A method of distortion stretch spectral separation and related apparatus

Abstract

The application discloses a distortion overlapping spectrum separation method and a related device, relates to the technical field of spectrum separation, and can be applied to fiber sensing spectrum, Raman spectrum and X-ray diffraction spectrum. The method comprises the following steps: obtaining and preprocessing distortion overlapping spectrum data through simulation, constructing a training data set through data enhancement and splitting simulation, extracting features and calculating a contrast learning loss by using a double-encoder network, reconstructing spectrum components by using a double-decoder network and calculating a reconstruction loss, training a model by combining the contrast learning loss and the reconstruction loss, fine-tuning a target model through actual measurement data, and finally performing spectrum separation by using the target model and outputting physical parameters. The method does not require a large amount of manually labeled data, can improve the processing capacity for spectrum distortion and overlapping through self-supervised contrast learning and physical constraints, can enhance the adaptability to complex environments, can greatly reduce strain measurement errors, can effectively solve the demodulation problem of spectrum distortion and overlapping, and can meet the high-precision requirement of precise structure health monitoring.

Description

Technical Field

[0001] This application relates to the field of spectral separation technology, and in particular to a method and related apparatus for separating distorted overlapping spectra, applicable to various spectral types, such as FBG sensor network spectroscopy, Raman spectroscopy, X-ray diffraction spectroscopy, etc. Background Technology

[0002] As an advanced distributed fiber optic sensing technology, FBG sensor networks have become one of the core technologies for large-scale structural health monitoring (SHM) due to their advantages such as high precision, miniaturization, and resistance to harsh environments (e.g., high temperature and corrosion). FBG sensors measure physical parameters through wavelength encoding. The change in the center wavelength of their reflected spectrum directly corresponds to changes in physical quantities such as strain and temperature at the sensing point. Therefore, the accuracy of spectral demodulation determines the monitoring performance.

[0003] To meet the needs of multi-point, high-density monitoring, FBG sensor networks typically employ wavelength division multiplexing (WDM) and time division multiplexing (TDM) technologies. However, with the increase in sensor density, the limitations of traditional multiplexing technologies are becoming increasingly apparent: under the influence of non-uniform physical fields (such as residual strain in composite materials and hot spots in high-temperature superconductors), FBG sensors at different locations are subjected to differential physical effects, resulting in asymmetric distortion, peak shift, and width changes in their reflection spectra, which in turn leads to excessive spectral overlap—a problem that severely interferes with wavelength localization and reduces measurement accuracy.

[0004] Existing solutions have significant drawbacks: Firstly, traditional optimization algorithms rely on physical constraints such as wavelength range and spectral spacing, leading to a significant increase in computation time as the number of sensors increases, and they also exhibit poor adaptability to complex distorted spectra. Secondly, machine learning methods based on CNNs and LSTMs require a large amount of manually labeled data and struggle to handle the combined problem of "spectral distortion + overlap" simultaneously, resulting in limited generalization ability. In actual monitoring, these issues cause strain measurement errors to reach ±37.2 με, far from meeting the high-precision requirements of precision structural health monitoring.

[0005] The aforementioned challenges of spectral distortion and overlap are not limited to FBG sensing; they are also prevalent in other spectroscopic techniques such as Raman spectroscopy and X-ray diffraction spectroscopy, and effective universal solutions are lacking. Therefore, there is an urgent need for a spectral separation technique that can effectively handle distorted and overlapping spectra, reduce dependence on labeled data, and improve adaptability to complex environments, in order to solve the demodulation challenges of various spectra in complex environments. Summary of the Invention

[0006] The purpose of this application is to provide a method and related apparatus for separating distorted and overlapping spectra, which can effectively solve the demodulation problem of spectral data under distorted and overlapping conditions.

[0007] To achieve the above objectives, this application provides the following solution:

[0008] In a first aspect, this application provides a method for separating distorted and overlapping spectra, comprising the following steps:

[0009] By adjusting the grating parameters to generate narrow linewidth spectra, spectral distortion under different physical fields and noise interference is simulated to obtain distorted and overlapping spectral data, and the distorted and overlapping spectral data is preprocessed. The distorted and overlapping spectral data contains spectral distortion and spectral overlap caused by uneven distribution of physical fields. The preprocessing includes standardizing the spectral intensity values ​​to the range of [0,1] and removing DC bias and system noise.

[0010] Data augmentation and split simulation are performed on the preprocessed distorted overlapping spectral data to construct a training dataset. The training dataset includes several single-channel distorted overlapping spectra and corresponding dual-channel separation components. Data augmentation involves adding Gaussian noise to the distorted overlapping spectral data, extracting wavelength segment sequences for time slicing, performing time distortion by random nonlinear stretching or compression of wavelengths, and reversing the spectral sequence order. The split simulation is based on the physical properties of the distorted overlapping spectra to separate the preprocessed distorted overlapping spectral data, resulting in dual-channel separation components that provide positive sample references for self-supervised comparative learning. The dual-channel separation components include a distorted spectral component and a Gaussian spectral component.

