A differential grating laser interferometer signal denoising method for impact test
By combining CEEMD and CNN-LSTM, the problem of poor low-frequency noise processing of differential grating laser interferometer signals in impact testing was solved, achieving efficient and accurate signal noise reduction and improving measurement accuracy and signal-to-noise ratio.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2022-09-20
- Publication Date
- 2026-06-19
Smart Images
Figure CN115470823B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a signal denoising method for a differential grating laser interferometer used for impact testing, and more particularly to a signal denoising method for a differential grating laser interferometer used for impact testing based on CEEMD and CNN-LSTM, belonging to the field of impact test signal processing. Background Technology
[0002] Laser interferometers offer advantages such as non-contact measurement, fast dynamic response, wide measurement range, and high precision, making them widely used in various fields. In impact testing, differential laser interferometers are typically used in conjunction with gratings to measure the velocity changes of impacted objects.
[0003] Demodulating the Doppler signal output by a differential grating laser interferometer yields the frequency signal. Substituting the grating parameters allows calculation of the target's velocity, and further integration and differentiation can be used to obtain changes in displacement and acceleration. However, in impact testing, the Doppler signal output by the differential grating laser interferometer is inevitably contaminated by noise from various sources, including micro-vibrations of the cooperative grating and environmental noise caused by the impact. This noise makes signal demodulation difficult, thus reducing the measurement accuracy of the impact test.
[0004] Existing methods for denoising differential grating laser interferometer signals mainly include traditional methods such as wavelet transform, which have limited denoising accuracy. Currently, there are few methods using deep learning for denoising differential grating laser interferometer signals used in impact testing. Deep learning techniques such as CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory) can achieve high-precision denoising. However, CNN ignores temporal information, and LSTM has limited feature extraction capabilities. The denoising ability of deep learning is limited by the amount of training data; it struggles to learn due to insufficient information in the low-frequency band of the Doppler signal, resulting in poor denoising performance in the low-frequency band of differential grating laser interferometer signals. Therefore, it is essential to develop a method that can accurately and effectively reduce the noise of differential grating laser interferometer signals used in impact testing. Summary of the Invention
[0005] To address the problem of insufficient signal noise reduction accuracy in existing differential grating laser interferometers used for impact testing, the main objective of this invention is to provide a signal noise reduction method for differential grating laser interferometers used for impact testing. This method can achieve efficient and accurate noise reduction processing of signals from differential grating laser interferometers used for impact testing, thereby improving the measurement accuracy of impact testing.
[0006] The objective of this invention is achieved through the following technical solutions.
[0007] This invention discloses a method for signal denoising of a differential grating laser interferometer used in impact testing. Based on impact test data parameters, multiple sets of clean differential grating laser interferometer signals are simulated, and multiple sets of noise are also simulated. The clean signals and noise are combined within a predetermined signal-to-noise ratio range to obtain the noisy differential grating laser interferometer signal. This noisy signal is divided into training data and test data. A CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network) neural network is constructed. The training data and the corresponding clean differential grating laser interferometer signals are used as the training set to train the neural network, obtaining weight parameters. These weight parameters are then used to process the test data to obtain the CNN-LSTM denoising result. Finally, the test data is processed using CEEMD (Complementary Set Empirical Mode Decomposition). The process involves obtaining the IMF (Intrinsic Mode Function) components; grouping the IMF components according to their permutation entropy; zeroing the data of the IMF components in the processing group within a predetermined range before the peak value; recording the zeroing point position closest to the predetermined permutation entropy; summing the processed IMF components with the retained group's IMF components to obtain a partial zeroing result of the test data; extracting the partial zeroing result before the recorded zeroing point position closest to the predetermined permutation entropy as the front-end signal; extracting the CNN-LSTM denoising result after the recorded zeroing point position closest to the predetermined permutation entropy as the back-end signal; and concatenating the front-end and back-end signals to obtain the denoised signal of the differential grating laser interferometer signal. This achieves efficient and accurate denoising processing of the differential grating laser interferometer signal used for impact testing, thereby improving the measurement accuracy of impact testing. CNN-LSTM can more accurately and effectively reduce noise in differential grating laser interferometer signals compared to wavelet thresholding. However, its noise reduction effect is limited by the amount of training data. It struggles to learn due to insufficient information in the low-frequency band of the Doppler signal, resulting in poor noise reduction performance in this band. Without changing the amount of training data, using CEEMD to process the low-frequency Doppler signal, which CNN-LSTM is not good at, and replacing the front part of the CNN-LSTM noise reduction result with the part of the zero-setting processing result, while retaining the high-frequency Doppler signal noise reduction result that CNN-LSTM excels at, can further reduce noise and improve the signal-to-noise ratio. In other words, combining CEEMD and CNN-LSTM, compared to using CNN-LSTM alone, can effectively reduce noise in the frequency band signal and achieve high-precision noise reduction.
