Full-connection-convolutional self-encoding network structure response noise reduction method based on noise self-supervision

Abstract

The invention discloses a full-connection-convolutional self-encoding network structure response noise reduction method based on noise self-supervision, and aims to solve the problem that a traditional convolutional self-encoding network is weak in low-frequency noise suppression capability and depends on pure signal samples. Inputting the structure response signal into a full-connection neural network comprising a filling layer, a dimension raising layer and a dimension reducing layer, improving the high-dimensional feature expression capability of the input signal through nonlinear mapping, and enhancing the discrimination of low-frequency random noise; and S2, inputting the feature data into a one-dimensional convolutional self-encoding network, extracting principal component features by using multi-layer convolution, deconvolution, pooling and anti-pooling operations, performing time domain reconstruction, and outputting a structure response signal after noise reduction. According to the method, on the premise of not depending on a pure sample, low-frequency random noise interference can be effectively suppressed, the characteristic resolution capability and the signal-to-noise ratio of the signal after noise reduction are improved, and an efficient data preprocessing means is provided for weak modal recognition of large structures such as ocean engineering under environmental excitation.

Description

TECHNICAL FIELD

[0001] The present application relates to a full-connection-convolutional auto-encoder network structure response noise reduction method based on noise self-supervision, which is used to solve the problem of low-frequency noise insensitivity of traditional self-supervised convolutional noise reduction methods, and provides an effective data preprocessing means for weak modal feature identification of large structures under environmental excitation conditions. The present application relates to the field of noise reduction processing of ocean engineering structure dynamic response data. BACKGROUND

[0002] In the actual operation process of ocean engineering structures, the structure response signal is often affected by multiple sources of noise such as environmental interference and measurement error. Especially under the excitation condition of unstable ocean environment, the acquired field dynamic response data is usually accompanied by significant random noise interference. These noises not only mask the true response characteristics of the structure, but also have an adverse effect on subsequent modal parameter identification and health monitoring. Therefore, how to effectively remove noise while preserving the weak modal characteristics of the original signal is one of the key problems in the processing of ocean engineering structure response signals.

[0003] Traditional signal noise reduction methods are mainly based on physical modeling or mathematical transformation principles, such as wavelet transform, empirical mode decomposition, and spectral subtraction. Although these methods have certain advantages in dealing with stationary or specific types of noise, they often have insufficient generalization ability, rely on human experience, and have poor low-frequency noise suppression ability when faced with non-stationary, non-Gaussian distributed complex ocean environment signals, making it difficult to adapt to complex and changing actual environmental requirements.

[0004] Under the background of ocean engineering big data, machine learning methods based on deep neural networks have been gradually introduced into the noise reduction processing of structure response signals. Especially the convolutional neural network (CNN) and auto-encoder (AE) structures, under the premise of proper design and sufficient training samples, have strong feature extraction and nonlinear mapping capabilities, and can achieve adaptive suppression of complex background noise. Such methods mainly construct network models through the following training paradigms:

[0005] 1. NCT (Noisy-Clean Training): Supervised training using noisy signals and corresponding clean signals;

[0006] 2. NNT (Noisy-Noisy Training): Adding different noise samples to the same clean signal to establish a mutual mapping relationship;

[0007] 3. NerNT (Noisier-Noisy Training): Adding more noise to the existing noisy signal and then training;

[0008] 4.ONT(Only-Noisy Training): A self-supervised training method that relies solely on a single noisy sample.

[0009] Among them, NCT and NNT methods generally have strong noise reduction ability and generalization ability, but the premise is to obtain pure signal samples. However, in the actual measurement of ocean engineering structures, pure signals are often not available. Although the ONT method has the advantage of self-supervision, its modeling effect is highly dependent on the independence distribution assumption of noise, and the actual noise reduction ability is limited.