[0011] The dual encoder network of the spectral separation model receives a single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, outputs a first embedding feature and a second embedding feature respectively, and calculates a contrastive learning loss function. The dual encoder network includes an overlapping spectral encoder and a separating spectral encoder. The overlapping spectral encoder is used to receive the single-channel distorted overlapping spectrum and output the first embedding feature. The separating spectral encoder is used to receive the dual-channel separation component and output the second embedding feature.

[0012] The second embedded feature is reconstructed into separate distorted spectral components and Gaussian spectral components through a dual-decoder network of the spectral separation model, and the reconstruction loss function is calculated based on mean square error and physical constraints.

[0013] The joint loss function is determined based on the contrastive learning loss function and the reconstruction loss function, and the spectral separation model is trained based on the training dataset.

[0014] Based on the measured single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, the parameters of the trained spectral separation model are fine-tuned to obtain the target spectral separation model; the target spectral separation model is used to separate the spectral components of the measured distorted overlapping spectral data.

[0015] The target spectral separation model is used to separate the spectral components of the single-channel distorted and overlapping spectrum under test, and the corresponding physical parameters are output based on the separated spectral components.

[0016] Optionally, both the overlapping spectral encoder and the separate spectral encoder employ a one-dimensional convolutional neural network structure, which includes residual connections and an attention mechanism; the mathematical expression for the overlapping spectral encoder is: The mathematical expression for the split spectral encoder is: ,in, Indicates overlapping spectral samples, Indicates the separation of component samples. and These represent the first embedding feature and the second embedding feature, respectively.

[0017] Optionally, a temperature coefficient τ=0.07 is introduced into the contrastive learning loss function to adjust the similarity calculation scale. The contrastive learning loss function is a bidirectional contrastive loss function used to calculate the similarity loss between the first embedded feature and the second embedded feature; the contrastive learning loss function is shown in the following formula:

[0018] .

[0019] Among them, among them, L contrast To compare the learning loss function, N For batch size, subscript i and j These are all indexes of the samples. i=j When it is a positive sample, i ≠ j When the sample is negative, log() is the logarithmic function with base e, exp() is the exponential function with base e, and sim() is the cosine similarity function. and The first embedded feature is respectively and the second embedded feature The low-dimensional representation is obtained after projection head and L2 normalization.

[0020] Optionally, the reconstruction loss function can be calculated according to the following formula:

[0021] .

[0022] in, L recon To reconstruct the loss function, L MSE Let the mean squared error loss function be . L physics The physical constraint loss function, λ phys The weights are the physical constraint loss function weights.

[0023] Alternatively, the joint loss function is shown in the following equation:

[0024] .

[0025] in, L total For the joint loss function, λ contrast To compare the weights of the learning loss function, L contrast To compare the learning loss function, λ recon To reconstruct the weights of the loss function, L recon To reconstruct the loss function.

[0026] Optionally, when training the spectral separation model based on the training dataset, the AdamW optimizer is used to update the parameters of the spectral separation model, with a learning rate set to 0.001, a training period of 200 epochs, and a batch size of 32; when fine-tuning the parameters of the trained spectral separation model, the learning rate is reduced to 1×10. -4 The training cycle is set to 10, and an early stop strategy is adopted. Training is stopped when the validation loss does not decrease for 5 consecutive cycles.

[0027] Secondly, this application provides a distortion-overlapping spectral separation system, including the following functional modules:

[0028] The distortion and overlap spectrum simulation acquisition module is used to generate narrow linewidth spectra by adjusting grating parameters, simulate spectral distortion under different physical fields and noise interference, acquire distortion and overlap spectrum data, and preprocess the distortion and overlap spectrum data; the distortion and overlap spectrum data includes spectral distortion and spectral overlap phenomena caused by uneven distribution of physical fields; the preprocessing includes standardizing the spectral intensity values ​​to the range of [0,1] and removing DC bias and system noise.

[0029] The training dataset construction module is used to perform data augmentation and split simulation on the preprocessed distorted overlapping spectral data to construct the training dataset. The training dataset includes several single-channel distorted overlapping spectra and corresponding dual-channel separation components. Data augmentation involves adding Gaussian noise to the distorted overlapping spectral data, extracting wavelength segment sequences for time slicing, performing time distortion by random nonlinear stretching or compression of wavelengths, and reversing the spectral sequence order. The split simulation is based on the physical properties of the distorted overlapping spectra to separate the preprocessed distorted overlapping spectral data, obtaining dual-channel separation components, which provide positive sample references for self-supervised contrastive learning. The dual-channel separation components include a distorted spectral component and a Gaussian spectral component.