[0008] This invention discloses a signal noise reduction method for a differential grating laser interferometer used in impact testing, comprising the following steps:
[0009] S1. Based on the impact test data parameters, simulate multiple sets of pure differential grating laser interferometer signals, and then simulate multiple sets of noise. Combine the pure signals and noise of the differential grating laser interferometer with a signal-to-noise ratio within a predetermined range to obtain the noisy signal of the differential grating laser interferometer. Divide the noisy signal of the differential grating laser interferometer into training data and test data.
[0010] S11, based on the velocity and other parameters of the impact test data, select the range of parameters such as the starting frequency and the ending frequency, and generate multiple sets of random parameters with the frequency of the S-shaped function as the simulated pure differential grating laser interferometer signal. The instantaneous frequency is shown in expression (1):
[0011]
[0012] Where: f s It is the starting frequency, f e This is the termination frequency; a and b are parameters that adjust the rise time and rise speed of the frequency.
[0013] The synthesized pure differential grating laser interferometer signal is shown in expression (2):
[0014]
[0015] S12 generates multiple sets of Gaussian white noise as simulated noise;
[0016] S13, combine the pure differential grating laser interferometer signal and noise generated in steps S11 and S12 according to a predetermined signal-to-noise ratio to obtain multiple sets of differential grating laser interferometer noisy signals, and divide the differential grating laser interferometer noisy signals into training data and test data according to a ratio.
[0017] S2. Build a CNN-LSTM neural network. Use the training data and the corresponding pure signal from the differential grating laser interferometer as the training set to train the neural network and obtain the weight parameters. Use the weight parameters to process the test data and obtain the CNN-LSTM noise reduction result.
[0018] S21, a CNN-LSTM neural network is constructed using convolutional layers, pooling layers, LSTM layers, and fully connected layers. The output of the l-th layer of the CNN is shown in expression (3):
[0019]
[0020] 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;
[0021] The formulas for LSTM are shown in expressions (4)-(8):
[0022] i t =σ(x t W ix +h t-1 W ih +b i (4)
[0023] f t =σ(x t W fx +h t-1 W fh +b f (5)
[0024] o t =σ(x t W ox +h t-1 W oh +b o (6)
[0025] c t =f t *c t-1 +i t *g t (7)
[0026] h t =o t *tanh(c t (8)
[0027] Where: i t f represents the input gate. t Represents the Gate of Oblivion, o t Represents the output gate, c t Represents memory cells, h t W represents the hidden state. ix W is the weight of the input term x of the input gate. ih W is the weight of the input term h of the input gate. fx W is the weight of the forget gate input x. fh W is the weight of the forget gate input term h. ox W is the weight of the input term x of the output gate. oh The weights of the output gate input term h, b i It is the deviation of the input gate, b f It's a deviation from the forgetting gate, b o It is the deviation of the output gate, x t It is the extracted feature vector, g t It is a candidate memory cell for improving memory cells;
[0028] S22, use the training data obtained in step S13 and the corresponding pure signal of the differential grating laser interferometer obtained in step S11 to train the constructed neural network and obtain the weight parameters.
[0029] S23. Use the weight parameters obtained in step S22 to process the test data obtained in step S13 to obtain the CNN-LSTM noise reduction result.
[0030] S3. Perform CEEMD processing on the test data to obtain each IMF component.
[0031] S31, copy each test data obtained in step S13 into two groups. Add Gaussian white noise with a set standard deviation to one group, and add Gaussian white noise with the opposite amplitude to the first group to the other group, as shown in expression (9):
[0032]
[0033] Where: M1 is the sum of the test data and the added positive amplitude Gaussian white noise, M2 is the sum of the test data and the added negative amplitude Gaussian white noise, S is the test data, and N is the added Gaussian white noise;
[0034] S32, repeat step S31 n times to obtain 2n groups of signals with added Gaussian white noise. The signal obtained after adding Gaussian white noise to the test data is defined as the synthesized signal. The envelope of the synthesized signal is obtained. Each IMF component is determined by judging whether the difference between the test data and the average value of the upper and lower envelopes meets the conditions.
