Optical fiber sensing mixed signal separation method based on convolution demixing and deep non-negative tensor decomposition
By using convolutional demixing and deep nonnegative tensor decomposition to dynamically generate frequency templates and time activation matrices, and combining time-frequency mask optimization and phase fine-tuning of the convolutional demixing module, the problems of nonlinear mixing, noise interference and insufficient generalization ability in fiber optic sensing signal separation are solved, and efficient signal separation in complex environments is achieved.
Patent Information
- Application Number
- CN202510973981.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-15
- Publication Date
- 2025-11-14
AI Technical Summary
Existing fiber optic sensing signal separation methods are insufficient in handling nonlinear synchronous mixed signals, noise robustness, and generalization ability, and are difficult to adapt to the separation of multi-source signals in complex environments.
An adaptive signal separation method based on convolutional demixing and deep nonnegative tensor decomposition is adopted. The frequency template matrix and time activation matrix are dynamically generated by deep neural network. Combined with the time-frequency mask optimization and phase fine-tuning of the convolutional demixing module, the signal is adaptively separated.
It significantly improves the separation performance of multi-source signals, enhances the ability to resist noise interference, has strong adaptability, and can efficiently separate mixed signals in complex environments.
Smart Images
Figure CN120950953A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of distributed optical fiber sensing signal processing, specifically relating to a method for separating mixed signals based on convolutional demixing and deep non-negative tensor decomposition, which is suitable for separating multi-source disturbance signals in optical fiber sensing under complex environments. Background Technology
[0002] Distributed fiber optic sensing systems, with their high sensitivity and resistance to electromagnetic interference, are widely used in industrial monitoring and perimeter security. However, in practical applications, sensor signals are often affected by the coupling of environmental vibrations, equipment noise, and multi-source disturbances, forming mixed signals. This leads to the annihilation of target signal features, low signal recognition accuracy, and a high false alarm rate, thus requiring signal separation. Existing fiber optic sensing signal separation methods, such as independent component analysis, require signals to meet a non-Gaussian distribution and are only applicable to linear instantaneous mixing scenarios, making it difficult to handle convolutional or nonlinear mixing. Methods such as wavelet transform rely on prior basis function selection and lack the ability to explicitly demix the mixing relationships of multi-source signals, resulting in poor separation performance. Traditional non-negative tensor decomposition methods are limited by the linear low-rank assumption and are difficult to adapt to dynamic mixing environments. In summary, existing fiber optic sensing signal separation technologies generally suffer from three problems:
[0003] 1) Poor separation performance for nonlinear synchronous mixed signals, such as multi-source signal convolutional mixing;
[0004] 2) Insufficient noise robustness;
[0005] 3) The generalization ability is insufficient to meet the needs of dynamic monitoring.
[0006] Therefore, developing a hybrid signal separation method for fiber optic sensing with strong noise resistance and high precision has significant practical value. Summary of the Invention
[0007] This invention addresses the shortcomings of existing technologies by proposing a method for separating mixed signals in fiber optic sensing based on convolutional demixing and deep non-negative tensor decomposition. The feature matrix and activation matrix of traditional non-negative tensor decomposition are transformed into frequency template matrices and time activation matrices dynamically generated by a deep neural network, thus enabling adaptation to the characteristics of different signals without manual iteration. These improvements preserve the interpretability of non-negative tensor decomposition and enhance its expressive power. Combined with real-time optimization of the time-frequency mask and phase fine-tuning in the convolutional demixing module, as well as dynamically generated source-dependent convolution kernel parameters, this method solves the problems of multi-source signal convolutional mixing, noise interference, and poor generalization ability, achieving efficient separation of multi-source signals in complex environments.
[0008] To achieve the above objectives, the present invention adopts the following technical solution:
[0009] A method for separating mixed signals in fiber optic sensing based on convolutional demixing and depth nonnegative tensor decomposition (C-DNTF) includes the following steps:
[0010] Step 1: Divide the mixed signal into a training set and a test set;
[0011] Step 2: Perform short-time Fourier transform on the mixed signals in the training set to extract the phase spectrum, amplitude spectrum, and logarithmic amplitude spectrum;
[0012] Step 3: Using the C-DNTF's deep nonnegative tensor decomposition module, the logarithmic amplitude spectrum is converted into an initial mask under the guidance of the frequency template W.
