Distributed optical fiber sensing mass target data screening method based on image processing

Abstract

The invention discloses a distributed optical fiber sensing mass target data screening method based on image processing, and the method comprises the steps: mapping a space-time two-dimensional vibration signal into a gray image, employing a multi-mode fusion technology, combining a lightweight convolutional neural network to achieve the efficient classification, introducing a dynamic incremental learning mechanism and a space propagation attenuation correction algorithm, and carrying out the screening of mass target data. The problems that a traditional method is single in feature representation, lagged in model updating and insufficient in positioning precision are solved. The method is suitable for the fields of perimeter security and protection, pipeline monitoring, geological disaster early warning and the like, and can perform real-time screening and target identification on mass vibration data under complex environment interference. The method breaks through the limitation of traditional time-frequency domain analysis through signal imaging representation, remarkably improves the distinguishing capability of similar events, reduces the false alarm rate and the operation and maintenance cost through a dynamic model optimization technology, effectively inhibits the influence of optical fiber attenuation on the positioning precision in combination with a space correction algorithm, and improves the positioning accuracy. And reliable technical support is provided for intelligent upgrading of a distributed optical fiber sensing system.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of optical fiber sensing and image processing, and specifically designs a lightweight convolutional neural network based on image processing and multi - modal feature fusion, as well as a method for screening massive data of distributed optical fiber sensing with dynamic incremental learning. Background Technique

[0002] Distributed optical fiber sensing technology is based on the sensing principle of a phase - sensitive optical time - domain reflectometer (Φ - OTDR), and realizes long - distance vibration monitoring in fields such as perimeter security and pipeline monitoring. However, the massive data generated by it still faces the bottleneck of real - time processing. Traditional methods rely on time - domain or frequency - domain for single - dimensional analysis, and it is difficult to distinguish the spatio - temporal coupling characteristics of environmental interference and real events, resulting in a relatively high false alarm rate. Existing shallow machine learning algorithms have the defect of feature confusion in multi - target scenarios, while the cloud - based offline processing mode generates a delay of minutes, which cannot meet the real - time warning requirements in high - risk scenarios.

[0003] Current deep - learning applications have significant limitations: Conventional convolutional neural network (CNN) models have too high a computational load and are difficult to adapt to edge devices; Static models cannot cope with data distribution shifts caused by environmental changes and require frequent manual iterations; Existing signal image - forming methods are limited to time - frequency spectrum conversion and lack the multi - modal fusion of texture, spatial topology, and dynamic behavior, making it difficult to identify similar signals (such as construction vibration and geological disasters). In view of the above defects, there is an urgent need to propose a technical solution that integrates lightweight networks, dynamic incremental learning, and collaborative analysis of multi - modal features to overcome the technical problems of real - time screening of massive data, adaptive modeling in dynamic environments, and high - precision target recognition, and to improve the reliability and timeliness of optical fiber sensing systems in the monitoring of critical infrastructure. Summary of the Invention

[0004] Based on the above problems, the present invention provides a method for screening massive target data of distributed optical fiber sensing based on image processing, which solves the technical problems of high false alarm rate, poor real - time performance, and weak environmental adaptability caused by single - feature representation, lagging model update, and insufficient collaboration of multi - modal information in traditional methods.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A method for screening massive target data of distributed optical fiber sensing based on image processing, comprising the following steps:

[0007] Step 1: Collect spatio - temporal two - dimensional vibration signals through a distributed optical fiber sensing system, and convert them into grayscale images after normalizing the signal amplitude.

