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

Abstract

The invention provides a high-precision distributed optical fiber sensing monitoring method and device based on deep learning, and relates to the technical field of optical fiber sensing, and the method comprises the steps: obtaining an original signal, carrying out the polarization fading compensation and morphological filtering preprocessing of the original signal, and obtaining a preprocessed signal; inputting the preprocessed signal into an auto-encoder for feature dimension reduction to obtain low-dimensional feature representation; performing multi-scale convolution feature extraction and phase sensitive feature extraction on the low-dimensional feature representation in parallel to obtain a multi-scale convolution feature and a frequency shift-phase joint feature respectively; fusing the multi-scale convolution features and the frequency shift-phase joint features to generate fused features; and inputting the fusion features into a multi-task output network, and executing anomaly detection, anomaly type classification and position positioning tasks at the same time to obtain a monitoring result. According to the invention, the signal processing precision, the real-time response capability and the environmental adaptability of the monitoring system can be improved, and the false alarm rate in a complex environment is reduced.

Description

Technical Field

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

[0002] Distributed fiber optic sensing technology is a new type of 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 sound waves. With its advantages of distribution, long distance, and anti-electromagnetic interference, this technology has been widely used in 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 fiber optic sensing has become an important research direction to improve system performance. Traditional fiber optic 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 fiber optic signals to achieve more accurate anomaly detection and positioning. At present, convolutional neural networks, recursive neural networks, etc. have been applied to fiber optic acoustic wave detection, temperature anomaly recognition, etc.

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

[0005] The present invention proposes a high-precision distributed fiber optic sensing monitoring method and device based on deep learning. By integrating deep learning with the fiber optic sensing physical model, the signal processing accuracy, real-time response capability and environmental adaptability of the monitoring system are improved, and the false alarm rate in complex environments is reduced. At the same time, accurate detection, precise classification and accurate positioning of abnormal events are achieved, providing more reliable technical support for the security monitoring of various types of infrastructure.

[0006] The technical solution of the present invention is achieved in this way: On the one hand, the present invention provides a high-precision distributed optical fiber sensing monitoring method based on deep learning, comprising: S1, obtaining the Brillouin scattering original signal of the distributed optical fiber sensing system, performing polarization fading compensation and morphological filtering preprocessing on the original signal to obtain a preprocessed signal; S2, input the preprocessed signal into the autoencoder for feature dimensionality reduction to obtain low-dimensional feature representation; S3, performing multi-scale convolution feature extraction and phase-sensitive feature extraction on the low-dimensional feature representation in parallel to obtain multi-scale convolution features and frequency shift-phase joint features respectively; S4, weights are assigned through a dynamic weight generation network, and the multi-scale convolution features and the frequency shift-phase joint features are fused to generate fused features in combination with the feature interaction function based on physical constraints; S5. Input the fused features into the multi-task output network, and simultaneously perform anomaly detection, anomaly type classification and location positioning tasks to obtain monitoring results.

[0007] Preferably, polarization fading compensation includes: Collecting a first Brillouin scattering signal and a second Brillouin scattering signal in orthogonal polarization directions; Performing complex domain synthesis on the first Brillouin scattering signal and the second Brillouin scattering signal to obtain a polarization-insensitive synthesized signal; Morphological filtering includes: An elliptical structure element is used for morphological operations. The temporal radius of the structure element is 3-7 sampling points, and the spatial radius is 1-3 spatial sampling points. 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. The results of the opening operation and the closing operation are weighted averaged to obtain a preprocessed signal.

[0008] Preferably, step S1 also includes a pulse parameter optimization process of the distributed optical fiber sensing system, the steps comprising: Dynamically adjust the optical pulse width according to the target spatial resolution requirements, spatial resolution Calculation formula: ; Where c is the speed of light, 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, the pulse repetition frequency The upper limit value is set up; Select the signal averaging times based on the monitoring mode, where: reduce the averaging times in fast response mode; increase the averaging times in high precision mode to improve the signal-to-noise ratio; Dynamically adjust pulse power based on fiber line loss measurement results to generate optical power distribution curves; The optimal pulse parameter combination is determined based on the spatial resolution, monitoring distance and response time requirements.

[0009] Preferably, the autoencoder comprises: The encoder part consists of at least three layers of fully connected networks, which compress the preprocessed signal into a low-dimensional feature representation layer by layer, with nonlinear activation functions and regularization operations in each layer; The decoder part reconstructs the original signal through a fully connected layer structure symmetrical to the encoder and is only used in the pre-training stage; Among them: in the pre-training stage, the signal reconstruction mean square error is used as the loss function to optimize the encoder and decoder parameters; after the pre-training is completed, the encoder part is retained for feature dimensionality reduction and the low-dimensional feature representation is output.