[0013] Step 4: Input the amplitude spectrum, phase spectrum, logarithmic amplitude spectrum and initial mask into the convolution demixing module of C-DNTF to achieve signal separation;
[0014] Step 5: By jointly optimizing multiple loss functions, the optimal separation model C-DNTF is obtained;
[0015] Step 6: Process the mixed signal of the test set according to steps 2-4, and output the separated signal;
[0016] Step 7: Apply the best matching function to the separated signal to find the original signal that is most similar to it, evaluate the signal separation effect, and visualize the separation results of multi-source signals.
[0017] Preferably, in step 3, the steps of the depth nonnegative tensor decomposition module are as follows:
[0018] 1) Initialize the learnable frequency template matrix W to explicitly model the non-negative tensor decomposition basis vectors of the signal frequency domain features, and guide the H encoder to perform feature extraction;
[0019] 2) Input the logarithmic magnitude spectrum into an H encoder implemented with a multi-layer convolutional network containing convolutional layers, batch normalization, and ReLU activation functions to generate an H latent representation;
[0020] 3) Define an independent H decoder for each signal source, which includes convolutional layers, batch normalization, and the Softplus activation function to generate the temporal activation matrix for each signal source;
[0021] 4) Stack and normalize the time activation matrices of each signal source to obtain the initial mask.
[0022] Preferably, in step 4, the steps of the convolution demixing module are as follows:
[0023] 1) Input the initial mask and amplitude spectrum into a mask refinement network (one dedicated network for each source) containing convolutional layers, batch normalization, and sigmoid activation function to generate a refined mask;
[0024] 2) Input the phase spectrum and amplitude spectrum into the phase correction network (one dedicated network for each source), constrain the phase adjustment amount through the Tanh activation function, and obtain the corrected phase;
[0025] 3) Combine the refined mask with the corrected phase to obtain the complex spectrum signal of each signal source in the preliminary reconstruction, and then reconstruct the time-domain separated signal in the preliminary reconstruction by the inverse short-time Fourier transform.
[0026] 4) Input the logarithmic magnitude spectrum into the convolution kernel estimation network (fully connected network) to generate the convolution kernel parameters for each source;
[0027] 5) Convolve the initially reconstructed temporal-domain separated signal using the generated convolution kernel parameters to obtain the reconstructed temporal-domain separated signal;
[0028] 6) After normalizing the reconstructed time-domain separated signal, output the separated signal.
[0029] Preferably, in step 5, multiple loss functions, including mean square error, spectral consistency loss, non-negativity loss, sparsity loss, and orthogonality constraint loss, are combined to adjust the model parameters and iterate continuously to retain the model with the minimum total loss, which is the optimal separation model.
[0030] Preferably, in step 7, the indicator of the signal separation effect is the correlation coefficient or the scale-invariant signal distortion ratio.
[0031] Compared with the prior art, the present invention has the following obvious and prominent substantive features and significant advantages:
[0032] 1. This invention explicitly models the distribution characteristics of signals in the frequency and time dimensions through a deep non-negative tensor decomposition module, enabling dynamic and accurate capture of the time-frequency features of different events; through real-time optimization of the time-frequency mask (CNN correction of the short-time Fourier transform amplitude spectrum) and phase fine-tuning (Tanh-constrained residual compensation) of the convolution demixing module, combined with dynamically generated source-dependent convolution kernel parameters (equivalent to frequency domain adaptive filtering), the signal separation performance in complex mixed scenarios is significantly improved.
[0033] 2. This invention combines spectral consistency loss and phase correction to ensure that the separated signal matches the real signal in both the frequency and time domains, avoiding the time-domain signal distortion problem caused by traditional methods when separating signals in the frequency domain;
[0034] 3. This invention has wide adaptability and can successfully separate mixed signals collected by different sensing systems, demonstrating strong generalization ability. Attached Figure Description
[0035] Figure 1This is a flowchart of the signal separation method of the present invention.
[0036] Figure 2 This is a schematic diagram of the structure of the depth non-negative tensor decomposition module in a preferred embodiment of the present invention.
[0037] Figure 3 This is a schematic diagram of the structure of the convolution demixing module in a preferred embodiment of the present invention.
