Brillouin optical fiber sensing noise reduction method based on dual-stage DnCNN
By segmenting the Brillouin gain curve into near-end and far-end segments, and amplifying the far-end segment and training a DnCNN model, the problem of low noise level of Brillouin fiber optic sensors in long-distance measurements was solved, resulting in a significant improvement in signal-to-noise ratio and measurement accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-08
AI Technical Summary
Existing Brillouin fiber optic sensors have excessively low noise levels in long-distance measurements, resulting in poor noise reduction performance of DnCNN and affecting measurement accuracy.
The Brillouin gain curve is divided into a proximal segment and a distal segment. The distal segment is amplified and the DnCNN model is trained separately. The proximal and distal pre-trained models are used for denoising to improve the noise level to above the effective denoising threshold of DnCNN.
It significantly improves the denoising effect of DnCNN on long-distance Brillouin fiber sensors, enhances the signal-to-noise ratio, and reduces the uncertainty of polarization-related measurements.
Smart Images
Figure CN121996928A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of distributed optical fiber sensing technology, and in particular to a Brillouin optical fiber sensing noise reduction method based on a two-stage DnCNN. Background Technology
[0002] Brillouin optical time-domain analysis (BOTDA) is a high-precision and high-stability distributed temperature and strain sensing technology used in many practical applications. The measurement accuracy of BOTDA fiber optic sensors ultimately depends on the signal-to-noise ratio (SNR) of the measured Brillouin gain. However, in long-distance BOTDA systems, some unavoidable random noise is introduced during the acquisition, transmission, amplification, and filtering processes due to external environmental factors, which degrades the system's measurement accuracy and affects the sensing effect.
[0003] Currently, many methods have been proven to improve SNR, such as pulse coding, distributed Raman amplification, concentrated amplification, and image denoising. Among them, image denoising has been proven to be very effective in enhancing the SNR of BOTDA fiber optic sensors. A relatively effective image denoising technique is the convolutional denoising neural network (DnCNN) based on residual learning and batch normalization. It can provide an SNR improvement of more than 10dB without introducing time delay and edge blurring. Moreover, the data processing time of the denoising network accelerated by GPU is negligible.
[0004] However, the denoising effectiveness of DnCNN is highly dependent on the noise level of the original data, but due to the long sensing distance of Brillouin fiber, the denoising effect of using DnCNN is poor. Summary of the Invention
[0005] This invention provides a Brillouin fiber optic sensing noise reduction method based on a two-stage DnCNN, which can solve the problems existing in the prior art.
[0006] This invention provides a Brillouin fiber sensor noise reduction method based on a two-stage DnCNN, comprising the following steps: During the transmission of scattered light in the Brillouin fiber optic sensing device, signal data containing the original noise signal of the scattered light during transmission is acquired, and the gain curve of Brillouin gain as a function of distance is obtained based on the signal data. The gain curve is divided into a proximal segment and a distal segment based on the gain change rate. The gain curve of the distal segment is amplified so that its Brillouin gain is within the normalized gain range, so as to improve the noise level represented by the original noise signal of the distal segment to above the effective noise reduction threshold of the convolutional noise reduction neural network DnCNN. The DnCNN model is trained using the original noise signal of the near end segment to obtain the near end pre-trained DnCNN model; the DnCNN model is trained using Gaussian noise that matches the amplified noise level to obtain the far end pre-trained DnCNN model; the near end pre-trained DnCNN model and the far end pre-trained DnCNN model are used respectively to denoise the gain curves of the near end segment and the far end segment to obtain the denoised gain curves of the near end segment and the far end segment. The denoised gain curves of the near-end and far-end segments are merged into a Brillouin gain denoising curve. The temperature and stress distribution along the fiber are then analyzed based on the Brillouin gain denoising curve.
[0007] Preferably, when dividing the gain curve into a proximal segment and a distal segment, the selection of the segmentation range includes: The segmentation range is determined based on the original noise level and the Brillouin gain attenuation with distance; When the noise level of the normalized gain curve is 1 / 255 when the measured Brillouin gain as a function of distance is found to be 1 / 255, the noise level is increased to above 5 / 255, and the gain curve is split when the gain value decays to 0.2.