[0030] The contrastive learning computation module is used to receive single-channel distorted overlapping spectra and corresponding dual-channel separation components through the dual encoder network of the spectral separation model, output the first embedding feature and the second embedding feature respectively, and calculate the contrastive learning loss function; the dual encoder network includes an overlapping spectral encoder and a separating spectral encoder; the overlapping spectral encoder is used to receive single-channel distorted overlapping spectra and output the first embedding feature; the separating spectral encoder is used to receive dual-channel separation components and output the second embedding feature.

[0031] The reconstruction loss calculation module is used to reconstruct the second embedded feature into separate distorted spectral components and Gaussian spectral components through the dual decoder network of the spectral separation model, and calculate the reconstruction loss function based on the mean square error and physical constraints.

[0032] The model training module is used to determine the joint loss function based on the contrastive learning loss function and the reconstruction loss function, and to train the spectral separation model based on the training dataset.

[0033] The parameter fine-tuning module is used to fine-tune the parameters of the trained spectral separation model based on the measured single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, so as to obtain the target spectral separation model; the target spectral separation model is used to separate the spectral components of the measured distorted overlapping spectral data.

[0034] The spectral separation and output module is used to separate the spectral components of the single-channel distorted and overlapping spectrum under test using the target spectral separation model, and output the corresponding physical parameters based on the separated spectral components.

[0035] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the distortion overlap spectral separation method described above.

[0036] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the distortion overlap spectral separation method described above.

[0037] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the distortion overlap spectral separation method described above.

[0038] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0039] This application provides a method and related apparatus for separating distorted overlapping spectra. The method first generates narrow-linewidth spectra by adjusting grating parameters to simulate spectral distortion under different physical fields and noise interference, acquiring distorted overlapping spectral data and preprocessing it. The spectral intensity values ​​are standardized to the range [0,1], eliminating spectral intensity differences and interference signals from different devices and environments, avoiding interference from noise and bias on subsequent feature extraction, and providing clean and uniform input data for the model. Then, data augmentation operations expand the scale and diversity of the training dataset, simulating random interference and spectral nonlinear distortion in real monitoring scenarios. Based on the physical characteristics of the distorted overlapping spectra, the data is split and simulated, eliminating the need for manual labeling of spectral components. Positive samples required for self-supervised learning are automatically generated through physical mechanisms, completely eliminating the dependence on large amounts of manually labeled data and reducing data acquisition and labeling costs. Subsequently, a dual-encoder network is used to specifically learn the mixed features of the distorted overlapping spectra and the independent features of the separated components, and a contrastive learning loss function is used to promote the aggregation of similar features and the separation of dissimilar features, effectively capturing the complex relationship between spectral deformation and overlap. During reconstruction using a dual-decoder network, mean squared error ensures a high numerical match between the reconstructed spectral components and the actual separated components. Physical constraints guarantee the reconstruction results conform to the physical validity of distorted overlapping spectra, avoiding false spectra that deviate from actual physical meaning. Subsequently, the contrastive learning loss and reconstruction loss are jointly determined as a joint loss function and used to train the model. This approach simultaneously considers the discriminative nature of feature learning and the accuracy of spectral reconstruction, avoiding model bias caused by a single loss function. The trained model can accurately extract key spectral features and output high-quality separated spectra. After fine-tuning the model parameters based on measured spectral data, accurate separation of spectral components in distorted overlapping spectral data can be achieved. Finally, the target spectral separation model is used to separate spectral components from the single-channel distorted overlapping spectrum under test, and the corresponding physical parameters are output based on the separated spectral components. This significantly reduces the strain measurement error from ±37.2με in existing technologies, meeting the high-precision physical parameter measurement requirements for health monitoring of precision structures such as bridges and aircraft, and completely solving the demodulation problem of multiple spectra in complex environments. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of a distortion-overlapping spectral separation method provided in an embodiment of this application.

[0042] Figure 2 This is a schematic diagram of the functional modules of a distortion overlap spectrum separation system provided in an embodiment of this application.

[0043] Figure 3 This is a schematic diagram of the functional modules of a distortion overlap spectral separation system provided in another embodiment of this application.

[0044] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] This application provides a method for separating distorted overlapping spectra. In one exemplary embodiment, such as... Figure 1 As shown, it includes the following steps:

[0048] A1. By adjusting the grating parameters to generate a narrow linewidth spectrum, spectral distortion under different physical fields and noise interference is simulated to obtain distorted overlapping spectral data, and the distorted overlapping spectral data is preprocessed. The distorted overlapping spectral data includes spectral distortion and spectral overlap caused by the non-uniform distribution of the physical field. Specifically, in this embodiment, the distorted overlapping spectrum of the FBG sensor network is separated. In the FBG sensor network, spectral distortion refers to the chirped distortion of the spectrum caused by the FBG sensor network sensing a non-uniform physical field, and spectral overlap refers to the different degrees of spectral overlap caused by the different wavelength offsets between two monitoring points of the FBG sensor network. The preprocessing includes standardizing the spectral intensity values ​​to the range of [0,1] and removing DC bias and system noise.