[0035] S33, average the multiple IMF components obtained in step S32 according to the decomposition order, and obtain the final IMF components as shown in expression (10):
[0036]
[0037] Among them: IMF j It is the j-th IMF component, F ij It is the j-th IMF component of the i-th signal in 2n synthesized signals.
[0038] S4. Group each IMF component according to the permutation entropy, set the data of the IMF component in the processing group to zero within a predetermined range before the peak, and record the zero point position closest to the predetermined permutation entropy.
[0039] S41, use the permutation entropy algorithm to group the IMF components obtained in step S33. The IMF components with permutation entropy values greater than a specified value are processed, and the rest are retained. The standardized permutation entropy values are shown in expression (11):
[0040]
[0041] Where: m is the size of the phase space dimension, and P is the probability of each size relationship permutation;
[0042] S42, set the data of the IMF components divided into processing groups in step S41 to zero within a predetermined range before the peak, and record the zeroing point position closest to the predetermined permutation entropy.
[0043] S5. The processed IMF component is summed with the IMF component of the retained group to obtain the partial zeroing result of the test data. The part of the zeroing result before the recorded position closest to the predetermined permutation entropy is taken as the front segment signal. The CNN-LSTM denoising result after the recorded position closest to the predetermined permutation entropy is taken as the back segment signal. The front segment signal and the back segment signal are spliced together to obtain the denoising result signal of the differential grating laser interferometer signal. This achieves efficient and accurate denoising processing of the differential grating laser interferometer signal used for impact testing, thereby improving the measurement accuracy of impact testing.
[0044] S51, the processed IMF component obtained in step S42 is summed with all IMF components in the retained group obtained in step S41 to obtain the partial zeroing result of the test data.
[0045] S52, extract the partial zeroing processing result obtained in step S51 before the zeroing point position closest to the predetermined permutation entropy recorded in step S42 as the front-end signal.
[0046] S53, the CNN-LSTM denoising result obtained in step S23 after the position closest to the predetermined permutation entropy zeroing point recorded in step S42 is taken as the subsequent signal. The preceding signal obtained in step S52 is concatenated with the subsequent signal to obtain the denoised result signal of the differential grating laser interferometer signal. This achieves efficient and accurate denoising processing of the differential grating laser interferometer signal used for impact testing. The denoised differential grating laser interferometer signal can provide higher precision measurement values for impact testing and improve the measurement accuracy of impact testing.
[0047] Beneficial effects:
[0048] 1. To address the insufficient accuracy of existing signal denoising methods for differential grating laser interferometers used in impact testing, this invention discloses a signal denoising method for differential grating laser interferometers used in impact testing. The method involves performing CEEMD processing on the test data to obtain individual IMF components; grouping the IMF components according to their permutation entropy; zeroing the data of the processed group's IMF components within a predetermined range before the peak value and recording the zero-point position closest to the predetermined permutation entropy; summing the processed IMF components with the retained group's IMF components to obtain a partial zero-point processing result of the test data; extracting the portion of the zero-point processing result before the recorded zero-point position closest to the predetermined permutation entropy as the front-end signal; and extracting the CNN-LSTM denoising result after the recorded zero-point position closest to the predetermined permutation entropy as the back-end signal; and concatenating the front-end and back-end signals to obtain the denoised signal of the differential grating laser interferometer signal. This method achieves efficient and accurate denoising processing of differential grating laser interferometer signals used in impact testing, thereby improving the measurement accuracy of impact testing.
[0049] 2. This invention discloses a signal noise reduction method for differential grating laser interferometers used in impact testing. The CNN-LSTM method combines the advantages of CNN and LSTM, and can more accurately and effectively reduce noise in differential grating laser interferometer signals compared to wavelet thresholding. However, the noise reduction effect is limited by the amount of training data, and it is difficult to learn due to insufficient information in the low-frequency band of the Doppler signal, resulting in poor noise reduction performance in the low-frequency band of the differential grating laser interferometer signal. Without changing the amount of training data, CEEMD is used to process the low-frequency Doppler signal, which CNN-LSTM is not good at. The front part of the partially zeroed processing result replaces the front part of the CNN-LSTM noise reduction result, while retaining the high-frequency Doppler signal noise reduction result that CNN-LSTM excels at. This further reduces noise and improves the signal-to-noise ratio. In other words, the combination of CEEMD and CNN-LSTM, compared to using CNN-LSTM alone, can effectively reduce noise in the frequency band signal, achieving high-precision noise reduction. Attached Figure Description
[0050] Figure 1 This is a flowchart illustrating the signal processing procedure of a differential grating laser interferometer used for impact testing using the noise reduction method described in this invention.