[0038] Figure 4 This is a typical signal from the public dataset of Beijing Jiaotong University in a preferred embodiment of the present invention.
[0039] Figure 5 This is a typical signal from a laboratory-made dataset in a preferred embodiment of the present invention.
[0040] Figure 6 In a preferred embodiment of the present invention, the evaluation metrics of the C-DNTF model of the present invention for separating two-source, three-source, and four-source convolutional mixed signals of public datasets are compared with those of the existing TFA-DRNN and CRNN-A models.
[0041] Figure 7 This is a schematic diagram illustrating the separation results of the C-DNTF model of the present invention and the existing TFA-DRNN and CRNN-A models on a public dataset, in a preferred embodiment of the present invention.
[0042] Figure 8 This is a schematic diagram illustrating the separation results of the C-DNTF model of the present invention and the existing TFA-DRNN and CRNN-A models on a public dataset in a preferred embodiment of the present invention.
[0043] Figure 9 This is a schematic diagram illustrating the separation results of the C-DNTF model of the present invention and the existing TFA-DRNN and CRNN-A models on a public dataset in a preferred embodiment of the present invention.
[0044] Figure 10 In a preferred embodiment of the present invention, the evaluation metrics of the C-DNTF model of the present invention for separating two-source, three-source, and four-source convolutional mixed signals of a self-made dataset are compared with those of the existing TFA-DRNN and CRNN-A models.
[0045] Figure 11 This is a schematic diagram illustrating the separation results of the C-DNTF model of the present invention and the existing TFA-DRNN and CRNN-A models on the two-source convolutional mixed signal of the self-made dataset in a preferred embodiment of the present invention.
[0046] Figure 12This is a schematic diagram showing the separation results of the C-DNTF model of the present invention and the existing TFA-DRNN and CRNN-A models on the three-source convolutional mixed signal of the self-made dataset in a preferred embodiment of the present invention.
[0047] Figure 13 This is a schematic diagram showing the separation results of the C-DNTF model of the present invention and the existing TFA-DRNN and CRNN-A models on the four-source convolutional mixed signal of the self-made dataset in a preferred embodiment of the present invention. Detailed Implementation
[0048] The following preferred embodiments, in conjunction with the accompanying drawings, further illustrate the above method:
[0049] Example 1
[0050] In this example, see Figure 1 A method for separating mixed signals in fiber optic sensing based on convolutional demixing and depth nonnegative tensor decomposition (C-DNTF) includes the following steps:
[0051] Step 1: Divide the mixed signal into a training set and a test set;
[0052] Step 2: Perform short-time Fourier transform on the mixed signals in the training set to extract the phase spectrum, amplitude spectrum, and logarithmic amplitude spectrum;
[0053] Step 3: Using the C-DNTF's deep nonnegative tensor decomposition module, the logarithmic amplitude spectrum is converted into an initial mask under the guidance of the frequency template W.
[0054] Step 4: Input the amplitude spectrum, phase spectrum, logarithmic amplitude spectrum and initial mask into the convolution demixing module of C-DNTF to achieve signal separation;
[0055] Step 5: By jointly optimizing multiple loss functions, the optimal separation model C-DNTF is obtained;
[0056] Step 6: Process the mixed signal of the test set according to steps 2-4, and output the separated signal;
[0057] Step 7: Apply the best matching function to the separated signal to find the original signal that is most similar to it, evaluate the signal separation effect, and visualize the separation results of multi-source signals.
[0058] This embodiment successfully separates multi-source convolutional mixed signals. By using deep non-negative tensor decomposition and convolution demixing, it significantly improves the signal separation performance in complex mixed scenarios and meets the requirements for separating multi-source perturbation signals in complex environments.
[0059] Example 2
[0060] This embodiment is basically the same as Embodiment 1, except that:
[0061] In this embodiment, see Figure 2 In step 3, the steps of the depth nonnegative tensor decomposition module are as follows:
[0062] 1) Initialize the learnable frequency template matrix W to explicitly model the non-negative tensor decomposition basis vectors of the signal frequency domain features, and guide the H encoder to perform feature extraction;
[0063] 2) Input the logarithmic magnitude spectrum into an H encoder implemented with a multi-layer convolutional network containing convolutional layers, batch normalization, and ReLU activation functions to generate an H latent representation;
[0064] 3) Define an independent H decoder for each signal source, which includes convolutional layers, batch normalization, and the Softplus activation function to generate the temporal activation matrix for each signal source;
[0065] 4) Stack and normalize the time activation matrices of each signal source to obtain the initial mask.
