A high-precision distributed optical fiber sensing monitoring method and device based on deep learning

Abstract

The present invention provides a high-precision distributed optical fiber sensing monitoring method and device based on deep learning, which relates to the technical field of optical fiber sensing, and includes: acquiring an original signal, performing polarization fading compensation and morphological filtering preprocessing on the original signal to obtain a preprocessed signal; inputting the preprocessed signal into an autoencoder for feature dimensionality reduction to obtain a low-dimensional feature representation; parallelly performing multi-scale convolutional feature extraction and phase-sensitive feature extraction on the low-dimensional feature representation to respectively obtain multi-scale convolutional features and frequency shift-phase joint features; fusing the multi-scale convolutional features and the frequency shift-phase joint features to generate a fused feature; inputting the fused feature into a multi-task output network to simultaneously perform anomaly detection, anomaly type classification, and location positioning tasks to obtain a monitoring result. The present invention can improve the signal processing accuracy, real-time response ability, and environmental adaptability of the monitoring system, and reduce the false alarm rate in complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field of optical fiber sensing, and in particular, to a high-precision distributed optical fiber sensing monitoring method and device based on deep learning. Background Art

[0002] Distributed optical fiber sensing technology is a new sensing technology that uses optical fiber as both a sensing element and a signal transmission channel. By analyzing Brillouin scattering, Raman scattering, or Rayleigh scattering signals in the optical fiber, it can achieve distributed real-time monitoring of physical quantities such as temperature, strain, and acoustic waves. With its advantages of being distributed, long-distance, and anti-electromagnetic interference, this technology has been widely applied in fields such as oil and gas pipeline leakage monitoring, power line fault detection, railway track health monitoring, and structural health monitoring.

[0003] With the rapid development of deep learning technology, combining it with distributed optical fiber sensing has become an important research direction for improving system performance. Traditional optical fiber sensing signal processing mainly relies on manually designed feature extraction and threshold judgment, which is difficult to adapt to complex and changing environments. Deep learning, through its powerful feature learning and pattern recognition capabilities, can automatically extract high-order features from raw optical fiber signals to achieve more accurate anomaly detection and localization. Currently, convolutional neural networks, recurrent neural networks, etc. have achieved certain applications in optical fiber acoustic wave detection, temperature anomaly recognition, etc.

[0004] However, the existing technologies face serious challenges in practical applications: traditional signal processing methods have insufficient recognition accuracy for weak signals and complex interferences, especially performing poorly in low signal-to-noise ratio environments; conventional algorithms have poor real-time performance when processing large-scale data in long-distance and dynamic environments, and it is difficult to meet the millisecond-level response requirements; when multiple damages or events occur simultaneously, the existing feature extraction and classification methods cannot effectively distinguish different types of events, resulting in high false alarm rates and missed alarm rates. Summary of the Invention

[0005] The present invention proposes a high-precision distributed optical fiber sensing monitoring method and device based on deep learning. By integrating deep learning with the physical model of optical fiber sensing, it improves the signal processing accuracy, real-time response ability, and environmental adaptability of the monitoring system, reduces the false alarm rate in complex environments, and at the same time realizes the precise detection, accurate classification, and accurate positioning of abnormal events, providing more reliable technical support for the safety monitoring of various infrastructure facilities.

[0006] The technical solution of the present invention is realized as follows:

[0007] On the one hand, the present invention provides a high-precision distributed optical fiber sensing monitoring method based on deep learning, including:

[0008] S1. Obtain the Brillouin scattering original signal of the distributed optical fiber sensing system, perform polarization fading compensation and morphological filtering preprocessing on the original signal to obtain a preprocessed signal;

[0009] S2. Input the preprocessed signal into an autoencoder for feature dimensionality reduction to obtain a low-dimensional feature representation;

[0010] S3. Parallelly perform multi-scale convolutional feature extraction and phase-sensitive feature extraction on the low-dimensional feature representation to respectively obtain multi-scale convolutional features and frequency shift-phase joint features;

[0011] S4. Assign weights through a dynamic weight generation network, and combine a feature interaction function based on physical constraints to fuse the multi-scale convolutional features and the frequency shift-phase joint features to generate fused features;

[0012] S5. Input the fused features into a multi-task output network, and simultaneously perform anomaly detection, anomaly type classification, and location positioning tasks to obtain monitoring results.

[0013] Preferably, the polarization fading compensation includes:

[0014] Collect the first Brillouin scattering signal and the second Brillouin scattering signal in the orthogonal polarization directions;

[0015] Perform complex domain synthesis on the first Brillouin scattering signal and the second Brillouin scattering signal to obtain a polarization-insensitive synthesized signal;

[0016] The morphological filtering includes:

[0017] Perform morphological operations using an elliptical structuring element, where the time domain radius of the structuring element is 3 - 7 sampling points, and the spatial radius is 1 - 3 spatial sampling points;

[0018] Perform morphological opening and closing operations on the synthesized signal in sequence, where: the opening operation removes isolated noise points through erosion followed by dilation; the closing operation fills small holes inside the signal through dilation followed by erosion;

[0019] Perform weighted averaging on the results of the opening and closing operations to obtain a preprocessed signal.

[0020] Preferably, step S1 further includes a pulse parameter optimization process for the distributed optical fiber sensing system, and the steps include:

[0021] Dynamically adjust the optical pulse width according to the target spatial resolution requirement, and the spatial resolution Calculation formula:

[0022] ;

[0023] where c is the speed of light and n is the refractive index of the optical fiber, is the pulse width;

[0024] Automatically adjust the pulse repetition frequency according to the monitoring distance L, and the upper limit value of the pulse repetition frequency is set according to ;

[0025] Select the signal averaging times based on the monitoring mode, where: reduce the averaging times in the fast response mode; increase the averaging times in the high-precision mode to improve the signal-to-noise ratio;

[0026] Dynamically adjust the pulse power according to the measurement result of the optical fiber line loss to generate an optical power distribution curve;

[0027] Determine the optimal pulse parameter combination by integrating the spatial resolution, monitoring distance, and response time requirements.

