1064nm fiber raman distributed temperature measurement signal denoising method and system based on variational mode

Abstract

The present application relates to the technical field of distributed optical fiber temperature sensing, and in particular to a 1064nm optical fiber Raman distributed temperature measurement signal denoising method and system based on variational mode, a temperature measurement system of a 1064nm pulse laser light source is built and a data set is constructed; a multi-scale attention fusion residual denoising model is constructed to perform primary denoising on the original Raman scattering signal; the temperature demodulation of the denoised signal obtains an initial temperature signal; an adaptive VMD bimodal network is constructed to perform secondary denoising on the initial temperature signal, and through variational mode decomposition, adaptive mode number optimization, mode classification and soft gating weighted reconstruction, accurate denoising is achieved; and a high-precision temperature result is output. Through the cascaded denoising architecture, the present application deeply fuses the strong representation capability of deep learning and the physical constraint advantage of variational mode decomposition, significantly improves the temperature measurement accuracy, robustness and result interpretability under complex working conditions, and solves the problems of serious noise in the temperature signal under extremely low signal-to-noise ratio and insufficient information utilization in single signal domain denoising.

Description

Technical Field

[0001] This invention relates to the field of distributed optical fiber temperature sensing technology, specifically to a method and system for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes. Background Technology

[0002] Distributed fiber optic temperature sensing technology has significant application value in energy extraction, downhole oil exploration, and other fields due to its ability to perform long-distance, continuous, and real-time temperature monitoring. Among them, Raman scattering-based temperature measurement technology is particularly suitable for complex downhole environments because it is only sensitive to temperature.

[0003] However, spontaneous Raman scattering signals are weak and susceptible to various types of noise interference, resulting in a low signal-to-noise ratio and limiting the accuracy and reliability of temperature measurement. Existing technologies, such as wavelet transform and moving average, rely heavily on manual parameter settings and have limited adaptability to different noise types. Some deep learning methods (such as one-dimensional convolutional neural networks and downsampling denoising methods) only denoise the original Raman signals (Stokes and anti-Stokes signals) without performing secondary processing on the demodulated temperature signal. Furthermore, these are purely data-driven "black box" models lacking physical constraints and exhibiting low interpretability.

[0004] Furthermore, current engineering systems mostly use 1550nm wavelength light sources. In contrast, the 1064nm wavelength exhibits stronger Raman scattering signals and lower transmission losses in extreme environments such as high temperature, high pressure, and hydrogen-rich conditions downhole, improving the long-term stability and measurement accuracy of the system, and making it more suitable for continuous temperature measurement downhole.

[0005] Therefore, it is necessary to propose a new denoising technique to address issues such as extremely low signal-to-noise ratio, insufficient utilization of denoising information in a single signal domain, lack of physical constraints, and strong parameter dependence of traditional variational mode decomposition (VMD). Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes. Through a cascaded denoising architecture, it deeply integrates the powerful representation capabilities of deep learning with the physical constraints of variational mode decomposition, significantly improving the temperature measurement accuracy, robustness, and interpretability of results under complex operating conditions. It aims to solve problems such as severe noise in temperature signals with extremely low signal-to-noise ratios, insufficient utilization of denoising information in a single signal domain, lack of physical constraints in traditional methods, and strong parameter dependence.

[0007] This invention is achieved through the following technical solution:

[0008] A method for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes is provided, comprising the following steps:

[0009] Step (1): Build a 1064nm fiber Raman distributed temperature measurement system, collect raw Raman signals to construct a dataset and perform preprocessing;

[0010] Step (2): Construct the multi-scale attention fusion residual denoising model MSAFR-D, train the model using the dataset, and obtain a denoising model for the original scattering signal;

[0011] Step (3): Demodulate the Stokes signal and anti-Stokes signal after noise reduction by the MSAFR-D model to obtain the initial temperature signal;

[0012] Step (4): Construct an adaptive VMD bimodal network AVDMNet, input the initial temperature signal into the network for training, and obtain a noise reduction model for the temperature demodulated signal;

[0013] Step (5): After the test set signal is successively denoised by MSAFR-D model, temperature demodulated and denoised by AVDMNet model, the final temperature measurement result is output.

[0014] Furthermore, in step (1), the 1064nm fiber Raman distributed temperature measurement system includes: a 1064nm pulsed laser source, a sensing fiber, a wavelength division multiplexer, a detector, a data acquisition card, and a computer; the data acquisition card is connected to the 1064nm pulsed laser source to provide a synchronous trigger signal to the source; the pulsed light generated by the 1064nm pulsed laser source is input to the wavelength division multiplexer and coupled into the sensing fiber by the wavelength division multiplexer; the backscattered Raman light generated in the sensing fiber returns to the wavelength division multiplexer, and after being separated into Stokes light and anti-Stokes light, it is converted into electrical signals by the corresponding detectors; the data acquisition card is connected to the detector to acquire electrical signals and is connected to the computer to transmit data.