[0008] Preferably, obtaining the near-end pre-trained DnCNN model and the far-end pre-trained DnCNN model includes: When training DnCNN models with different noise levels, the relationship between the signal-to-noise ratio (SNR) improvement of normalized data and the noise level is obtained. Based on the relationship between the SNR improvement and the noise level, it is measured that when the noise level represented by the original noise signal is less than 1 / 255, the SNR improvement brought by image denoising is 0. When the noise level is increased from 1 / 255 to 5 / 255, the signal-to-noise ratio increases from 0dB to 12dB. Then, the convolutional denoising neural network DnCNN is trained using the gain curve data with a noise level of 5 / 255 to obtain a pre-trained DnCNN model for the far end that matches the noise level of the far end segment. The pre-trained DnCNN model is obtained by training the convolutional denoising neural network DnCNN using the gain curve data of the proximal segment.
[0009] Preferably, the gain curve of the distal segment is amplified, and the amplification factor is dynamically adjusted according to the target noise level so that its Brillouin gain is within the normalized gain range of [0,1], so as to raise the noise level of the distal segment to between 5 / 255 and 10 / 255, which is the effective noise reduction threshold of the convolutional denoising neural network DnCNN.
[0010] This invention provides a Brillouin fiber sensing noise reduction method based on a two-stage DnCNN, which has the following advantages compared with the prior art: This invention divides the gain curve into a proximal segment and a distal segment. The gain curve of the distal segment is amplified so that its Brillouin gain is within the normalized gain range of [0,1]. This operation amplifies the noise level at the point where the original gain data has the lowest signal-to-noise ratio, thereby raising the noise level of the distal segment to above the effective denoising threshold of the Convolutional Denoising Neural Network (DnCNN). This ensures that the noise levels of both the proximal and distal segments reach the noise level of DnCNN denoising. Then, pre-trained DnCNN models matching the noise levels of the proximal and distal segments are used to denoise the two segments respectively. This two-stage DnCNN denoising method significantly improves the denoising effect of DnCNN. Attached Figure Description
[0011] Figure 1 This is a schematic diagram of the overall process of a Brillouin fiber sensing noise reduction method based on a two-stage DnCNN provided in an embodiment of the present invention. Figure 2 A schematic diagram illustrating the relationship between the signal-to-noise ratio improvement and noise level of normalized data in a Brillouin fiber sensing noise reduction method based on a two-stage DnCNN provided in an embodiment of the present invention. Figure 3 The diagram shows the original Brillouin gain versus time / distance curve (blue curve) with a noise level of 1 / 255 and its 5x magnified replica (red curve) of a Brillouin fiber sensing noise reduction method based on a two-stage DnCNN provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the data function relationship of a Brillouin fiber sensing denoising method based on a two-stage DnCNN provided in an embodiment of the present invention; wherein (a) is the function relationship between the noise standard deviation (STD) of the original data (black curve), the data after first-stage image denoising (blue curve) and second-stage image denoising (red curve) and distance; (b) is the function relationship between the signal-to-noise ratio of the original and denoised Brillouin gain data and distance. Detailed Implementation
[0012] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0013] Current DnCNN-based image denoising requires certain conditions. Actual measurements show that when the noise level is below 1 / 255, the signal-to-noise ratio (SNR) improvement from image denoising is close to zero. This means that for some high-precision measurements, image denoising methods are ineffective for BOTDA fiber optic sensors when the original noise level is below 1 / 255. Only when the noise level of the original data is greater than 5 / 255 can a significant denoising effect be achieved, improving the SNR by approximately 8dB. Therefore, the key to solving this problem is how to improve the noise level of the original image without introducing new noise.
[0014] To address the issue of poor denoising performance of DnCNN due to excessively low original noise levels, this invention provides a Brillouin fiber sensing denoising method based on a two-stage DnCNN, such as... Figure 1 As shown, at the far end of the fiber, where the original gain data has the lowest signal-to-noise ratio, the noise level is amplified to improve it. This enhances the noise removal performance of DnCNN for BOTDA fiber sensors without amplifying the original noise or introducing new noise, thus significantly reducing the uncertainty of polarization-dependent measurements. Specifically, the steps include:
[0015] Step S1: Start the BOTDA fiber optic sensing device and obtain the Brillouin gain as a function of distance.
[0016] Step S2: Divide the gain curve according to the gain change rate.