[0028] Preferably, the autoencoder includes:

[0029] An encoder part, including at least three layers of fully connected networks, which compress the preprocessed signal layer by layer into a low-dimensional feature representation, and each layer is provided with a non-linear activation function and a regularization operation;

[0030] A decoder part, which reconstructs the original signal through a fully connected layer structure symmetric to the encoder and is only used in the pre-training stage;

[0031] Among them: in the pre-training stage, the mean squared error of signal reconstruction is used as the loss function to optimize the parameters of the encoder and decoder; after the pre-training is completed, the encoder part is retained for feature dimensionality reduction, and a low-dimensional feature representation is output.

[0032] Preferably, the multi-scale convolutional feature extraction includes:

[0033] Reshape the low-dimensional feature representation into a three-dimensional tensor to adapt to the convolutional operation;

[0034] Extract features of different time scales through multiple parallel convolutional paths, and each path uses convolutional kernels of different sizes;

[0035] Concatenate the features output by each path in the channel dimension to form a joint feature;

[0036] Perform channel dimensionality reduction on the joint feature, and through the global average pooling operation, convert the dimensionality-reduced feature into a multi-scale convolutional feature.

[0037] Preferably, the phase-sensitive feature extraction includes:

[0038] Extract the signal phase feature of the low-dimensional feature representation through a phase extraction sub-network, where the phase extraction sub-network includes a phase perception layer, which extracts the phase feature through frequency domain transformation;

[0039] The frequency shift extraction sub-network performs frequency shift feature extraction on the low-dimensional feature representation. Among them, the frequency shift extraction sub-network includes a frequency shift perception layer, which extracts frequency shift characteristics through time-frequency analysis;

[0040] The outputs of the phase extraction sub-network and the frequency shift extraction sub-network are concatenated and fused through a fully connected layer to generate a frequency shift-phase joint feature.

[0041] Preferably, the dynamic weight generation network includes:

[0042] The first weight generation sub-network receives multi-scale convolutional features as input, and outputs the multi-scale convolutional feature weight coefficient α through a fully connected layer and a Sigmoid activation function;

[0043] The second weight generation sub-network receives the frequency shift-phase joint feature as input, and outputs the frequency shift-phase joint feature weight coefficient β through a fully connected layer and a Sigmoid activation function;

[0044] The third weight generation sub-network receives and the concatenated vector of as input, and outputs the interaction weight coefficient γ through a fully connected layer and a Sigmoid activation function;

[0045] Among them, the weight coefficients of each sub-network satisfy .

[0046] Preferably, the feature interaction function based on physical constraints is defined as follows:

[0047] ;

[0048] In the formula, is a learnable Brillouin physical constraint matrix, initialized to a value close to the identity matrix; is a learnable parameter vector of the frequency shift-phase sensitivity; is the Brillouin frequency shift center offset coefficient; is the Brillouin gain spectrum width parameter; represents the Hadamard product; is the hyperbolic tangent function; represents the natural exponential operation;

[0049] The calculation formula of the fused feature is as follows:

[0050] ;

[0051] Among them, is the fused feature, and the final feature representation is obtained by adjusting the dimension through the output transformation layer.

[0052] Preferably, the multi-task output network includes:

[0053] A shared feature layer that receives the fused feature as input and extracts a general feature representation through a fully connected layer;

[0054] An anomaly detection branch that includes a fully connected layer and a non-linear activation function and outputs the probability of an anomaly occurring;

[0055] An anomaly type classification branch that includes a fully connected layer and a probability normalization function and outputs the probability distribution of different anomaly types;

[0056] A location positioning branch that includes a fully connected layer and outputs a quantized distance value of the anomaly location.

[0057] On the other hand, the present invention also provides a high-precision distributed optical fiber sensing monitoring device based on deep learning. The device is used to execute the method described in any one of the above, and the device includes:

[0058] A signal acquisition module: including a narrow linewidth laser, a distributed optical fiber sensing network, a photodetector, and a high-speed data acquisition card. The narrow linewidth laser is configured to generate an optical signal with an adjustable pulse width. The distributed optical fiber sensing network is coupled to the optical fiber to be monitored. The photodetector receives the Brillouin scattering signal and converts it into a raw electrical signal through the high-speed data acquisition card;

[0059] A preprocessing module: including a polarization fading compensation unit and a morphological filtering unit. The polarization fading compensation unit is configured to eliminate the signal fluctuations of the orthogonal polarization states. The morphological filtering unit is configured to perform opening and closing operations using an elliptical structural element to suppress noise and retain the spatio-temporal features of events;

[0060] A feature extraction module: including an autoencoder, a multi-scale convolutional feature extraction unit, and a phase-sensitive feature extraction unit. The autoencoder is used to reduce the dimensionality of the preprocessed signal. The multi-scale convolutional feature extraction unit and the phase-sensitive feature extraction unit are used to perform feature extraction in parallel to obtain multi-scale convolutional features and frequency shift-phase joint features;

[0061] A feature fusion module that is used to assign weights through a dynamic weight generation network and combine a feature interaction function based on physical constraints to fuse the multi-scale convolutional features and the frequency shift-phase joint features to generate a fused feature;

[0062] A multi-task output module that is used to input the fused feature into the multi-task output network and simultaneously perform anomaly detection, anomaly type classification, and location positioning tasks to obtain monitoring results;

[0063] A processor that is used to execute computer program instructions to control each module to perform corresponding operations;

[0064] A memory for storing the computer program instructions, parameters of each module, and monitoring results;

[0065] Wherein, the device further includes a parameter optimization module integrated in the signal acquisition module, which is used to dynamically adjust the optical pulse width, power, and repetition frequency according to the monitoring distance, target spatial resolution, and response time requirements.

