A method for denoising accelerometer shock signals based on underconvergent neural networks

Abstract

The present invention discloses an accelerometer impact signal denoising method based on an underconverged neural network, which belongs to the field of accelerometer impact signal processing. The present invention is implemented by: using CEEMD to process the original accelerometer impact signal to obtain IMF components and residuals, calculating the overclocking frequency, and adding the IMF components and residuals whose principal component frequencies are less than the overclocking frequency to obtain an over-denoised accelerometer signal; building a CNN-LSTM neural network, using a small number of training samples and a low number of training rounds, training step size and learning rate to train the underconverged neural network to obtain weight parameters; performing denoising on the accelerometer impact signal to be processed, and using the CNN-LSTM neural network with a limited convergence rate to maintain the denoising result between the original accelerometer impact signal and the over-denoised accelerometer signal obtained by using CEEMD processing, thereby obtaining a denoising result signal that is closer to a pure, noise-free accelerometer impact signal.

Description

Technical Field

[0001] The present invention relates to an accelerometer impact signal denoising method based on an underconvergence neural network, and in particular to an accelerometer impact signal denoising method based on CNN-LSTM and CEEMD, belonging to the field of accelerometer impact signal processing. Background Art

[0002] High-G accelerometers are widely used in various high-shock measurements, but they require shock testing and calibration to improve measurement accuracy. High-energy shocks are required to stimulate the performance of high-G accelerometers, but these high-energy shocks introduce significant noise into the accelerometer's shock signal. This noise significantly interferes with shock test accuracy, necessitating the development of a method that can accurately and effectively reduce noise in accelerometer shock signals.

[0003] Neural networks are widely used for noise reduction in various signals due to their ability to automatically extract features, demonstrating excellent noise reduction capabilities. However, they have two drawbacks: they require pure, noise-free signals for training, and they require a large amount of training data and time to ensure convergence. For accelerometer shock signals, obtaining a pure, noise-free signal is difficult, and large amounts of data are difficult to obtain. Summary of the Invention

[0004] The existing accelerometer impact signal noise reduction method is not accurate enough. Excessive noise will seriously affect the calibration of the accelerometer's dynamic characteristics such as frequency response, while excessive denoising will seriously affect the calibration of the accelerometer's static characteristics such as sensitivity. Neural networks are excellent noise reduction methods, but they require pure noise-free signals and a large amount of training data and training time, and are difficult to be applied to accelerometer impact signal noise reduction. In response to the above problems, the main purpose of the present invention is to provide an accelerometer impact signal noise reduction method based on an underconvergent neural network, apply the neural network to accelerometer impact signal noise reduction, and achieve efficient and accurate noise reduction of the accelerometer impact signal, thereby improving the accuracy of the impact test and improving the calibration accuracy of parameters such as the sensitivity of the accelerometer.

[0005] The objectives of the present invention are achieved through the following technical solutions.