[0049] The grating reflection spectrum is determined by factors such as grating period, grating length, and refractive index modulation depth. By adjusting these parameters, narrow linewidth spectra with different reflectivities and full width at half maximum (FWHM) can be obtained. Strain field types include linear, quadratic, cubic, and random strain fields. The degree of overlap is controlled by adjusting the spacing between the original and distorted spectra.

[0050] In an exemplary embodiment, distorted and overlapping spectral data of an FBG sensor network under a non-uniform physical field are acquired using a spectral analyzer (such as a Yokogawa AQ6370D). This data exhibits spectral distortion (asymmetry, peak shift, and width variation) and spectral overlap due to the non-uniform strain distribution sensed by the sensors. A computer device maps the spectral intensity values ​​of the acquired spectral data to the [0,1] range using a Min-Max normalization method to eliminate intensity differences between different devices and environments. Simultaneously, a wavelet threshold denoising algorithm is used to remove system noise, and the DC bias is subtracted to ensure data purity.

[0051] A2. Data augmentation and split simulation are performed on the preprocessed distorted overlapping spectral data to construct a training dataset. The training dataset includes several single-channel distorted overlapping spectra and corresponding dual-channel separation components. Data augmentation involves adding Gaussian noise to the distorted overlapping spectral data, extracting wavelength segment sequences for time slicing, performing random nonlinear stretching or compression of wavelengths, time distortion, and reversing the spectral sequence order. The split simulation is based on the physical characteristics of the FBG sensor network spectrum to separate the preprocessed distorted overlapping spectral data, obtaining dual-channel separation components to provide positive sample references for self-supervised comparative learning. The dual-channel separation components include a distorted spectral component and a Gaussian spectral component.

[0052] Specifically, to address the issues of insufficient training data and poor model generalization ability, the computer equipment performs multi-dimensional data augmentation operations on the preprocessed spectral data: Gaussian noise with a mean of 0 and a standard deviation of 0.01-0.05 is added to the spectral data to simulate random electromagnetic interference in actual monitoring; subsequences (1 / 2-3 / 4 the length of the original sequence) within the wavelength range (e.g., 1530nm-1560nm) are truncated to enhance the model's ability to learn local spectral features; the wavelengths are randomly stretched or compressed nonlinearly (with a distortion range of ±10%) through cubic spline interpolation to simulate nonlinear distortion of the spectrum under non-uniform physical fields; and the wavelength order of the spectral sequence is reversed to break the inherent order dependency of the data and improve the robustness of the model.

[0053] A3. The dual encoder network of the spectral separation model receives the single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, outputs the first embedding feature and the second embedding feature respectively, and calculates the contrastive learning loss function; the dual encoder network includes an overlapping spectral encoder and a separating spectral encoder; the overlapping spectral encoder is used to receive the single-channel distorted overlapping spectrum and output the first embedding feature; the separating spectral encoder is used to receive the dual-channel separation component and output the second embedding feature.

[0054] In an exemplary embodiment, both the overlapping spectral encoder and the separate spectral encoder employ a one-dimensional convolutional neural network structure, which includes residual connections and an attention mechanism. The mathematical expression for the overlapping spectral encoder is as follows: The mathematical expression for the split spectral encoder is: ,in, Indicates overlapping spectral samples, Indicates the separation of component samples. and These represent the first embedding feature and the second embedding feature, respectively.

[0055] Specifically, the input to the overlapping spectral encoder is a single-channel distorted overlapping spectrum (dimension 1×N, where N is the number of wavelength sampling points). It employs a 3-layer one-dimensional CNN structure: Layer 1 has a kernel size of 3, a stride of 1, and 32 output channels; Layer 2 has a kernel size of 5, a stride of 1, and 64 output channels; Layer 3 has a kernel size of 3, a stride of 1, and 128 output channels. Each convolutional layer is followed by a ReLU activation function and a batch normalization layer. Residual connections are established between layers 1-2 and 2-3 to prevent gradient vanishing. Finally, an attention mechanism (such as a channel attention module) is used to strengthen the weights of key spectral features (e.g., peak regions), outputting 1×128 embedded features. .

[0056] The input to the split spectral encoder is dual-channel split component data (asymmetric distortion component + Gaussian ideal component, dimension 2×N). Its structure is identical to the overlapping spectral encoder, but the number of input channels is adapted to 2, and the output is embedded features. (Dimension 1×128).

[0057] In this embodiment, a temperature coefficient τ=0.07 is introduced into the contrastive learning loss function to adjust the similarity calculation scale. The contrastive learning loss function is a bidirectional contrastive loss function used to calculate the similarity loss between the first embedded feature and the second embedded feature.

[0058] To achieve discriminative feature learning on unlabeled data, this embodiment designs a bidirectional contrastive loss function based on temperature-scale cross-entropy. The contrastive learning loss function is shown in the following equation:

[0059] .