[0015] Furthermore, in step (2), the multi-scale attention fusion residual denoising model MSAFR-D includes a multi-scale feature extraction module, an attention enhancement module, a deep fusion module, and a residual output module connected in sequence. The multi-scale feature extraction module is used to extract local detail features of the signal through multiple sets of one-dimensional dilated convolution kernels with different dilation rates set in parallel, and to capture long-distance dependency features of the signal through global context branches, and to fuse local detail features with long-distance dependency features to generate hybrid perception features.

[0016] Furthermore, the attention enhancement module includes a channel attention unit and a spatial attention unit; the channel attention unit is used to adaptively weight the hybrid sensory features in the channel dimension; the spatial attention unit is used to adaptively weight the channel-weighted features in the spatial dimension to obtain refined features.

[0017] Furthermore, the deep fusion module is implemented through a 1x1 convolutional layer, which is used to perform channel compression and cross-channel information fusion on the weighted features.

[0018] Furthermore, the residual output module includes at least one residual block and a global residual connection; the residual block is used for deep feature extraction; the global residual connection is used to predict the noise residual and remove the predicted noise residual from the original input signal to obtain the denoised output signal.

[0019] Furthermore, in step (2), the loss function used to train the multi-scale attention fusion residual denoising model MSAFR-D is the joint loss function; the joint loss function is composed of a weighted sum of a signal domain loss term and a temperature domain loss term; wherein, the signal domain loss term is used to constrain the difference between the denoised Stokes signal and the anti-Stokes signal and the corresponding pure signal, respectively; the temperature domain loss term is used to constrain the difference between the temperature obtained by demodulating the denoised signal and the pure temperature.

[0020] Furthermore, in step (4), constructing the adaptive VMD bimodal network AVDMNet includes:

[0021] The initial temperature signal is subjected to variational mode decomposition, and an adaptive optimization strategy is used to determine the optimal number of modes M.

[0022] The intrinsic mode functions (IMFs) obtained from the decomposition are classified into signal-dominant and noise-dominant states.

[0023] The adaptive weights of each IMF are calculated based on the Pearson correlation coefficient, time-domain correlation, and information entropy of each IMF with the initial temperature signal.

[0024] The temperature signal is finally reduced by noise by using adaptive weights to reconstruct each IMF.

[0025] Furthermore, determining the optimal number of modes M using an adaptive optimization strategy includes:

[0026] Set the initial maximum number of modes M. max ;

[0027] Calculate the Euclidean distance between the center frequencies of each adjacent intrinsic mode function (IMF) after decomposition;

[0028] If there are IMF pairs with a center frequency interval less than a preset threshold, then the mode number M is reduced to M=M-1 and decomposed again until there are no IMF pairs with a center frequency interval less than the preset threshold. The mode number at this point is taken as the optimal mode number M.

[0029] A variational mode-based 1064nm fiber Raman distributed temperature measurement signal noise reduction system for implementing the method includes:

[0030] The data acquisition and preprocessing module is used to acquire raw Raman signals using a 1064nm pulsed laser source and sensing fiber, and to preprocess the signals to construct a dataset.

[0031] The first-level noise reduction module is used to run the trained multi-scale attention fusion residual denoising model MSAFR-D to denoise the original Raman signal.

[0032] The temperature demodulation module is used to demodulate the noise-reduced Stokes signal and the anti-Stokes signal to obtain the initial temperature signal.

[0033] The second-stage noise reduction module is used to run the trained adaptive VMD bimodal network AVDMNet to perform noise reduction processing on the initial temperature signal;

[0034] The result output module is used to output the final temperature measurement result after two stages of noise reduction processing.

[0035] The beneficial effects of this invention are:

[0036] This application designs a cascaded architecture. The first-stage MSAFR-D model operates in the original signal domain (Stokes and anti-Stokes signals), aiming to remove high-frequency random noise introduced during photoelectric conversion and transmission. The second-stage AVDMNet model operates in the temperature demodulation domain, specifically handling non-stationary disturbances and residual noise in the demodulated temperature signal. This divide-and-conquer strategy achieves domain-specific noise reduction. The MSAFR-D model ensures that the original signal input to the temperature demodulation formula is as pure as possible, laying the foundation for accurate demodulation; the AVDMNet model refines the noise that may be amplified or introduced during demodulation. The two stages work together, fully utilizing the complementary information of different signal domains, achieving better noise reduction results and higher temperature signal fidelity than single-domain processing. Through two-stage domain-specific processing, the noise reduction accuracy and signal fidelity are significantly improved.

[0037] This application explicitly uses a 1064nm pulsed laser source. The 1064nm wavelength exhibits stronger Raman scattering signals and lower transmission loss in the high-temperature, high-pressure, and hydrogen-rich environment of downhole drilling, which physically improves the intrinsic quality of the signal and provides better input conditions for subsequent algorithm processing. The core of the AVDMNet model is variational mode decomposition. VMD is a signal processing method with a solid mathematical foundation; its eigenmode functions have practical physical meaning (quasi-orthogonal components at specific center frequencies). Embedding VMD into a deep learning framework is equivalent to introducing a physical prior to the model that the signal should be decomposed into different frequency bands.