[0066] In this embodiment, the frequency template matrix W is explicitly modeled as a learnable parameter to adaptively process data; the time activation matrix is generated through a deep neural network rather than directly optimized, which is more conducive to the model's self-learning. Through these improvements to the traditional non-negative tensor decomposition, its interpretability is preserved and its expressive power is enhanced.
[0067] Example 3
[0068] This embodiment is basically the same as the above embodiments, except that:
[0069] In this embodiment, see Figure 3 In step 4, the steps of the convolution demixing module are as follows:
[0070] 1) Input the initial mask and amplitude spectrum into a mask refinement network (one dedicated network for each source) containing convolutional layers, batch normalization, and sigmoid activation function to generate a refined mask;
[0071] 2) Input the phase spectrum and amplitude spectrum into the phase correction network (one dedicated network for each source), constrain the phase adjustment amount through the Tanh activation function, and obtain the corrected phase;
[0072] 3) Combine the refined mask with the corrected phase to obtain the complex spectrum signal of each signal source in the preliminary reconstruction, and then reconstruct the time-domain separated signal in the preliminary reconstruction by the inverse short-time Fourier transform.
[0073] 4) Input the logarithmic magnitude spectrum into the convolution kernel estimation network (fully connected network) to generate the convolution kernel parameters for each source;
[0074] 5) Convolve the initially reconstructed temporal-domain separated signal using the generated convolution kernel parameters to obtain the reconstructed temporal-domain separated signal.
[0075] 6) After normalizing the reconstructed time-domain separated signal, output the separated signal.
[0076] This embodiment uses a convolutional demixing module to refine the mask, correct the phase, and dynamically generate convolutional kernels to handle complex mixed scenes. The mask refinement network is used to refine the separation mask of each signal source, making the mask more refined; the phase correction network constrains the phase adjustment amount, corrects the phase, and improves the quality of the reconstructed signal; the convolutional kernel estimation network is used to learn the convolutional relationship between the source signals.
[0077] Example 4
[0078] This embodiment is basically the same as the above embodiments, except that:
[0079] In this embodiment, in step 5, multiple loss functions, including mean square error, spectral consistency loss, non-negativity loss, sparsity loss, and orthogonality constraint loss, are combined to adjust the model parameters and iterate continuously to retain the model with the minimum total loss, which is the optimal separation model.
[0080] In this embodiment, in step 7, the indicator of the signal separation effect is the correlation coefficient or the scale-invariant signal distortion ratio.
[0081] like Figures 1-3As shown, the fiber optic sensing mixed signal separation method based on convolutional demixing and deep non-negative tensor decomposition in Examples 1 to 4 above performs a short-time Fourier transform on the mixed signal in the training set to extract the phase spectrum, amplitude spectrum, and logarithmic amplitude spectrum. The logarithmic amplitude spectrum is converted into an initial mask using a deep non-negative tensor decomposition module. A learnable frequency template matrix W guides the H encoder for feature extraction. The H encoder uses a multi-layer convolutional network to encode the input logarithmic amplitude spectrum, generating an H latent representation. An independent H decoder is defined for each signal source to generate the time activation matrix for each signal source, and these matrices are stacked and... Normalization yields an initial mask; signal separation is achieved through a convolutional demixing module. A preliminary complex spectrum signal is obtained using a mask refinement network (one dedicated network per source) and a phase correction network (one dedicated network per source). The temporal-domain separated signal is then reconstructed using inverse short-time Fourier transform. Convolution kernel parameters are generated using a convolution kernel estimation network (fully connected layer network). These parameters are then used to convolve the preliminary reconstructed temporal-domain separated signal to obtain the reconstructed temporal-domain separated signal. After normalization, the separated signal is output. The optimal separation model, C-DNTF, is obtained by jointly optimizing multiple loss functions. By processing the mixed signal of the test set according to the aforementioned steps, the separated signal can be obtained. This invention can dynamically and accurately capture the time-frequency characteristics of different events, has strong generalization ability, and is suitable for separating multi-source convolutional mixed signals in fiber optic perimeter defense systems under complex scenarios.