[0010] In order to improve the performance of the noise reduction model, researchers have tried to introduce data preprocessing or feature enhancement modules to improve the model's expression ability of the original signal. Common strategies include:

[0011] 1. Preprocessing method based on domain transformation: Use Fourier transform, wavelet transform, Laplace transform, etc. to convert one-dimensional time series signal to frequency domain or transform domain to extract more representative spectral features;

[0012] 2. Dimension transformation method: Convert time signal to two-dimensional time-frequency graph, and then use graph neural network or convolutional network for image learning to capture time and frequency coupling features;

[0013] 3. High-dimensional feature transformation method: Introduce deep feature modules such as convolutional layers or fully connected networks to map input data to high-dimensional space, making the feature space have stronger discrimination ability. In this method, preprocessing modules and post-processing modules usually exist in pairs to ensure the consistency of network input and output dimensions.

[0014] Although the above methods improve the expression ability of feature information to some extent, the existing convolution-based noise reduction network still has the following problems:

[0015] The feature extraction stage is easily disturbed by random noise, especially in the case of shallow network or limited convolution kernel receptive field, weak modal features are easily covered;

[0016] There is a lack of effective mechanism to distinguish between useful features and irrelevant noise, especially when dealing with low-frequency random noise, the suppression ability is insufficient;

[0017] The coupling between the feature enhancement module and the noise reduction network is poor, and end-to-end optimization cannot be achieved, which limits the effectiveness of the self-supervised training strategy.

[0018] Therefore, a new structure that combines feature enhancement and deep network representation ability is urgently needed, especially without the need for pure signal, to improve the self-supervised noise reduction method's ability to suppress low-frequency noise and preserve weak modal features, to meet the efficient preprocessing needs of ocean engineering structure dynamic response data. SUMMARY

[0019] The present application aims to solve the problem that the existing self-supervised convolutional denoising network is not sensitive when processing low-frequency random noise, and proposes a full connection-convolutional autoencoder network (FCAE) structure response denoising method based on noise self-supervision. The method introduces a set of full connection structure for dimensionality reduction and dimensionality reduction as a feature enhancement preprocessing module of the convolutional autoencoder network (CAE), which significantly improves the high-dimensional feature expression capability of the original signal, so that the low-frequency random noise can be effectively distinguished and suppressed, overcoming the problem of low-frequency noise retention caused by the band-pass characteristic of the traditional convolutional network.

[0020] The proposed FCAE method adopts a noise self-supervised training strategy, realizes end-to-end training without pure signal samples, is suitable for large-scale structure dynamic response signal denoising tasks under complex environmental excitation conditions, improves the ability to maintain weak modal characteristics and overall denoising performance, and provides an efficient and reliable data preprocessing means for ocean engineering structure health monitoring and modal identification.

[0021] The technical scheme adopted by the present application is:

[0022] A full connection-convolutional autoencoder network structure response denoising method based on noise self-supervision, comprising:

[0023] Step S1: inputting the structure response signal into a feedforward FC neural network comprising a padding layer, a first full connection layer (FC) and a second FC, the padding layer being used for dimension expansion of the input signal to simulate frequency domain or complex domain features, the first FC and the second FC both using a nonlinear activation function to map the padded signal to a higher dimension and then to a lower dimension, forming a time domain feature expression of the input signal while maintaining its original information structure;

[0024] Step S2: inputting the feature data processed in S1 into a one-dimensional convolutional autoencoder network (1D-CAE), the convolutional autoencoder network comprising multiple convolutional layers, deconvolutional layers, pooling layers and de-pooling layers, extracting effective features of the input signal through layer-by-layer convolution, deconvolution, pooling and de-pooling operations, and realizing time domain restoration, and outputting the denoised structure response signal.

[0025] Further, in the aforementioned FC neural network:

[0026] The first FC includes a dimensionality increasing layer for dimensionality increasing of the signal output by the padding layer to improve the ability to distinguish low-frequency random noise features;

[0027] The second FC includes a dimensionality reducing layer for compressing the dimensionality of the dimensionality increased features to the same dimension as the original signal to maintain the consistency of the network input and output.

[0028] Further, the 1D-CAE comprises an encoding layer, a feature layer and a decoding layer, wherein:

[0029] The encoding layer is used to extract the time-domain principal component features of the input signal;

[0030] The feature layer is a compression bottleneck structure, used to represent the key information after dimension reduction;

[0031] The decoding layer is used to reconstruct the signal and realize noise reduction output;

[0032] The encoding layer and the decoding layer are both composed of multiple convolution layers, deconvolution layers, pooling layers and de-pooling layers, to realize layer-by-layer feature extraction and reconstruction of the signal.