[0038] The choice of a 1064nm light source makes the system more suitable for extreme industrial environments in terms of hardware. The introduction of physical priors constrains the solution space of the model, making it less likely to learn false noise patterns that violate physical laws. As a result, it exhibits stronger generalization ability and stability when facing complex working conditions not covered by training data (such as new noise types and intensity changes), avoiding the problem of overfitting or failure of pure data-driven models. It has excellent robustness and stability in extreme environments.

[0039] This application employs an adaptive optimization strategy in AVDMNet: setting a larger initial M max Then, by calculating the Euclidean distance between the center frequencies of adjacent IMFs, the parameters are dynamically adjusted until the optimal value M is found. This adaptive strategy reduces reliance on human experience and automates parameter settings. It effectively avoids mode aliasing or loss of effective signal caused by fixing the value of M, ensuring the accuracy of signal decomposition and laying a reliable foundation for subsequent mode classification and weighted reconstruction, thereby improving the integrity of the final temperature signal.

[0040] The MSAFR-D model's structural design is highly targeted: Multi-scale feature extraction: By using dilated convolutions with different dilation rates in parallel, the model can simultaneously capture local details of the signal (such as sharp temperature abrupt changes) and broader contextual information (such as slow baseline drift). Dual attention mechanism: Channel attention allows the model to autonomously learn which feature channels are more important; spatial attention allows the model to focus on key segments of the signal in the time dimension. This mechanism enables the model to adaptively focus on the information most beneficial to the denoising task and suppress irrelevant noise. This design allows the model to comprehensively and meticulously handle mixed high-frequency noise and low-frequency drift in Raman signals, significantly improving its adaptability to complex noise structures and feature extraction efficiency, thereby more accurately separating noise from the real signal and accurately extracting multi-scale features and key information.

[0041] In training the MSAFR-D model, this invention incorporates not only signal domain errors but also temperature domain errors into the loss function. This means the model optimization objective is not only to make the output signal appear clean but also to ensure the accuracy of the temperature demodulated from this signal. This introduces the physical process of temperature demodulation as a constraint into model training. AVDMNet does not simply retain or discard the individual IMFs after VMD decomposition; instead, it calculates a continuous adaptive weight (0~1.2) based on their correlation coefficient with the original signal, information entropy, etc., for weighted reconstruction. The joint loss function ensures that the noise reduction result is physically reliable, not just a mathematical fit. Soft-gated weighting is a gentler approach, avoiding false positives on suspected noise components, better preserving weak but effective temperature signals, further improving the discernibility of temperature trends and measurement accuracy, and enhancing the reliability and physical consistency of the results.

[0042] In summary, this invention establishes a hardware advantage through a 1064nm light source, achieves refined processing through a cascaded noise reduction architecture, ensures robustness and reliability through physically constrained deep learning, and addresses core challenges through specific technical means such as adaptive VMD and multi-scale attention models. It successfully combines the powerful representation capabilities of deep learning with the physical interpretability advantages of traditional signal processing, providing an effective solution for high-precision and high-reliability distributed fiber optic temperature measurement, and possesses significant engineering application value. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the 1064nm fiber Raman distributed temperature measurement system of the present invention.

[0044] Figure 2 This is a flowchart illustrating the structure of the multi-scale attention fusion residual denoising model MSAFR-D of this invention.

[0045] Figure 3 This is a flowchart illustrating the structure of the adaptive VMD dual-modal network AVDMNet of this invention.

[0046] As shown in the figure:

[0047] 1-Computer, 2-Data acquisition card, 3-Detector, 4-Wavelength division multiplexer, 5-Sensing fiber, 6-1064nm pulsed laser source. Detailed Implementation

[0048] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0049] Example 1:

[0050] This embodiment details the implementation process of the noise reduction method.

[0051] A noise reduction method for 1064nm fiber Raman distributed temperature measurement signals based on variational modes includes the following steps:

[0052] Step (1): System setup and data preparation.

[0053] Building such Figure 1 The diagram shows a 1064nm fiber Raman distributed temperature measurement system. The system includes: a computer 1, a data acquisition card 2, a detector 3, a wavelength division multiplexer 4, a sensing fiber 5, and a 1064nm pulsed laser source 6. The pulsed light generated by the 1064nm pulsed laser source 6 is input to the wavelength division multiplexer 4 and coupled into the sensing fiber 5. The backscattered Raman light generated in the sensing fiber 5 returns to the wavelength division multiplexer 4, where it is separated into Stokes light and anti-Stokes light, which are then converted into electrical signals by the corresponding detectors 3. The data acquisition card 2 is connected to the detector 3 to acquire the electrical signals and to the computer 1 to transmit data. Simultaneously, the data acquisition card 2 is connected to the 1064nm pulsed laser source 6 to provide a synchronization trigger signal to the source, ensuring strict synchronization between data acquisition and laser pulse emission.