[0082] Example 5
[0083] This embodiment is basically the same as the above embodiments, except that:
[0084] In this embodiment, we first start with phase-sensitive optical temporal reflectance from Beijing Jiaotong University. From the publicly available dataset of the distributed fiber optic sensing system, 2-4 signal sources are randomly selected to construct two-source, three-source, and four-source convolutional mixed signals. A total of six signal classes are selected, including (1) background, (2) digging, (3) tapping, (4) watering, (5) shaking, and (6) walking. Typical signals are as follows: Figure 4 As shown, it has 12 spatial channels and 10,000 sampling points. One-dimensional time-series data from its sixth channel is selected for convolutional mixing. Then, from a dataset created in the laboratory based on a linear Sagnac interferometric distributed fiber optic sensing system, 2-4 signal sources are randomly selected to construct two-source, three-source, and four-source convolutional mixed signals. Four types of signals are selected: (a) background, (b) shaking, (c) tapping, and (d) walking. Typical signals are shown below. Figure 5As shown, both datasets have 20,000 sampling points in their one-dimensional time series data. Signal processing for both datasets followed the same steps, implemented using PyCharm software on a computer with a 12th Gen Intel(R) Core(TM) i5-12600KF@3.70GHz 10-core processor, 32GB RAM, an NVIDIA GeForce RTX 4060 graphics card, and a Windows 11 operating system. The only difference was in the data reading process; subsequent processing did not differentiate between the two datasets, but simply displayed the separation results of the mixed signals from the two different datasets. The cross-correlation coefficient between the separated signal and the original signal, as well as the scale-invariant signal distortion ratio, were used to measure the effectiveness of the mixed signal separation.
[0085] The correlation coefficient is
[0086]
[0087] In the formula, X i Y i These are the i-th data points of the original signal and the separated signal, respectively. These are the average values of the original signal and the separated signal, respectively, where n represents the number of data points in the signal.
[0088] The formula for calculating the scale-invariant signal distortion ratio is:
[0089]
[0090] In the formula, s is the amplitude scaling factor. t The original signal, For the separated signal, T represents the transpose operation, and |||| represents the L2 norm operation.
[0091] Perform multi-source convolutional mixed signal separation according to the following steps.
[0092] Step 1: Split the 200 mixed signals into a training set and a test set in a 4:1 ratio;
[0093] Step 2: Apply a 512-point Hanning window short-time Fourier transform to the mixed signals in the training set, with a step size of 128 points, which is 1 / 4 of the window length, balancing time resolution and computational efficiency; extract the phase spectrum, amplitude spectrum, and logarithmic amplitude spectrum of the time spectrum. The logarithmic amplitude spectrum, by compressing the dynamic range, suppresses the dominance of high-energy components and highlights weak signal features, providing a high-resolution frequency domain representation basis for deep feature extraction;
[0094] Step 3: Convert the logarithmic magnitude spectrum into an initial mask through depthwise nonnegative tensor decomposition; for example... Figure 2As shown, a learnable frequency template matrix W (rank 16) is initialized to explicitly model the non-negative tensor decomposition basis vectors of the signal frequency domain features. Its non-negativity is guaranteed by the Softplus function, replacing the fixed basis dictionary of traditional non-negative matrix factorization (NMF). The H encoder is then guided to extract features. The logarithmic amplitude spectrum is input into the H encoder, which contains three convolutional layers (5×5 kernels), batch normalization, and ReLU activation functions, to extract deep time-frequency features. A single-layer latent space compression module (3×3 convolutional kernels, 32 channels) compresses the feature dimension from 64 to 32, generating a low-dimensional latent representation. For each signal source, a specific decoder containing convolutional layers (3×3 kernels), batch normalization, and Softplus activation functions is used to generate the temporal activation matrix for each signal source. The temporal activation matrices of each signal source are stacked and normalized using Softmax to obtain an initial mask, which serves as the basis for subsequent fine-tuning.