[0066] The present invention has the following beneficial effects compared with the prior art:

[0067] (1) The high-precision distributed optical fiber sensing monitoring method and device based on deep learning proposed by the present invention establish a complete monitoring system from signal acquisition, preprocessing, feature extraction to multi-task output by integrating deep learning with the physical model of optical fiber sensing. Compared with traditional methods, it can improve the monitoring accuracy, enhance the real-time response speed, reduce the false alarm rate, and at the same time improve in terms of monitoring distance, spatial resolution, and abnormal type recognition ability, enhancing the adaptability and reliability of the system in complex environments;

[0068] (2) The present invention adopts a signal preprocessing method combining polarization fading compensation and morphological filtering. By acquiring Brillouin scattering signals in two orthogonal polarization directions and performing complex domain synthesis, it effectively reduces the problem of random polarization state changes caused by optical fiber micro-bending, stress changes, etc., and reduces the signal fluctuation amplitude; at the same time, using morphological opening and closing operations with elliptical structural elements can effectively filter out noise while retaining the spatio-temporal characteristics of the signal, improving the signal-to-noise ratio;

[0069] (3) The adaptive pulse parameter optimization strategy of the present invention dynamically adjusts the optical pulse width, power, and repetition frequency according to the monitoring distance, target spatial resolution, and response time requirements, realizing flexible configuration of monitoring performance on the premise of ensuring the signal-to-noise ratio, enabling the system to achieve corresponding spatial resolution within different monitoring distances, and the response time can also be flexibly adjusted, which is beneficial to improving the applicable range and engineering adaptability of the system;

[0070] (4) The phase-sensitive feature extraction module of the present invention extracts complementary features from the phase domain and frequency domain respectively through the phase extraction sub-network and the frequency shift extraction sub-network, and generates frequency shift-phase joint features through the joint coding layer, making full use of the frequency shift and phase information of the Brillouin scattering signal, enabling the system to distinguish the frequency shift differences caused by temperature changes and strain changes;

[0071] (5) The feature interaction function based on physical constraints proposed by the present invention establishes a non-linear mapping relationship between convolutional features and frequency shift-phase features, integrating the physical mechanism of Brillouin scattering into the feature fusion process, enabling the deep learning model to not only learn the statistical laws of data but also follow the physical constraints of optical fiber sensing. Description of the Drawings

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0073] Figure 1 It is a flowchart of the method of the present invention;

[0074] Figure 2 It is a schematic diagram of the technical implementation of the present invention;

[0075] Figure 3 It is a schematic diagram of the device of the present invention. Specific embodiments

[0076] The following will combine the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0077] As Figure 1 shown, the present invention provides a high-precision distributed optical fiber sensing monitoring method based on deep learning, including:

[0078] S1. Obtain the Brillouin scattering original signal of the distributed optical fiber sensing system, perform polarization fading compensation and morphological filtering preprocessing on the original signal to obtain a preprocessed signal;

[0079] S2. Input the preprocessed signal into an autoencoder for feature dimensionality reduction to obtain a low-dimensional feature representation;

[0080] S3. Parallelly perform multi-scale convolutional feature extraction and phase-sensitive feature extraction on the low-dimensional feature representation to obtain multi-scale convolutional features and frequency shift-phase joint features respectively;

[0081] S4. Allocate weights through a dynamic weight generation network, and combine a feature interaction function based on physical constraints to fuse the multi-scale convolutional features and the frequency shift-phase joint features to generate fused features;

[0082] S5. Input the fused features into a multi-task output network to simultaneously perform anomaly detection, anomaly type classification, and location positioning tasks to obtain monitoring results.

[0083] As Figure 2As shown, in view of the problems such as insufficient signal processing accuracy, poor real-time performance and low environmental adaptability in distributed optical fiber sensing monitoring, the present invention proposes a high-precision monitoring method and device that integrates deep learning with the physical model of optical fiber sensing. This method first preprocesses the original Brillouin scattering signal through polarization fading compensation and morphological filtering to suppress noise and retain key features; then uses an autoencoder to reduce the dimensionality of the preprocessed signal and reduce the computational complexity; then parallelly executes multi-scale convolutional feature extraction and phase-sensitive feature extraction to capture the time characteristics and frequency shift-phase characteristics of the signal respectively; subsequently, organically fuses the two types of features through a dynamic weight generation network and a feature interaction function based on physical constraints, making full use of their complementarity; finally, inputs the fused features into a multi-task output network to simultaneously realize functions such as anomaly detection, anomaly type classification and location positioning. By combining the feature extraction ability of deep learning with the physical mechanism of Brillouin scattering, a complete monitoring system from signal acquisition to result output is established, enabling the system to have good real-time performance and environmental adaptability while maintaining high precision.

[0084] Specifically, in an embodiment of the present invention, a distributed optical fiber sensing system is used for signal acquisition. The system architecture includes a narrow linewidth laser, a distributed optical fiber sensing network, a photodetector and a high-speed data acquisition card.

[0085] This system adopts coherent detection technology to eliminate common-mode noise through a balanced detector and improve the signal reception sensitivity. In long-distance monitoring scenarios, an optical amplifier (EDFA) is set every 10 - 20 km to compensate for transmission losses, and Raman amplification technology is used to improve the signal quality at the far end.

[0086] Specifically, in an embodiment of the present invention, considering that pulse parameters have a significant impact on monitoring performance and the traditional pulse parameters are fixed, it is difficult to balance the spatial resolution and signal-to-noise ratio of long-distance optical fiber monitoring. Therefore, this embodiment designs an adaptive pulse parameter optimization strategy based on monitoring requirements, adjusts the parameters of the distributed optical fiber sensing system according to the real-time monitoring environment and requirements, and balances the monitoring distance, spatial resolution and response time to collect high-quality original signals.

[0087] Specifically, the four core acquisition parameters optimized in this embodiment are respectively:

[0088] Pulse width , this parameter determines the spatial resolution of the system. A narrower pulse provides higher spatial resolution, but the signal intensity is weaker; a wider pulse has a higher signal intensity, but the spatial resolution is reduced. The optimized range of this parameter is set to 50 - 200 ns, corresponding to a spatial resolution of about 5 - 20 m.