[0054] In practical downhole oil and gas field monitoring scenarios, sensing fiber optic cable 5 is deployed along the wellbore. Stokes and anti-Stokes Raman raw scattering signals are acquired under different temperature (e.g., 30℃ to 150℃), different noise conditions (simulated by adjusting laser power and amplifier gain), and different pressures (e.g., 5MPa to 20MPa). Simultaneously, simulated signals are generated based on the physical principles of Raman scattering, and these, together with the actual acquired signals, constitute the dataset. The dataset undergoes dual-channel synchronous signal alignment, outlier removal, and data augmentation (e.g., adding random noise and time offsets). Standardization employs a joint normalization strategy: the mean μ and standard deviation σ of the Stokes and anti-Stokes channels are calculated on the training set, and the shared set (μ, σ) for both channels is normalized. Then, the acquired pure temperature data is used as labels to annotate the dataset, which is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0055] Step (2): Train the first-level noise reduction model MSAFR-D.

[0056] Build as Figure 2 The multi-scale attention fusion residual denoising model MSAFR-D is shown. This model processes dual-channel input signals. ;in Represents the original stokes signal. This represents the original Anti-Stokes signal, and L represents the signal length.

[0057] Multi-scale feature extraction module: The input signal is first mapped to a high-dimensional feature space through shallow embedding. Subsequently, the features are split into two parallel branches.

[0058] Multi-scale convolutional branches: utilizing four sets of dilation rates Different one-dimensional dilated convolutional kernels are used in parallel to extract local detail features. To control model complexity and ensure balanced contributions across the four scale branches, the total number of feature channels is C=64, and each branch outputs C / 4 channels. Local features F local The calculation is as follows:

[0059] ;

[0060] in, Indicates the void ratio The dilated convolution operation, in which Dual-channel input, convolution kernel In the channel dimension Perform joint convolution to achieve cross-channel feature fusion. This indicates a batch normalization operation. It is a non-linear activation function.

[0061] Global context branch: Calculated by the query matrix Q, key matrix K, and value matrix V, as shown in the following formula:

[0062] ;

[0063] Calculate the global relevance matrix using scaled dot product attention to generate global features F. global The formula is as follows:

[0064] ;

[0065] in: Let K be the transpose of the key matrix. This is a scaling factor, representing the dimension of the key vector. (Using...) Scaling the attention score is to prevent the dot product result from becoming too large, which would cause the gradient of the Softmax function to vanish.

[0066] Finally, the local features F local With global feature F global The channels are concatenated, and a 1x1 convolutional layer is used for cross-channel information fusion and dimensionality reduction to generate hybrid sensing features F. mix The formula is as follows:

[0067] ;

[0068] Attention Enhancement Module: For hybrid sensory features F mixPerform adaptive weighting.

[0069] Channel attention: Channel description vectors are obtained through global average pooling.

[0070] ;

[0071] in, This represents the response value of the c-th channel in the multi-scale feature map at position t, reflecting the activation intensity of this channel for the current spatial location features. This represents a summation operation in the spatial dimension (signal length direction), used to aggregate statistical information across the entire signal range. This is the normalization coefficient, used to eliminate the influence of signal length on statistical results. This represents the global descriptor scalar of the c-th channel, which characterizes the average response intensity of that channel over the entire signal range.

[0072] The channel weights are then generated via a gating network consisting of two fully connected layers. :

[0073] ;

[0074] in, This represents the Sigmoid function. This represents the first fully connected layer, used to reduce the dimensionality and compress the channel features, reduce the number of parameters, and introduce nonlinear modeling capabilities. This indicates the second fully connected layer. This step effectively suppresses noise-dominated invalid channels.

[0075] Channel-wise weighting is applied to the features to obtain the channel-enhanced features F. ca .

[0076] Spatial attention: Max pooling and average pooling are performed separately along the channel dimension. The results are concatenated and then convolution with a large receptive field is used to extract spatial context associations to generate a spatial mask. :

[0077] ;

[0078] Finally, the refined characteristic F is obtained. ref :

[0079] ;

[0080] in, Indicates a spatial mask. This represents the Sigmoid function. This represents a one-dimensional convolution operation used to model the correlation between adjacent spatial locations and introduce local contextual information; Max Pooling is maximum pooling, and Average Pooling is average pooling.

[0081] Deep fusion module: Uses a 1x1 convolutional layer as the bottleneck layer to integrate features F ref The number of channels is compressed to the original dimension to generate the fused feature F. fuse :

[0082] ;

[0083] in, Indicates batch normalization. Represents a 1×1 convolution. Convolution bias term.

[0084] Residual output module: Several residual blocks are stacked after the fusion layer for deep feature extraction. The operation of each residual block is as follows:

[0085] ;

[0086] in This represents the output characteristics of the k-th layer residual block. Indicates batch normalization. These represent the weights of the two convolutional kernels in the k-th residual block, corresponding to the first and second convolutional layers, respectively. This indicates a convolution operation.

[0087] The network ultimately predicts the noise residual component R by outputting the convolutional layer:

[0088] ;

[0089] The predicted noise is then removed from the original input signal through a global residual connection, resulting in the denoised output signal Y.

