Method for identifying coal mine shaft crack in non-illumination environment based on optical fiber sensing grating

Abstract

The invention discloses a coal mine shaft crack identification method in a non-illumination environment based on an optical fiber sensing grating, and belongs to the field of coal mine shaft disease detection. According to the method, distributed optical fibers form an optical fiber sensing grating, the optical fiber sensing grating is fixed on the surface of a well wall, an optical fiber demodulator is used for collecting strain along the optical fibers, a blank blind area of the optical fiber sensing grating is filled through an interpolation algorithm, and a strain field image data set of the well wall is obtained; and a CCD industrial camera is matched to shoot a well wall crack image sample at the optical fiber sensing grating, binarization processing is carried out, and a well wall crack image data set is established. And establishing a well wall crack identification model based on deep learning, introducing two parallel feature extraction branches, respectively receiving an optical fiber strain field image data set and a well wall crack image data set, and performing feature extraction and fusion of the two different types of data sets through training. Because the strain data is sensed through the optical fiber, the illumination environment is not needed, and after the model training is completed, the identification and early warning of the well wall crack in the illumination-free environment can be realized only by inputting the optical fiber strain data acquired in real time into the trained model.

Description

Technical Field

[0001] The invention relates to the field of coal mine safety detection, and in particular to a method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensing grid. Background Art

[0002] The shaft of a coal mine is the throat of a mine. Cracks formed in the shaft during its service life can cause water inrush, changes in the direction of underground wind flow, gas accumulation and other disasters, posing a major threat to the shaft structure and the safety of internal workers. Therefore, it is necessary to effectively monitor the development of shaft cracks. The difficulty of shaft crack monitoring lies in the fact that the distribution of shaft cracks is highly random and the crack shapes are complex. Traditional manual detection technology is inefficient, requires a lot of manpower and time costs, and is prone to missed detection, making it impossible to efficiently detect all cracks. Therefore, it is urgent to develop effective shaft crack detection technology and methods.

[0003] With the continuous development of computer vision technology, more and more image recognition algorithms are being applied to crack detection in bridges, road structures, etc. Although these classic algorithms can achieve good results under good lighting conditions, they often have problems of missed detection and false detection in low-light environments such as coal mine shafts with poor lighting, resulting in low crack recognition accuracy. Due to the limitations of these image recognition algorithms in terms of lighting conditions and application environment, it is difficult to achieve ideal results when directly applied to crack detection in coal mine shafts. Summary of the invention

[0004] The present invention aims to provide a method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber strain field, so as to improve the crack identification accuracy in a coal mine shaft in a non-lighting environment.

[0005] Technical solution: A method for identifying cracks in a coal mine shaft in a light-free environment based on an optical fiber sensing grid, comprising the following steps:

[0006] S1, construct a fiber optic sensor grid for coal mine shaft stress detection, collect the strain data of the optical fiber through a distributed optical fiber demodulator, and obtain the optical fiber strain field data set strain1;

[0007] S2, collect the well wall crack image samples at the optical fiber sensor grid through the CCD industrial camera, and perform binarization processing to establish the well wall crack image data set cracks1;

[0008] S3, interpolating the blank to-be-interpolated area in the optical fiber sensing grid obtained in step S1 according to the well wall crack image data set obtained in step S2, to obtain a two-dimensional strain field of the well wall;

[0009] S4, establishing a wellbore crack recognition model based on deep learning, and associating the optical fiber strain field image features in the optical fiber strain field data set strain1 and the crack image features in the wellbore crack data set cracks1 through training;

[0010] S5. After the model training is completed, the real-time optical fiber strain data can be input into the trained model. Then, only the strain data is needed to detect the cracks on the well wall, and the detected well wall cracks can be visualized to realize the identification and early warning of the well wall cracks.

[0011] Furthermore, in the optical fiber sensing grid constructed in step S1, the sensing layer corresponding to the optical fiber sensing network is located between the surface layer and the base layer, the base layer and the surface layer are composed of polyamide fibers, and the sensing layer is a longitude-latitude grid structure composed of longitudinal and transverse sensing optical fibers. A cavity is left between the base layer and the surface layer, and then the sensing optical fiber is placed in the cavity, and the molten polyamide material is injected into the cavity by injection molding, and after cooling, an overall structure of clamping the sensing optical fiber is formed.