[0060] in, L contrast To compare the learning loss function, N For batch size, subscript i and j These are all indexes of the samples. i=j When it is a positive sample, i ≠ jWhen the sample is negative, log() is the logarithmic function with base e, exp() is the exponential function with base e, and sim() is the cosine similarity function. and The first embedded feature is respectively and the second embedded feature The resulting low-dimensional representation is obtained after projection and L2 normalization. This contrastive learning loss function enables the clustering of features from similar samples and the separation of features from dissimilar samples.

[0061] A4. The second embedded feature is reconstructed into separate distorted spectral components and Gaussian spectral components using a dual-decoder network based on the spectral separation model. The reconstruction loss function is calculated based on mean squared error and physical constraints. The dual-decoder network includes a distorted spectral decoder and a Gaussian spectral decoder. The distorted spectral decoder employs a 3-layer one-dimensional deconvolution structure, corresponding to the convolution kernel parameters of the overlapping spectral encoder, and outputs an asymmetric distorted spectral component. The Gaussian spectral decoder has the same structure as the distorted spectral decoder and outputs an ideal Gaussian spectral component. The input to both decoders is the second embedded feature. h s The first embedded feature can also be combined if necessary. h o As input to each decoder.

[0062] In this embodiment, the reconstruction loss function is calculated according to the following formula:

[0063] .

[0064] in, L recon To reconstruct the loss function, L MSE Let the mean squared error loss function be . L physics The physical constraint loss function, λ phys The weights of the physical constraint loss function, λ phys Set it to 0.5 to balance the two losses.

[0065] The mean squared error loss function is calculated using the following formula:

[0066] .

[0067] Among them, y i These are the true spectral values. To reconstruct spectral values.

[0068] The physical constraint loss function is calculated using the following formula:

[0069] .

[0070] A5. Determine the joint loss function based on the contrastive learning loss function and the reconstruction loss function, and train the spectral separation model based on the training dataset. As an exemplary embodiment, the joint loss function is shown in the following equation:

[0071] .

[0072] in, L total For the joint loss function, λ contrast To compare the weights of the learning loss function, L contrast To compare the learning loss function, λ recon To reconstruct the weights of the loss function, L recon To reconstruct the loss function, the optimal weights are determined through grid search. λ contrast and λ recon The values ​​are 0.2 and 0.8 respectively.

[0073] In this embodiment, when training the spectral separation model based on the training dataset, the AdamW optimizer is used to update the model parameters. The learning rate is set to 0.001, the training duration is 200 epochs, and the batch size is 32. During training, stochastic gradient descent (SGD) is used to update the model parameters, and TensorBoard is used to monitor the training loss and validation loss in real time to ensure model convergence. The purpose of this step is to ensure both the discriminative nature of feature learning and the accuracy of spectral reconstruction through multi-objective optimization.

[0074] A6. Based on the measured single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, the parameters of the trained spectral separation model are fine-tuned to obtain the target spectral separation model; the target spectral separation model is used to separate the spectral components of the measured distorted overlapping spectral data of the FBG sensor network.

[0075] To adapt to the spectral characteristics of real FBG sensors (e.g., manufacturing errors such as bottom ripples can occur when femtosecond lasers inscribe FBGs, leading to increased spectral noise and affecting wavelength demodulation), this embodiment uses experimentally acquired FBG spectral data (such as 1000 sets of distorted overlapping spectral data obtained on a non-uniform strain test bench in the laboratory, as well as measured distorted spectral components and Gaussian spectral components) to fine-tune the trained model parameters. Specifically, in this embodiment, when fine-tuning the parameters of the trained spectral separation model, the learning rate is reduced to 1×10⁻⁶. -4The training cycle is set to 10, and an early stop strategy is adopted. Training is stopped when the validation loss does not decrease for 5 consecutive cycles.

[0076] A7. Using the target spectral separation model, the spectral components of the single-channel distorted overlapping spectrum under test are separated, and the corresponding physical parameters are output according to the FBG sensing principle. The separated spectral components are output in the form of a one-dimensional light intensity-wavelength data sequence (Gaussian spectrum and distorted spectrum). According to the FBG sensing principle, the precise physical parameters (such as strain and temperature) experienced by each FBG sensor can be calculated. For example, based on the wavelength-strain sensitivity coefficient (such as 1.2 pm / με) and wavelength-temperature sensitivity coefficient (such as 10 pm / ℃) of the FBG sensor, the wavelength is converted into strain and temperature values ​​and output.

[0077] The embodiments described above mainly take the distorted overlapping spectrum of an FBG sensor network as an example. It will be apparent to those skilled in the art that the core steps of this application—including spectral data simulation and preprocessing, constructing a dataset through data augmentation and splitting simulations, performing contrastive learning and reconstruction using a dual encoder-decoder network, and optimizing the model through a joint loss function—constitute a general framework for overlapping spectrum separation.

