Pressure signal reconstruction method based on deep learning
By combining deep learning technology with the time-frequency feature fusion filling of LSTM and GAIN and the noise reduction method of CEEMDAN-wavelet packet fusion LSTM, the problem of missing dynamic pressure signals and noise interference in extreme environments is solved, and high-precision signal reconstruction and clarity improvement are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-03-31
AI Technical Summary
In extreme environments, dynamic pressure signals are prone to loss and noise interference, which existing technologies struggle to effectively address, leading to inaccurate signal reconstruction and analysis.
We employ a deep learning-based approach, combining time-frequency feature fusion filling of LSTM and GAIN with noise reduction techniques of CEEMDAN-wavelet packet fusion LSTM. Through time-frequency transformation, adaptive decomposition, and wavelet packet analysis, we achieve signal missing filling and noise suppression.
It achieves high-precision signal reconstruction, improves signal-to-noise ratio and goodness of fit, and can maintain signal integrity and clarity in extreme environments, adapting to various noise and missing modes.
Smart Images
Figure CN121765310A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of signal processing and deep learning technology, specifically relating to a reconstruction method for missing filling and noise suppression of dynamic pressure signals under extreme environments, which is particularly suitable for dynamic pressure signal processing under high impact and high noise environments. Background Technology
[0002] In high-speed dynamic testing, dynamic pressure signals are key parameters for evaluating system performance, analyzing dynamic characteristics, and optimizing structural design. This signal records the dynamic data of pressure changes over time during system operation. It is a typical complex nonlinear time-series signal, exhibiting a rapid rise to a peak followed by a slow decay, with a duration typically between 5 and 8 milliseconds and an effective frequency band covering the range of 0 to 5 kHz. However, in actual testing, due to extremely harsh environments such as high temperature, high pressure, strong impact, and severe vibration, testing instruments are frequently damaged or malfunction, resulting in significant gaps or noise interference in the acquired dynamic pressure signals.
[0003] Currently, research on dynamic pressure signals largely focuses on the design and development of testing devices. These devices typically employ high-precision pressure sensors to capture changes in dynamic pressure and convert analog signals into digital signals using high-speed data acquisition equipment. However, data processing and analysis remain a relatively weak area, particularly in signal missing data filling and noise suppression. Traditional signal processing methods, such as linear interpolation, spline interpolation, and mean imputation, perform poorly when dealing with complex and unique time-series data like dynamic pressure signals. They struggle to effectively capture the dynamic characteristics and temporal features of the signal, leading to significant differences between the reconstructed signal and the original signal. Furthermore, machine learning methods based on K-nearest neighbors and random forests may fail to adequately capture this information when processing missing data with complex temporal relationships, resulting in reduced accuracy in predicting long-term trends or periodic patterns.
[0004] In recent years, deep learning technology has made significant progress in the field of time-series signal processing. Its powerful feature learning and time-series data processing capabilities make deep learning models an ideal choice for handling complex time-series data. Models such as convolutional neural networks, recurrent neural networks, autoencoders, and generative adversarial networks have demonstrated good performance in signal missing data filling and noise reduction. However, research on dynamic pressure signals with complex temporal characteristics remains insufficient. Existing deep learning methods often fail to fully integrate the time and frequency domain features of dynamic pressure signals, are not accurate enough in capturing long-term temporal dependencies, and exhibit significant performance degradation in noisy environments. Therefore, there is an urgent need for a dynamic pressure signal reconstruction method that can effectively combine the advantages of signal processing techniques and deep learning to solve the problems of signal missing data and noise interference in extreme environments. Summary of the Invention
[0005] To address the issues of signal loss and noise interference in extreme environments, this invention provides a deep learning-based stress signal reconstruction method. This method fully utilizes the time and frequency domain characteristics of the signal to accurately capture long-term temporal dependencies, maintaining excellent performance even in extreme environments, and providing reliable data support for system performance evaluation and structural optimization.
[0006] This invention is achieved through the following technical solution: a pressure signal reconstruction method based on deep learning, comprising the following steps:
[0007] S1. Classify the input dynamic pressure signal and determine whether the signal is complete;
[0008] S2. For incomplete signals, first use the time-frequency feature fusion filling method based on LSTM and GAIN to fill in the missing data, and then apply the noise reduction method based on CEEMDAN-wavelet packet fusion LSTM to filter out noise.