[0089] Pulse Repetition Frequency (PRF), this parameter affects the data acquisition rate and the maximum monitoring distance of the system. If the PRF is too high, aliasing of reflected signals from different positions will occur; if the PRF is too low, the data update rate will be reduced. The optimized range of this parameter is 5 - 20 kHz.

[0090] Number of Signal Averaging N, this parameter is related to the signal-to-noise ratio and the system response time. The more the number of averaging, the higher the signal-to-noise ratio, but the longer the response time.

[0091] Pulse power, this parameter affects the signal strength and the monitoring distance. Excessive power may cause non-linear effects; too low power may cause distal signals to be undetectable.

[0092] In this embodiment, the adaptive pulse parameter optimization strategy is as follows:

[0093] Dynamically adjust the optical pulse width according to the target spatial resolution requirement, the spatial resolution Calculation formula:

[0094] ;

[0095] where c is the speed of light, n is the refractive index of the optical fiber, is the pulse width; for example, a 100 ns pulse corresponds to a spatial resolution of approximately 10 m.

[0096] Automatically adjust the pulse repetition frequency according to the monitoring distance L, the upper limit value of the pulse repetition frequency is set according to ; for example, the maximum PRF of a 10 km optical fiber is approximately 10 kHz.

[0097] Select the number of signal averaging N based on the monitoring mode, where: reduce the number of averaging in the fast response mode; increase the number of averaging in the high-precision mode to improve the signal-to-noise ratio; the relationship between the number of averaging N and the improvement of the signal-to-noise ratio: , N = 100 - 1000 in the fast response mode, and N = 1000 - 10000 in the high-precision mode.

[0098] Dynamically adjust the pulse power according to the measurement result of the optical fiber line loss to generate an optical power distribution curve; specifically, first use the optical time domain reflectometry technology to measure the actual loss distribution of the optical fiber line, and calculate the total round-trip loss L total (z) from the transmitter to a distance z, including the inherent attenuation of the optical fiber, connector loss, and micro-bending loss, etc. Based on these loss measurement results, the system dynamically calculates the power distribution curve at different distances through the following power adjustment algorithm:

[0099] ;

[0100] where, is the power distribution curve, is the maximum output power of the laser (50 mW), is the target signal-to-noise ratio (10 - 15 dB), is the noise power spectral density at the receiving end.

[0101] This algorithm ensures that the signal-to-noise ratio of the returned signal at any position z is not lower than the target value. Based on the power distribution curve, the peak power of the pulse is dynamically adjusted through an electro-optic modulator (EOM): , where rect is the rectangular pulse function.

[0102] Considering the requirements of spatial resolution, monitoring distance, and response time, the optimal pulse parameter combination is determined.

[0103] In this embodiment, the above parameters are adjusted according to the real-time monitoring environment and requirements to balance the monitoring distance, spatial resolution, and response time. The spatial resolution formula is derived based on the fiber refractive index and the speed of light, and the pulse width (50 - 200 ns) and repetition frequency (5 - 20 kHz) are dynamically adjusted. The adaptive averaging strategy quickly responds / high-precision adjusts the number of averaging times according to the monitoring mode, and adjusts the pulse power based on the line loss to optimize the pulse parameters based on the monitoring requirements.

[0104] Specifically, in an embodiment of the present invention, considering that the fiber signal generates fading noise due to the random change of the polarization state, the traditional threshold method cannot effectively suppress it. Therefore, a preprocessing method combining polarization fading compensation and morphological filtering is adopted.

[0105] Specifically, the polarization diversity technology is combined with the elliptical interpolation algorithm to eliminate the signal fluctuations of the orthogonal polarization states. The polarization fading compensation includes:

[0106] Collect the first Brillouin scattering signal in the orthogonal polarization direction and the second Brillouin scattering signal ;

[0107] Perform complex domain synthesis on the first Brillouin scattering signal and the second Brillouin scattering signal to obtain a polarization-insensitive synthesized signal; the calculation formula of the synthesized signal:

[0108] ;

[0109] Reduce the signal fluctuations caused by the change of the polarization state.

[0110] The morphological filtering includes:

[0111] Perform morphological operations using an elliptical structuring element. The time domain radius of the structuring element is 3 - 7 sampling points, and the spatial radius is 1 - 3 spatial sampling points; use an asymmetric elliptical structuring element B(a, b), where a is the time domain radius, which is 3 - 7 sampling points, and b is the spatial radius, which is 1 - 3 spatial sampling points.

[0112] The morphological opening and closing operations are performed on the synthetic signal in sequence, where the opening operation removes isolated noise points by first corroding and then dilating; the closing operation fills small holes inside the signal by first dilating and then corroding.

[0113] The results of the opening operation and the closing operation are weighted averaged to obtain a preprocessed signal.

[0114] Specifically, the morphological opening operation is erosion followed by dilation, which is used to remove peak noise. The expression is:

[0115] ;

[0116] The morphological closing operation is dilation followed by erosion, which is used to fill the signal valley value:

[0117] ;

[0118] The final filtering result is the weighted average of the opening and closing operations:

[0119] ;

[0120] The weight α is set to 0.5-0.7 and can be adjusted dynamically according to the signal characteristics.

[0121] This filtering method can effectively suppress environmental interference and polarization changes while preserving the spatiotemporal characteristics of the event signal.

[0122] Specifically, in one embodiment of the present invention, an integrated deep learning model is constructed to perform feature dimension reduction, feature extraction, fusion, anomaly detection, etc. on the preprocessed signal. The model performs high-dimensional signal dimension reduction and denoising based on the autoencoder, and at the same time, the multi-scale convolution feature extraction module captures the signal change characteristics at different time scales, and the phase-sensitive feature extraction module is used to obtain the physical characteristics of Brillouin frequency shift and phase. The two complement each other to form a comprehensive feature representation. The feature fusion module realizes the adaptive integration of different types of features based on the interaction function of physical constraints, especially focusing on the embedding of physical laws in deep learning. Finally, the multi-task output module simultaneously performs anomaly detection, classification and location positioning in a shared feature and dedicated task layer manner, so as to achieve the purpose of balancing efficiency and accuracy.