[0008] Step 2: Extract multi-modal features from the grayscale image in parallel; specifically, use the local binary pattern algorithm to extract the texture features of the grayscale image, calculate the shape features of the event region based on the Hu invariant moment; calculate the motion vector field of adjacent frames through the Lucas-Kanade optical flow method and then extract the spatio-temporal dynamic features;

[0009] Step 3: Concatenate the texture features, shape features and dynamic features to generate a joint feature vector;

[0010] Step 4: Use the MobileNetV3 network architecture to process the joint feature vector to obtain the classification confidence;

[0011] Step 5: Set the trigger for dynamic incremental learning. When the classification confidence < θ, trigger the online hard example mining strategy and update the model parameters with the hybrid loss function;

[0012] Step 6: Perform spatial propagation attenuation correction and compensate for the optical fiber attenuation of the classification result;

[0013] Step 7: When the continuous trigger times K of incremental learning ≥ 3, generate pseudo-labels based on historical high-confidence samples and expand them to the training dataset to achieve autonomous iterative optimization of the model.

[0014] Further, the conversion of the signal amplitude to a grayscale image after normalization is specifically as follows:

[0015] G (x,t) = floor(255 * (S max - S min ) / (S (x,t) - S min ))

[0016] where G (x,t) represents the grayscale image, x is the spatial position coordinate of the sensing optical fiber, t is the time variable; floor means rounding down; S min , S max are the minimum and maximum values of the signal amplitude respectively; S (x,t) is the spatio-temporal two-dimensional vibration signal.

[0017] Further, the use of the local binary pattern algorithm to extract the texture features of the grayscale image is specifically as follows:

[0018]

[0019] where F texture is the texture feature, g c is the pixel value of the central pixel, g p is the pixel value of the p-th sampling point on the neighborhood, and s(·) is the step function.

[0020] Furthermore, the calculation of Hu invariant moments is as follows:

[0021] The main invariant moment combinations M1, M2, M3, M4, M5, M6, and M7 calculate the shape feature F of the event region shape :

[0022] M1 = (η 20 + η 02 ) 2

[0023] M2 = (η 20 + η 02 ) 2 + 4η 11 2

[0024] Among them, the first-order moment M1 is used to describe the pixel distribution range, and the second-order moment M2 is used to capture the directional feature; the calculation formulas of Hu invariant moments M3 to M7 are as follows:

[0025] M3 = (η 30 - 3η 12 ) 2 +(3η 21 - η 03 ) 2

[0026] M4 = (η 30 + η 12 ) 2 +(η 21 + η 03 ) 2

[0027] M5 = (η 30 + 3η 12 )[(η 21 + η 03 ) 2 - 3(η 21 + η 03 )] 2

[0028] M6 = (η 20 - η 02 )[(η 30 + η 12 ) 2 -(η 21 + η 03 ) 2 + 4η 11 (η 30 + η 12 )(η 21 + η 03 )]

[0029] M7 = (3η21 +η 03 )(3η 21 +η 03 )[(η 30 +η 12 ) 2 -3(η 21 +η 03 ) 2 -(η 30 +3η 12 )(η 21

[0030] +η 03 )[3(η 30 +η 12 ) 2 -(η 21 +η 03 ) 2

[0031] Among them, M3 describes symmetry, and M4, M5, M6, and M7 describe curvature; in the formula, η pq is the normalized central moment, and its calculation formula is as follows:

[0032]

[0033] Among them, u pq represents the geometric moment of the image relative to the centroid, which is used to represent the elimination of translational effects; u 00 represents the zero-order central moment.

[0034] Furthermore, the dynamic features satisfy the brightness constancy constraint equation:

[0035]

[0036] Among them, x, y, and t represent spatial coordinates and time stamps, Gu, Gv, and G represent the horizontal displacement, vertical displacement, and initial and final linear displacement of pixels, represents the partial derivative with respect to the variables x, y, t, Gu, Gv, and G.

[0037] Furthermore, step 3 adopts an attention weighting mechanism, and the calculation formula of the weight coefficient is as follows:

[0038]

[0039] Among them, w ij represents the similarity between each modal feature, q is a learnable query vector, and F i and F j are the modal feature vectors. Furthermore, the process of the MobileNetV3 network architecture for processing the joint feature vector is represented by the following formula: ​

[0040] FCNN = DWConv(ReLU(BN(F joint )))

[0041] Wherein, DWConv is a depthwise separable convolution, and F joint represents the joint feature; ReLU represents the rectified activation function, BN represents the batch normalization function, and the output feature map is FCNN; finally, the feature map FCNN uses the normalization function to obtain the maximum probability score of the target object, and this is used as the classification confidence P c Output.