[0012] Furthermore, step S3 specifically includes:

[0013] S31, in the optical fiber strain field data set strain1 collected by the optical fiber sensor grid based on the longitude and latitude grid structure, the blank area between the longitudinal and transverse optical fibers is the point to be interpolated, denoted as x0, and the strain estimation value corresponding to the point to be interpolated is Z * (x0); There are n sampling data points along the sensing optical fiber around the interpolation point x0, denoted as 1, 2..., n, and the distance from x0 is x i , x j The corresponding strain sampling values ​​of the optical fiber sampling points are Z(x i ), Z(x j ), the mean of the strain sampling values ​​of all sampling points is recorded as m;

[0014] S32, strain estimation value Z corresponding to the interpolation point * The calculation formula for (x0) is:

[0015]

[0016] Among them, λ i is assigned to the i-th strain sampling value Z(x i )’s residual weights;

[0017] When the variance of the error of the strain estimate is the smallest, the first-order partial derivative of each weight is calculated and set to 0, and we can get:

[0018]

[0019] Among them, C(...) represents the covariance;

[0020] Solving the problem yields the residual weights λ1,λ2,...λ n Substituting the residual weight into equation (1) we can get the strain estimate Z corresponding to the interpolation point: * (x0).

[0021] Furthermore, step S4 specifically includes:

[0022] S41, interpolate the strain data of the optical fiber sensor grid through step S3 to obtain a continuous strain field, make a distribution diagram of the strain field, and use the four farthest corner points of the upper, lower, left, and right of the strain change influence range as boundaries to determine the area D where the strain change occurs on the optical fiber sensor grid due to the cracking of the well wall i (i=1,2...,n), D i The strain field image d i (i=1,2...,n) is made into optical fiber strain field image data set strain0;

[0023] S42, using a CCD explosion-proof industrial camera, in the area D of the well wall described in step S41 i (i=1,2...,n), take a photo of the cracks in the well wall c i (i=1,2...,n), perform binarization on the collected photos of wellbore cracks, separate the cracks from the wellbore background, mark the cracks in black, and mark the wellbore background in white, and obtain the photos of wellbore cracks after binarization ε i (i=1,2...,n), and establish the crack data set cracks0 to facilitate the subsequent feature extraction and analysis;

[0024] S43, using SIFT algorithm to pick up feature points of corresponding images in the optical fiber strain field image dataset strain0 and the wellbore crack dataset cracks0;

[0025] Spatially register and synchronize the optical fiber strain field image in strain0 with the photos of wellbore cracks in cracks0 to ensure the spatial consistency between the two.

[0026] S44, enhancing the image data sets strain0 and cracks0 by cropping, mirroring, flipping, scaling and translating, etc., to generate the expanded data sets strain1 and cracks1;

[0027] S45. Label the fracture dataset. In the labelimg environment, create a polygonal labeling format to label the well wall fractures in the well wall fracture dataset cracks1 one by one.

[0028] Furthermore, step S5 specifically includes:

[0029] S51, constructing a wellbore crack recognition model, which includes two parallel branches, responsible for processing the optical fiber strain field data set strain1 and the wellbore crack data set cracks1 respectively;

[0030] Each branch consists of multiple downsampling convolutional layers, which gradually extract high-level features of the image. The pre-trained CNN model is used as the basic network to improve the feature extraction capability.

[0031] After the feature extraction of the two branches is completed, their feature maps are spliced ​​in the depth direction to form a new feature map. The splicing operation is performed at a deeper layer of the network to ensure that the fused features are sufficiently abstract.

[0032] S52, performing feature fusion and decoding of two types of data, including:

[0033] After feature fusion, the feature map can be decoded by upsampling through a series of convolutional layers, gradually restoring it to the same spatial resolution as the input image;

[0034] The convolution operation of the decoding layer is designed with a ReLU activation function, and the upsampling operation uses bilinear interpolation;

[0035] S53, define the semantic segmentation output layer, the specific steps include:

[0036] In the last layer of the decoding network, a semantic segmentation output layer is defined to output the segmented crack image;

[0037] After the output layer, a softmax layer is applied to obtain the class probability distribution of each pixel;

[0038] S54. Define the loss function of the model, the specific steps of which include:

[0039] In order to improve the sensitivity of the model to the crack area, the weighted cross entropy loss Weighted CE Loss is used as the loss function, and its expression is:

[0040]

[0041] Where N is the number of samples, w yi It is judged as crack type y i The weight of the crack area is higher than that of the non-crack area, p yi is the model for crack category y i The predicted probability of

[0042] S55, training the established model, the specific steps of which include:

[0043] Use the labeled training data sets strain1 and cracks1 to train the model, ensuring that each sample has a corresponding label, where the crack area is marked as 1 and the non-crack area is marked as 0;

[0044] During the training process, mini-batch gradient descent or Adam optimization algorithm is used to update parameters, and the performance indicators on the validation set are regularly evaluated. Based on the evaluation results, the model parameters and training strategies are adjusted to improve the accuracy and robustness of target detection.