[0123] In this embodiment, the autoencoder is located at the front end of the model, and its structure is:

[0124] The encoder part contains at least three layers of fully connected networks, which compress the preprocessed signal into a low-dimensional feature representation layer by layer. Each layer is equipped with a nonlinear activation function and regularization operation.

[0125] Specifically, the encoder is used to compress the high-dimensional Brillouin scattering raw signal into a low-dimensional feature representation. In a specific example, the encoder includes an input layer and three fully connected layers:

[0126] The input layer receives the raw Brillouin scattering signal , where is the number of sampling points. The first fully connected layer: , where , , applying L2 regularization: , . The second fully connected layer: , where , , applying Dropout, with a dropout rate . The third fully connected layer: , where , , applying batch normalization. The feature representation: , where is the low-dimensional feature representation.

[0127] The decoder part reconstructs the original signal through a fully connected layer structure symmetric to the encoder, and is only used in the pre-training stage.

[0128] Specifically, the decoder reconstructs the features output by the encoder into the original signal for the pre-training of the autoencoder. In a specific example, the decoder also includes an input layer and three fully connected layers:

[0129] The input layer receives the features output by the encoder . The fourth fully connected layer: , where , . The fifth fully connected layer: , where , , applying Dropout, with a dropout rate . The sixth fully connected layer: , where , , is the Sigmoid activation function. The reconstructed signal: .

[0130] Among them: During the pre-training stage, the mean square error of signal reconstruction is used as the loss function to optimize the parameters of the encoder and decoder; after the pre-training is completed, the encoder part is retained for feature dimensionality reduction, and the low-dimensional feature representation is output.

[0131] In this embodiment, the pre-training process of the autoencoder is as follows:

[0132] Design loss function: mean squared error The optimizer uses Adam and the learning rate is set to 0.001. Early stopping strategy: stop if the validation set loss does not improve for 5 consecutive rounds.

[0133] In this embodiment, the multi-scale convolutional feature extraction module adopts a parallel multi-scale convolutional structure to capture feature patterns at different time scales. The multi-scale convolutional feature extraction includes:

[0134] Reshape the low-dimensional feature representation into a three-dimensional tensor to adapt to the convolution operation.

[0135] In one example, the feature vector output by the autoencoder encoder is reshaped into a three-dimensional tensor suitable for the convolution operation.

[0136] Extract features at different time scales through multiple parallel convolution paths, and each path uses convolution kernels of different sizes.

[0137] In one example, there are three groups of parallel convolutional layers that use different convolution kernel sizes to capture multi-scale features. Small-scale convolution path: , where is the convolution kernel with the shape of , the stride , and the padding . Medium-scale convolution path: , where is the convolution kernel with the shape of , the stride = 1, and the padding = 1. Large-scale convolution path: , where is the convolution kernel with the shape of , the stride = 1, and the padding = 1.

[0138] Concatenate the features output by each path in the channel dimension to form a joint feature.

[0139] In one example, the concatenation operation: , obtaining .

[0140] Perform channel dimensionality reduction on the joint feature, and through global average pooling operation, convert the dimensionality-reduced feature into a multi-scale convolutional feature.

[0141] In one example, use 1×1 convolution for dimensionality reduction: , where is the convolution kernel, obtaining . Global average pooling: , obtaining . Flatten operation: , obtaining the final multi-scale convolutional feature .

[0142] In this embodiment, the phase-sensitive feature extraction module is used to extract the phase and frequency shift features in the Brillouin scattering signal. The phase-sensitive feature extraction includes:

[0143] Performing signal phase feature extraction on the low-dimensional feature representation through a phase extraction sub-network, where the phase extraction sub-network includes a phase perception layer, which extracts phase features through frequency domain transformation.

[0144] In one example, the phase extraction sub-network includes an input layer, two fully connected layers, and a phase perception layer: The input layer receives the autoencoder features . The first fully connected layer: , where , . The phase perception layer: , where The operation applies the Hilbert transform and the Fourier transform to extract phase information. The second fully connected layer: , where , .

[0145] It should be noted that the present invention can also generate an analytic signal through a learnable complex convolution kernel and calculate its phase to extract phase features.

[0146] Performing frequency shift feature extraction on the low-dimensional feature representation through a frequency shift extraction sub-network, where the frequency shift extraction sub-network includes a frequency shift perception layer, which extracts frequency shift characteristics through time-frequency analysis.

[0147] In one example, the frequency shift extraction sub-network includes an input layer, two fully connected layers, and a frequency shift perception layer: The input layer receives the autoencoder features . The first fully connected layer: , where , . The frequency shift perception layer: , where The operation applies the short-time Fourier transform to extract frequency shift features. The second fully connected layer: , where , .

[0148] It should be noted that the present invention can also dynamically learn frequency shift features through a differentiable time-frequency convolution layer.

[0149] Concatenate the outputs of the phase extraction sub-network and the frequency shift extraction sub-network and fuse them through a fully connected layer to generate a frequency shift-phase joint feature.

[0150] In one example, generating a frequency shift-phase joint feature in the joint encoding layer, specifically including:

[0151] Feature splicing: , obtaining . Joint feature extraction: , where , . Output: Frequency shift-phase joint feature .

[0152] In this embodiment, the feature fusion module fuses multi-scale convolutional features and frequency shift-phase joint features, considering the physical characteristics of Brillouin scattering. It includes a dynamic weight generation network and a feature interaction function based on physical constraints. Specifically, the dynamic weight generation network includes:

[0153] The first weight generation sub-network, which receives the multi-scale convolutional features as input, and outputs the weight coefficient α of the multi-scale convolutional features through a fully connected layer and a Sigmoid activation function.

[0154] The second weight generation sub-network, which receives the frequency shift-phase joint features as input, and outputs the weight coefficient β of the frequency shift-phase joint features through a fully connected layer and a Sigmoid activation function;

[0155] The third weight generation sub-network, which receives and the concatenated vector of

[0156] as input, and outputs the interaction weight coefficient γ through a fully connected layer and a Sigmoid activation function; .