[0017] When selecting the segmentation range, it is necessary to determine it based on the original noise level and the Brillouin gain attenuation with distance. For example, if the noise level of the normalized gain curve is measured to be 1 / 255, a more significant denoising effect needs to be achieved by increasing the noise level to above 5 / 255. Therefore, segmentation can be performed when the gain attenuates to 0.2. Normalizing the far end results in a 5-fold increase in noise level, leading to an 8dB signal-to-noise ratio (SNR) improvement. For different BOTDA gain curves, the segmentation point can be slightly modified to achieve a more significant SNR enhancement.
[0018] Among them, the noise level and signal-to-noise ratio of the gain curve were measured using Matlab offline processing; Step S3: Magnify the gain curve of the far end so that its Brillouin gain is within the normalized gain range of [0,1].
[0019] Step S4: Use the DnCNN model trained with the original noise level and the amplified noise level to denoise the near-end curve and far-end curve of the optical fiber, respectively, to obtain the denoised Brillouin curve, and analyze the temperature and stress distribution along the sensing fiber.
[0020] DnCNN denoising uses a PC-based neural network model for noise reduction and GPU acceleration. During model training at different noise levels, the relationship between the signal-to-noise ratio (SNR) improvement of normalized data and the noise level was measured. The data showed that when the noise level was below 1 / 255, the SNR improvement from image denoising was 0. Subsequently, as the noise level increased from 1 / 255 to 50 / 255, the SNR improvement increased from 0 dB to 12 dB. Specifically, when the noise level of the original data was greater than 5 / 255 (i.e., an SNR improvement exceeding 8 dB), DnCNN significantly improved sensing performance.
[0021] like Figure 2 The figure shows the relationship between the calculated signal-to-noise ratio (SNR) improvement of normalized data and the noise level. As can be seen from the figure, when the noise level is below 1 / 255, image denoising cannot be effectively performed. When the noise level is between 1 / 255 and 10 / 255, the SNR improvement follows a logarithmic function trend, increasing significantly, and gradually slowing down after 10 / 255, with a maximum SNR improvement of about 12dB. Therefore, in order to maximize noise reduction and set a reasonable amplification factor, setting the noise level to 5 / 255~10 / 255 can achieve an SNR improvement of about 8dB, significantly improving sensing performance.
[0022] like Figure 3 The figure shows that the Brillouin gain measured at 10 meters is 1% when the pump-probe frequency difference is 10.854 GHz. This means that when the original data is normalized to the [0 1] range for single-stage image denoising, the maximum magnification is only 100 times. The blue curve is the curve of the normalized Brillouin gain with a noise level of 1 / 255 at a magnification of 100 times as a function of distance. Figure 2 It is known that the low noise level makes it difficult to improve the signal-to-noise ratio after image denoising; due to the distance loss, the gain in the later stages is relatively small. Therefore, the method of this invention is used to amplify the later stages of the blue curve by a factor of 5, that is, to perform a [0,1] normalization operation on the 25-50km fiber range, resulting in a noise level that is also amplified by a factor of 5 to 6 / 255; further, pre-trained noise models with noise levels of 1.2 / 255 and 6 / 255 are used to denoise these two parts of the data respectively, and the noise level is then reduced by a factor of 5. Figure 2 It can be seen that although the Brillouin gain signal-to-noise ratio cannot be improved in the 0-25km fiber section, it is expected to achieve a signal-to-noise ratio improvement of up to 8dB in the later 25-50km fiber section where more measurements are needed.
[0023] like Figure 4(a) shows the relationship between the noise standard deviation (STD) and distance for the original data (black curve), the data after first-level image denoising (blue curve), and the data after second-level image denoising (red curve). Since the first-level image denoising has a denoising effect of 0 at a noise level of 1 / 255, its noise standard deviation completely coincides with the standard deviation of the original image. The second-level image uses the same denoising model as the first-level image in the first half, so its curve also coincides with the original image. However, in the second half, due to the use of a model with a higher denoising level, the noise standard deviation increases from 4.5 × 10⁻⁶ in the 50-60km fiber optic region. -5 Reduced to 6×10 -6 This indicates that the SNR improved by 8.75 dB, compared to Figure 1 The theoretically predicted 9dB SNR enhancement from a noise level of 1.2 / 255 to 6 / 255 is in very good agreement; Figure 4 (b) It can also be seen that the SNR of the first half of the fiber is almost the same, while the dual-stage DnCNN initially provides a 4dB SNR improvement in the second half, and eventually achieves a SNR improvement of more than 7dB, effectively removing noise.