[0010] Preferably, multi-scale convolution feature extraction includes: Reshape the low-dimensional feature representation into a three-dimensional tensor to adapt to the convolution operation; Features at different time scales are extracted through multiple parallel convolution paths, each using convolution kernels of different sizes; The features output by each path are concatenated in the channel dimension to form a joint feature; The joint features are subjected to channel dimension reduction, and the reduced features are converted into multi-scale convolutional features through global average pooling operation.

[0011] Preferably, phase sensitive feature extraction includes: Extracting signal phase features from low-dimensional feature representation through a phase extraction subnetwork, wherein the phase extraction subnetwork includes a phase perception layer, which extracts phase features through frequency domain transformation; The frequency shift feature is extracted from the low-dimensional feature representation through a frequency shift extraction subnetwork, wherein the frequency shift extraction subnetwork includes a frequency shift perception layer, which extracts frequency shift characteristics through time-frequency analysis; The outputs of the phase extraction subnetwork and the frequency shift extraction subnetwork are concatenated and fused through a fully connected layer to generate a frequency shift-phase joint feature.

[0012] Preferably, the dynamic weight generation network includes: The first weight generation subnetwork receives multi-scale convolutional features As input, the multi-scale convolution feature weight coefficient α is output through the fully connected layer and the Sigmoid activation function; The second weight generation subnetwork receives the frequency shift-phase joint feature As input, the frequency shift-phase joint feature weight coefficient β is output through the fully connected layer and the Sigmoid activation function; The third weight generation sub-network receives and The concatenated vector of is taken as input, and the interaction weight coefficient γ is output through the fully connected layer and the Sigmoid activation function; Among them, the weight coefficient of each sub-network satisfies .

[0013] Preferably, feature interaction functions based on physical constraints The definition is as follows: ; In the formula, is the learnable Brillouin physical constraint matrix, initialized to a value close to the identity matrix; is the learnable parameter vector of 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; The calculation formula of fusion features is as follows: ; in, To fuse features, the dimension is adjusted through the output transformation layer to obtain the final feature representation.

[0014] Preferably, the multi-task output network includes: The shared feature layer receives the fused features as input and extracts the common feature representation through the fully connected layer; The anomaly detection branch includes a fully connected layer and a nonlinear activation function, and outputs the probability of anomaly occurrence; The anomaly type classification branch includes a fully connected layer and a probability normalization function, which outputs the probability distribution of different anomaly types. The location branch contains a fully connected layer and outputs the quantized distance value of the abnormal location.