[0157] In a specific example, all three weight generation sub-networks include 2 layers of fully connected layers, and the specific settings are as follows:

[0158] The first weight generation sub-network, that is, the generation network. The input is . The first fully connected layer reduces 64 dimensions to 32 dimensions, using the ReLU activation function where , . The second fully connected layer reduces 32 dimensions to 1 dimension, using the Sigmoid activation function where , , is the Sigmoid function. The output is a scalar .

[0159] The second weight generation sub-network, that is, the generation network. The input is The first fully connected layer reduces the 64 dimensions to 32 dimensions and uses the ReLU activation function where , The second fully connected layer reduces the 32 dimensions to 1 dimension and uses the Sigmoid activation function where , The output is a scalar .

[0160] The third weight generation sub-network, that is the generation network. The input is , representing the concatenation of two feature vectors. The first fully connected layer reduces the 128 dimensions to 32 dimensions and uses the ReLU activation function where , The second fully connected layer reduces the 32 dimensions to 1 dimension and uses the Sigmoid activation function where , The output is a scalar .

[0161] The feature interaction function based on physical constraints incorporates the physical characteristics of Brillouin scattering and is defined as follows:

[0162] ;

[0163] In the formula, is a learnable Brillouin physical constraint matrix, initialized to a value close to the identity matrix; is a learnable parameter vector of the frequency shift-phase sensitivity; is the Brillouin frequency shift center offset coefficient; is the Brillouin gain spectrum width parameter; represents the Hadamard product; is the hyperbolic tangent function; represents the natural exponential operation.

[0164] The calculation formula of the fused feature is as follows:

[0165] ;

[0166] where is the fused feature, and the final feature representation is obtained by adjusting the dimension through the output transformation layer.

[0167] To enhance the feature expression ability and adjust the dimension, the fused feature undergoes the following transformation:

[0168] ;

[0169] Among them , the output .

[0170] In this embodiment, the multi-task output module is responsible for simultaneously outputting the prediction results of three tasks based on the fused features: anomaly detection, anomaly type classification, and location positioning. This module is the multi-task output network, which includes:

[0171] A shared feature layer that receives the fused features as input and extracts a general feature representation through a fully connected layer.

[0172] Specifically, the input of this shared layer is the fused feature . Shared fully connected layer: , among which , , applying Dropout, with a dropout rate .

[0173] The anomaly detection branch, which includes a fully connected layer and a non-linear activation function, and outputs the probability of an anomaly occurring.

[0174] Specifically, this branch includes 2 layers of fully connected layers. The first fully connected layer: , among which , . The second fully connected layer: , among which , , is the Sigmoid activation function. The output is the anomaly probability .

[0175] The anomaly type classification branch, which includes a fully connected layer and a probability normalization function, and outputs the probability distribution of different anomaly types.

[0176] Specifically, this branch includes 2 layers of fully connected layers. The first fully connected layer: , among which , . The second fully connected layer: , among which , , is the number of anomaly types. The output is the probability distribution of each anomaly type , satisfying .

[0177] The location positioning branch, which includes a fully connected layer and outputs a quantized distance value of the anomaly location.

[0178] Specifically, this branch includes 2 layers of fully connected layers. The first fully connected layer: , among which , . Second fully connected layer: , where , , using a linear activation function. The output is the abnormal position distance value .

[0179] In an embodiment of the present invention, the overall training process of the deep learning model is as follows:

[0180] Autoencoder pre-training: Use unlabeled data to train the autoencoder to reconstruct the original signal, fix the encoder weights, and extract feature vectors.

[0181] End-to-end model training: Load the pre-trained encoder weights, use labeled data to train the complete multi-task model, monitor the validation metrics and overall loss of each task, and apply an early stopping strategy to prevent overfitting.

[0182] Model fine-tuning: Unfreeze the encoder weights, perform fine-tuning using a smaller learning rate, and focus on optimizing the physical parameters in the physical perception fusion module.

[0183] Among them, when training the end-to-end model, a multi-task loss function is adopted, and the expression is: , where , , are the weights of the losses of each task.

[0184] is the loss of the anomaly detection task:

[0185] ;

[0186] In the formula, N is the number of training samples, is the true anomaly label of the i-th sample, is the probability that the model predicts the i-th sample as an anomaly.

[0187] is the loss of the anomaly type classification task:

[0188] ;

[0189] In the formula, N is the number of training samples, is the number of anomaly categories, is the true label that the i-th sample belongs to category c, is the probability that the model predicts the i-th sample belongs to category c.

[0190] is the loss of the position localization task:

[0191] ;

[0192] Where N is the number of training samples, is the position label of the i-th abnormal sample, is the position of the i-th abnormal sample predicted by the model.

[0193] It should be noted that when the signal of the present invention is collected, the data is usually a series of measurement points along the length of the optical fiber. The position information of each sampling point has been recorded by the system. The position information of each sampling point represents the distance from the starting end of the optical fiber. These position labels are obtained together with the signal data. The position label mentioned in the present invention is a relative distance, that is, the distance of the abnormal point from the starting end of the optical fiber system. Therefore, the position positioning branch outputs a quantized distance value.

[0194] The training strategy during training is as follows: Optimizer: Adam, learning rate = 0.0005. Learning rate scheduling: If the validation loss does not improve after every 10 epochs, the learning rate decays to 0.5 of the original. Early stopping strategy: Stop training if the total validation loss does not improve for 15 consecutive epochs.

[0195] Specifically, in an embodiment of the present invention, to improve the generalization ability of the model, a data augmentation strategy is designed for the characteristics of the fiber optic sensing signal:

[0196] Signal-to-noise ratio transformation: Add different intensities of noise to the training samples. Gaussian white noise: SNR ranges from -30 dB to 0 dB. Colored noise: 1 / f noise simulates the environmental vibration background. Impulse noise: Simulates electrical interference.

[0197] Time-frequency domain transformation: Time stretching / compression (±20%): Adapt to events with different propagation speeds. Frequency shift (±5%): Cope with frequency changes caused by temperature drift. Random attenuation (0.5 - 1.5 times): Simulate different distances and attenuation conditions.