[0033] Further, the method adopts a self-supervised training strategy based on the noisy signal, and trains the network model in an end-to-end manner without relying on pure signal samples.

[0034] Further, the structural response signal is an acceleration signal or a strain signal collected under environmental excitation of the ocean engineering structure.

[0035] The beneficial effects of the present application are:

[0036] 1) The present application avoids the problem that the traditional one-dimensional time sequence denoising network has poor denoising effect on real measurement data due to the need for pure signal as a supervision target or the requirement that the noise component meets the independent and identically distributed feature as a prerequisite.

[0037] 2) The present application improves the problem that the traditional convolutional auto-encoding network is not sensitive to low-frequency random noise. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The FCAE network processing steps and structure diagram constructed for the method of the present application;

[0039] Figure 2 The aforementioned FC network structure constructed for the method of the present application;

[0040] Figure 3 The time-frequency comparison of the pure harmonic signal and the noisy signal of the present application;

[0041] Figure 4 The time-frequency comparison of the CAE network signal denoising of the present application (SNR=-10);

[0042] Figure 5 The time-frequency comparison of the FCAE network signal denoising of the present application (SNR=-10). DETAILED DESCRIPTION

[0043] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0044] The application provides a noise self-supervised FCAE network structure response noise reduction method, which is used for solving the problem that a traditional self-supervised convolutional noise reduction network is not sensitive to low-frequency random noise.

[0045] The method comprises the following steps:

[0046] Step S1: inputting a structure response signal into the aforementioned FC neural network comprising a padding layer, a first FC and a second FC. The padding layer is used to expand the dimension of the input signal to simulate the frequency domain or complex domain feature; the first full connection layer and the second full connection layer both adopt a nonlinear activation function (such as ReLU) to perform dimension increasing and dimension reducing mapping on the signal, so as to enhance the time domain feature representation of the input signal and improve the distinguishing ability to low-frequency noise, while maintaining the original information structure thereof.

[0047] Step S2: inputting the feature data processed in step S1 into a 1D-CAE. The convolutional auto-encoding network comprises multiple convolutional layers, deconvolutional layers, pooling layers and de-pooling layers, and through layer-by-layer feature extraction and time domain reconstruction operations, the principal component extraction and noise reduction output of the signal are realized, and finally the structure response signal after noise reduction is generated.

[0048] To further illustrate the feature enhancement mechanism of the full connection structure, as shown in Figure 1 The application introduces a padding layer in the aforementioned FC network structure, which is used to simulate the expansion dimension required for complex or frequency domain mapping. Traditional neural networks often add a preprocessing module at the input end and introduce a postprocessing module at the output end to ensure that the input and output dimensions are consistent, but the introduction of the postprocessing module will significantly increase the model depth and parameter size, reduce the network efficiency and cause overfitting problems. The application directly completes the dimension alignment at the front end through the dimension increasing-dimension reducing symmetric structure, avoids the design burden of the postprocessing module, and improves the generalization ability and computational efficiency of the overall network.

[0049] Specifically, Figure 1 The feature enhancement structure shown in the figure is composed of 5 layers, including: 1 layer of padding layer, 2 layers of dimension increasing FC (FC#1), and 2 layers of dimension reducing FC (FC#2), wherein each full connection layer adopts a nonlinear activation function, the number of neurons of FC#1 is the same as that of the padding layer, and the number of neurons of FC#2 is the same as that of the input data, so as to ensure the consistency of the input and output data dimensions. The connection of FC#1 and FC#2 realizes the dimension reducing calculation of data features.

[0050] As shown in Figure 2As shown, the 1D-CAE part includes three sub-modules of encoding layer, feature layer and decoding layer: the encoding layer is composed of 3 layers of convolution-activation-pooling structure; the feature layer is a bottleneck structure output by the encoder; the decoding layer is composed of 3 layers of deconvolution-activation-depooling structure, and is used for restoring the signal. The convolution structure adopts ReLU as the activation function, and the pooling layer adopts maximum pooling (Max-pooling) with a step of 2 and an input signal length of 2048 sampling points.