[0045] Beneficial effects: Compared with the prior art, the method of the present invention has substantial characteristics and significant improvements in:

[0046] 1. No need for lighting conditions: After the model training is completed, the method adopted by the present invention does not require external light source assistance when detecting cracks in the shaft wall. The sensing optical fiber can be carried out in the dark environment of the coal mine shaft and is not restricted by lighting conditions, thereby improving the applicability in mine safety detection.

[0047] 2. Improve detection efficiency: The method provided by the present invention can sense the changes of cracks in the well wall through optical fiber in real time. There is no need to take manual inspection methods such as going down the well to take pictures. The data in the optical fiber can be read by an optical fiber demodulator on the ground, which improves the detection efficiency and avoids the safety hazards of manual inspections down the well.

[0048] 3. Reduce false detection rate: By combining strain field data and crack image information, high-precision detection of well wall cracks can be achieved even in the absence of light, reducing the false detection rate of traditional single image detection methods and improving detection reliability.

[0049] 4. Long-term stability: Fiber optic sensors have long-term stability and can be used for a long time in harsh coal mine environments, reducing the frequency of maintenance and replacement.

[0050] 5. Environmental adaptability: The method of the present invention can adapt to the complex environment in the coal mine, including the high humidity, dust and other unfavorable factors underground, which will not affect the data of the optical fiber sensor, thus ensuring the stable operation of the monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a schematic diagram of the optical fiber sensing grid in the present invention;

[0052] Figure 2 It is a schematic diagram of the fixing of the optical fiber sensing grid on the well wall and the strain data collection in the present invention;

[0053] Figure 3 It is a schematic diagram of the interpolation algorithm based on the optical fiber strain field;

[0054] Figure 4 This is a schematic diagram of the binarization processing of the well wall crack photo;

[0055] Figure 5 This is a schematic diagram of the wellbore fracture identification model based on deep learning;

[0056] The reference numerals in the figure are: 101-fiber optic sensing grid, 102-base layer, 103-surface layer, 104-sensing layer, 105-longitudinal sensing fiber, 106-lateral sensing fiber, 107-on the wall of the coal mine shaft, 108-communication optical cable, 109-communication optical cable, distributed optical fiber demodulator, 110-point to be interpolated, 111-optical fiber sampling point 1, 112-optical fiber sampling point 2, 113-strain field distribution map, 114-area where strain changes occur, 115-photo of well wall cracks, 116-photo of well wall cracks after binary processing, 117-model branch, 118-downsampling convolution layer, 119-feature map, 120-upsampling convolution layer, 121-semantic segmentation layer, 122-softmax layer. DETAILED DESCRIPTION

[0057] To explain the technical solution disclosed by the present invention in detail, further introduction is given below in conjunction with the accompanying drawings.

[0058] The present invention provides a method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensor grid, which belongs to the field of coal mine shaft disease detection. The method forms an optical fiber sensor grid with distributed optical fibers, fixes it on the surface of the shaft wall, uses an optical fiber demodulator to collect strains along the optical fiber, and fills the blank blind area of ​​the optical fiber sensor grid through an interpolation algorithm to obtain a strain field image data set of the shaft wall; cooperates with a CCD industrial camera to shoot the shaft wall crack image samples at the optical fiber sensor grid, and performs binarization processing to establish a shaft wall crack image data set. A shaft wall crack recognition model based on deep learning is established, and two parallel feature extraction branches are introduced to receive the optical fiber strain field image data set and the shaft wall crack image data set respectively, and feature extraction and fusion of two different types of data sets are performed through training. Since the optical fiber sensing strain data does not require a lighting environment, after the model training is completed, it is only necessary to input the real-time optical fiber strain data into the trained model to realize the recognition and early warning of shaft wall cracks in a non-lighting environment.