[0090] ;

[0091] in, This represents the output convolutional layer, used to map high-dimensional features to a noisy residual space. This represents the final residual block output characteristics. This represents the original noisy input signal.

[0092] When training this MSAFR-D model, a joint loss function is used. This function consists of a signal domain loss term. and temperature domain loss term The weighted summation is used, where the weight coefficient λ can be adjusted according to the actual situation (e.g., set to 1.0).

[0093] ;

[0094] ;

[0095] ;

[0096] in, , This represents the Stokes and anti-Stokes signals after denoising. , Indicates clean Stokes and anti-Stokes signals. Take 10 -3 . This indicates temperature demodulation. Indicates the pure temperature.

[0097] Using the Adam optimizer with a learning rate of 1e-4, iteratively train on the training set until the loss on the validation set no longer decreases significantly.

[0098] Step (3), temperature demodulation.

[0099] The MSAFR-D model trained in step (2) is applied to the test set to obtain the denoised Stokes signal. and anti-Stokes signal Subsequently, based on the principle of Raman scattering temperature demodulation, the denoised dual-channel signal was temperature demodulated using a temperature demodulation formula.

[0100] ;

[0101] in, It is Planck's constant. For Raman frequency shift, Boltzmann's constant, For anti-Stokes photon frequency, For Stokes photon frequency, , These are the anti-Stokes and Stokes scattering coefficients, respectively. , The transmission losses for anti-Stokes and Stokes signals are respectively. The length of the optical fiber. This is a reference temperature.

[0102] Obtain the initial temperature signal .

[0103] Step (4): Train the second-level noise reduction model AVDMNet.

[0104] Build as Figure 3 The adaptive VMD bimodal network AVDMNet is shown.

[0105] Adaptive VMD decomposition: for initial temperature signal The mathematical expression for variational mode decomposition (VMD) is shown below:

[0106] ;

[0107] in Let M represent the k-th modal component, and M be the number of modes.

[0108] To avoid over- or under-decomposition caused by a fixed number of modes, an adaptive optimization strategy is adopted to determine the optimal number of modes: an initial maximum number of modes M is set. max =10, calculate the Euclidean distance between the center frequencies of each adjacent intrinsic mode function (IMF) after decomposition; if there are IMF pairs with a center frequency interval less than a preset threshold (e.g., 50Hz), reduce the number of modes (M=M-1) and decompose again until there are no IMF pairs with a center frequency interval less than the threshold. At this time, M is the optimal number of modes M.

[0109] Modal classification and soft-gated weighting: Calculating each IMF component Compared with the original temperature signal Pearson correlation coefficient Based on a preset correlation coefficient threshold (e.g., 0.3), the IMF is classified into signal dominant states. ≥threshold) and noise-dominant state ( <Threshold). For each IMF, its time-domain correlation with the temperature signal is also considered. and information entropy Its adaptive weights are calculated through a learnable gating network. The weight ranges from 0 to 1.2. A weight close to 1.2 indicates enhancement of the IMF component, while a weight close to 0 indicates suppression of the component. This soft gating mechanism effectively avoids information loss that may be caused by hard thresholding. The specific calculation formula is as follows:

[0110] ;

[0111] ;

[0112] ;

[0113] in, This represents the Pearson correlation coefficient. Let represent the intrinsic mode function (IMF) obtained from the k-th VMD decomposition. This indicates the temperature signal input to AVDMNet. Describes the covariance operator. They represent The standard deviation of the temperature signal T. This represents the gate weight (continuous value) of the k-th IMF. This represents the time-domain correlation between the k-th mode and the temperature signal. This represents the information entropy of the k-th mode. express M represents the optimal number of modes. This represents the k-th IMF.

[0114] Signal reconstruction: using calculated adaptive weights For all IMF components Weighted reconstruction is performed to obtain the final denoised temperature signal. .

[0115] AVDMNet is trained using an optimization strategy similar to that in step (2), and the loss function can be either mean absolute error (MAE) or mean squared error (MSE).

[0116] Step (5), testing and output of results.

[0117] The test set data is input into the trained MSAFR-D model to obtain the denoised dual-channel signal. The initial temperature is obtained by temperature demodulation of the signal. Then, it is input into the trained AVDMNet model. After the above two-stage noise reduction processing, the system outputs the final high-precision and high-robust temperature measurement result. .

[0118] Example 2: System Implementation and Practical Application Case

[0119] This embodiment illustrates the system for implementing the above method and its deployment in a practical application scenario.

[0120] A variational mode-based 1064nm fiber Raman distributed temperature measurement signal noise reduction system for implementing the method, which can be integrated into computer 1, includes:

[0121] Data acquisition and preprocessing module: Through the control interface of data acquisition card 2, it acquires Stokes and anti-Stokes electrical signals from detector 3 and performs data preprocessing (such as alignment and normalization).

[0122] The first-level noise reduction module loads and runs the trained MSAFR-D model to reduce the noise of the original Raman scattering signal.