[0051] The detailed parameters of each full connection, convolution and deconvolution structure in the FCAE network of the application are shown in Table 1. The adaptive moment estimation (Adam) optimizer is selected as the gradient optimization function for network training, and the MSELoss function, i.e. the L2 norm, is used to define the loss function of the network input and the supervised target, which is represented as:

[0052]

[0053] Table 1 FCAE network structure parameter table

[0054]

[0055] To verify the effect and advantage of the FCAE noise reduction network, a harmonic-noise numerical model is constructed to test and analyze the noise reduction processing effectiveness of the FCAE. By comparing the noise reduction effects of FCAE and CAE networks on the same noisy signal, the effectiveness of the feature-enhanced full connection preprocessing structure in improving the noise reduction performance of the CAE network is tested.

[0056] A noisy harmonic signal s of length n is constructed:

[0057] s(n)=x pure (n)+x noise (n)

[0058] wherein, x pure (n) represents a pure harmonic signal, and x noise (n) is a zero-mean Gaussian signal. The expression of the pure signal x pure (n) is shown in (2.20), and the formula parameters are shown in Table 2.

[0059]

[0060] Table 2 Undamped harmonic signal parameters

[0061]

[0062] Based on the aforementioned harmonic signal model, four different levels of random Gaussian noise were added to construct the training and testing datasets for the FCAE and CAE networks. The noise level of the noisy signal is represented by the signal-to-noise ratio (SNR), and its formula is:

[0063]

[0064] The signal-to-noise ratios (SNRs) of the four noisy signals are SNR = 10, 5, -5, and -10, respectively. Figure 3 This section compares the time and frequency of a clean harmonic signal with a noisy signal, with a signal-to-noise ratio (SNR) of -10. The red line represents the clean signal, and the blue line represents the noisy signal, which is formed by adding random Gaussian noise with an SNR of -10 to the signal represented by the red line. The comparison of the two signals indicates the degree of noise interference on the clean signal in both the time and frequency domains.

[0065] Depend on Figure 3 It can be seen that when the signal-to-noise ratio is -10, the peak energy of some random noise is greater than the energy of the harmonic component at a frequency of 0.6 Hz, that is, the noise spectrum masks the harmonic spectrum.

[0066] To meet the data training requirements of the denoising network, 500 data samples of length 2048 were constructed for each of the four signal-to-noise ratio signals mentioned above to build the total dataset for the denoising network analysis. The ratio of training set to validation set data samples was 5:1, and the test set data samples were 20. The data sampling frequency was 100Hz.

[0067] Furthermore, traditional denoising methods based on singular values ​​and time-frequency transforms often struggle to handle data denoising under noise-masked conditions. Therefore, the comparative test will focus on analyzing the denoising network's performance on low signal-to-noise ratio signals to verify the denoising advantages of the FCAE network. The denoising results of the CAE and FCAE networks on the test set data (SNR = -10) are compared as follows: Figure 4 , 5 As shown.

[0068] Figure 4 The results of time-frequency comparison of CAE network denoising of a signal with a signal-to-noise ratio of -10 are shown. Figure 4 (a) shows the time-domain comparison results. Figure 4 (b) shows the frequency domain comparison results. The red line represents the clean signal; the blue line represents the input signal to the denoising network, which is based on the signal represented by the red line but with added random Gaussian noise of a signal-to-noise ratio of -10; the green line represents the network output signal, which is the signal after denoising by the CAE network. The comparison of these three signals is used to indicate the degree of noise interference on the clean signal in the time and frequency domains, and the denoising effect of the CAE network output data.

[0069] Depend onFigure 4 It can be seen that the CAE network still has obvious low-pass filtering characteristics when processing low SNR data, with a cutoff frequency of about 30 Hz and a relatively wide filter cutoff transition band.

[0070] Figure 5 Fig. 6 is a time-frequency comparison result of the FCAE network for denoising a signal with an SNR of -10, Figure 5 (a) is a time-domain comparison result, Figure 5 (b) is a frequency-domain comparison result. The red line represents the pure signal; the blue line represents the input signal of the denoising network, which is the pure signal with an SNR of -10 added; and the green line represents the output signal of the network, which is the signal after denoising by the FCAE network. The comparison of the three signals is used to represent the degree of interference of the pure signal in the time domain and the frequency domain by the noise, and the denoising effect of the FCAE network output data.