[0009] S3. For the complete signal, directly apply the noise reduction method based on CEEMDAN-wavelet packet fusion LSTM to remove noise interference;
[0010] S4. Outputs a complete and clear dynamic pressure signal after reconstruction.
[0011] As a further improvement to the technical solution of the present invention, the time-frequency feature fusion and filling method based on LSTM and GAIN is specifically as follows:
[0012] S21. Construct a network structure that integrates LSTM and GAIN, wherein the generator contains a three-layer LSTM network and a serial-parallel dual-branch structure;
[0013] S22. Extract the frequency domain features of the signal using the time-frequency conversion module;
[0014] S23. Use a serial-parallel dual-branch network to fuse the time-domain and frequency-domain characteristics of a signal;
[0015] S24. Introduce masking and hint matrices to accurately locate missing values;
[0016] S25. Optimize the generator and discriminator through adversarial training.
[0017] As a further improvement to the technical solution of the present invention, the time-frequency conversion module uses short-time Fourier transform to convert the input signal from the time domain to the frequency domain to obtain the spectrum of the dynamic pressure signal, and then performs power-law compression on the spectral characteristics to enhance the small amplitude changes in the spectrum.
[0018] As a further improvement to the technical solution of the present invention, the serial-parallel dual-branch network includes a Res-GRU branch and a Transformer branch, wherein:
[0019] The Res-GRU branch mines the temporal dependencies of signals and captures local features in a serial manner;
[0020] The Transformer branch extracts the frequency domain features of the signal in parallel and learns global features;
[0021] The complementary fusion of time-domain features and frequency-domain features is achieved through a feature fusion unit.
[0022] As a further improvement to the technical solution of the present invention, the noise reduction method based on CEEMDAN-wavelet packet fusion LSTM includes:
[0023] S26. Use CEEMDAN to adaptively decompose the signal and obtain multiple intrinsic mode components;
[0024] S27. Screening for noisy components using correlation coefficient and spectral analysis;
[0025] S28. Perform wavelet packet denoising on the noisy components;
[0026] S29. Input all components into an LSTM network for deep denoising;
[0027] S30. Reconstruct the noise-reduced signal.
[0028] As a further improvement to the technical solution of the present invention, the wavelet packet denoising adopts the db8 wavelet basis in the Daubechies wavelet series, with a decomposition level of 6 layers, and the threshold selection adopts the unbiased likelihood estimation principle.
[0029] The stress signal reconstruction method based on deep learning provided by this invention has the following advantages compared with the prior art:
[0030] High-precision filling capability: By fusing time-frequency features with long-term temporal dependencies and combining the adversarial training mechanism of the GAIN network, it can deeply explore the complex patterns and potential distribution characteristics inside the signal, and the filled signal is highly consistent with the real signal. The goodness of fit reaches 0.9736 and 0.9968 at 15dB and 30dB signal-to-noise ratios, respectively.
[0031] Strong noise suppression effect: Combining CEEMDAN's adaptive decomposition, wavelet packet multi-resolution analysis and LSTM's deep denoising capability, noise and signal can be effectively separated. The signal-to-noise ratio of the denoised signal can reach 97.4493, and the cross-correlation coefficient is close to 1.
[0032] Comprehensive feature extraction: By using a serial-parallel dual-branch network structure, the system simultaneously captures the local time-domain features and global frequency-domain features of the signal, achieving effective fusion of multi-scale features and enhancing the model's ability to understand the characteristics of complex signals.
[0033] Strong generalization and robustness: It performs stably in various noise environments and missing modes, and can adapt to different test conditions and signal characteristics, providing reliable protection for applications in real-world environments.
[0034] End-to-end reconfiguration capability: Provides a complete signal processing flow, capable of simultaneously resolving signal loss and noise interference issues, and is suitable for the entire process of dynamic pressure signal processing in extreme environments. Attached Figure Description
[0035] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0037] Figure 1 This is the overall technical roadmap of the dynamic pressure signal reconstruction method of the present invention.