[0015] 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 any of the above methods, the device comprising: Signal acquisition module: including narrow linewidth laser, distributed fiber optic sensor network, photodetector and high-speed data acquisition card, the narrow linewidth laser is configured to generate an optical signal with adjustable pulse width, the distributed fiber optic sensor 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; Preprocessing module: including a polarization fading compensation unit and a morphological filtering unit. The polarization fading compensation unit is configured to eliminate the fluctuation of orthogonal polarization state signals. The morphological filtering unit is configured to use an elliptical structural element to perform an opening and closing operation to suppress noise and retain the spatiotemporal characteristics of the event. Feature extraction module: including an autoencoder, a multi-scale convolution feature extraction unit and a phase-sensitive feature extraction unit. The autoencoder is used to perform feature dimension reduction on the preprocessed signal. The multi-scale convolution feature extraction unit and the phase-sensitive feature extraction unit are used to perform feature extraction in parallel to obtain multi-scale convolution features and frequency shift-phase joint features. The feature fusion module is used to assign weights through a dynamic weight generation network and fuse multi-scale convolution features and frequency shift-phase joint features in combination with a feature interaction function based on physical constraints to generate fused features; The multi-task output module is used to input the fused features into the multi-task output network, and simultaneously perform anomaly detection, anomaly type classification and location positioning tasks to obtain monitoring results; A processor, configured to execute computer program instructions to control each module to perform corresponding operations; A memory, used to store the computer program instructions, parameters of each module and monitoring results; The device also 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.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The high-precision distributed optical fiber sensing monitoring method and device based on deep learning proposed in the present invention integrates deep learning with the physical model of optical fiber sensing to establish a complete monitoring system from signal acquisition, preprocessing, feature extraction to multi-task output. Compared with traditional methods, it can improve monitoring accuracy, increase real-time response speed, and reduce false alarm rate. At the same time, it has made improvements in monitoring distance, spatial resolution, and abnormal type recognition capabilities, thereby enhancing the adaptability and reliability of the system in complex environments. (2) The present invention adopts a signal preprocessing method that combines polarization fading compensation with morphological filtering. By acquiring Brillouin scattering signals in two orthogonal polarization directions and performing complex domain synthesis, the problem of random changes in polarization states caused by optical fiber micro-bending, stress changes, etc. is effectively alleviated, and the amplitude of signal fluctuations is reduced. At the same time, the morphological opening and closing operations of elliptical structural elements are used to effectively filter out noise while retaining the spatiotemporal characteristics of the signal, thereby improving the signal-to-noise ratio. (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, and realizes flexible configuration of monitoring performance under the premise of ensuring the signal-to-noise ratio, so that the system can achieve corresponding spatial resolution within different monitoring distances, and the response time can also be flexibly adjusted, which is conducive to improving the scope of application and engineering adaptability of the system; (4) The phase-sensitive feature extraction module of the present invention extracts complementary features from the phase domain and the frequency domain respectively through the phase extraction subnetwork and the frequency shift extraction subnetwork, and generates a frequency shift-phase joint feature through the joint coding layer, making full use of the frequency shift and phase information of the Brillouin scattering signal, so that the system can distinguish the frequency shift differences caused by temperature changes and strain changes; (5) The feature interaction function based on physical constraints proposed in the present invention establishes a nonlinear mapping relationship between convolution features and frequency shift-phase features, and integrates the physical mechanism of Brillouin scattering into the feature fusion process, so that the deep learning model can not only learn the statistical laws of data, but also follow the physical constraints of fiber optic sensing. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a schematic diagram of the technical implementation of the present invention; Figure 3 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with 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 part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, the present invention provides a high-precision distributed optical fiber sensing monitoring method based on deep learning, comprising: S1, obtaining the Brillouin scattering original signal of the distributed optical fiber sensing system, performing polarization fading compensation and morphological filtering preprocessing on the original signal to obtain a preprocessed signal; S2, input the preprocessed signal into the autoencoder for feature dimensionality reduction to obtain low-dimensional feature representation; S3, performing multi-scale convolution feature extraction and phase-sensitive feature extraction on the low-dimensional feature representation in parallel to obtain multi-scale convolution features and frequency shift-phase joint features respectively; S4, weights are assigned through a dynamic weight generation network, and the multi-scale convolution features and the frequency shift-phase joint features are fused to generate fused features in combination with the feature interaction function based on physical constraints; S5. Input the fused features into the multi-task output network, and simultaneously perform anomaly detection, anomaly type classification and location positioning tasks to obtain monitoring results.

[0021] like Figure 2 As shown, the present invention aims at the problems of insufficient signal processing accuracy, poor real-time performance and low environmental adaptability in distributed optical fiber sensing monitoring, and proposes a high-precision monitoring method and device that integrates deep learning with the physical model of optical fiber sensing. The method first pre-processes the original Brillouin scattering signal through polarization fading compensation and morphological filtering to suppress noise and retain key features; then uses the autoencoder to reduce the feature dimension of the pre-processed signal to reduce the computational complexity; then performs multi-scale convolution feature extraction and phase-sensitive feature extraction in parallel to capture the time characteristics and frequency shift-phase characteristics of the signal respectively; then, the two types of features are organically integrated through a dynamic weight generation network and a feature interaction function based on physical constraints to make full use of their complementarity; finally, the fused features are input into a multi-task output network to simultaneously realize anomaly detection, anomaly type classification and location positioning functions. By combining the feature extraction capability of deep learning with the physical mechanism of Brillouin scattering, a complete monitoring system from signal acquisition to result output is established, so that the system has good real-time performance and environmental adaptability while maintaining high accuracy.

[0022] Specifically, in one embodiment of the present invention, a distributed optical fiber sensing system is used to collect signals, and the system architecture includes a narrow linewidth laser, a distributed optical fiber sensing network, a photodetector, and a high-speed data acquisition card.

[0023] The system uses coherent detection technology to eliminate common mode noise through balanced detectors and improve signal reception sensitivity. In long-distance monitoring scenarios, an optical amplifier (EDFA) is set up every 10-20km to compensate for transmission loss, and Raman amplification technology is used to improve remote signal quality.

[0024] Specifically, in one embodiment of the present invention, considering that pulse parameters have a significant impact on monitoring performance, it is difficult to balance the spatial resolution and signal-to-noise ratio of long-distance optical fiber monitoring with fixed traditional pulse parameters. 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, balances the monitoring distance, spatial resolution and response time, and collects high-quality original signals.

[0025] Specifically, this embodiment optimizes four core acquisition parameters, namely: Pulse Width , which determines the spatial resolution of the system. Narrower pulses provide higher spatial resolution, but weaker signal strength; wider pulses provide higher signal strength, but reduced spatial resolution. The optimal range of this parameter is set to 50-200ns, corresponding to a spatial resolution of about 5-20m.