[0006] The present invention discloses an accelerometer impact signal denoising method based on an underconvergence neural network. The method uses CEEMD (complementary set empirical mode decomposition) to process the original accelerometer impact signal to obtain IMF (intrinsic mode function) components and residuals, obtains the principal component frequency of each IMF component, calculates the overclocking coefficient according to the sampling rate, the principal component frequency of IMF1 and the principal component frequency of IMF2, multiplies the principal component frequency of IMF1 by the overclocking coefficient to obtain the overclocking frequency, and adds the IMF components with a principal component frequency less than the overclocking frequency and the residuals to obtain the over-denoised accelerometer signal; builds a CNN-LSTM (convolutional neural network-long short-term memory network) neural network, and converts a small number of over-denoised components obtained into a denoised accelerometer signal. The accelerometer impact signal with noise is used as output and the corresponding original accelerometer impact signal as input, and a low number of training rounds, training step size and learning rate are used to train an underconverged neural network to obtain weight parameters; the weight parameters of the obtained underconverged neural network are used to perform denoising on the accelerometer impact signal to be processed, and a CNN-LSTM neural network with limited convergence speed is used to keep the denoising result between the original accelerometer impact signal and the over-denoised accelerometer signal obtained by CEEMD processing, so as to obtain a denoising result signal that is closer to the pure noise-free accelerometer impact signal, thereby achieving efficient and accurate denoising of the accelerometer impact signal, and thus improving the accuracy of the impact test. Compared with traditional denoising methods, CNN-LSTM can more accurately and effectively reduce the noise in the accelerometer impact signal, but requires a pure noise-free signal for training, and requires a large amount of training data and training time to ensure the convergence speed. The present invention uses CEEMD to process the original accelerometer impact signal to obtain an over-denoised accelerometer signal, and uses the over-denoised accelerometer signal as the input of CNN-LSTM for training, which can solve the problem of not being able to obtain a pure noise-free signal. When using a small amount of training data, a lower number of training rounds, training step size and learning rate are used to control the convergence speed of the CNN-LSTM neural network. By setting the over-denoised signal as input, a neural network that stops at the convergence process is trained, which is then applied to denoising to obtain ideal noise reduction results.

[0007] The present invention discloses a method for reducing noise of an accelerometer impact signal based on an underconvergent neural network, comprising the following steps:

[0008] S1, use CEEMD to process the original accelerometer impact signal to obtain IMF components and residuals, and obtain the principal component frequency of each IMF component. The overclocking coefficient is calculated based on the sampling rate, the principal component frequency of IMF1, and the principal component frequency of IMF2. The overclocking frequency is obtained by multiplying the principal component frequency of IMF1 by the overclocking coefficient. The IMF components with principal component frequencies less than the overclocking frequency and the residuals are added to obtain the over-denoised accelerometer signal.

[0009] S11, adding a small number of collected original accelerometer impact signals to Gaussian white noise with opposite amplitude to obtain a noisy signal, obtaining an envelope of the noisy signal, determining multiple groups of IMF components by judging whether the difference between the original accelerometer impact signal and the average values ​​of the upper and lower envelopes meets a condition, averaging the obtained multiple groups of IMF components in the decomposition order to obtain final IMF components, and subtracting the sum of all the obtained IMF components from the original accelerometer impact signal to obtain a residual;

[0010] S12, obtaining the frequency spectrum of each IMF component obtained in step S11, and determining the main component frequency of each IMF component;

[0011] S13, calculate the overclocking coefficient according to the sampling rate, the main component frequency of IMF1 and the main component frequency of IMF2. The overclocking coefficient is shown in expression (1):

[0012]

[0013] Where: cp is the overclocking coefficient, f1 is the main component frequency of IMF1, f2 is the main component frequency of IMF2, and fs is the sampling rate;

[0014] S14, multiplying the main component frequency of IMF1 by the overclocking coefficient obtained in step S13 to obtain the overclocking frequency, removing all IMF components whose main component frequency is greater than the overclocking frequency, and adding the remaining IMF components and the residual to obtain the over-denoised accelerometer signal.

[0015] S2: Build a CNN-LSTM neural network, use a small number of over-denoised accelerometer signals as output and the corresponding original accelerometer impact signals as input, and use a lower number of training rounds, training step size, and learning rate than normal neural network training to train an underconverged neural network and obtain weight parameters.

[0016] S21, build a CNN-LSTM neural network using 2 convolutional layers, 2 pooling layers, 2 LSTM layers and 1 fully connected layer;

[0017] S22, using a small number of over-denoised accelerometer signals obtained in step S14 as output and the corresponding original accelerometer impact signals as input, and training the CNN-LSTM neural network constructed in step S21 using a lower number of training rounds, training step size, and learning rate to obtain weight parameters of the underconverged neural network;

[0018] S3 uses the weight parameters of the obtained under-converged neural network to perform denoising on the accelerometer impact signal to be processed, and utilizes the CNN-LSTM neural network with limited convergence speed to maintain the denoising result between the original accelerometer impact signal and the over-denoised accelerometer signal obtained by CEEMD processing, thereby obtaining a denoising result signal that is closer to a pure and noise-free accelerometer impact signal, achieving efficient and accurate denoising of the accelerometer impact signal, and thereby improving the accuracy of the impact test.