[0051] Figure 2 This is a comparison between the CNN-LSTM denoising result obtained through step S2 of the denoising method described in this invention and the noisy signal and the clean signal.
[0052] Figure 3 This refers to a portion of the IMF component of the CEEMD processing result obtained through step S3 of the noise reduction method described in this invention.
[0053] Figure 4 These are the IMF1 components before and after processing obtained through step S4 of the noise reduction method described in this invention.
[0054] Figure 5 This is a comparison of the denoised signal of the differential grating laser interferometer obtained by step S5 of the denoising method described in this invention with the wavelet thresholding result signal and the result signal obtained by using only CNN-LSTM. Detailed Implementation
[0055] To illustrate the technical problems solved by the present invention and its beneficial effects, the invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0056] Reference Figure 1 This embodiment discloses a signal noise reduction method for a differential grating laser interferometer used in impact testing, the specific steps of which are as follows:
[0057] S1. Based on the impact test data parameters, simulate multiple sets of pure differential grating laser interferometer signals, and then simulate multiple sets of noise. Combine the pure signals and noise of the differential grating laser interferometer with a signal-to-noise ratio within a predetermined range to obtain the noisy signal of the differential grating laser interferometer. Divide the noisy signal of the differential grating laser interferometer into training data and test data.
[0058] S11, based on the velocity and other parameters of the impact test data, select the range of parameters such as the starting frequency and the ending frequency, and generate a sweep frequency signal with 100 sets of random parameters, 480,000 points, S-shaped frequency, and a sampling rate of 50MHz as the simulated pure differential grating laser interferometer signal. The instantaneous frequency is shown in expression (1):
[0059]
[0060] Where: f s It is the starting frequency, f e This is the termination frequency; a and b are parameters that adjust the rise time and rise speed of the frequency.
[0061] The synthesized pure differential grating laser interferometer signal is shown in expression (2):
[0062]
[0063] S12, generate 100 groups of 480,000 points of Gaussian white noise with a sampling rate of 50MHz as analog noise;
[0064] S13, combine the first 90 sets of pure differential grating laser interferometer signals and noise generated in steps S11 and S12 according to the signal-to-noise ratio in the range of 1dB to 10dB to obtain 90 sets of noisy differential grating laser interferometer signals as training data; combine the last 10 sets of pure differential grating laser interferometer signals and noise according to 10 signal-to-noise ratios from 1dB, 2dB to 10dB respectively to obtain 100 sets of noisy laser Doppler velocimeter signals as test data.
[0065] S2. Build a CNN-LSTM neural network. Use the training data and the corresponding pure signal from the differential grating laser interferometer as the training set to train the neural network and obtain the weight parameters. Use the weight parameters to process the test data and obtain the CNN-LSTM noise reduction result.
[0066] S21, a CNN-LSTM neural network is constructed using 2 convolutional layers, 2 pooling layers, 2 LSTM layers, and 1 fully connected layer. The output of the l-th layer of the CNN is shown in expression (3):
[0067]
[0068] 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;
[0069] The formulas for LSTM are shown in expressions (4)-(8):
[0070] i t =σ(x t W ix +h t-1 W ih +b i (4)
[0071] f t =σ(x t W fx +h t-1 W fh +b f (5)
[0072] o t =σ(x t W ox +h t-1 W oh +b o (6)
[0073] c t =f t *c t-1 +i t *g t (7)
[0074] h t =o t *tanh(c t (8)
[0075] Where: i t f represents the input gate. t Represents the Gate of Oblivion, o t Represents the output gate, c t Represents memory cells, h t W represents the hidden state. ix W is the weight of the input term x of the input gate. ih W is the weight of the input term h of the input gate. fx W is the weight of the forget gate input x. fh W is the weight of the forget gate input term h. ox W is the weight of the input term x of the output gate. oh The weights of the output gate input term h, b i It is the deviation of the input gate, b f It's a deviation from the forgetting gate, b o It is the deviation of the output gate, x t It is the extracted feature vector, g t It is a candidate memory cell for improving memory cells;
[0076] S22, use the training data obtained in step S13 and the corresponding pure signal of the differential grating laser interferometer obtained in step S11 to train the constructed neural network and obtain the weight parameters.