[0026] Pulse repetition frequency (PRF) affects the data acquisition rate and maximum monitoring distance of the system. A PRF that is too high will cause aliasing of reflected signals from different locations; a PRF that is too low will reduce the data update rate. The optimization range of this parameter is 5-20kHz.

[0027] The signal average number N is related to the signal-to-noise ratio and system response time. The more average times, the higher the signal-to-noise ratio, but the longer the response time.

[0028] Pulse power, this parameter affects the signal strength and monitoring distance. Too high power may cause nonlinear effects; too low power may make the remote signal undetectable.

[0029] In this embodiment, the adaptive pulse parameter optimization strategy is as follows: Dynamically adjust the optical pulse width according to the target spatial resolution requirements, spatial resolution Calculation formula: ; Where c is the speed of light, n is the refractive index of the optical fiber, is the pulse width; for example, a 100ns pulse corresponds to about 10m spatial resolution.

[0030] Automatically adjust the pulse repetition frequency according to the monitoring distance L, the pulse repetition frequency The upper limit value is Settings; for example, the maximum PRF of a 10km optical fiber is about 10kHz.

[0031] Select the signal averaging times N based on the monitoring mode, where: reduce the averaging times in fast response mode; increase the averaging times in high precision mode to improve the signal-to-noise ratio; the relationship between the averaging times N and the signal-to-noise ratio improvement is: , in fast response mode N=100-1000, in high precision mode N=1000-10000.

[0032] Dynamically adjust the pulse power according to the optical fiber line loss measurement results to generate an optical power distribution curve; specifically, first use the optical time domain reflection technology to measure the actual loss distribution of the optical fiber line, and calculate the total round-trip loss L from the transmitter to the distance z. total (z), including inherent fiber attenuation, connector loss, and microbending loss. Based on these loss measurement results, the system dynamically calculates the power distribution curve at different distances through the following power adjustment algorithm: ; in, is the power distribution curve, is the maximum output power of the laser (50mW), is the target signal-to-noise ratio (10-15dB), is the noise power spectral density at the receiving end.

[0033] The 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 the electro-optic modulator (EOM): , rect is a rectangular pulse function.

[0034] The optimal pulse parameter combination is determined based on the spatial resolution, monitoring distance and response time requirements.

[0035] This embodiment adjusts the above parameters according to the real-time monitoring environment and needs, balancing 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 (50-200ns) and repetition frequency (5-20kHz) are dynamically adjusted. The adaptive averaging strategy adjusts the average times according to the monitoring mode for fast response / high precision, and adjusts the pulse power based on the line loss to achieve pulse parameter optimization based on monitoring needs.

[0036] Specifically, in one embodiment of the present invention, considering that the optical fiber signal generates fading noise due to random changes in polarization state, the traditional threshold method cannot effectively suppress it. Therefore, a preprocessing method combining polarization fading compensation and morphological filtering is adopted.

[0037] Specifically, polarization diversity technology is combined with elliptic interpolation algorithm to eliminate the fluctuation of orthogonal polarization state signals. Polarization fading compensation includes: Collect the first Brillouin scattering signal in the orthogonal polarization direction and the second Brillouin scattering signal ; The first Brillouin scattering signal and the second Brillouin scattering signal are synthesized in the complex domain to obtain a polarization-insensitive synthesized signal; the synthesized signal calculation formula is: ; Reduce signal fluctuations caused by changes in polarization state.

[0038] Morphological filtering includes: An elliptical structure element is used for morphological operations, with a time domain radius of 3-7 sampling points and a spatial radius of 1-3 spatial sampling points. An asymmetric elliptical structure element B(a,b) is used, 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.

[0039] 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.

[0040] The results of the opening operation and the closing operation are weighted averaged to obtain a preprocessed signal.

[0041] Specifically, the morphological opening operation is erosion followed by dilation, which is used to remove peak noise. The expression is: ; The morphological closing operation is dilation followed by erosion, which is used to fill the signal valley value: ; The final filtering result is the weighted average of the opening and closing operations: ; The weight α is set to 0.5-0.7 and can be adjusted dynamically according to the signal characteristics.

[0042] This filtering method can effectively suppress environmental interference and polarization changes while preserving the spatiotemporal characteristics of the event signal.

[0043] 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.

[0044] In this embodiment, the autoencoder is located at the front end of the model, and its structure is: 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.

[0045] Specifically, the encoder is used to compress the high-dimensional Brillouin scattering original signal into a low-dimensional feature representation. In a specific example, the encoder includes an input layer and three fully connected layers: The input layer receives the original Brillouin scattering signal ,in is the number of sampling points. The first fully connected layer: ,in , , apply L2 regularization: , . Second fully connected layer: ,in , , apply Dropout, dropout rate . The third fully connected layer: ,in , , apply batch normalization. Feature representation: ,in That is, low-dimensional feature representation.