[0019] S31, using the weight parameters of the underconverged neural network obtained in step S22 to perform noise reduction processing on the accelerometer impact signal to be processed, to obtain a noise reduction result signal of the accelerometer impact signal;

[0020] S32 uses a CNN-LSTM neural network with a limited convergence rate to maintain the denoising result between the original accelerometer shock signal and the over-denoised accelerometer signal obtained by CEEMD processing, obtaining a denoising result signal that is closer to a pure, noise-free accelerometer shock signal, thereby achieving efficient and accurate denoising of the accelerometer shock signal. The denoised accelerometer shock signal can provide higher-precision measurement values ​​for shock testing, thereby improving the accuracy of shock testing.

[0021] Beneficial effects:

[0022] 1. Aiming at the problem that the existing accelerometer impact signal denoising method is not accurate enough and seriously affects the accuracy of accelerometer calibration, the present invention discloses an accelerometer impact signal denoising method based on underconvergence neural network, which uses CEEMD to process the original accelerometer impact signal to obtain IMF components and residuals, and obtains the principal component frequency of each IMF component. The overclocking coefficient is calculated according to the sampling rate, the principal component frequency of IMF1 and the principal component frequency of IMF2, and the principal component frequency of IMF1 is multiplied by the overclocking coefficient to obtain the overclocking frequency. The IMF components with a principal component frequency less than the overclocking frequency and the residuals are added to obtain the over-denoised accelerometer signal; a CNN-LSTM neural network is built to obtain a small number of The over-denoised accelerometer signal is used as output and the corresponding original accelerometer shock signal is used as input. A low number of training rounds, training step size and learning rate are used to train an under-converged neural network to obtain weight parameters. The weight parameters of the under-converged neural network are used to perform denoising on the accelerometer shock signal to be processed. A CNN-LSTM neural network with a limited convergence rate is used to maintain the denoising result between the original accelerometer shock signal and the over-denoised accelerometer signal obtained by CEEMD processing, so as to obtain a denoising result signal that is closer to a pure and noise-free accelerometer shock signal, thereby achieving efficient and accurate denoising of the accelerometer shock signal and improving the accuracy of the shock test.

[0023] 2. The present invention discloses a method for denoising accelerometer impact signals based on an underconverged neural network. The processed accelerometer impact signals are relatively complex, making it impossible to obtain a pure, noise-free signal and difficult to construct through simulation. In addition, the complex accelerometer impact test process makes it difficult to obtain a large amount of accelerometer impact signal data. The CNN-LSTM neural network used can more accurately and effectively reduce the noise in the accelerometer impact signal than traditional denoising methods, thereby enabling more accurate calibration of parameters such as accelerometer sensitivity. However, a pure, noise-free signal is required for training, and a large amount of training data is required to ensure convergence speed. The characteristics of CEEMD are used to obtain over-denoised accelerometer signals, and the CNN-LSTM neural network with limited convergence speed is combined to obtain ideal denoising results. That is, the combination of CEEMD and CNN-LSTM can make up for the disadvantage of not being able to obtain a pure, noise-free signal for neural network training, and can obtain accurate and effective denoising results under the conditions of a small amount of training data. The neural network with strong noise reduction performance is successfully applied to the denoising of accelerometer impact signals, improving the accuracy of impact testing, that is, improving the calibration accuracy of parameters such as accelerometer sensitivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The figure is a flow chart of the process of processing the accelerometer impact signal by the noise reduction method of the present invention.

[0025] Figure 2 It is a part of the IMF components obtained by the CEEMD processing in step S1 of the noise reduction method of the present invention.