[0078] When applying this framework to other types of distorted overlapping spectra, only the specific parameters and physical constraints in the steps need to be adjusted:

[0079] In the data simulation step (step A1): "Adjust grating parameters" can be replaced with "Adjust characteristic parameters of the corresponding spectrum" (such as peak width, peak position and intensity of Raman spectrum; diffraction angle, peak shape function and relative intensity of XRD spectrum) to generate narrow linewidth spectra typical of this field and simulate their distortion and overlap.

[0080] In the physical constraint loss function (step A4): the constraints of the FBG spectrum can be replaced with the physical prior knowledge of the corresponding spectrum (such as the Lorentz peak shape constraint of the Raman spectrum, the diffraction angle position constraint of the XRD spectrum, etc.).

[0081] This adaptation is a routine operation for those skilled in the art, and once the adaptation is completed, the self-supervised learning paradigm proposed in this method can effectively overcome the bottleneck of scarce labeled data in this field and improve the demodulation accuracy of overlapping spectra.

[0082] The distortion overlap spectral separation method provided in the above embodiments of this application has the following significant advantages compared with the prior art:

[0083] High spectral separation accuracy: Through a dual encoder architecture and contrastive learning mechanism, the spectral structure similarity index (SSIM) reaches 0.9074, which is significantly better than the GAN method (0.8279) and the DCNN method (0.8677); the wavelength detection accuracy reaches 1.6388 pm, laying the foundation for high-precision measurement of physical parameters.

[0084] The strain measurement error is significantly reduced: precise spectral separation eliminates overlapping interference, and the strain measurement error is reduced from ±37.2με in the prior art to 1.35με, which meets the error requirements (≤5με) for precision structural health monitoring.

[0085] Strong noise robustness: The contrastive learning framework and attention mechanism enhance the key feature recognition capability. Even in a low signal-to-noise ratio environment of 15dB, the wavelength detection accuracy remains at 4.1pm, making it suitable for complex industrial environments.

[0086] Label-free data dependence: It adopts a self-supervised contrastive learning paradigm, which uses the inherent features of the data (such as spectral similarity) as supervision signals. It eliminates the need for manual labeling of spectral components, reduces data acquisition costs, and improves engineering practicality.

[0087] Based on the same inventive concept, this application also provides a system for implementing the distortion overlap spectral separation method described above. The solution provided by this system is similar to the solution described in the above method. In an exemplary embodiment, such as... Figure 2 As shown, a distortion-overlapping spectral separation system is provided, including the following functional modules:

[0088] The distortion and overlap spectrum simulation acquisition module is used to generate narrow linewidth spectra by adjusting grating parameters, simulate spectral distortion under different physical fields and noise interference, acquire distortion and overlap spectrum data of FBG sensor network under non-uniform physical field, and preprocess the distortion and overlap spectrum data; the distortion and overlap spectrum data includes spectral distortion and spectral overlap phenomena caused by non-uniform physical field distribution; the preprocessing includes standardizing the spectral intensity values ​​to the range of [0,1] and removing DC bias and system noise.

[0089] The training dataset construction module is used to perform data augmentation and split simulation on the preprocessed distorted overlapping spectral data to construct the training dataset. The training dataset includes several single-channel distorted overlapping spectra and corresponding dual-channel separation components. Data augmentation involves adding Gaussian noise to the distorted overlapping spectral data, extracting wavelength segment sequences for time slicing, performing time distortion by randomly nonlinearly stretching or compressing wavelengths, and reversing the spectral sequence order. The split simulation is based on the physical characteristics of the FBG sensor network spectrum to separate the preprocessed distorted overlapping spectral data, obtaining dual-channel separation components, which provide positive sample references for self-supervised comparative learning. The dual-channel separation components include a distorted spectral component and a Gaussian spectral component.

[0090] The contrastive learning computation module is used to receive single-channel distorted overlapping spectra and corresponding dual-channel separation components through the dual encoder network of the spectral separation model, output the first embedding feature and the second embedding feature respectively, and calculate the contrastive learning loss function; the dual encoder network includes an overlapping spectral encoder and a separating spectral encoder; the overlapping spectral encoder is used to receive single-channel distorted overlapping spectra and output the first embedding feature; the separating spectral encoder is used to receive dual-channel separation components and output the second embedding feature.

[0091] The reconstruction loss calculation module is used to reconstruct the second embedded feature into separate distorted spectral components and Gaussian spectral components through the dual decoder network of the spectral separation model, and calculate the reconstruction loss function based on the mean square error and physical constraints.

[0092] The model training module is used to determine the joint loss function based on the contrastive learning loss function and the reconstruction loss function, and to train the spectral separation model based on the training dataset.

[0093] The parameter fine-tuning module is used to fine-tune the parameters of the trained spectral separation model based on the measured single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, so as to obtain the target spectral separation model. The target spectral separation model is used to separate the spectral components of the measured distorted overlapping spectral data of the FBG sensor network.

[0094] The spectral separation and output module is used to separate the spectral components of the single-channel distorted overlapping spectrum under test using the target spectral separation model, and output the corresponding physical parameters according to the FBG sensing principle.