[0046] The decoder part reconstructs the original signal through a fully connected layer structure symmetrical to the encoder and is only used in the pre-training stage.

[0047] Specifically, the decoder reconstructs the features output by the encoder into the original signal for pre-training of the autoencoder. In a specific example, the decoder also includes an input layer and 3 fully connected layers: The input layer receives the encoder output features . The fourth fully connected layer: ,in , . Fifth fully connected layer: ,in , , apply Dropout, dropout rate . Sixth fully connected layer: ,in , , is the Sigmoid activation function. Reconstruct the signal: .

[0048] Among them: in the pre-training stage, the signal reconstruction mean square error is used as the loss function to optimize the encoder and decoder parameters; after the pre-training is completed, the encoder part is retained for feature dimensionality reduction and the low-dimensional feature representation is output.

[0049] In this embodiment, the pre-training process of the autoencoder is as follows: Design loss function: mean square 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.

[0050] In this embodiment, the multi-scale convolution feature extraction module uses a parallel multi-scale convolution structure to capture feature patterns at different time scales. The multi-scale convolution feature extraction includes: Reshape the low-dimensional feature representation into a three-dimensional tensor to adapt to the convolution operation.

[0051] In one example, the feature vector output by the autoencoder Reshape into a 3D tensor , suitable for convolution operation.

[0052] Features at different time scales are extracted through multiple parallel convolution paths, each using convolution kernels of different sizes.

[0053] In one example, there are three sets of parallel convolutional layers, using different convolution kernel sizes to capture multi-scale features. Small-scale convolution path: ,in The shape is The convolution kernel, step size ,filling .Mesoscale convolution path: ,in The shape is Convolution kernel, stride = 1, padding = 1. Large-scale convolution path: ,in The shape is The convolution kernel is , stride=1, and padding=1.

[0054] The features output by each path are concatenated in the channel dimension to form a joint feature.

[0055] In an example, the concatenation operation: ,get .

[0056] The joint features are subjected to channel dimension reduction, and the reduced features are converted into multi-scale convolutional features through global average pooling operation.

[0057] In one example, 1×1 convolution is used for dimensionality reduction: ,in is the convolution kernel, and we get . Global Average Pooling: ,get . Flattening operation: , and obtain the final multi-scale convolutional features .

[0058] 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: Signal phase feature extraction is performed on the low-dimensional feature representation through a phase extraction subnetwork, wherein the phase extraction subnetwork includes a phase perception layer, which extracts phase features through frequency domain transformation.

[0059] In one example, the phase extraction subnetwork includes an input layer, two fully connected layers, and a phase perception layer: The input layer receives the autoencoder features . First fully connected layer: ,in , . Phase perception layer: ,in The operation applies Hilbert transform and Fourier transform to extract phase information. Second fully connected layer: ,in , .

[0060] It should be noted that the present invention can also generate an analytical signal through a learnable complex convolution kernel and calculate its phase to extract phase features.

[0061] The frequency shift feature is extracted from the low-dimensional feature representation through the frequency shift extraction subnetwork, wherein the frequency shift extraction subnetwork includes a frequency shift perception layer, which extracts the frequency shift characteristics through time-frequency analysis.

[0062] In one example, the frequency shift extraction subnetwork includes an input layer, two fully connected layers, and a frequency shift perception layer: The input layer receives the autoencoder features . First fully connected layer: ,in , Frequency shift perception layer: ,in The operation applies short-time Fourier transform to extract frequency shift features. The second fully connected layer: ,in , .

[0063] It should be noted that the present invention can also dynamically learn frequency shift features through a differentiable time-frequency convolutional layer.

[0064] The outputs of the phase extraction subnetwork and the frequency shift extraction subnetwork are concatenated and fused through a fully connected layer to generate a frequency shift-phase joint feature.

[0065] In one example, generating a frequency shift-phase joint feature at a joint coding layer specifically includes: Feature stitching: ,get . Joint feature extraction: ,in , Output: frequency shift-phase joint feature .

[0066] In this embodiment, the feature fusion module fuses multi-scale convolution features and frequency shift-phase joint features, taking into account 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: The first weight generation subnetwork receives multi-scale convolutional features As input, the multi-scale convolution feature weight coefficient α is output through a fully connected layer and a Sigmoid activation function.

[0067] The second weight generation subnetwork receives the frequency shift-phase joint feature As input, the frequency shift-phase joint feature weight coefficient β is output through the fully connected layer and the Sigmoid activation function; The third weight generation sub-network receives and The concatenated vector of is taken as input, and the interaction weight coefficient γ is output through the fully connected layer and the Sigmoid activation function; Among them, the weight coefficient of each sub-network satisfies .