[0095] Step 4: Separate the signals through convolutional demixing; for example... Figure 3 As shown, the initial mask and amplitude spectrum are input into the fine-tuning mask network (5×5→1×1 convolution kernel, batch normalization, and Sigmoid activation function) of each source to generate a fine-tuning mask; the phase spectrum and amplitude spectrum are input into the dual-channel phase correction network (16→8 channels) of each source, and the phase adjustment amount is constrained by the Tanh activation function to obtain the corrected phase; the fine-tuning mask and the corrected phase are combined to obtain the preliminary reconstructed complex spectrum signal of each signal source; a 512-point Hanning window inverse short-time Fourier transform is performed on the reconstructed complex spectrum signal with a stride of 128 points to obtain the preliminary reconstructed time-domain separated signal; the logarithmic amplitude spectrum is input into the convolution kernel estimation network, and the adaptive convolution kernel parameters of 32 sampling points for each source are dynamically generated through a 256→128→32 unit fully connected structure and the ReLU activation function; the generated convolution kernel parameters are used to convolve the time-domain signal to obtain the reconstructed time-domain separated signal; after normalizing the reconstructed time-domain separated signal, the separated signal is output.
[0096] Step 5 utilizes mean squared error loss to ensure the accuracy of time-domain signal reconstruction, spectral consistency loss to constrain frequency-domain feature matching, nonnegativity loss to force the frequency template matrix W to be nonnegative, sparsity loss to promote the sparsity of the W matrix, enhance the discriminative power of basis vectors, and orthogonality constraint loss to encourage orthogonality of W column vectors and reduce basis vector redundancy. The model parameters are adjusted iteratively, and the AdamW optimizer (initial learning rate 0.0005, weight decay 1e-4) combined with gradient pruning (threshold 1.0) ensures model stability. The model with the minimum total loss is retained, thus generating the optimal separation model C-DNTF.
[0097] Step 6: Process the mixed signal of the test set according to steps 2-4, and output the separated signal;
[0098] Step 7: Apply the best matching function to the separated signal to find the original signal that is most similar to it, and evaluate the signal separation effect. In this embodiment, two indicators are used: correlation coefficient and scale-invariant signal distortion ratio, and the separation results of multi-source signals are visualized.
[0099] For the separation of convolutional mixed signals from the Beijing Jiaotong University public dataset, the separation performance metrics are as follows: Figure 6 As shown, the visualization separation results are as follows: Figures 7-9 For the separation of convolutional mixed signals from a self-made laboratory dataset, the metrics for separation performance are as follows: Figure 10 As shown, the visualization separation results are as follows: Figures 11-13 Regarding the metrics for separating the mixed signals from the two datasets, the C-DNTF method proposed in this embodiment outperforms the other two comparative methods in both correlation coefficient and scale-invariant signal distortion ratio. In the visualization results of the mixed signal separation from the two datasets, due to the disordered nature of the separated signals, the separated signals and original signals do not correspond one-to-one in sub-figure numbering. However, the original signals corresponding to the separated signals are still visually distinguishable, indicating successful separation. Furthermore, the signals separated by the method of this invention are clearer and of higher quality. The separation results of the mixed signals from the two datasets also show that the results on the self-made laboratory dataset are better than those on the publicly available dataset from Beijing Jiaotong University. This is because the characteristics of the self-made laboratory dataset are more pronounced than those of the Beijing Jiaotong University dataset. Even though the publicly available dataset from Beijing Jiaotong University is more complex, the results of this embodiment also demonstrate successful separation of multi-source mixed signals.
[0100] This embodiment presents a method for separating mixed signals from fiber optic sensing based on convolutional demixing and deep non-negative tensor decomposition. The mixed signal is divided into a training set and a test set. A short-time Fourier transform is performed on the mixed signal in the training set to extract the phase spectrum, amplitude spectrum, and logarithmic amplitude spectrum. Guided by a frequency template W, the logarithmic amplitude spectrum is converted into an initial mask using a deep non-negative tensor decomposition module. The amplitude spectrum, phase spectrum, logarithmic amplitude spectrum, and initial mask are then input into a convolutional demixing module to achieve signal separation. An optimal separation model is obtained by jointly optimizing multiple loss functions. The mixed signal in the test set is processed to output the separated signal. The signal separation performance is evaluated using metrics, and finally, the separation results of the multi-source signals are visualized. This embodiment can dynamically and accurately capture the time-frequency characteristics of different events, exhibits strong generalization ability, and is suitable for separating multi-source convolutional mixed signals in fiber optic perimeter defense systems under complex scenarios.