[0077] S23, using the weight parameters obtained in step S22, process the test data obtained in step S13 to obtain the CNN-LSTM noise reduction result, such as... Figure 2 As shown.
[0078] S3. Perform CEEMD processing on the test data to obtain each IMF component.
[0079] S31, copy each test data obtained in step S13 into two groups. Add Gaussian white noise of 0.2 times the standard deviation of the noise reduction result signal to one group, and add Gaussian white noise with the opposite amplitude to the first group to the other group, as shown in expression (9):
[0080]
[0081] Where: M1 is the sum of the test data and the added positive amplitude Gaussian white noise, M2 is the sum of the test data and the added negative amplitude Gaussian white noise, S is the test data, and N is the added Gaussian white noise;
[0082] S32, repeat step S31 20 times to obtain 40 sets of signals with added Gaussian white noise. The signal obtained after adding Gaussian white noise to the test data is defined as the synthetic signal. The envelope of the synthetic signal is obtained by cubic spline interpolation. Each IMF component is determined by judging whether the difference between the test data and the average value of the upper and lower envelopes meets the conditions.
[0083] S33, average the 40 IMF components obtained in step S32 according to the decomposition order, and obtain the final IMF components as shown in expression (10):
[0084]
[0085] Among them: IMF j It is the j-th IMF component, F ij It is the j-th IMF component of the i-th signal in the 40 synthesized signals; the IMF components of the CEEMD processing result are obtained, IMF1, IMF2, IMF3 and the remaining parts are as follows: Figure 3 As shown.
[0086] S4. Group each IMF component according to the permutation entropy, set the data of the IMF component in the processing group to zero within a predetermined range before the peak, and record the zero point position closest to the predetermined permutation entropy.
[0087] S41, the IMF components obtained in step S33 are grouped using the permutation entropy algorithm. The IMF components with permutation entropy values greater than 0.4 are processed, and the rest are retained. The standardized permutation entropy values are shown in expression (11):
[0088]
[0089] Where: m is the size of the phase space dimension, and P is the probability of each size relationship permutation;
[0090] S42, calculate the absolute value of the IMF components divided into processing groups in step S41, obtain the average of every 201 points of the absolute value data, and then calculate the average of the maximum and minimum values of the average data as the judgment threshold. When the absolute value data reaches the judgment threshold, record the 5000th point before that point as the zero point, and set the data within the range before the zero point of the IMF component to zero, such as... Figure 4 As shown, the position of the zero point where the permutation entropy is closest to 0.6 is recorded among all IMF components in the processing group.
[0091] S5. The processed IMF component is summed with the IMF component of the retained group to obtain the partial zeroing result of the test data. The part of the zeroing result before the recorded position closest to the predetermined permutation entropy is taken as the front segment signal. The CNN-LSTM denoising result after the recorded position closest to the predetermined permutation entropy is taken as the back segment signal. The front segment signal and the back segment signal are spliced together to obtain the denoising result signal of the differential grating laser interferometer signal. This achieves efficient and accurate denoising processing of the differential grating laser interferometer signal used for impact testing, thereby improving the measurement accuracy of impact testing.
[0092] S51, the processed IMF component obtained in step S42 is summed with all IMF components in the retained group obtained in step S41 to obtain the partial zeroing result of the test data.
[0093] S52, extract the partial zeroing processing result obtained in step S51 before the zeroing point position closest to the predetermined permutation entropy recorded in step S42 as the front-end signal.
[0094] S53, the CNN-LSTM denoising result obtained in step S23 after the position closest to the predetermined permutation entropy zeroing point recorded in step S42 is taken as the subsequent signal. The preceding signal obtained in step S52 is concatenated with the subsequent signal to obtain the denoised signal of the differential grating laser interferometer signal, such as... Figure 5 As shown, efficient and accurate noise reduction processing is achieved for the differential grating laser interferometer signal used for impact testing. The noise-reduced differential grating laser interferometer signal can provide higher accuracy velocity values for impact testing, thereby improving the measurement accuracy of impact testing.
