of the neural network; 5. A random inactivation layer that randomly assigns "0" to the convolution result according to the set parameters, and prevents the model from overfitting by inactivating the corresponding elements; 6. A fully connected layer that outputs data of a predetermined dimension. The final output is the angular acceleration correction

[0031] Step 5: Correct the angular acceleration data. First, design the angular acceleration sensor error model, the formula is as follows:

[0032] β n =Cβ′ n +b n +η n (1)

[0033] Among them, β n is the measured value of the angular acceleration sensor, β′ n is the true value of angular acceleration, C is the error coefficient matrix, b n is the random walk error, η n is Gaussian white noise with zero mean.

[0034] The error coefficient matrix C is calculated by formula (2):

[0035] C=S n M n (2) Among them, S n is the non-orthogonality error, M n is the proportional error.

[0036] Since C and b in formula (1) n They will change over time and are difficult to measure. It is difficult to conduct quantitative analysis by building a mathematical model, but they affect the output accuracy of the angular acceleration sensor. In order to improve the output accuracy, the learning ability of the neural network is used here to design an angular acceleration sensor error compensation model based on deep learning. The definition of the angular acceleration sensor error compensation model is shown in formula (3):

[0037]

[0038] Among them, γ nis the three-axis angular acceleration measured by the n sensor at the current moment, is the angular acceleration correction, is the corrected angular acceleration value, is the learnable error coefficient matrix.

[0039] From formula (3), we can know that the angular acceleration correction value using neural network is Perform training, and the training model is shown in formula (4):

[0040]

[0041] in, is the angular acceleration correction output by the neural network, f(·) is the function defined by the neural network structure, β′ n-N ,…,β′ c is the input of the neural network, which represents the angular acceleration data at N time points before the time point n. N is determined by the neural network structure. The network parameters will be continuously optimized in step 7 until the optimal value is reached.

[0042] Learnable error coefficient matrix Initialization is performed at the beginning of training. The initialization method is shown in formula (5):

[0043]

[0044] Among them, R is the matrix of the corresponding dimension of the random assignment, and I is the unit matrix. This parameter will be continuously optimized in step 7 until the optimal value is reached.

[0045] The angular acceleration correction and angular acceleration γ n Input the error compensation model, that is, formula (3), to obtain the corrected angular acceleration value

[0046] Step 6: Loss value and gradient calculation. Using the corrected angular acceleration value Perform posture calculation to obtain the posture change between two frames of data. The calculation method is as follows:

[0047]

[0048] R n =R n-1 exp(ω n dt) (7)

[0049] Among them, ω0 represents the angular velocity at the initial moment, ω n represents the angular velocity at the current moment n, R n Represents the rotation matrix of the current moment n, R n-1It represents the rotation matrix at the n-1 moment, that is, the previous moment, and dt is the time difference between the two moments.

[0050] Using formula (6), we can get the angular velocity ω at the corresponding moment: n , and calculate formula (7) to get the attitude change from time n-1 to time n. The attitude true value is collected by motion capture or high-precision inertial navigation and processed into the attitude change between two frames, that is, the attitude change true value, and the attitude change R calculated by the corrected angular acceleration value. n Compare and pass the obtained difference into the loss function to calculate the loss value. The loss function selected by the present invention is the Huber loss function, and the specific expression is as follows:

[0051]

[0052] Wherein, y represents a reference value, which is usually a value close to 0, and is set to 0 in this embodiment. Δ is the difference between the true value of the posture change and the estimated value of the posture change, and δ is a set threshold hyperparameter, which determines the specific expression of the formula.

[0053] According to the loss value calculated by the loss function, the gradient is calculated and back-propagated to train the neural network.

[0054] Step 7: Parameter optimization. The parameters of the dilated convolutional neural network and the learnable parameters of the error compensation model are updated by the selected optimizer. The Adam optimizer is selected in the present invention, which can significantly reduce the training time. The cosine annealing learning rate adjustment algorithm is used during the training process to adjust the optimizer learning rate and improve the learning efficiency.

[0055] Step 8: Determine whether the number of training times has been reached. If so, proceed to step 9; if not, jump to step 4.

[0056] Step 9: Output the trained model. The trained model can be used to denoise the input angular acceleration data to obtain angular acceleration data with higher accuracy.