[0038] Figure 2 This is a block diagram of a time-frequency feature fusion and filling method based on LSTM and GAIN.
[0039] Figure 3 This is a schematic diagram of a parallel dual-branch network structure.
[0040] Figure 4 This is a diagram of the Res-GRU module structure.
[0041] Figure 5 This is the structure diagram of the Trans module.
[0042] Figure 6 This is a structural diagram of the feature fusion unit.
[0043] Figure 7 This is a block diagram of the dynamic pressure testing device.
[0044] Figure 8 This is a diagram showing the reconstruction results of the measured dynamic pressure signal.
[0045] Figure 9 This is a graph showing the reconstruction result of the 15dB measured dynamic pressure signal.
[0046] Figure 10 This is a diagram showing the reconstruction result of the measured dynamic pressure signal at 30dB. Detailed Implementation
[0047] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.
[0048] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.
[0049] The specific embodiments of the present invention will be described in detail below.
[0050] A deep learning-based method for reconstructing stress signals includes the following steps:
[0051] S1. Classify the input dynamic pressure signal to determine its integrity, specifically:
[0052] First, an experimental dataset was created. Based on relevant data from laboratory simulated dynamic pressure test signals, 200 sets of data with the same shape characteristics were fitted, each containing 990 time points. To enrich the dynamic pressure signal characteristics, Gaussian white noise with a signal-to-noise ratio of 15dB and 30dB was added respectively, resulting in a total of 400 sets of noisy simulated dynamic pressure signal data.
[0053] The data in the experimental dataset were randomly combined into 100 matrices of size 10×990. The experiment simulated a situation where the measurement data was missing due to sensor damage. The time intervals from 260 to 990 in columns 3, 5, 8, and 10 were processed to remove missing values, which were represented by NaN. The remaining columns contained complete signal data.
[0054] The experimental dataset was divided into training and test sets in a 2:1 ratio. The infilling model was trained using the training set data, and the trained model was tested using the test set data to verify its infilling performance.
[0055] S2. For incomplete signals, first use the time-frequency feature fusion filling method based on LSTM and GAIN to fill in the missing data, and then apply the noise reduction method based on CEEMDAN-wavelet packet fusion LSTM to filter out noise.
[0056] The time-frequency feature fusion and filling method based on LSTM and GAIN is as follows:
[0057] S21. Construct a network structure that integrates LSTM and GAIN, wherein the generator contains a three-layer LSTM network and a serial-parallel dual-branch structure.
[0058] The serial-parallel dual-branch network consists of an embedding layer, a dual-branch structure, a feature fusion unit, and a linear layer for output. The dual-branch structure includes a Res-GRU branch and a Transformer branch. The Res-GRU branch contains a GRU layer, an activation function layer, and a LayerNorm layer for feature normalization and stabilizing the training process. The Trans module in the Transformer branch contains a multi-head self-attention mechanism, a feedforward neural network, and a LayerNorm layer, where the feedforward neural network uses the ReLU activation function for nonlinear transformation.
[0059] The Res-GRU branch mines the temporal dependencies of signals and captures local features in a serial manner;
[0060] The Transformer branch extracts the frequency domain features of the signal in parallel and learns global features;
[0061] The complementary fusion of time-domain features and frequency-domain features is achieved through a feature fusion unit.
[0062] S22. Extract the frequency domain features of the signal through the time-frequency conversion module; the time-frequency conversion module uses short-time Fourier transform to convert the input signal from the time domain to the frequency domain to obtain the spectrum of the dynamic pressure signal, and then performs power-law compression on the spectral features to enhance the small amplitude changes in the spectrum.
[0063] S23. Use a serial-parallel dual-branch network to fuse the time-domain and frequency-domain characteristics of a signal;
[0064] S24. Introduce masking and hint matrices to accurately locate missing values;
[0065] S25. Optimize the generator and discriminator through adversarial training.
[0066] The discriminator architecture consists of a multi-layer fully connected neural network, with each neuron fully connected to the layers above and below it. The hidden layers use the LeakyReLU activation function, and the output layer uses the sigmoid activation function to restrict the output value to the range of 0 to 1.