[0059] Example 1

[0060] A method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensing grid comprises the following steps:

[0061] S1. Assemble a fiber optic sensor grid 101, please refer to Figure 1The method comprises a base layer 102 directly in contact with the well wall, a surface layer 103 for protecting the sensing optical fiber, and a sensing layer 104 composed of the sensing optical fiber. The method comprises the following sub-steps:

[0062] S11, wherein the surface layer 103 is at the top of the optical fiber sensor grid 101, the base layer 102 is at the bottom of the optical fiber sensor grid 101, and the sensing layer 104 is placed between the surface layer 103 and the base layer 102;

[0063] S12, the base layer 102 and the surface layer 103 are composed of polyamide fibers, and the sensing layer 104 is a longitude-latitude grid structure composed of longitudinal sensing optical fibers 105 and transverse sensing optical fibers 106. Preferably, the spacing between the longitudinal sensing optical fibers 105 and the transverse sensing optical fibers 106 is 50 mm.

[0064] S13, leaving a cavity between the base layer 102 and the surface layer 103, and then placing the sensing layer 104 in the cavity. Subsequently, using injection molding technology, the molten polyamide material is injected into the cavity, and after cooling, an overall structure of clamping the sensing optical fiber is formed, namely, the optical fiber sensing grid 101.

[0065] S2, please refer to Figure 2 , a fiber optic sensor grid 101 is fixed on the wall 107 of a coal mine shaft, and a communication optical cable 108 is used to connect the sensing optical fiber in the sensing layer 104 of the fiber optic sensor grid 101 and a distributed optical fiber demodulator 109 to collect strain data generated on the fiber optic sensor grid 101;

[0066] S3, exporting the strain data collected by the distributed optical fiber demodulator 109, and interpolating the blank interpolation area in the optical fiber sensor grid 101 according to the collected strain data, to obtain the two-dimensional strain field of the well wall. The method includes the following sub-steps:

[0067] S31, reference Figure 3 , the blank area between the longitudinal and transverse optical fibers is the interpolation point 110, denoted as x0, and the strain estimation value corresponding to the interpolation point 110 is Z * (x0); there are n sampling data points along the sensing optical fiber around the interpolation point 110, which are recorded as 1, 2..., n;

[0068] The distance from x0 is x i is the optical fiber sampling point 111, and the distance from x0 is x j is the optical fiber sampling point 112, and the corresponding strain sampling values ​​are Z(x i ), Z(x j ); the mean of the strain sampling values ​​of all sampling points is recorded as m;

[0069] S32, strain estimation value Z corresponding to the interpolation point 110 *The calculation formula for (x0) is:

[0070]

[0071] Among them, λ i is assigned to the i-th strain sampling value Z(x i ). When the variance of the error of the strain estimate is the smallest, the first-order partial derivative of each weight is calculated and set to 0, and we can get:

[0072]

[0073] Among them, C(...) represents the covariance;

[0074] Solving the problem yields the residual weights λ1,λ2,...λ n Substituting the residual weight into equation (1), we can obtain the strain estimate Z corresponding to the interpolation point 110: * (x0);

[0075] S33. According to steps S31 and S32, strain data is interpolated for all blank areas of the optical fiber sensor grid to obtain a continuous strain field and make a distribution diagram 113 of the strain field.

[0076] S4. Establishing the optical fiber strain field and wellbore fracture image training data set. The specific steps are as follows:

[0077] S41. Reference Figure 3 In the strain field distribution diagram 113, the four most distal corner points of the strain change influence range are used as boundaries to determine the region 114 where strain changes occur on the optical fiber sensor grid due to the cracking of the well wall. The strain field image d in the region 114 is i (i=1,2...,n) is made into optical fiber strain field image data set strain0;

[0078] S42. Reference Figure 4 , using a CCD explosion-proof industrial camera, in the well wall area 114 described in S41, taking a photo 115 of the well wall cracks, performing binarization processing on the collected well wall crack photos, separating the cracks from the well wall background, marking the cracks with black, and marking the well wall background with white, obtaining a binarized well wall crack photo 116, and establishing a crack data set cracks0 based on the binarized crack photo 116 to facilitate subsequent feature extraction and analysis;

[0079] S43, using SIFT algorithm to pick up feature points of each corresponding image in the optical fiber strain field image dataset strain0 and the well wall crack dataset cracks0. Perform spatial registration and synchronization on the optical fiber strain field image in strain0 and the well wall crack photo in cracks0 to ensure spatial consistency between the two.

[0080] S44, enhancing the image data sets strain0 and cracks0 by cropping, mirroring, flipping, scaling and translating, etc., to generate the expanded data sets strain1 and cracks1;

[0081] S45, labeling the fracture data set, creating a polygonal labeling form in the labelimg environment to label the well wall fractures in cracks1 one by one;

[0082] S5. Establish a deep learning-based wellbore crack recognition model, and associate the optical fiber strain field image features in the optical fiber strain field dataset strain1 with the crack image features in the wellbore crack dataset cracks1 through training. The specific steps are as follows:

[0083] S51 . Construct a wellbore crack recognition model. The model includes two parallel branches 117 . Each branch is responsible for processing one type of image data (corresponding to image data sets strain1 and cracks1 ).