[0026] Figure 3 It is the over-denoised accelerometer signal obtained by the CEEMD processing in step S1 of the noise reduction method of the present invention.

[0027] Figure 4 This is the result obtained by training the CNN-LSTM neural network in step S2 of the noise reduction method described in the present invention.

[0028] Figure 5 It is the noise reduction result signal of the accelerometer impact signal obtained through step S3 of the noise reduction method of the present invention. DETAILED DESCRIPTION

[0029] In order to illustrate the technical problems solved by the present invention and its beneficial effects, the present invention is further described below with reference to the accompanying drawings and embodiments.

[0030] Reference Figure 1 This embodiment discloses a method for reducing noise of an accelerometer impact signal based on an underconvergent neural network, and the specific steps are as follows:

[0031] S1, use CEEMD to process the original accelerometer impact signal to obtain IMF components and residuals, and obtain the principal component frequency of each IMF component. The overclocking coefficient is calculated based on the sampling rate, the principal component frequency of IMF1, and the principal component frequency of IMF2. The overclocking frequency is obtained by multiplying the principal component frequency of IMF1 by the overclocking coefficient. The IMF components with principal component frequencies less than the overclocking frequency and the residuals are added to obtain the over-denoised accelerometer signal.

[0032] S11, each original accelerometer shock signal is copied into two groups, one group is added with Gaussian white noise of 0.2 times the standard deviation of the original accelerometer shock signal, and the other group is added with Gaussian white noise with the opposite amplitude to the previous group, as shown in expression (1):

[0033]

[0034] Where: M1 is the sum of the original accelerometer impact signal and the added positive amplitude Gaussian white noise, M2 is the sum of the original accelerometer impact signal and the added negative amplitude Gaussian white noise, S is the original accelerometer impact signal, and N is the added Gaussian white noise;

[0035] Repeat the above operation 20 times to obtain 40 sets of signals with Gaussian white noise added. The signal obtained by adding Gaussian white noise to the original accelerometer impact signal is defined as the composite signal. The envelope of the composite signal is obtained by cubic spline interpolation. The IMF components are determined by judging whether the difference between the original accelerometer impact signal and the average value of the upper and lower envelopes meets the conditions.

[0036] The 40 IMF components are averaged in the order of decomposition, and the following is obtained: Figure 2 The IMF components of the final result are shown in expression (2):

[0037]

[0038] Of which: IMF j is the jth IMF component, F ij is the jth IMF component of the i-th signal among the 40 synthetic signals;

[0039] S12, using fast Fourier transform to obtain the frequency spectrum of each IMF component obtained in step S11, and selecting the frequency with the largest intensity in the frequency spectrum as the main component frequency of each IMF component;

[0040] S13, the sampling rate is 1MHz, the main component frequency of IMF1 is 193kHz, the main component frequency of IMF2 is 66kHz, and the overclocking coefficient is calculated as 0.043 by substituting into expression (3);

[0041]

[0042] Where: cp is the overclocking coefficient, f1 is the main component frequency of IMF1, f2 is the main component frequency of IMF2, and fs is the sampling rate;

[0043] S14, multiply the main component frequency 193kHz of IMF1 by the overclocking coefficient 0.043 obtained in step S13 to obtain an overclocking frequency of 8.3kHz, remove all IMF components whose main component frequency is greater than the overclocking frequency 8.3kHz, and add the remaining IMF components and the residual to obtain the over-denoised accelerometer signal, such as Figure 3 shown.

[0044] S2, build a CNN-LSTM neural network, use a small number of over-denoised accelerometer signals as output and the corresponding original accelerometer impact signals as input, use a low number of training rounds, training step size, and learning rate to train an underconverged neural network and obtain weight parameters.