[0057] The present invention has carried out experiments to verify the algorithm. Considering: 1. In the algorithm, the attitude information of motion capture or high-precision inertial navigation is used as the true value, and the attitude value is used as the benchmark to perform denoising of the angular acceleration sensor; 2. There is no angular acceleration data with high accuracy as the control object, and it is difficult to select experiments with angular acceleration as the comparison result. Therefore, this experiment will select the attitude value as the comparison result, and compare the attitude angle calculated by the original angular acceleration, the attitude angle calculated by the corrected angular acceleration, and the attitude angle calculated by the angular velocity collected by the gyroscope with the high-precision attitude angle.

[0058] During the experiment, a handheld data acquisition device containing an angular acceleration sensor was first built to collect angular velocity data. At the same time, the motion capture system was used to obtain the motion information of the data acquisition device to provide the true value of the experimental attitude angle. The initial learning rate adopted in the experiment was 1e-5, the minimum learning rate was 1e-6, T_0=600, T_mult=2, and the learning rate was adjusted according to the above parameters and the cosine annealing learning rate adjustment algorithm. The overall number of training times was set to 4200 times. In the network layer, the padding layer parameter is 520, which is related to the characteristics of convolution; the input dimensions of the convolution layer are [3,16,32,64], the output dimensions are [16,32,64,128], the convolution kernel is set to [7,7,7,7], and the expansion parameter is [1,4,16,64]; the normalization layer parameter is set to 0.1; the random dropout layer parameter is set to 0.3; the input dimension of the fully connected layer is 128, the output dimension is 3, the convolution kernel is set to 1, and the expansion parameter is 1. The hyperparameter δ in the Huber function is selected as 1e-3. The posture solution method adopted in the experiment is as follows:

[0059]

[0060] Among them, q0, q1, q2, q3 are carrier quaternions, Δt is the time difference between two states, [ω x ω y ω z ] is the three-axis angular velocity of the carrier.

[0061] The quaternion is obtained by using formula (9), and the quaternion is converted into the attitude angle parameter to complete the attitude solution. The specific error indicators of this experiment are as follows:

[0062]

[0063]

[0064] In the above table, IMU represents attitude calculation using data obtained from the gyroscope, CNN-AC represents attitude calculation after correcting angular acceleration using the method proposed by the present invention, and traditional attitude calculation uses the original angular acceleration data for attitude calculation. The experimental results show that compared with the attitude calculation using uncorrected angular acceleration data, the attitude calculation using the accurate angular acceleration data corrected by this method has higher accuracy, and the average errors of the three axes are reduced by 62.5%, 26.07%, and 22.98%, respectively.

[0065] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for denoising an angular acceleration sensor based on a dilated convolutional neural network, characterized in that: include: The angular acceleration sensor is used to collect the three-axis angular acceleration β at the current time n n , and the three-axis angular accelerations at multiple moments before, are input into the dilated convolutional neural network to obtain the output value of the network, i.e., the angular acceleration correction The corrected angular acceleration value is obtained according to the defined angular acceleration sensor error compensation model Among them, β n is the three-axis angular acceleration measured by the n sensor at the current moment, is the angular acceleration correction, is the corrected angular acceleration value, is the learnable error coefficient matrix; Using the corrected angular acceleration Perform posture calculation to obtain the posture change R between two frames of data n ; The true value of the attitude change is obtained, and the attitude change R calculated by the corrected angular acceleration value is obtained. n Compare and pass the difference into the loss function to calculate the loss value. The specific expression is as follows: Among them, y represents the reference value, Δ is the difference between the true value of the posture change and the estimated value of the posture change, and δ is the set threshold hyperparameter; According to the loss value calculated by the loss function, the gradient is calculated and back-propagated to update the parameters of the dilated convolutional neural network and the learnable parameters of the error compensation model; When the training stop condition is reached, the matrix based on the training completion The dilated convolutional neural network uses the input angular acceleration data to obtain the angular acceleration correction The angular acceleration sensor is denoised according to formula (1).

2. The angular acceleration sensor denoising method based on a dilated convolutional neural network as claimed in claim 1, characterized in that: The structure of the dilated convolutional neural network includes the following structures in sequence: a padding layer that performs padding operations on input data; a convolution layer that convolves the data with a set convolution kernel; a normalization layer; a nonlinear layer; a random inactivation layer; and a fully connected layer that outputs data of a predetermined dimension.

3. The angular acceleration sensor denoising method based on a dilated convolutional neural network as claimed in claim 1 or 2, characterized in that: Learnable error coefficient matrix Initialize at the beginning of training: Among them, R is the matrix of corresponding dimension of random assignment, and I is the unit matrix.

4. The angular acceleration sensor denoising method based on a dilated convolutional neural network as claimed in claim 1 or 2, characterized in that: The Adam optimizer is used to update the parameters of the dilated convolutional neural network.

Images