[0071] Figure 5 The results show that the FCAE network can effectively retain the harmonic frequency information when the random noise energy is greater than the harmonic frequency energy, and at the same time, effectively suppress the random noise in the full frequency band, eliminating the masking of the high-energy noise in the original input data to the harmonic frequency, proving that the FCAE network still has strong suppression ability to high-energy noise, and has obvious advantages compared with traditional filtering methods based on singular value and frequency domain transformation. However, the FCAE network still has a certain degree of error in the frequency energy of the output signal compared with the pure signal. Figure 5 As can be seen from the enlarged subgraph in (b), although the output signal of the FCAE network does not lose the harmonic frequency information, there is a certain difference in the frequency energy of the denoised signal compared with the pure signal. Among them, the spectral energy of 0.6 Hz and 2.2 Hz is reduced compared with the pure signal, and the energy of 0.35 Hz is increased. The above results show that the FCAE network with an added fully connected preprocessing structure has a greater degree of improvement in the suppression ability of random noise than the CAE network, and still has a good denoising effect for high-energy noise signals.

[0072] To verify the robustness of the denoising effect of the FCAE network, the average SNR of the denoising results of the FCAE and CAE networks for all test data is counted, and the statistical results are shown in Table 3.

[0073] Table 3 Denoising results of FCAE and CAE networks for four test data sets (dB)

[0074]

[0075] From the comparison results of the above table, the noise reduction effect of the FCAE network is better than that of the CAE network under four signal-to-noise ratio conditions, and the signal-to-noise ratio of the four sample data after noise reduction is increased by 10.14 dB, 8.62 dB, 9.85 dB and 11.22 dB respectively, which proves that the FCAE network has good noise reduction stability.

[0076] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for part or all of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solution deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A noise self-supervised based fully connected-convolutional auto-encoding network structure response denoising method, characterized in that, The method comprises the following steps: Step S1: inputting a structural response signal into a feedforward fully connected layer neural network comprising a padding layer, a first fully connected layer and a second fully connected layer, wherein the padding layer is used for dimension expansion of the input signal to simulate frequency domain or complex domain features, the first fully connected layer and the second fully connected layer are both used for dimension increasing and dimension reducing mapping of the padded signal by using a nonlinear activation function, thereby forming a time domain feature expression of the input signal while keeping the original information structure thereof; Step S2: inputting the feature data processed in S1 into a one-dimensional convolutional auto-encoding network, wherein the convolutional auto-encoding network comprises multiple convolutional layers, deconvolutional layers, pooling layers and de-pooling layers, and the effective features of the input signal are extracted by layer-by-layer convolution, deconvolution, pooling and de-pooling operations, and the time domain restoration of the input signal is realized, thereby outputting a denoised structural response signal.

2. The method of claim 1, wherein, In the feedforward fully connected layer neural network: The first fully connected layer comprises a dimension increasing layer for dimension increasing of the signal output by the padding layer, so as to improve the resolution of low-frequency random noise features; The second fully connected layer comprises a dimension reducing layer for compressing the dimension-increased features to the same dimension as the original signal, so as to keep the consistency between the network input and output.

3. The method of claim 1, wherein, The one-dimensional convolutional auto-encoding network comprises an encoding layer, a feature layer and a decoding layer, wherein: The encoding layer is used for extracting time domain principal component features of the input signal; The feature layer is a compression bottleneck structure and is used for representing key information after dimension reduction; The decoding layer is used for reconstructing the signal and realizing denoising output; The encoding layer and the decoding layer are both composed of multiple convolutional layers, deconvolutional layers, pooling layers and de-pooling layers, so as to realize layer-by-layer feature extraction and reconstruction of the signal.

4. The method of claim 1, wherein, The method adopts a self-supervised training strategy based on a noisy signal, and the network model is trained in an end-to-end manner without relying on pure signal samples.

5. The method of claim 1, wherein, The structural response signal is an acceleration signal or a strain signal collected by an ocean engineering structure under environmental excitation.

Images