[0095] The average signal-to-noise ratio (SNR) generated in this embodiment is 5.500 dB, and the average root mean square error (RMSE) is 0.395. The average SNR after denoising using the wavelet thresholding method is 13.405 dB, and the average RMSE is 0.156. The average SNR after denoising using the CNN-LSTM method is 20.287 dB, and the average RMSE is 0.069. The average SNR after processing with CEEMD is 12.570 dB, and the average RMSE is 0.176. The average SNR after denoising using the CEEMD and CNN-LSTM combined splicing method proposed in this invention is 23.855 dB, and the average RMSE is 0.048. The signal-to-noise ratio obtained by the method of the present invention is significantly better than that of the method using CNN-LSTM alone (17.59% improvement) and far superior to the wavelet thresholding method (77.96% improvement). Furthermore, the root mean square error obtained by the present invention is significantly lower than that of the CNN-LSTM method (30.43% reduction) and far lower than that of the wavelet thresholding method (69.23% reduction), further demonstrating that the method of the present invention can achieve efficient and accurate noise reduction processing of the differential grating laser interferometer signal used for impact testing.
[0096] After solving for the instantaneous frequency using Hilbert transform, the frequency is converted into velocity based on the grating parameters. Using the velocity obtained from the pure differential grating laser interferometer signal as the standard, the average root mean square error (RMS) of the velocity value after noise reduction using the wavelet thresholding method is calculated to be 4.652 m / s; the average RMS of the velocity value after noise reduction using the CNN-LSTM method is 0.559 m / s; and the average RMS of the velocity value after noise reduction using the proposed method combining CEEMD and CNN-LSTM is 0.489 m / s. The RMS of the velocity values obtained by the method of this invention are significantly smaller than those obtained using the CNN-LSTM method (reduced by 12.52%) and significantly lower than those obtained using the wavelet thresholding method (reduced by 89.49%), further demonstrating that this invention can improve the measurement accuracy of impact testing.
[0097] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. 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 within the scope of protection of the present invention.
Claims
1. A method for noise reduction of a differential grating laser interferometer signal for shock testing, characterized in that, Includes the following steps, S1. Based on the impact test data parameters, simulate multiple sets of pure differential grating laser interferometer signals, and then simulate multiple sets of noise. Combine the pure signals and noise of the differential grating laser interferometer with a signal-to-noise ratio within a predetermined range to obtain the noisy signal of the differential grating laser interferometer. Divide the noisy signal of the differential grating laser interferometer into training data and test data. S2, Build a CNN-LSTM neural network, use the training data and the corresponding pure signal of the differential grating laser interferometer as the training set to train the neural network, obtain the weight parameters, use the weight parameters to process the test data, and obtain the CNN-LSTM noise reduction result; S3, perform CEEMD processing on the test data to obtain each IMF component; S4. Group each IMF component according to the permutation entropy, set the data of the IMF component in the processing group within a predetermined range before the peak to zero, and record the zero point position closest to the predetermined permutation entropy. S5. The processed IMF component is summed with the IMF component of the retained group to obtain the partial zeroing result of the test data. The part of the zeroing result before the recorded position closest to the predetermined permutation entropy is taken as the front segment signal. The CNN-LSTM denoising result after the recorded position closest to the predetermined permutation entropy is taken as the back segment signal. The front segment signal and the back segment signal are spliced together to obtain the denoising result signal of the differential grating laser interferometer signal. This achieves efficient and accurate denoising processing of the differential grating laser interferometer signal used for impact testing, thereby improving the measurement accuracy of impact testing.
2. A differential grating laser interferometer signal denoising method for shock testing as claimed in claim 1, wherein, The method for implementing step S1 is as follows: S11, based on the velocity parameters of the impact test data, select the range of the starting frequency and the ending frequency parameters, and generate multiple sets of random parameters with frequencies that are S-shaped functions as simulated pure differential grating laser interferometer signals. The instantaneous frequency is shown in expression (1): （1） where: f s is the starting frequency, f e is the ending frequency, a, b are parameters for adjusting the frequency rise time and the rise speed; The synthesized pure differential grating laser interferometer signal is shown in expression (2): （2） S12 generates multiple sets of Gaussian white noise as simulated noise; S13, combine the pure differential grating laser interferometer signal and noise generated in steps S11 and S12 according to a predetermined signal-to-noise ratio to obtain multiple sets of differential grating laser interferometer noisy signals, and divide the differential grating laser interferometer noisy signals into training data and test data according to a ratio.