[0101] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent substitutions. As long as they meet the purpose of the invention and do not deviate from the technical principle and inventive concept of the present invention, they shall fall within the protection scope of the present invention.
Claims
1. A method for separating mixed signals in fiber optic sensing based on convolutional demixing and deep nonnegative tensor decomposition, characterized in that, Includes the following steps: Step 1: Divide the mixed signal into a training set and a test set; Step 2: Perform short-time Fourier transform on the mixed signals in the training set to extract the phase spectrum, amplitude spectrum, and logarithmic amplitude spectrum; Step 3: Using the C-DNTF's deep nonnegative tensor decomposition module, the logarithmic amplitude spectrum is converted into an initial mask under the guidance of the frequency template W. Step 4: Input the amplitude spectrum, phase spectrum, logarithmic amplitude spectrum and initial mask into the convolution demixing module of C-DNTF to achieve signal separation; Step 5: By jointly optimizing multiple loss functions, the optimal separation model C-DNTF is obtained; Step 6: Process the mixed signal of the test set according to steps 2-4, and output the separated signal; Step 7: Apply the best matching function to the separated signal to find the original signal that is most similar to it, evaluate the signal separation effect, and visualize the separation results of multi-source signals.
2. The method for separating mixed signals of fiber optic sensing based on convolutional demixing module and depth nonnegative tensor decomposition according to claim 1, characterized in that, In step 3, the steps of the depth nonnegative tensor decomposition module are as follows: 1) Initialize the learnable frequency template matrix W to explicitly model the non-negative tensor decomposition basis vectors of the signal frequency domain features, and guide the H encoder to perform feature extraction; 2) Input the logarithmic magnitude spectrum into an H encoder implemented with a multi-layer convolutional network containing convolutional layers, batch normalization, and ReLU activation functions to generate an H latent representation; 3) Define an independent H decoder for each signal source, which includes convolutional layers, batch normalization, and the Softplus activation function to generate the temporal activation matrix for each signal source; 4) Stack and normalize the time activation matrices of each signal source to obtain the initial mask.
3. The method for separating mixed signals of fiber optic sensing based on convolutional demixing module and depth nonnegative tensor decomposition according to claim 1, characterized in that, In step 4, the steps of the convolution demixing module are as follows: 1) Input the initial mask and amplitude spectrum into a mask refinement network (one dedicated network for each source) containing convolutional layers, batch normalization, and sigmoid activation function to generate a refined mask; 2) Input the phase spectrum and amplitude spectrum into the phase correction network (one dedicated network for each source), constrain the phase adjustment amount through the Tanh activation function, and obtain the corrected phase; 3) Combine the refined mask with the corrected phase to obtain the complex spectrum signal of each signal source in the preliminary reconstruction, and then reconstruct the time-domain separated signal in the preliminary reconstruction by the inverse short-time Fourier transform. 4) Input the logarithmic magnitude spectrum into the convolution kernel estimation network (fully connected network) to generate the convolution kernel parameters for each source; 5) Convolve the initially reconstructed temporal-domain separated signal using the generated convolution kernel parameters to obtain the reconstructed temporal-domain separated signal; 6) After normalizing the reconstructed time-domain separated signal, output the separated signal.
4. The method for separating mixed signals of fiber optic sensing based on convolutional demixing module and deep nonnegative tensor decomposition according to claim 1, characterized in that, In step 5, multiple loss functions, including mean square error, spectral consistency loss, non-negativity loss, sparsity loss, and orthogonality constraint loss, are combined to adjust the model parameters and iterate continuously. The model with the minimum total loss is retained, which is the optimal separation model.
5. The method for separating mixed signals of fiber optic sensing based on convolutional demixing module and depth nonnegative tensor decomposition according to claim 1, characterized in that, In step 7, the indicators for the signal separation effect are the correlation coefficient or the scale-invariant signal distortion ratio.
Citation Information
Cited By
Satellite magnetic field data earthquake anomaly detection method and system
CN122218798A