[0095] certainly, Figure 2 The architecture shown is merely exemplary; it can be omitted as needed when implementing different functionalities. Figure 2 One or at least two components of the system shown. As in another exemplary embodiment of this application, such as Figure 3As shown, this embodiment provides a distortion-overlapping spectrum separation system, including the following functional modules: a dataset construction module, used to construct a training dataset through data simulation and to construct a measured dataset based on measured data; a self-supervised contrastive learning pre-training module, used to pre-train the spectral separation model based on the simulated training dataset according to steps A3 to A5 in the above method embodiment, the structure of which is the same as the spectral separation model in the aforementioned embodiment; a semi-supervised fine-tuning module, used to fine-tune the parameters of the dual encoder network in the pre-trained spectral separation model based on the measured dataset; and a spectral separation and output module, used to separate the spectral components of the single-channel distortion-overlapping spectrum under test using the target spectral separation model and output the corresponding physical parameters according to the FBG sensing principle.

[0096] The core idea of ​​the distortion and overlap spectral separation system proposed in this application lies in resolving mixed spectral signals through simulated data construction, self-supervised contrastive learning, and physical constraint reconstruction. Although the above embodiments are mainly described in detail using FBG sensor networks as a typical example, those skilled in the art should understand that the system has high versatility and scalability. Without departing from the core principles of this application, by adapting to the physical characteristics of specific spectra (e.g., peak position and intensity of Raman spectra, diffraction angle and peak shape of XRD spectra), the system can also be applied to solve overlap and distortion separation problems in other types of spectra (including but not limited to Raman spectroscopy, X-ray diffraction spectroscopy, fluorescence spectroscopy, etc.).

[0097] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it can implement the distortion overlap spectral separation method provided in the previous embodiment.

[0098] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0099] In one exemplary embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. The memory may be DDR4 memory (capacity ≥ 8GB), and the processor may be an Intel Core i7 or an ARM processor of equivalent performance to ensure the computational power requirements for model training and inference.

[0100] In one exemplary embodiment, a computer-readable storage medium is provided, storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments. The storage medium may be an SSD (capacity ≥ 256GB) or an HDD, supporting high-speed data read / write and adapting to the large-capacity storage requirements of spectral data.

[0101] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments. This computer program can be packaged as a Python script (relying on the PyTorch or TensorFlow framework), supports operation on operating systems such as Windows and Linux, and facilitates integration with existing spectral analysis systems.

[0102] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0103] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0104] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0105] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0106] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for separating distorted and overlapping spectra, characterized in that, include: By adjusting the grating parameters to generate narrow linewidth spectra, spectral distortion under different physical fields and noise interference is simulated to obtain distorted and overlapping spectral data, and the distorted and overlapping spectral data is preprocessed; the distorted and overlapping spectral data includes spectral distortion and spectral overlap phenomena caused by uneven distribution of physical fields. The preprocessing includes normalizing the spectral intensity values ​​to the range of [0,1] and removing DC bias and system noise; Data augmentation and splitting simulation are performed on the preprocessed distorted overlapping spectral data to construct a training dataset. The training dataset includes several single-channel distorted overlapping spectra and corresponding dual-channel separation components. Data augmentation involves at least one of the following: adding Gaussian noise to the distorted overlapping spectral data; extracting wavelength segments for time slicing; applying random nonlinear stretching or compression to the wavelengths; and reversing the spectral sequence order. The splitting simulation, based on the physical properties of the distorted overlapping spectra, separates the preprocessed distorted overlapping spectral data to obtain dual-channel separation components, providing positive sample references for self-supervised comparative learning. The dual-channel separation components include a distorted spectral component and a Gaussian spectral component. A dual encoder network using a spectral separation model receives a single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, outputs a first embedding feature and a second embedding feature respectively, and calculates a contrastive learning loss function; the dual encoder network includes an overlapping spectral encoder and a separating spectral encoder; the overlapping spectral encoder is used to receive the single-channel distorted overlapping spectrum and output the first embedding feature; The separation spectral encoder is used to receive the dual-channel separation component and output a second embedded feature; Both the overlapping spectral encoder and the separate spectral encoder employ a one-dimensional convolutional neural network structure, which includes residual connections and an attention mechanism; the mathematical expression for the overlapping spectral encoder is: The mathematical expression for the separated spectral encoder is: ,in, Indicates overlapping spectral samples, Indicates the separation of component samples. and These represent the first embedding feature and the second embedding feature, respectively. The second embedded feature is reconstructed into a separated distorted spectral component and a Gaussian spectral component through a dual decoder network of a spectral separation model, and the reconstruction loss function is calculated based on the mean square error and physical constraints. The joint loss function is determined based on the contrastive learning loss function and the reconstruction loss function, and the spectral separation model is trained based on the training dataset. Based on the measured single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, the parameters of the trained spectral separation model are fine-tuned to obtain the target spectral separation model; the target spectral separation model is used to separate the spectral components of the measured distorted overlapping spectral data; The target spectral separation model is used to separate the spectral components of the single-channel distorted and overlapping spectrum under test, and the corresponding physical parameters are output based on the separated spectral components.