[0123] Temperature demodulation module: Based on the principle of Raman scattering, it realizes the decomposition from the noise-reduced dual-channel signal to the initial temperature signal.

[0124] The second-level noise reduction module loads and runs the trained AVDMNet model to perform secondary noise reduction on the initial temperature signal.

[0125] The results output module displays the final temperature distribution curve in real time on the computer interface and stores the data in the database. At the same time, it can transmit alarm information (such as temperature abnormalities) to the monitoring center through the network interface.

[0126] Practical Application Case: Downhole Oil and Gas Field Temperature Profile Monitoring

[0127] In a shale gas production well, this system is used for distributed temperature monitoring to monitor the formation temperature recovery after fracturing and the fluid flow state inside the wellbore.

[0128] Deployment: A multimode sensing fiber optic cable 5, approximately 2000 meters long, is lowered into the target well section along with the production tubing. On the surface, the fiber optic cable is connected to the computer 1, data acquisition card 2, detector 3, wavelength division multiplexer 4, and 1064nm pulsed laser source 6, which are used to deploy this system.

[0129] Data Acquisition and Processing: The system operates continuously at a sampling frequency of 1Hz. The acquired raw Raman signal has an extremely low signal-to-noise ratio (SNR) (approximately 5dB), resulting in drastic fluctuations in the directly demodulated temperature curve. The system feeds the acquired data into the MSAFR-D model in real time for the first stage of noise reduction, improving the SNR to approximately 15dB. Temperature demodulation is then performed to obtain the initial temperature profile. This initial temperature signal still exhibits non-stationary disturbances. Next, the AVDMNet model performs a second stage of noise reduction on the temperature signal, effectively smoothing out random fluctuations while preserving true temperature steps (such as the low-temperature anomalies corresponding to gas-producing zones).

[0130] After two stages of noise reduction, the signal-to-noise ratio of the final output temperature profile was improved to over 25 dB, the temperature measurement accuracy reached ±1°C, and the temperature resolution was better than 0.1°C. The system clearly revealed the precise location and relative gas production of multiple gas-producing layers within the wellbore and successfully monitored the slow recovery trend of formation temperature over time. The system operated stably continuously for over three months, significantly outperforming the previously used traditional wavelet transform-based method, providing crucial data support for production optimization.

[0131] Example 3: Ablation Experiment and Performance Analysis

[0132] To verify the effectiveness of each innovative module in the technical solution of this application and the superiority of the overall solution, we designed systematic ablation experiments and comparative experiments. All experiments were conducted on the same dataset and under the same experimental environment to ensure the comparability of the results.

[0133] 4.1 Ablation Experiment of Cascaded Noise Reduction Architecture

[0134] This experiment aims to verify the necessity of the cascaded noise reduction architecture (i.e., the combination of MSAFR-D and AVDMNet). By setting different module combinations, its impact on the final temperature measurement accuracy is evaluated. The test set includes Raman scattering signals of various noise intensities and types. Evaluation metrics include the mean absolute error (MAE), root mean square error (RMSE), temperature resolution, and peak signal-to-noise ratio (PSNR) of the temperature measurement.

[0135] The experimental results are shown in Table 1 below:

[0136] Table 1. Ablation Experiment Results of Cascaded Noise Reduction Architecture

[0137]

[0138] Verification of the effectiveness of the cascaded noise reduction architecture: Comparison of data from the present invention (No. 1) and single-stage noise reduction models (Nos. 2 and 3) shows that the present invention significantly outperforms any single-stage noise reduction scheme in all metrics. Compared to "noise reduction only in the first stage," the complete scheme reduces MAE by 46.4% and RMSE by 46.5%. This fully demonstrates that the cascaded, collaborative noise reduction architecture proposed in the present invention, which performs noise reduction in the original signal domain and the temperature demodulation domain, can fully utilize the complementarity of information in both domains. This is key to achieving high-precision temperature measurement and effectively solves the problem of "insufficient utilization of noise reduction information in a single signal domain" mentioned in the background technology.

[0139] The necessity of multi-scale feature extraction in the MSAFR-D model: Comparing the data of this invention (No. 1) with the data of replacing MSAFR-D with a single convolutional kernel model (No. 4), the performance of the model of this invention is superior across the board. This shows that using multiple sets of different hole rates The dilated convolution parallel extraction of multi-scale features (corresponding to the core design of the MSAFR-D model) can more effectively capture local details and global context information in Raman signals compared with traditional single-scale convolution, thereby improving the quality of the first-level noise reduction and laying a better foundation for subsequent temperature demodulation and second-level noise reduction.

[0140] Superiority of VMD physical constraints in AVDMNet: Comparing the model data of this invention (No. 1) with the model data of replacing AVDMNet with a common one-dimensional denoising network (No. 5), the model of this invention also performs better. This verifies that introducing variational mode decomposition (VMD) as a physical constraint in the second-stage denoising can enhance the interpretability of the model and improve the generalization ability to unseen noise modes compared to a purely data-driven black-box network, thereby obtaining a more accurate and stable temperature output.