[0042] Furthermore, the calculation process of the hybrid loss function is as follows:

[0043] L = λ·LCE + (1 - λ)·L triplet

[0044] Wherein, L represents the regularization loss, LCE is the cross-entropy loss, and L triplet is the triplet loss, and λ is the regularization parameter.

[0045] Furthermore, the correction calculation formula in step 6 is as follows:

[0046]

[0047] Wherein, P final (x) represents the final power at position x, and P c (x) represents the initial power at the reference point x0, α is the fiber attenuation coefficient, x0 is the initial position of the event, and |x - x0| represents the absolute distance from the reference point x0 to the position x of the fiber optic sensor.

[0048] The advantages and technical effects of the present invention are as follows:

[0049] The present invention is applicable to the fields of perimeter security, oil and gas pipeline monitoring, rail transit status perception, and geological disaster warning. Based on the image-based signal representation and multi-modal feature fusion technology, the present invention can efficiently screen a large amount of vibration data in complex scenarios such as wind and rain interference, mechanical vibration, and geological deformation, and the recognition accuracy of the target event is increased to more than 95%.

[0050] The present invention does not require the use of high-cost hardware acceleration devices or frequent manual annotation, is suitable for the deployment of edge computing nodes, relies on a lightweight convolutional neural network (the CPU computing load is reduced by 70%) and a dynamic incremental learning mechanism to improve the generalization ability of traditional static models, establishes an environment-adaptive data screening system through the collaborative analysis of spatio-temporal-texture-shape multi-modal features, realizes millisecond-level real-time response at a sampling frequency of 10 kHz, and at the same time supports the autonomous evolution of the model in long-term scenarios (such as seasonal changes, geological subsidence), and the false alarm rate is stably lower than 5% for a long time. Description of the Drawings

[0051] Figure 1 It is a schematic diagram of the composition of each functional module provided by the present invention.

[0052] Figure 2 It is a flowchart of the method of the present invention. Detailed Embodiments

[0053] Please refer to Figure 1 , Figure 1 which is a schematic diagram of signal imaging conversion provided by an example of the present invention; please refer to Figure 2 , Figure 2 which is a flowchart of an embodiment of the present invention. A method for screening a large amount of target data of distributed optical fiber sensing based on image processing includes the following steps:

[0054] Step 1: Use the signal conversion imaging module to realize the imaging conversion of spatio-temporal vibration signals, collect the spatio-temporal two-dimensional vibration signal S (x,t) (such as the ground vibration caused by vehicle driving), normalize the signal amplitude and convert it into a grayscale image G (x,t) , where x is the spatial position coordinate of the sensing optical fiber and t is the time variable. The conversion formula is as follows:

[0055] G (x,t) = floor(255 * (S max - S min ) / (S (x,t) - s min ))

[0056] where floor means rounding down; S min , S max are respectively the minimum and maximum values of the signal amplitude; the sampling frequency fs of the spatio-temporal two-dimensional vibration signal is ≥ 10 kHz, and the spatial resolution Δx ≤ 1 m;

[0057] Step 2: Use the multi-modal feature extraction module to extract multi-modal features from the grayscale image in parallel. Specifically, use the local binary pattern (LBP) algorithm to extract the texture feature F texture of the grayscale image G(x, t); calculate the shape feature F shape by extracting the event area based on the Hu invariant moment; calculate the motion vector field of adjacent frames through the Lucas-Kanade optical flow method (the relevant algorithm module can be directly called) and then extract the spatio-temporal dynamic feature F motion ;