[0045] S21, using 2 convolutional layers, 2 pooling layers, 2 LSTM layers and 1 fully connected layer to build a CNN-LSTM neural network. The output of the 1st layer of CNN is shown in expression (4):

[0046]

[0047] Where: f(.) is the activation function, m is the number of feature maps, k is the number of convolution kernels, * is matrix multiplication, and b is the bias matrix;

[0048] The formula of LSTM is shown in expressions (5)-(9):

[0049] i t =σ(x t W ix +h t-1 W ih +b i ) (5)

[0050] f t =σ(x t W fx +h t-1 W fh +b f ) (6)

[0051] o t =σ(x t W ox +h t-1 W oh +b o ) (7)

[0052] c t =ft *c t-1 +i t *g t (8)

[0053] h t =o t *tanh(c t ) (9)

[0054] Where: i t represents the input gate, f t represents the forget gate, o t represents the output gate, c t Represents memory cells, h t represents the hidden state, W ix is the weight of the input gate input x, W ih is the weight of the input gate input h, W fx is the weight of the forget gate input x, W fh is the weight of the forget gate input h, W ox is the weight of the output gate input x, W oh is the weight of the output gate input h, b i is the bias of the input gate, b f is the bias of the forget gate, b o is the bias of the output gate, x t is the extracted feature vector, g t It is a candidate memory cell for improving memory cells;

[0055] S22, taking the 10 over-denoised accelerometer signals obtained in step S14 as output and the corresponding 10 original accelerometer impact signals as input, and training the CNN-LSTM neural network constructed in step S21, wherein the maximum number of training rounds is set to 10, the minimum training step size is set to 2, and the initial learning rate is set to 0.005, to obtain the weight parameters of the underconverged neural network. The results obtained by training the CNN-LSTM neural network are as follows Figure 4 As shown;

[0056] S3, using the obtained weight parameters to perform noise reduction processing on the accelerometer impact signal to be processed, to obtain a noise reduction result signal of the accelerometer impact signal, thereby achieving efficient and accurate noise reduction processing of the accelerometer impact signal, thereby improving the accuracy of the impact test.

[0057] S31, using the weight parameters obtained in step S22 to perform noise reduction processing on the five accelerometer impact signals to be processed, to obtain noise reduction result signals of the accelerometer impact signals, one of the results is as follows Figure 5 As shown;

[0058] S32, using the obtained weight parameters of the under-converged neural network to perform denoising on the accelerometer impact signal to be processed, using the CNN-LSTM neural network with limited convergence speed to keep the denoising result between the original accelerometer impact signal and the over-denoised accelerometer signal obtained by CEEMD processing, so as to obtain a denoising result signal that is closer to the pure noise-free accelerometer impact signal, thereby achieving efficient and accurate denoising of the accelerometer impact signal, thereby improving the accuracy of the impact test. The high-precision impact test can calibrate the accelerometer more accurately, obtain more accurate accelerometer parameters such as sensitivity, etc., thereby improving the measurement accuracy of the accelerometer.

[0059] Five accelerometer impact signals to be processed were processed using the CEEMD processing of the over-denoised accelerometer signal obtained in step S14 and the method of the present invention. The processing results show that the changes in the peak values ​​of the accelerometer signals using the method of the present invention were -13.96%, -16.78%, -4.42%, -4.42%, and -4.96%, respectively, which are superior to the -33.89%, -32.77%, -25.00%, -24.19%, and -26.83% obtained by CEEMD processing. The average peak value change obtained by the method of the present invention was -8.91%, which was 68.78% lower than the -28.54% obtained by CEEMD processing. This further demonstrates that the method of the present invention can effectively reduce noise while ensuring that the peak value is not changed too much, thereby achieving efficient and accurate noise reduction processing of accelerometer impact signals. When calibrating an accelerometer's dynamic characteristics, such as frequency response, it's necessary to compare the peak values ​​of the accelerometer signal and the excitation signal in the frequency domain, which is primarily affected by noise. The method of the present invention effectively reduces the noise of the accelerometer signal, ensuring the effectiveness of calibrating the accelerometer's dynamic characteristics. When calibrating an accelerometer's static characteristics, such as sensitivity, it's necessary to compare the peak values ​​of the accelerometer signal and the excitation signal in the time domain. The sensitivity is calculated by dividing the peak value of the accelerometer signal by the peak value of the excitation signal. However, the method of the present invention does not significantly alter the peak value of the accelerometer signal, thus avoiding signal distortion. This ensures the effectiveness of calibrating the accelerometer's static characteristics, improves the accuracy of impact testing, and, consequently, improves the measurement accuracy of the accelerometer.