[0067] During training, the network hyperparameters were set as follows: BatchSize was 100, the initial learning rate was 0.01, the Adam algorithm was used to update the learning rate, and the maximum number of iterations was 1000.
[0068] The noise reduction method based on CEEMDAN-wavelet packet fusion LSTM includes:
[0069] S26. Adaptively decompose the signal using CEEMDAN to obtain multiple intrinsic mode components; specifically, perform CEEMDAN decomposition on the padded complete signal to obtain a series of intrinsic mode components. Set the CEEMDAN algorithm to add Gaussian white noise with a standard deviation of 0.2, add it 100 times, and set the maximum envelope count to 500.
[0070] After CEEMDAN decomposition, the original dynamic pressure signal was decomposed into 14 intrinsic mode components.
[0071] S27. Screening noisy components through correlation coefficient and spectrum analysis; specifically: using the numpy.corrocef function to calculate the correlation coefficient between each modal component and the original signal, and according to the evaluation criteria of Pearson correlation coefficient, screening components with correlation coefficients less than 0.14 as noisy components.
[0072] S28. Perform wavelet packet denoising on the noisy components; perform Fast Fourier Transform analysis on components IMF1 to IMF14 to observe the frequency characteristics of each component. Combined with correlation coefficient analysis, it is determined that components IMF1 to IMF9 contain a large amount of interference signals.
[0073] Wavelet packet denoising was performed on the components IMF1 to IMF9. The db8 wavelet basis was selected, the decomposition level was 6, and the threshold was selected by unbiased likelihood estimation.
[0074] S29. Input all components into the LSTM network for deep denoising; input the IMF components after wavelet packet denoising and the unprocessed IMF components IMF10~IMF14 into the LSTM network for training, and use the memory mechanism and gating structure of LSTM to remove deep noise in the signal while retaining the key features of the signal.
[0075] S30. Reconstruct the denoised signal. Finally, the IMFs after LSTM denoising are linearly summed to reconstruct the clean denoised signal.
[0076] S3. For complete signals, directly apply the noise reduction method based on CEEMDAN-wavelet packet fusion LSTM to remove noise interference; classify the acquired signals to determine if there are any missing signals. For signals with missing signals, first use the above-mentioned missing signal filling method to complete the signal, and then apply the above-mentioned noise reduction method to filter the noise. For complete signals, directly apply the above-mentioned noise reduction method to remove noise interference.
[0077] S4. Outputs a complete and clear dynamic pressure signal after reconstruction.
[0078] To verify the practicality of the reconstruction method, the measured dynamic pressure signals were reconstructed. A set of dynamic pressure signals measured at a test site was selected, but signal loss occurred during the test due to a power outage of the sensor. To ensure the accuracy of the reconstruction, complete signal data obtained under the same test conditions were selected as a reference group.
[0079] Goodness of fit, signal-to-noise ratio, and root mean square error are used as evaluation metrics, where the formula for calculating goodness of fit is:
[0080]
[0081] In the formula: This represents the actual value of the real dynamic pressure signal at time point i. This represents the predicted value of the reconstructed dynamic pressure signal at time point i. This represents the average value of the actual dynamic pressure signal at all points in time.
[0082] The formula for calculating the signal-to-noise ratio is:
[0083]
[0084] In the formula: This represents the value of the signal at the i-th sampling point after noise reduction.
[0085] The formula for calculating the root mean square error is:
[0086]
[0087] In the formula: This represents the total number of time points in the signal.
[0088] Experimental results show that the reconstructed dynamic pressure signal waveform exhibits good integrity. The entire signal time series is continuous and smooth, without obvious discontinuities or missing parts, ensuring the signal's continuity in the time domain. The reconstructed waveform retains the key features of the original signal and maintains consistency with the original signal in timing, without any time shift or distortion. Compared to the original signal, noise in the reconstructed waveform is effectively suppressed; high-frequency noise and interference originally present in the signal are significantly reduced, resulting in a clearer waveform that is easier for subsequent analysis and processing.