[0068] In a specific example, the three weight generation subnetworks each include 2 fully connected layers, and the specific settings are as follows: The first weight generation subnetwork, i.e. Generate a network. The input is The first fully connected layer reduces the 64-dimensional space to 32-dimensional space and uses the ReLU activation function. in , The second fully connected layer reduces the 32-dimensional space to 1-dimensional space and uses the Sigmoid activation function. in , , is the Sigmoid function. The output is a scalar .

[0069] The second weight generation subnetwork, i.e. Generate a network. The input is The first fully connected layer reduces the 64-dimensional space to 32-dimensional space and uses the ReLU activation function. in , The second fully connected layer reduces the 32-dimensional space to 1-dimensional space and uses the Sigmoid activation function. in , The output is a scalar .

[0070] The third weight generation subnetwork, i.e. Generate a network. The input is , which represents the concatenation of two feature vectors. The first fully connected layer reduces the 128-dimensional vector to 32-dimensional vectors and uses the ReLU activation function. in , The second fully connected layer reduces the 32-dimensional space to 1-dimensional space and uses the Sigmoid activation function. in , The output is a scalar .

[0071] Feature interaction functions based on physical constraints Incorporating the physical characteristics of Brillouin scattering, it is defined as follows: ; In the formula, is the learnable Brillouin physical constraint matrix, initialized to a value close to the identity matrix; is the learnable parameter vector of 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 exponentiation operation.

[0072] The calculation formula of fusion features is as follows: ; in, To fuse features, the dimension is adjusted through the output transformation layer to obtain the final feature representation.

[0073] In order to enhance the feature expression ability and adjust the dimension, the fusion features are transformed as follows: ; in , output .

[0074] In this embodiment, the multi-task output module is responsible for simultaneously outputting the prediction results of three tasks based on the fusion features: anomaly detection, anomaly type classification, and position positioning. This module is the multi-task output network, which includes: The shared feature layer receives the fused features as input and extracts common feature representations through the fully connected layer.

[0075] Specifically, the shared layer input is the fusion feature . Shared fully connected layer: ,in , , apply Dropout, dropout rate .

[0076] The anomaly detection branch contains a fully connected layer and a nonlinear activation function, and outputs the probability of anomaly occurrence.

[0077] Specifically, this branch includes 2 fully connected layers. The first fully connected layer is: ,in , . Second fully connected layer: ,in , , is the Sigmoid activation function. The output is the abnormal probability .

[0078] The anomaly type classification branch includes a fully connected layer and a probability normalization function, which outputs the probability distribution of different anomaly types.

[0079] Specifically, this branch includes 2 fully connected layers. The first fully connected layer is: ,in , . Second fully connected layer: ,in , , is the number of abnormal types. The output is the probability distribution of each abnormal type ,satisfy .

[0080] The location branch contains a fully connected layer and outputs the quantized distance value of the abnormal location.

[0081] Specifically, this branch includes 2 fully connected layers. The first fully connected layer is: ,in , . Second fully connected layer: ,in , , using a linear activation function. The output is the distance value of the abnormal position .

[0082] In one embodiment of the present invention, the overall training process of the deep learning model is: Autoencoder pre-training: Use unlabeled data to train the autoencoder to reconstruct the original signal, fix the encoder weights, and extract the feature vector.

[0083] End-to-end model training: load pre-trained encoder weights, train a full multi-task model using labeled data, monitor validation metrics and overall loss for each task, and apply early stopping to prevent overfitting.

[0084] Model fine-tuning: Unfreeze the encoder weights and use a smaller learning rate for fine-tuning, focusing on optimizing the physical parameters in the physical perception fusion module.

[0085] Among them, the multi-task loss function is used in end-to-end model training, and the expression is: ,in , , is the weight of the loss of each task.

[0086] Loss for the anomaly detection task: ; In the formula, N is the number of training samples, is the true abnormality label of the i-th sample, is the probability that the model predicts the i-th sample as abnormal.

[0087] Loss for the anomaly type classification task: ; In the formula, N is the number of training samples, is the number of abnormal categories, is the true label of the i-th sample belonging to category c, Predict the probability that the i-th sample belongs to category c for the model.

[0088] Loss for the location positioning task: ; In the formula, N is the number of training samples, is the location label of the i-th abnormal sample, The position of the i-th abnormal sample predicted by the model.

[0089] 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 tags are obtained together with the signal data. The position tag mentioned in the present invention is a relative distance, that is, the distance from the abnormal point to the starting end of the optical fiber system. Therefore, the position location branch outputs a quantized distance value.

[0090] The training strategy during the training process is: Optimizer: Adam, learning rate = 0.0005. Learning rate scheduling: If the validation loss does not improve every 10 rounds, the learning rate decays to the original 0.5. Early stopping strategy: Stop training if the total validation loss does not improve for 15 consecutive rounds.