[0060] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection 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 scope of protection of the present invention.

Claims

1. A method for denoising accelerometer impact signals based on an underconvergent neural network, characterized in that: The following steps are included: S1, adding a small number of collected original accelerometer impact signals to Gaussian white noise with opposite amplitude to obtain a noisy signal, obtaining the envelope of the noisy signal, determining whether the difference between the original accelerometer impact signal and the average value of the upper and lower envelopes meets the conditions to determine multiple groups of IMF components, averaging the obtained multiple groups of IMF components in the decomposition order to obtain the final IMF components, and subtracting the sum of all the obtained IMF components from the original accelerometer impact signal to obtain the residual; Obtain the frequency spectrum of each IMF component obtained in step S11, and determine the main component frequency of each IMF component; The overclocking coefficient is calculated based on the sampling rate, the main component frequency of IMF1 and the main component frequency of IMF2. The overclocking coefficient is shown in expression (1): Where: cp is the overclocking coefficient, f1 is the main component frequency of IMF1, f2 is the main component frequency of IMF2, and fs is the sampling rate; Multiplying the main component frequency of IMF1 by the overclocking coefficient obtained in step S13 to obtain the overclocking frequency, removing all IMF components whose main component frequency is greater than the overclocking frequency, and adding the remaining IMF components and the residual to obtain the over-denoised accelerometer signal; S2: Build a CNN-LSTM neural network, use a small number of over-denoised accelerometer signals as output and the corresponding original accelerometer impact signals as input, and train the underconverged neural network by setting the number of training rounds, training step size, and learning rate to obtain weight parameters; S3 uses the weight parameters of the obtained under-converged neural network to perform denoising on the accelerometer impact signal to be processed, and utilizes the CNN-LSTM neural network with limited convergence speed to maintain the denoising result between the original accelerometer impact signal and the over-denoised accelerometer signal obtained by CEEMD processing, thereby obtaining a denoising result signal that is closer to a pure and noise-free accelerometer impact signal, thereby achieving efficient and accurate denoising of the accelerometer impact signal and improving the accuracy of the impact test.

2. The method for denoising an accelerometer impact signal based on an underconvergent neural network according to claim 1, wherein: The implementation method of step S2 is: S21, build a CNN-LSTM neural network using 2 convolutional layers, 2 pooling layers, 2 LSTM layers and 1 fully connected layer; S22, taking a small number of over-denoised accelerometer signals obtained in step S14 as output and the corresponding original accelerometer impact signals as input, and training the CNN-LSTM neural network constructed in step S21 using a lower number of training rounds, training step size and learning rate to obtain the weight parameters of the underconverged neural network.

3. The accelerometer impact signal denoising method based on an underconvergent neural network according to claim 2, wherein: The implementation method of step S3 is: S31, using the weight parameters of the underconverged neural network obtained in step S22 to perform noise reduction processing on the accelerometer impact signal to be processed, to obtain a noise reduction result signal of the accelerometer impact signal; S32 uses a CNN-LSTM neural network with a limited convergence rate to maintain the denoising result between the original accelerometer shock signal and the over-denoised accelerometer signal obtained by CEEMD processing, obtaining a denoising result signal that is closer to a pure, noise-free accelerometer shock signal, thereby achieving efficient and accurate denoising of the accelerometer shock signal. The denoised accelerometer shock signal can provide higher-precision measurement values ​​for shock testing, thereby improving the accuracy of shock testing.

Images