[0058] Specifically, the spatial texture distribution is extracted through the LBP algorithm to obtain the texture feature F texture The calculation process is as follows:

[0059]

[0060] where g c is the pixel value of the central pixel, and g p is the pixel value of the p-th sampling point in the neighborhood, and s(·) is the step function;

[0061] The Hu invariant moment geometric transformation (translation, rotation, scaling), and the main invariant moment combinations M1, M2, M3, M4, M5, M6, M7 are used to calculate the shape feature F of the event region shape , and the calculation formula is as follows:

[0062] M1 = (η 20 + η 02 ) 2

[0063] M2 = (η 20 + η 02 ) 2 + 4η 11 2

[0064] where the first-order moment M1 is used to describe the pixel distribution range, and the second-order moment M2 is used to capture the directional feature (such as the tilt or the main axis direction of the shape). Further, the calculation formulas for the Hu invariant moments M3 to M7 are as follows:

[0065] M3 = (η 30 - 3η 12 ) 2 +(3η 21 - η 03 ) 2

[0066] M4 = (η 30 + η 12 ) 2 +(η 21 + η 03 ) 2

[0067] M5 = (η 30 + 3η 12 )[(η 21 + η 03 ) 2 - 3(η 21 + η 03 )] 2

[0068] M6 = (η 20 - η 02 )[(η 30 + η 12 ) 2 -(η 21 + η03 ) 2 + 4η 11 (η 30 + η 12 )(η 21 + η 03 )]

[0069] M7 = (3η 21 + η 03 )(3η 21 + η 03 )[(η 30 + η 12 ) 2 - 3(η 21 + η 03 ) 2 - (η 30 + 3η 12 )(η 21

[0070] + η 03 )[3(η 30 + η 12 ) 2 - (η 21 + η 03 ) 2 )

[0071] Among them, M3 describes symmetry, and M4, M5, M6, and M7 describe curvature. In addition, η pq in the formula is the normalized central moment, and its calculation formula is as follows:

[0072]

[0073] Among them, u pq represents the geometric moment of the image relative to the centroid, which is used to represent the elimination of translational influence. u 00 represents the zero-order central moment (i.e., the total number of pixels), which is used as the normalization standard in the formula to eliminate scale influence.

[0074] Calculate the motion vector field between adjacent frames based on the optical flow method to obtain the dynamic feature F motion , and this feature satisfies the brightness constancy constraint equation:

[0075]

[0076] Among them, x, y, and t represent spatial coordinates and time stamps, Gu, Gv, and G represent the horizontal displacement, vertical displacement, and the straight-line displacement from the beginning to the end of the pixel, represents the partial derivative of the variables x, y, t, Gu, Gv, and G.

[0077] Step 3: Combine the texture feature F texture and the shape feature Fshape and the dynamic feature F motion are concatenated to generate a combined feature vector F joint , realizing feature fusion; The multi-modal feature fusion adopts an attention weighting mechanism, and the calculation formula of the weight coefficient is as follows:

[0078]

[0079] where w ij represents the similarity between the features of each modality, q is a learnable query vector, and F i and F j are the feature vectors of each modality; The dimension of the query vector q is d = 128, and it is dynamically updated through backpropagation.

[0080] Step 4: Use the MobileNetV3 lightweight CNN network architecture to process the combined features, and obtain the maximum probability score of the target object in its fully connected layer as the classification confidence P c Output. The multi-modal features are input into a lightweight convolutional neural network (CNN) classification model. The CNN adopts the MobileNetV3 architecture. The depthwise separable convolution kernel size of this architecture is 3×3, the channel compression ratio is 1 / 4, and the number of neurons N fc in the last fully connected layer = 256; This architecture uses the depthwise separable convolutional layer to output the feature map FCNN, and the feature calculation process is as follows:

[0081] FCNN = DWConv(ReLU(BN(F joint )))

[0082] where DWConv is the depthwise separable convolution, and F joint represents the combined features; ReLU represents the rectified activation function, and BN represents the batch normalization function;

[0083] Step 5: Set the trigger for dynamic incremental learning. When the classification confidence P c < θ, trigger the online hard example mining strategy, and update the model parameters W with the hybrid loss function, where θ ∈ [0.6, 0.8]; The calculation process of the hybrid loss function is as follows:

[0084] L = λ·LCE + (1 - λ)·L triplet

[0085] where L represents the regularization loss, LCE is the cross-entropy loss, and L triplet is the triplet loss, and λ is the regularization parameter.