[0024] This invention uses a Brillouin gain curve with a sensing distance of 50 kilometers and a spatial resolution of 2 meters. The two-stage image denoising scheme used improves the signal-to-noise ratio at the far-end fiber optic end by about 7 dB, thereby reducing the uncertainty of BFS measurement by 3.5 times.
[0025] This invention improves the noise removal performance of DnCNN under low raw data noise levels. For the far-end fiber end, the average gain value is low due to fiber loss, and normalization can significantly improve the noise level, thereby significantly improving the signal-to-noise ratio of its Brillouin gain through DnCNN.
[0026] This invention significantly reduces the uncertainty of polarization-dependent measurements by improving the signal-to-noise ratio of the far-end fiber. The residual noise after image denoising is gain-dependent polarization noise, and further efforts can focus on optimizing the pump-probe interaction to minimize this type of noise.
[0027] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A Brillouin fiber optic sensing noise reduction method based on a two-stage DnCNN, characterized in that, Includes the following steps: During the transmission of scattered light in the Brillouin fiber optic sensing device, signal data containing the original noise signal of the scattered light during transmission is acquired, and the gain curve of Brillouin gain as a function of distance is obtained based on the signal data. The gain curve is divided into a proximal segment and a distal segment based on the gain change rate. The gain curve of the distal segment is amplified so that its Brillouin gain is within the normalized gain range, so as to improve the noise level represented by the original noise signal of the distal segment to above the effective noise reduction threshold of the convolutional noise reduction neural network DnCNN. The DnCNN model is trained using the original noise signal of the near end segment to obtain the near end pre-trained DnCNN model; the DnCNN model is trained using Gaussian noise that matches the amplified noise level to obtain the far end pre-trained DnCNN model; the near end pre-trained DnCNN model and the far end pre-trained DnCNN model are used respectively to denoise the gain curves of the near end segment and the far end segment to obtain the denoised gain curves of the near end segment and the far end segment. The denoised gain curves of the near-end and far-end segments are merged into a Brillouin gain denoising curve. The temperature and stress distribution along the fiber are then analyzed based on the Brillouin gain denoising curve.
2. The Brillouin fiber optic sensing noise reduction method based on a two-stage DnCNN according to claim 1, characterized in that, When dividing the gain curve into proximal and distal segments, the selection of the segmentation range includes: The segmentation range is determined based on the original noise level and the Brillouin gain attenuation with distance; When the noise level of the normalized gain curve is 1 / 255 when the measured Brillouin gain as a function of distance is found to be 1 / 255, the noise level is increased to above 5 / 255, and the gain curve is split when the gain value decays to 0.
2.
3. The Brillouin fiber sensing noise reduction method based on a two-stage DnCNN according to claim 1, characterized in that, The acquisition of the near-end pre-trained DnCNN model and the far-end pre-trained DnCNN model includes: When training DnCNN models with different noise levels, the relationship between the signal-to-noise ratio (SNR) improvement of normalized data and the noise level is obtained. Based on the relationship between the SNR improvement and the noise level, it is measured that when the noise level represented by the original noise signal is less than 1 / 255, the SNR improvement brought by image denoising is 0. When the noise level is increased from 1 / 255 to 5 / 255, the signal-to-noise ratio increases from 0dB to 12dB. Then, the convolutional denoising neural network DnCNN is trained using the gain curve data with a noise level of 5 / 255 to obtain a pre-trained DnCNN model for the far end that matches the noise level of the far end segment. The pre-trained DnCNN model is obtained by training the convolutional denoising neural network DnCNN using the gain curve data of the proximal segment.
4. The Brillouin fiber sensing noise reduction method based on a two-stage DnCNN according to claim 1, characterized in that, The gain curve of the distal segment is amplified, and the amplification factor is dynamically adjusted according to the target noise level so that its Brillouin gain is within the normalized gain range of [0,1], so as to raise the noise level of the distal segment to between 5 / 255 and 10 / 255, which is the effective noise reduction threshold of the convolutional denoising neural network DnCNN.