[0091] Specifically, in one embodiment of the present invention, in order to improve the generalization ability of the model, a data enhancement strategy is designed for the characteristics of optical fiber sensing signals: Signal-to-noise ratio transformation: Add noise of different intensities to the training samples. Gaussian white noise: SNR ranges from -30dB to 0dB. Colored noise: 1 / f noise simulates environmental vibration background. Impulse noise: simulates electrical interference.

[0092] 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.

[0093] Synthetic event generation: Event superposition: Linear combination of multiple event signals. Position randomization: Randomly place events at different positions on the fiber. Background mixing: Mix with the actual environmental noise background.

[0094] Through these data enhancement techniques, the original data set is expanded 10-20 times, improving the model's ability to adapt to various environmental conditions.

[0095] Specifically, in one embodiment, after the deep learning model is trained, the present invention deploys it in the monitoring device, and continuously optimizes the model performance through the built-in incremental learning mechanism. The incremental learning algorithm is built into the back end of the multi-task output module, sharing the feature extraction layer with the main model but having an independent set of fine-tuning parameters. After the system runs for a period of time, new labeled data will be collected, especially those samples that are misclassified and newly emerging anomaly types. At this point, the incremental learning process begins to execute, which is implemented by the following steps: First, for the newly collected dataset , the model is updated by combining knowledge distillation and elastic weight. Knowledge distillation updates the original model by The output of is used as a soft label, which together with the true label guides the learning of the new model. Its loss function is: ; in: is the cross entropy loss; is the KL divergence loss, is a balanced parameter. At the same time, to prevent catastrophic forgetting, the system applies elastic weight consolidation (EWC) regularization constraints when updating parameters to ensure that the parameters important to old tasks change less: ; in, is the new parameter, is the old parameter, It is the balance coefficient, which controls the weight of new and old tasks. is the Fisher information matrix, and the calculation formula is: ; Approximate calculation is performed by sampling old mission data. Important parameters are largely protected, and non-critical parameters can be updated freely.

[0096] The steps for incremental learning are as follows: 1. Train the basic model on the initial data set to obtain the parameters . 2. Calculate the Fisher information matrix F using the basic model and training data. 3. When new data arrives, use the extended loss function for training. 4. Periodically update the Fisher matrix to adapt to environmental changes.

[0097] In addition, if Figure 3 As 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 any of the above methods, and the device includes: Signal acquisition module: including narrow linewidth laser, distributed fiber optic sensor network, photodetector and high-speed data acquisition card, the narrow linewidth laser is configured to generate an optical signal with adjustable pulse width, the distributed fiber optic sensor 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; Preprocessing module: including a polarization fading compensation unit and a morphological filtering unit. The polarization fading compensation unit is configured to eliminate the fluctuation of orthogonal polarization state signals. The morphological filtering unit is configured to use an elliptical structural element to perform an opening and closing operation to suppress noise and retain the spatiotemporal characteristics of the event. Feature extraction module: including an autoencoder, a multi-scale convolution feature extraction unit and a phase-sensitive feature extraction unit. The autoencoder is used to perform feature dimension reduction on the preprocessed signal. The multi-scale convolution feature extraction unit and the phase-sensitive feature extraction unit are used to perform feature extraction in parallel to obtain multi-scale convolution features and frequency shift-phase joint features. The feature fusion module is used to assign weights through a dynamic weight generation network and fuse multi-scale convolution features and frequency shift-phase joint features in combination with a feature interaction function based on physical constraints to generate fused features; The multi-task output module is used to input the fused features into the multi-task output network, and simultaneously perform anomaly detection, anomaly type classification and location positioning tasks to obtain monitoring results; A processor, configured to execute computer program instructions to control each module to perform corresponding operations; A memory, used to store the computer program instructions, parameters of each module and monitoring results; The device also 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.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should 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: include: S1, obtaining the Brillouin scattering original signal of the distributed optical fiber sensing system, performing polarization fading compensation and morphological filtering preprocessing on the original signal to obtain a preprocessed signal; S2, input the preprocessed signal into the autoencoder for feature dimensionality reduction to obtain low-dimensional feature representation; S3, performing multi-scale convolution feature extraction and phase-sensitive feature extraction on the low-dimensional feature representation in parallel to obtain multi-scale convolution features and frequency shift-phase joint features respectively; S4, weights are assigned through a dynamic weight generation network, and the multi-scale convolution features and the frequency shift-phase joint features are fused to generate fused features in combination with the feature interaction function based on physical constraints; S5. Input the fused features into the multi-task output network, and simultaneously perform anomaly detection, anomaly type classification and location positioning tasks to obtain monitoring results.