[0086] Step 6: Perform spatial propagation attenuation correction and compensate for the optical fiber attenuation of the classification result. The correction calculation formula is as follows:

[0087]

[0088] Among them, P final (x) represents the final power at position x, and P c (x) represents the initial power at the reference point x0. α is the fiber attenuation coefficient (obtained through a calibration experiment), x0 is the initial position of the event (an event that causes attenuation of the optical signal, such as a physical perturbation), and |x - x0| represents the absolute distance from the reference point x0 to the position x of the fiber optic sensor.

[0089] Specifically, a spatial calibration model can be established as follows:

[0090] Based on the attenuation compensation result of step 6, combined with the fiber layout topology map, construct an event propagation attenuation mapping matrix A to achieve collaborative correction of the positioning error in the multi-event concurrent scenario. On this basis, construct an event propagation attenuation mapping matrix. The calculation formula of the event propagation attenuation mapping matrix is as follows:

[0091] A ij = α·L ij = α·|x i - d j |

[0092] α represents the power consumption per unit distance of the fiber, x i represents the position of the physical perturbation event, d j represents the signal receiving point, and L ij represents the path length. Further, when multiple events occur concurrently, distinguish the event types by methods such as separating the source signal, PCA analysis, and time-frequency analysis. Further, adopt a collaborative positioning algorithm to solve the event position through an optimization function. The calculation formula of this optimization function is as follows:

[0093]

[0094] Among them, min represents the minimum value, S j represents the signal at the jth detection point (sensor), A kj represents the signal attenuation coefficient from the jth event source to the kth detection point, represents the signal of the kth event source to be estimated (such as vibration amplitude, position parameters, etc.). represents the mean square error between the predicted signal and the actual signal.

[0095] Step 7: When the continuous trigger times K of incremental learning ≥ 3, generate pseudo-labels based on historical high-confidence samples (P c ≥ 0.9) and expand them to the training dataset to achieve autonomous iterative optimization of the model.

[0096] It will be understood that the present invention is described by way of some embodiments, and those skilled in the art will know that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Additionally, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. A method for screening massive target data of distributed optical fiber sensing based on image processing, characterized in that: The following steps are involved: Step 1: Collect the spatiotemporal two-dimensional vibration signal through the distributed optical fiber sensing system, normalize the signal amplitude and convert it into a grayscale image; Step 2: Extract multimodal features from grayscale images in parallel; specifically, use the local binary pattern algorithm to extract grayscale image texture features, calculate the event region based on Hu invariant moments to extract shape features; calculate the motion vector field of adjacent frames using the Lucas-Kanade optical flow method to extract spatiotemporal dynamic features; Step 3: Concatenate texture features, shape features, and dynamic features to generate a joint feature vector; Step 4: Use the MobileNetV3 network architecture to process the joint feature vector and obtain the classification confidence; Step 5: Set the dynamic incremental learning trigger. When the classification confidence <θ, trigger the online hard example mining strategy and update the model parameters with the hybrid loss function. Step 6: Perform spatial propagation attenuation correction and perform optical fiber attenuation compensation on the classification results; Step 7: When the incremental learning is triggered continuously for K ≥ 3 times, pseudo labels are generated based on historical high-confidence samples and expanded to the training data set to achieve autonomous iterative optimization of the model.

2. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 1 is characterized in that: The signal amplitude is normalized and converted into a grayscale image as follows: G (x,t) =floor(255*(S max -S min ) / (S (x,t) -S min )) Among them, G (x,t) represents a grayscale image, x is the spatial position coordinate of the sensing fiber, t is the time variable; floor refers to rounding down; S min , S max are the minimum and maximum values ​​of the signal amplitude respectively; S (x,t) It is a two-dimensional vibration signal in space and time.

3. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 2 is characterized in that: The local binary pattern algorithm is used to extract the grayscale image texture features as follows: Among them, F texture is the texture feature, g c is the pixel value of the center pixel, g p is the pixel value of the pth sampling point in the neighborhood, and s(·) is a step function.

4. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 3 is characterized in that: The calculation of Hu invariant moment is as follows: The main invariant moment combination M1, M2, M3, M4, M5, M6, M7 calculates the shape feature F of the event area shape : M1=(η 20 +n 02 ) 2 M2=(η 20 +n 02 ) 2 +4th 11 2 Among them, the first-order moment M1 is used to describe the pixel distribution range, and the second-order moment M2 is used to capture the directional characteristics; the calculation formula of Hu invariant moments M3~M7 is as follows: M3=(η 30 -3rd 12 ) 2 +(3rd 21 -or 03 ) 2 M4=(η 30 +n 12 ) 2 +(the 21 +n 03 ) 2 M5=(η 30 +3rd 12 )[(or 21 +n 03 ) 2 -3(h 21 +n 03 )] 2 M6=(η 20 -or 02 )[(or 30 +n 12 ) 2 -(or 21 +n 03 ) 2 ]+4th 11 (or 30 +n 12 )(or 21 +n 03 )] M7=(3rd 21 +n 03 (3rd) 21 +n 03 )[(or 30 +n 12 ) 2 -3(h 21 +n 03 ) 2 ]-(or 30 +3rd 12 )(or 21 +n 03 )[3(h 30 +n 12 ) 2 -(or 21 +n 03 ) 2 ] Among them, M3 describes the symmetry, M4, M5, M6 and M7 describe the curvature; where η pq is the normalized central moment, which is calculated as follows: Among them, u pq Represents the geometric moment of the image relative to the center of mass, which is used to eliminate the translation effect; u 00 represents the zero-order central moment.

5. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 4 is characterized in that: The dynamic characteristics satisfy the brightness constant constraint equation: Among them, x, y, t represent the spatial coordinates and timestamp, Gu, Gv, G represent the horizontal displacement, vertical displacement and start-end linear displacement of the pixel, It means to find partial derivatives of variables x, y, t, Gu, Gv, and G.

6. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 5 is characterized in that: The step 3 adopts the attention weighting mechanism, and the weight coefficient calculation formula is as follows: Among them, w ij represents the similarity between the features of each modality, q is the learnable query vector, and F i With F j is the eigenvector of each mode.

7. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 6, characterized in that: The process of processing the joint feature vector in the MobileNetV3 network architecture is expressed as follows: FCNN=DWConv(ReLU(BN(F joint ))) Among them, DWConv is a depth-separable convolution, F joint ReLU represents the joint feature; ReLU represents the modified activation function, BN represents the batch normalization function, and the output feature map FCNN; finally, the feature map FCNN uses the normalization function to obtain the maximum probability score of the target object, which is used as the classification confidence P c Output.

8. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 7, characterized in that: The hybrid loss function calculation process is as follows: L=λ·LCE+(1-λ)·L triplet Among them, L represents the regularization loss, LCE is the cross entropy loss, and L triplet is the triplet loss and λ is the regularization parameter.

9. The method for screening massive target data of distributed optical fiber sensing based on image processing according to claim 8, characterized in that: The correction calculation formula of step 6 is as follows: Among them, P final (x) represents the final power at position x, P c (x) represents the initial power at the reference point x0, α ​​is the fiber attenuation coefficient, x0 is the initial position of the event, and |x-x0| represents the absolute distance from the reference point x0 to the fiber sensor position x.

Images