2. The distortion overlap spectral separation method according to claim 1, characterized in that, The contrastive learning loss function incorporates a temperature coefficient τ=0.07 to adjust the similarity calculation scale. This contrastive learning loss function is a bidirectional contrastive loss function used to calculate the similarity loss between the first embedded feature and the second embedded feature. The contrastive learning loss function is shown in the following formula: ； in, L contrast To compare the learning loss function, N For batch size, subscript i and j These are all indexes of the samples. i=j When it is a positive sample, i ≠ j When the sample is negative, log() is the logarithmic function with base e, exp() is the exponential function with base e, and sim() is the cosine similarity function. and The first embedded feature is respectively and the second embedded feature The low-dimensional representation is obtained after projection head and L2 normalization.

3. The distortion overlap spectral separation method according to claim 1, characterized in that, The reconstruction loss function is calculated using the following formula: ； in, L recon To reconstruct the loss function, L MSE Let the mean squared error loss function be . L physics The physical constraint loss function, λ phys The weights are the physical constraint loss function weights.

4. The distortion overlap spectral separation method according to claim 1, characterized in that, The joint loss function is shown in the following equation: ； in, L total For the joint loss function, λ contrast To compare the weights of the learning loss function, L contrast To compare the learning loss function, λ recon To reconstruct the weights of the loss function, L recon To reconstruct the loss function.

5. The distortion-overlapping spectral separation method according to claim 1, characterized in that, When training the spectral separation model based on the training dataset, the AdamW optimizer is used to update the parameters of the spectral separation model, with a learning rate set to 0.001, a training period of 200 epochs, and a batch size of 32. When fine-tuning the parameters of the trained spectral separation model, the learning rate is reduced to 1×10. -4 The training cycle is set to 10, and an early stop strategy is adopted. Training is stopped when the validation loss does not decrease for 5 consecutive cycles.

6. A distortion-overlapping spectral separation system, characterized in that, For implementing the distortion-overlapping spectral separation method as described in any one of claims 1-5, the distortion-overlapping spectral separation system comprises: The distortion and overlap spectrum simulation acquisition module is used to generate narrow linewidth spectra by adjusting grating parameters, simulate spectral distortion under different strain fields and noise interference, acquire distortion and overlap spectrum data, and preprocess the distortion and overlap spectrum data; the distortion and overlap spectrum data includes spectral distortion and spectral overlap phenomena caused by uneven strain distribution; the preprocessing includes standardizing the spectral intensity values ​​to the range of [0,1] and removing DC bias and system noise; A training dataset construction module is used to perform data augmentation and split simulation on preprocessed distorted overlapping spectral data to construct a training dataset. The training dataset includes several single-channel distorted overlapping spectra and corresponding dual-channel separation components. Data augmentation involves at least one of the following: adding Gaussian noise to the distorted overlapping spectral data, extracting wavelength segments for time slicing, performing random nonlinear stretching or compression of wavelengths, and reversing the spectral sequence order. The split simulation, based on the physical properties of the distorted overlapping spectra, separates the preprocessed distorted overlapping spectral data to obtain dual-channel separation components, providing positive sample references for self-supervised comparative learning. The dual-channel separation components include a distorted spectral component and a Gaussian spectral component. The contrastive learning computation module is used to receive single-channel distorted overlapping spectra and corresponding dual-channel separation components through a dual encoder network of a spectral separation model, output a first embedding feature and a second embedding feature respectively, and calculate a contrastive learning loss function; the dual encoder network includes an overlapping spectral encoder and a separating spectral encoder; the overlapping spectral encoder is used to receive single-channel distorted overlapping spectra and output the first embedding feature; the separating spectral encoder is used to receive dual-channel separation components and output the second embedding feature; The reconstruction loss calculation module is used to reconstruct the second embedded feature into separate distorted spectral components and Gaussian spectral components through the dual decoder network of the spectral separation model, and calculate the reconstruction loss function based on the mean square error and physical constraints. The model training module is used to determine the joint loss function based on the contrastive learning loss function and the reconstruction loss function, and to train the spectral separation model based on the training dataset; The parameter fine-tuning module is used to fine-tune the parameters of the trained spectral separation model based on the measured single-channel distorted overlapping spectrum and the corresponding dual-channel separation component, to obtain the target spectral separation model; the target spectral separation model is used to separate the spectral components from the measured distorted overlapping spectral data; The spectral separation and output module is used to separate the spectral components of the single-channel distorted and overlapping spectrum under test using the target spectral separation model, and output the corresponding physical parameters based on the separated spectral components.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the distortion overlap spectral separation method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the distortion overlap spectrum separation method according to any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the distortion overlap spectrum separation method according to any one of claims 1-5.

Images