3. The signal noise reduction method for a differential grating laser interferometer used in impact testing as described in claim 2, characterized in that, The implementation method for step S2 is as follows: S21, a CNN-LSTM neural network is constructed using convolutional layers, pooling layers, LSTM layers, and fully connected layers. The output of the l-th layer of the CNN is shown in expression (3): （3） Where: f(.) is the activation function, m is the number of feature maps, k is the number of convolution kernels, and * represents matrix multiplication. It is the bias matrix; The formulas for LSTM are shown in expressions (4)-(8): （4） （5） （6） （7） （8） Where: i t f represents the input gate. t Represents the Gate of Oblivion, o t Represents the output gate, c t Represents memory cells, h t W represents the hidden state. ix W is the weight of the input term x of the input gate. ih W is the weight of the input term h of the input gate. fx W is the weight of the forget gate input x. fh W is the weight of the forget gate input term h. ox W is the weight of the input term x of the output gate. oh The weights of the output gate with input term h are the weights. It's the input gate's deviation. It's a deviation from the forgetting gate. It is the deviation of the output gate, x t It is the extracted feature vector, g t It is a candidate memory cell for improving memory cells; S22, use the training data obtained in step S13 and the corresponding pure signal of the differential grating laser interferometer obtained in step S11 to train the constructed neural network and obtain the weight parameters. S23. Use the weight parameters obtained in step S22 to process the test data obtained in step S13 to obtain the CNN-LSTM noise reduction result.
4. The signal noise reduction method for a differential grating laser interferometer used in impact testing as described in claim 3, characterized in that, The implementation method for step S3 is as follows: S31, copy each test data obtained in step S13 into two groups. Add Gaussian white noise with a set standard deviation to one group, and add Gaussian white noise with the opposite amplitude to the first group to the other group, as shown in expression (9): （9） Where: M1 is the sum of the test data and the added positive amplitude Gaussian white noise, M2 is the sum of the test data and the added negative amplitude Gaussian white noise, S is the test data, and N is the added Gaussian white noise; S32, repeat step S31 n times to obtain 2n groups of signals with added Gaussian white noise. The signal obtained after adding Gaussian white noise to the test data is defined as the synthesized signal. The envelope of the synthesized signal is obtained. Each IMF component is determined by judging whether the difference between the test data and the average value of the upper and lower envelopes meets the conditions. S33, average the multiple IMF components obtained in step S32 according to the decomposition order, and obtain the final result of each IMF component as shown in expression (10): （10） Among them: IMF j It is the j-th IMF component, F ij It is the j-th IMF component of the i-th signal in 2n synthesized signals.
5. A differential grating laser interferometer signal denoising method for shock testing as claimed in claim 4, wherein, The implementation method for step S4 is as follows: S41, use the permutation entropy algorithm to group the IMF components obtained in step S33. The IMF components with permutation entropy values greater than a specified value are processed, and the rest are retained. The standardized permutation entropy values are shown in expression (11): （11） Where: m is the size of the phase space dimension, and P is the probability of each size relationship permutation; S42, set the data of the IMF components in the processing group divided in step S41 to zero within a predetermined range before the peak value, and record the zero point position closest to the predetermined permutation entropy.
6. A differential grating laser interferometer signal denoising method for shock testing as claimed in claim 5, wherein, The implementation method for step S5 is as follows: S51, the processed IMF component obtained in step S42 is summed with all IMF components in the retained group obtained in step S41 to obtain the partial zeroing result of the test data. S52, extract the partial zeroing processing result obtained in step S51 before the zeroing point position closest to the predetermined permutation entropy recorded in step S42 as the front-end signal. S53, the CNN-LSTM denoising result obtained in step S23 after the position closest to the predetermined permutation entropy zeroing point recorded in step S42 is taken as the subsequent signal. The preceding signal obtained in step S52 is concatenated with the subsequent signal to obtain the denoised result signal of the differential grating laser interferometer signal. This achieves efficient and accurate denoising processing of the differential grating laser interferometer signal used for impact testing. The denoised differential grating laser interferometer signal can provide a higher accuracy velocity value for impact testing and improve the measurement accuracy of impact testing.