[0198] Synthetic event generation: Event superposition: Linear combination of multiple event signals. Position randomization: Randomly place events at different positions on the optical fiber. Background mixing: Mix with the actual environmental noise background.

[0199] Through these data augmentation techniques, the original dataset is expanded by 10 - 20 times, improving the model's adaptability to various environmental conditions.

[0200] Specifically, in an embodiment, after the deep learning model of the present invention is trained, it is deployed in the monitoring device, and the model performance is continuously optimized through the built-in incremental learning mechanism. The incremental learning algorithm is built into the backend of the multi-task output module, sharing the feature extraction layer with the main model but having an independent set of fine-tuning parameters. When the system runs for a period of time, new labeled data will be collected, especially those misclassified samples and newly emerging abnormal types. At this time, the incremental learning process starts to execute and is achieved through the following steps:

[0201] First, for the newly collected dataset , the model is updated by combining knowledge distillation and elastic weight. Knowledge distillation uses the output of the original model as the soft label, which together with the true label guides the learning of the new model. Its loss function is:

[0202] ;

[0203] where: is the cross-entropy loss; is the KL divergence loss, is the balance parameter. At the same time, to prevent catastrophic forgetting, the system applies elastic weight consolidation (EWC) regularization constraints during parameter update to ensure that the parameters important for the old tasks change less:

[0204] ;

[0205] where, is the new parameter, is the old parameter, is the balance coefficient, which controls the weights of the old and new tasks. is the Fisher information matrix, and its calculation formula is:

[0206] ;

[0207] It is approximately calculated by sampling the old task data. The important parameters are well protected, and the non-critical parameters can be updated freely.

[0208] The implementation steps of incremental learning are as follows: 1. Train the basic model on the initial dataset to obtain the parameters . 2. Calculate the Fisher information matrix F using the basic model and the training data. 3. When new data arrives, train using the extended loss function. 4. Periodically update the Fisher matrix to adapt to environmental changes.

[0209] In addition, as Figure 3 shown, the present invention also provides a high-precision distributed optical fiber sensing monitoring device based on deep learning. The device is used to execute the method described in any one of the above, and the device includes:

[0210] Signal acquisition module: It includes a narrow-linewidth laser, a distributed optical fiber sensing network, a photodetector, and a high-speed data acquisition card. The narrow-linewidth laser is configured to generate an optical signal with an adjustable pulse width. The distributed optical fiber sensing network is coupled to the optical fiber to be monitored. The photodetector receives the Brillouin scattering signal and converts it into an original electrical signal through the high-speed data acquisition card;

[0211] Preprocessing module: It includes a polarization fading compensation unit and a morphological filtering unit. The polarization fading compensation unit is configured to eliminate the signal fluctuations of orthogonal polarization states, and the morphological filtering unit is configured to perform opening and closing operations using an elliptical structural element to suppress noise and retain the spatio-temporal characteristics of events;

[0212] Feature extraction module: It includes an autoencoder, a multi-scale convolutional feature extraction unit, and a phase-sensitive feature extraction unit. The autoencoder is used to perform feature dimensionality reduction on the preprocessed signal, and the multi-scale convolutional feature extraction unit and the phase-sensitive feature extraction unit are used to perform feature extraction in parallel to obtain multi-scale convolutional features and frequency shift-phase joint features;

[0213] Feature fusion module, which is used to allocate weights through a dynamic weight generation network and combine a feature interaction function based on physical constraints to fuse the multi-scale convolutional features and the frequency shift-phase joint features to generate fused features;

[0214] Multi-task output module, which is used to input the fused features into a multi-task output network and perform anomaly detection, anomaly type classification, and location positioning tasks simultaneously to obtain monitoring results;

[0215] Processor, which is used to execute computer program instructions to control each of the above modules to perform corresponding operations;

[0216] Memory, which is used to store the computer program instructions, the parameters of each module, and the monitoring results;

[0217] Among them, the device further includes a parameter optimization module, which is integrated in the signal acquisition module and is used to dynamically adjust the optical pulse width, power, and repetition frequency according to the monitoring distance, target spatial resolution, and response time requirements.

[0218] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A high-precision distributed optical fiber sensing monitoring method based on deep learning, characterized in that Including: S1. Obtain the Brillouin scattering original signal of the distributed optical fiber sensing system, perform polarization fading compensation and morphological filtering preprocessing on the original signal to obtain a preprocessed signal; S2. Input the preprocessed signal into an autoencoder for feature dimensionality reduction to obtain a low-dimensional feature representation; S3. Parallelly perform multi-scale convolutional feature extraction and phase-sensitive feature extraction on the low-dimensional feature representation to respectively obtain multi-scale convolutional features and frequency shift-phase joint features; S4. Assign weights through a dynamic weight generation network, and combine a feature interaction function based on physical constraints to fuse the multi-scale convolutional features and frequency shift-phase joint features to generate fused features; S5. Input the fused features into a multi-task output network to simultaneously perform anomaly detection, anomaly type classification, and location positioning tasks to obtain monitoring results.

2. The high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, wherein The polarization fading compensation includes: Collect the first Brillouin scattering signal and the second Brillouin scattering signal in orthogonal polarization directions; Perform complex domain synthesis on the first Brillouin scattering signal and the second Brillouin scattering signal to obtain a polarization-insensitive synthesized signal; The morphological filtering includes: Perform morphological operations using an elliptical structuring element, where the time domain radius of the structuring element is 3 - 7 sampling points and the spatial radius is 1 - 3 spatial sampling points; Successively perform morphological opening and closing operations on the synthesized signal, where: the opening operation removes isolated noise points through erosion followed by dilation; the closing operation fills small holes inside the signal through dilation followed by erosion; Perform weighted averaging on the results of the opening and closing operations to obtain a preprocessed signal.