[0141] As can be seen from the analysis in Table 1, the ablation experimental data strongly demonstrate that the complete technical solution proposed in this application and its core modules (multi-scale design of MSAFR-D and VMD constraint of AVDMNet) are indispensable for achieving excellent noise reduction performance. The modules work together to solve the key technical problems mentioned in the background technology, such as extremely low signal-to-noise ratio, insufficient information utilization, and lack of physical constraints.

[0142] 4.2 Detailed Analysis of Multi-Scale Feature Extraction Module

[0143] This experiment further investigates the impact of different designs of the multi-scale feature extraction module in the MSAFR-D model on performance. The experimental results are shown in Table 2.

[0144] Table 2. Refinement Analysis Results of the Multi-Scale Feature Extraction Module

[0145]

[0146] As shown in Table 2, comparing numbers 1 and 4 in this embodiment, the introduction of the multi-scale module significantly improves performance, proving its necessity. Comparing numbers 2, 3, and 4, the complete model employing multiple void ratios in parallel exhibits the best performance. While a single void ratio (whether void ratio = 1 or void ratio = 4) is superior to that without a multi-scale module, it cannot fully utilize the feature information at different scales. This demonstrates the rationality of the parallel multi-scale structure design in this application, which can simultaneously capture local details and a wider range of contextual information of the signal.

[0147] 4.3 Ablation Experiment of Key AVDMNet Components

[0148] This experiment aims to verify the necessity of three key design elements in AVDMNet (adaptive VMD, modal classification, and weighting). The experimental results are shown in Table 3 below.

[0149] Table 3 Ablation Experiment Results of Key AVDMNet Components

[0150]

[0151] The analysis in Table 3 reveals the importance of VMD physical constraints: compared to numbers 1 and 4, removing VMD significantly degrades performance, increasing the mean absolute error by 48.3%. Furthermore, VMD also significantly improves temperature resolution at a distance of 8-10km, with a 42.2% improvement compared to numbers 1 and 4. This demonstrates that introducing VMD decomposition as a physical prior is crucial for enhancing noise reduction performance.

[0152] Advantages of adaptive modality number optimization: Compared with indices 2 and 4, the complete model using the adaptive strategy to determine the number of modalities exhibits better performance. This demonstrates that the adaptive optimization strategy can avoid modal aliasing or under-decomposition problems caused by fixed parameters, thus improving the robustness of the method.

[0153] Effectiveness of Modal Classification and Weighting Mechanism: Comparing No. 3 and No. 4, the performance improvement after adopting adaptive weighting is significant. This indicates that the correlation-based modal classification and soft-gated weighting mechanism can process each IMF component more precisely, effectively enhancing signal components and suppressing noise components.

[0154] 4.4 Performance Comparison with Existing Noise Reduction Methods

[0155] This experiment compares the complete solution of this application with existing mainstream noise reduction methods under two different average number of iterations (4000 and 8000) to evaluate its overall performance advantages. The experimental results are shown in Table 4 below.

[0156] Table 4 Performance Comparison Results of Different Noise Reduction Methods

[0157]

[0158] As shown in Table 4, the significant performance advantages are evident: regardless of whether the average is 4000 or 8000 cycles, the complete solution of this application (serial numbers 1 and 6) significantly outperforms other comparative methods in all evaluation metrics. For example, at an average of 8000 cycles, the MAE of this solution is reduced by 37.8% compared to the second-best RRCDNet, and the PSNR is improved by 4.88 dB, demonstrating its superior overall noise reduction performance.

[0159] Effective utilization of signal accumulation: As the average number of times increases from 4000 to 8000, the performance of all methods improves, but the improvement of the proposed solution is the largest (MAE is reduced by 31.1%), indicating that it can more effectively utilize the increased signal energy and has excellent noise adaptability and signal extraction efficiency.

[0160] Advantages over existing deep learning methods: This approach significantly outperforms existing deep learning methods such as RRCDNet, DSDN, and 1D-CNN, demonstrating the advanced nature of the design concept that combines cascaded noise reduction architecture with physically constrained deep learning.

[0161] Significant improvement over traditional methods: Compared with the traditional wavelet denoising (WD) method, this scheme shows an order-of-magnitude performance improvement, highlighting the breakthrough effect of this invention in solving the problem of distributed fiber optic temperature measurement under extremely low signal-to-noise ratio conditions.

[0162] The aforementioned series of experimental data fully and consistently demonstrate that the 1064nm fiber Raman distributed temperature measurement signal denoising method and system based on variational modes proposed in this application effectively solves the key technical problems mentioned in the background technology through its innovative multi-scale feature extraction, cascaded denoising architecture, and deep integration of physical constraints and data-driven approaches. It has significant advantages in denoising accuracy, stability, robustness, and adaptability, and possesses outstanding industrial application value.