2. According to the high-precision distributed optical fiber sensing monitoring method based on deep learning in claim 1, it is characterized in that: Polarization fading compensation includes: Collecting a first Brillouin scattering signal and a second Brillouin scattering signal in orthogonal polarization directions; Performing complex domain synthesis on the first Brillouin scattering signal and the second Brillouin scattering signal to obtain a polarization-insensitive synthesized signal; Morphological filtering includes: An elliptical structure element is used for morphological operations. The temporal radius of the structure element is 3-7 sampling points, and the spatial radius is 1-3 spatial sampling points. 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. The results of the opening operation and the closing operation are weighted averaged to obtain a preprocessed signal.

3. According to a high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1, it is characterized in that: Step S1 also includes a pulse parameter optimization process of the distributed optical fiber sensing system, the steps comprising: Dynamically adjust the optical pulse width according to the target spatial resolution requirements, spatial resolution Calculation formula: ； Where c is the speed of light, 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, the pulse repetition frequency The upper limit value is set up; Select the signal averaging times based on the monitoring mode, where: reduce the averaging times in fast response mode; increase the averaging times in high precision mode to improve the signal-to-noise ratio; Dynamically adjust pulse power based on fiber line loss measurement results to generate optical power distribution curves; The optimal pulse parameter combination is determined based on the spatial resolution, monitoring distance and response time requirements.

4. The high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1 is characterized in that: The autoencoders include: The encoder part consists of at least three layers of fully connected networks, which compress the preprocessed signal into a low-dimensional feature representation layer by layer, with nonlinear activation functions and regularization operations in each layer; The decoder part reconstructs the original signal through a fully connected layer structure symmetrical to the encoder and is only used in the pre-training stage; Among them: in the pre-training stage, the signal reconstruction mean square error is used as the loss function to optimize the encoder and decoder parameters; after the pre-training is completed, the encoder part is retained for feature dimensionality reduction and the low-dimensional feature representation is output.

5. The high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 1 is characterized in that: Multi-scale convolutional feature extraction includes: Reshape the low-dimensional feature representation into a three-dimensional tensor to adapt to the convolution operation; Features at different time scales are extracted through multiple parallel convolution paths, each using convolution kernels of different sizes; The features output by each path are concatenated in the channel dimension to form a joint feature; The joint features are subjected to channel dimension reduction, and the reduced features are converted into multi-scale convolutional features through global average pooling operation.

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

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

8. The high-precision distributed optical fiber sensing monitoring method based on deep learning according to claim 7 is characterized in that: Feature interaction functions based on physical constraints The definition is as follows: ； In the formula, is the learnable Brillouin physics constraint matrix; is the learnable parameter vector of 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; The calculation formula of fusion features is as follows: ； in, To fuse features, the dimension is adjusted through the output transformation layer to obtain the final feature representation.

9. The 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: The shared feature layer receives the fused features as input and extracts the common feature representation through the fully connected layer; The anomaly detection branch includes a fully connected layer and a nonlinear activation function, and outputs the probability of anomaly occurrence; The anomaly type classification branch includes a fully connected layer and a probability normalization function, which outputs the probability distribution of different anomaly types. The location branch contains a fully connected layer and outputs the quantized distance value of the abnormal location.

10. A high-precision distributed optical fiber sensing monitoring device based on deep learning, characterized in that: The device is used to perform the method according to any one of claims 1 to 9, and the device comprises: Signal acquisition module: including narrow linewidth laser, distributed fiber optic sensor network, photodetector and high-speed data acquisition card, the narrow linewidth laser is configured to generate an optical signal with adjustable pulse width, the distributed fiber optic sensor 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; Preprocessing module: including a polarization fading compensation unit and a morphological filtering unit. The polarization fading compensation unit is configured to eliminate the fluctuation of orthogonal polarization state signals. The morphological filtering unit is configured to use an elliptical structural element to perform an opening and closing operation to suppress noise and retain the spatiotemporal characteristics of the event. Feature extraction module: including an autoencoder, a multi-scale convolution feature extraction unit and a phase-sensitive feature extraction unit. The autoencoder is used to perform feature dimension reduction on the preprocessed signal. The multi-scale convolution feature extraction unit and the phase-sensitive feature extraction unit are used to perform feature extraction in parallel to obtain multi-scale convolution features and frequency shift-phase joint features. The feature fusion module is used to assign weights through a dynamic weight generation network and fuse multi-scale convolution features and frequency shift-phase joint features in combination with a feature interaction function based on physical constraints to generate fused features; The multi-task output module is used to input the fused features into the multi-task output network, and simultaneously perform anomaly detection, anomaly type classification and location positioning tasks to obtain monitoring results; A processor, configured to execute computer program instructions to control each module to perform corresponding operations; A memory, used to store the computer program instructions, parameters of each module and monitoring results; The device also 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