3. A high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, characterized in that Step S1 also includes a pulse parameter optimization process for the distributed optical fiber sensing system, and the steps include: Dynamically adjust the optical pulse width according to the target spatial resolution requirement, and the spatial resolution Calculation formula: ； where c is the speed of light and n is the refractive index of the optical fiber, is the pulse width; Automatically adjust the pulse repetition frequency according to the monitoring distance L, and the upper limit value of the pulse repetition frequency is set according to the setting; Select the signal averaging times based on the monitoring mode, where: reduce the averaging times in the fast response mode; increase the averaging times in the high-precision mode to improve the signal-to-noise ratio; Dynamically adjust the pulse power according to the measurement result of the optical fiber line loss to generate an optical power distribution curve; Based on the requirements of spatial resolution, monitoring distance, and response time, determine the optimal pulse parameter combination.

4. A high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, characterized in that, The autoencoder includes: An encoder part, including at least three layers of fully connected networks, gradually compressing the preprocessed signal into a low-dimensional feature representation, with a non-linear activation function and regularization operation set for each layer; A decoder part, reconstructing the original signal through a fully connected layer structure symmetric to the encoder, only used in the pre-training stage; Among them: in the pre-training stage, use the mean square error of signal reconstruction as the loss function to optimize the parameters of the encoder and decoder; after the pre-training is completed, retain the encoder part for feature dimensionality reduction and output the low-dimensional feature representation.

5. A high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, characterized in that, The multi-scale convolutional feature extraction includes: Reshape the low-dimensional feature representation into a three-dimensional tensor to adapt to convolutional operations; Extract features of different time scales through multiple parallel convolutional paths, and each path uses convolutional kernels of different sizes; Concatenate the features output by each path in the channel dimension to form a joint feature; Perform channel dimensionality reduction on the joint feature, and through global average pooling operation, convert the reduced-dimensional feature into multi-scale convolutional features.

6. A high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, characterized in that, The phase-sensitive feature extraction includes: The signal phase feature of the low-dimensional feature representation is extracted by a phase extraction sub-network, wherein the phase extraction sub-network includes a phase perception layer, which extracts the phase feature through frequency domain transformation; The frequency shift feature of the low-dimensional feature representation is extracted by a frequency shift extraction sub-network, wherein the frequency shift extraction sub-network includes a frequency shift perception layer, which extracts the frequency shift feature through time-frequency analysis; The outputs of the phase extraction sub-network and the frequency shift extraction sub-network are concatenated and fused through a fully connected layer to generate a frequency shift-phase joint feature.

7. A high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, characterized in that, The dynamic weight generation network includes: The first weight generation sub-network, which receives multi-scale convolutional features as inputs, outputs the multi-scale convolutional feature weight coefficient α through a fully connected layer and a Sigmoid activation function; The second weight generation sub-network, which receives the frequency shift-phase joint feature as input, and outputs the frequency shift-phase joint feature weight coefficient β through a fully connected layer and a Sigmoid activation function; The third weight generation sub-network receives the concatenated vector with as input, and outputs the interaction weight coefficient γ through a fully connected layer and a Sigmoid activation function; Among them, the weight coefficients of each sub-network satisfy .

8. A high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 7, characterized in that Feature interaction function based on physical constraints It is defined as follows: ； In the formula, is a learnable Brillouin physical constraint matrix; is a learnable parameter vector of frequency shift-phase sensitivity; is the Brillouin frequency shift center offset coefficient; is the Brillouin gain spectrum width parameter; denotes the Hadamard product; is the hyperbolic tangent function; denotes the natural exponential operation; The calculation formula of the fusion feature is as follows: ； Among them, is the fusion feature, and the final feature representation is obtained by adjusting the dimension through the output transformation layer.

9. A high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, characterized in that The multi-task output network includes: A shared feature layer that receives the fusion feature as input and extracts a general feature representation through a fully connected layer; An anomaly detection branch, including a fully connected layer and a non-linear activation function, and outputs the probability of anomaly occurrence; An anomaly type classification branch, including a fully connected layer and a probability normalization function, and outputs the probability distribution of different anomaly types; A location positioning branch, including a fully connected layer, and outputs a quantization distance value of the anomaly location.

10. A high-precision distributed optical fiber sensing monitoring device based on deep learning, characterized in that, The device is used to execute the method according to any one of claims 1-9, and the device includes: A signal acquisition module: including a narrow linewidth laser, a distributed optical fiber sensing network, a photodetector and a high-speed data acquisition card. The narrow linewidth laser is configured to generate an optical signal with an adjustable pulse width. The distributed optical fiber sensing network is coupled to the optical fiber to be monitored. The photodetector receives the Brillouin scattering signal and converts it into an original electrical signal through the high-speed data acquisition card; A preprocessing module: including a polarization fading compensation unit and a morphological filtering unit. The polarization fading compensation unit is configured to eliminate the signal fluctuation of the orthogonal polarization state. The morphological filtering unit is configured to perform opening and closing operations using an elliptical structural element to suppress noise and retain the spatio-temporal characteristics of the event; A feature extraction module: including an autoencoder, a multi-scale convolutional feature extraction unit and a phase-sensitive feature extraction unit. The autoencoder is used to reduce the dimension of the preprocessed signal. The multi-scale convolutional feature extraction unit and the phase-sensitive feature extraction unit are used to perform feature extraction in parallel to obtain a multi-scale convolutional feature and a frequency shift-phase joint feature; A feature fusion module, which is used to assign weights through a dynamic weight generation network and combine a feature interaction function based on physical constraints to fuse the multi-scale convolutional feature and the frequency shift-phase joint feature to generate a fusion feature; A multi-task output module, which is used to input the fusion feature into the multi-task output network and perform anomaly detection, anomaly type classification and location positioning tasks simultaneously to obtain a monitoring result; A processor, which is used to execute computer program instructions to control each module to perform corresponding operations; A memory, which is used to store the computer program instructions, the parameters of each module and the monitoring result; Wherein, the device further includes a parameter optimization module, which is integrated in the signal acquisition module and is used to dynamically adjust the optical pulse width, power and repetition frequency according to the monitoring distance, target spatial resolution and response time requirements.

Images