[0089] As attached Figure 9 and attached Figure 10 The reconstructed dynamic pressure signal, as shown in the figure, exhibits excellent performance across all evaluation metrics after processing using the method of this invention. When performing simple noise reduction on a simulated signal with 30dB Gaussian white noise, the signal-to-noise ratio (SNR) reaches 97.4493, demonstrating the extreme performance of the noise reduction module. When reconstructing a measured signal containing both missing data and complex noise, the goodness of fit reaches 0.9936, indicating a high degree of consistency between the reconstructed and original signals. The SNR still reaches 96.5568, significantly higher than the noise level of the original signal. The root mean square error (RMSE) is 1.78 × 10⁻³, indicating minimal difference between the reconstructed and original signals. This demonstrates the strong robustness and comprehensive performance of this method in complex real-world environments. The differences between the two methods are within a reasonable range and both significantly exceed the performance levels of traditional methods. These data fully validate the effectiveness and superiority of the method of this invention in practical applications.
[0090] The specific noise reduction performance comparison is shown in the table below:
[0091]
[0092] The performance comparison of the reconstruction method of this invention is shown in the table below:
[0093]
[0094] The dynamic pressure signal reconstruction method provided by this invention can be widely applied in fields such as high-speed dynamic testing, system performance evaluation, and structural optimization design. This method effectively solves the problems of signal loss and noise interference under extreme environments, providing reliable technical support for the research, development, testing, and performance evaluation of industrial equipment. It has significant industrial application value and broad market prospects.
[0095] This invention is not only applicable to dynamic pressure signal processing, but can also be extended to the processing of other dynamic test signals, such as vibration test signals and acceleration test signals, and has strong versatility and scalability.
[0096] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.
Claims
1. A method for reconstructing stress signals based on deep learning, characterized in that, Includes the following steps: S1. Classify the input dynamic pressure signal and determine whether the signal is complete; S2. For incomplete signals, first use the time-frequency feature fusion filling method based on LSTM and GAIN to fill in the missing data, and then apply the noise reduction method based on CEEMDAN-wavelet packet fusion LSTM to filter out noise. S3. For the complete signal, directly apply the noise reduction method based on CEEMDAN-wavelet packet fusion LSTM to remove noise interference; S4. Outputs a complete and clear dynamic pressure signal after reconstruction.
2. The method for reconstructing pressure signals based on deep learning according to claim 1, characterized in that, The time-frequency feature fusion and filling method based on LSTM and GAIN is as follows: S21. Construct a network structure that integrates LSTM and GAIN, wherein the generator contains a three-layer LSTM network and a serial-parallel dual-branch structure; S22. Extract the frequency domain features of the signal using the time-frequency conversion module; S23. Use a serial-parallel dual-branch network to fuse the time-domain and frequency-domain characteristics of a signal; S24. Introduce masking and hint matrices to accurately locate missing values; S25. Optimize the generator and discriminator through adversarial training.
3. The method for reconstructing stress signals based on deep learning according to claim 2, characterized in that, The time-frequency conversion module uses short-time Fourier transform to convert the input signal from the time domain to the frequency domain, obtaining the spectrum of the dynamic pressure signal. Then, power-law compression is applied to the spectral characteristics to enhance smaller amplitude changes in the spectrum.
4. The method for reconstructing stress signals based on deep learning according to claim 2, characterized in that, The serial-parallel dual-branch network includes a Res-GRU branch and a Transformer branch, wherein: The Res-GRU branch mines the temporal dependencies of signals and captures local features in a serial manner; The Transformer branch extracts the frequency domain features of the signal in parallel and learns global features; The complementary fusion of time-domain features and frequency-domain features is achieved through a feature fusion unit.
5. The pressure signal reconstruction method based on deep learning according to claim 1, characterized in that, The noise reduction method based on CEEMDAN-wavelet packet fusion LSTM includes: S26. Use CEEMDAN to adaptively decompose the signal and obtain multiple intrinsic mode components; S27. Screening for noisy components using correlation coefficient and spectral analysis; S28. Perform wavelet packet denoising on the noisy components; S29. Input all components into an LSTM network for deep denoising; S30. Reconstruct the noise-reduced signal.
6. The method for reconstructing stress signals based on deep learning according to claim 1, characterized in that, The wavelet packet denoising uses the db8 wavelet basis from the Daubechies wavelet series, with a decomposition level of 6 layers, and the threshold selection adopts the unbiased likelihood estimation principle.