[0163] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments and drawings are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A noise reduction method for 1064nm fiber Raman distributed temperature measurement signals based on variational modes, characterized in that: Includes the following steps: Step (1): Build a 1064nm fiber Raman distributed temperature measurement system, collect raw Raman signals to construct a dataset and perform preprocessing; Step (2): Construct the multi-scale attention fusion residual denoising model MSAFR-D, train the model using the dataset, and obtain a denoising model for the original scattering signal. The multi-scale attention fusion residual denoising model MSAFR-D includes a multi-scale feature extraction module, an attention enhancement module, a deep fusion module, and a residual output module connected in series. The multi-scale feature extraction module is used to extract the local detail features of the signal through multiple sets of one-dimensional dilated convolution kernels with different dilation rates set in parallel, and to capture the long-range dependency features of the signal through the global context branch, and to fuse the local detail features and the long-range dependency features to generate hybrid sensing features. Step (3): Demodulate the Stokes signal and anti-Stokes signal after noise reduction by the MSAFR-D model to obtain the initial temperature signal; Step (4): Construct an adaptive VMD bimodal network AVDMNet, input the initial temperature signal into the network for training, and obtain a noise reduction model for the temperature demodulated signal; The construction of the adaptive VMD bimodal network AVDMNet includes: The initial temperature signal is subjected to variational mode decomposition, and an adaptive optimization strategy is used to determine the optimal number of modes M. The intrinsic mode functions (IMFs) obtained from the decomposition are classified into signal-dominant and noise-dominant states. The adaptive weights of each IMF are calculated based on the Pearson correlation coefficient, time-domain correlation, and information entropy of each IMF with the initial temperature signal. The temperature signal after noise reduction is obtained by weighting and reconstructing each IMF using adaptive weights. Determining the optimal number of modes M using an adaptive optimization strategy includes: Set the initial maximum number of modes M. max ; Calculate the Euclidean distance between the center frequencies of each adjacent intrinsic mode function (IMF) after decomposition; If there are IMF pairs with a center frequency interval less than a preset threshold, then reduce the number of modes M, M=M-1 and decompose again until there are no more IMF pairs with a center frequency interval less than the preset threshold. The number of modes at this time is taken as the optimal number of modes M. Step (5): After the test set signal is successively denoised by MSAFR-D model, temperature demodulated and denoised by AVDMNet model, the final temperature measurement result is output.

2. The method for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes according to claim 1, characterized in that: In step (1), the 1064nm fiber Raman distributed temperature measurement system includes a computer, a data acquisition card, a detector, a wavelength division multiplexer, a sensing fiber, and a 1064nm pulsed laser source; the data acquisition card is connected to the computer and the 1064nm pulsed laser source to provide a synchronous trigger signal to the source. The pulsed light generated by the 1064nm pulsed laser source is input to the wavelength division multiplexer (WDM) and then coupled into the sensing fiber by the WDM. The backscattered Raman light generated in the sensing fiber returns to the WDM, where it separates the Stokes light and the anti-Stokes light, which are then converted into electrical signals by the corresponding detectors. The data acquisition card is connected to the detector to acquire the electrical signals and is connected to the computer to transmit data.

3. The method for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes according to claim 1, characterized in that: The attention enhancement module includes a channel attention unit and a spatial attention unit; the channel attention unit is used to adaptively weight the hybrid sensing features in the channel dimension; the spatial attention unit is used to adaptively weight the channel-weighted features in the spatial dimension to obtain refined features.

4. The method for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes according to claim 1, characterized in that: The deep fusion module is implemented through a 1x1 convolutional layer. The deep fusion module is used to perform channel compression and cross-channel information fusion on weighted features.

5. The method for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes according to claim 1, characterized in that: The residual output module includes at least one residual block and a global residual connection; the residual block is used for deep feature extraction; the global residual connection is used to predict noise residuals and remove the predicted noise residuals from the original input signal to obtain the denoised output signal.

6. The method for denoising 1064nm fiber Raman distributed temperature measurement signals based on variational modes according to claim 1, characterized in that: In step (2), the loss function used to train the multi-scale attention fusion residual denoising model MSAFR-D is the joint loss function. The joint loss function is composed of a weighted sum of a signal domain loss term and a temperature domain loss term. The signal domain loss term is used to constrain the difference between the denoised Stokes signal and the anti-Stokes signal and the corresponding clean signal, respectively. The temperature domain loss term is used to constrain the difference between the temperature obtained by demodulating the denoised signal and the clean temperature.

7. A variational mode-based 1064nm fiber Raman distributed temperature measurement signal noise reduction system for implementing the method of any one of claims 1 to 6, characterized in that: include: The data acquisition and preprocessing module is used to acquire raw Raman signals using a 1064nm pulsed laser source and sensing fiber, and to preprocess the signals to construct a dataset. The first-level noise reduction module is used to run the trained multi-scale attention fusion residual denoising model MSAFR-D to denoise the original Raman signal. The temperature demodulation module is used to demodulate the noise-reduced Stokes signal and the anti-Stokes signal to obtain the initial temperature signal. The second-stage noise reduction module is used to run the trained adaptive VMD bimodal network AVDMNet to perform noise reduction processing on the initial temperature signal; The result output module is used to output the final temperature measurement result after two stages of noise reduction processing.

Images