[0084] Further, step S51 includes:

[0085] Reference Figure 5 , using the VGG model as the basic network, each branch 117 is composed of multiple downsampling convolutional layers 118, and the high-level features of the image are gradually extracted through the convolutional layers to improve the ability of feature extraction;

[0086] After the feature extraction of the two branches is completed, their feature maps are spliced ​​in the depth direction to form a new feature map 119. The splicing operation is performed at a deeper layer of the network to ensure that the fused features are sufficiently abstract;

[0087] S52, performing feature fusion and decoding of two types of images. Further, step S52 includes:

[0088] After feature fusion, the feature map can be decoded by upsampling through a series of convolutional layers 120, gradually restoring it to the same spatial resolution as the input image;

[0089] The convolution operation of the decoding layer is designed with a ReLU activation function, and the upsampling operation uses bilinear interpolation;

[0090] Step S53: define the semantic segmentation output layer, the specific steps of which include:

[0091] In the last layer of the decoding network, a semantic segmentation output layer 121 is defined to output the segmented crack image;

[0092] After the output layer, a softmax layer 122 is applied to obtain the class probability distribution of each pixel;

[0093] Step S54: define the loss function of the model, the specific steps include:

[0094] In order to improve the sensitivity of the model to the crack area, the weighted cross entropy loss Weighted CE Loss is used as the loss function, and its expression is:

[0095]

[0096] Where N is the number of samples, w yi It is judged as crack type y i The weight of the crack area is higher than that of the non-crack area. yi is the model for crack category y i The predicted probability of .

[0097] Step S55: training the established model, the specific steps of which include:

[0098] Use the labeled training data sets strain1 and cracks1 to train the model. Make sure each sample has a corresponding label, where the crack area is marked as 1 and the non-crack area is marked as 0;

[0099] During the training process, mini-batch gradient descent or Adam optimization algorithm is used to update parameters, and the performance indicators on the validation set are regularly evaluated. Based on the evaluation results, the model parameters and training strategies are adjusted to improve the accuracy and robustness of target detection.

[0100] Step S6: After the model training is completed, the optical fiber strain data collected in real time can be input into the trained model to realize the recognition and early warning of well wall cracks in a dark environment.

[0101] The technical solution described in the above specific implementation steps is a specific implementation of the present invention. Without departing from the concept of the present invention, ordinary technicians in this field can make appropriate deformations and improvements according to specific application requirements, such as using other shapes of distributed optical fiber grids, strain field interpolation algorithms, crack identification models, etc., which all belong to the protection scope of the present invention. At the same time, the steps involved in the present invention can be flexibly adjusted and combined according to the specific application situation to achieve better coal mine shaft crack target detection effect.

Claims

1. A method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensing grid, characterized in that the steps include: S1, construct a fiber optic sensor grid for coal mine shaft stress detection, collect the strain data of the optical fiber through a distributed optical fiber demodulator, and obtain the optical fiber strain field data set strain1; S2, collect the well wall crack image samples at the optical fiber sensor grid through the CCD industrial camera, and perform binarization processing to establish the well wall crack image data set cracks1; S3, interpolating the blank to-be-interpolated area in the optical fiber sensing grid obtained in step S1 according to the well wall crack image data set obtained in step S2, to obtain a two-dimensional strain field of the well wall; S4, establishing a wellbore crack recognition model based on deep learning, and associating the optical fiber strain field image features in the optical fiber strain field data set strain1 and the crack image features in the wellbore crack data set cracks1 through training; S5. After the model training is completed, the real-time optical fiber strain data can be input into the trained model. Then, only the strain data is needed to detect the cracks on the well wall, and the detected well wall cracks can be visualized to realize the identification and early warning of the well wall cracks.

2. The method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensing grid according to claim 1 is characterized in that: In the optical fiber sensing grid constructed in step S1, the sensing layer corresponding to the optical fiber sensing network is located between the surface layer and the base layer, the base layer and the surface layer are composed of polyamide fibers, and the sensing layer is a longitude-latitude grid structure composed of longitudinal and transverse sensing optical fibers. A cavity is left between the base layer and the surface layer, and then the sensing optical fiber is placed in the cavity, and the molten polyamide material is injected into the cavity by injection molding, and after cooling, an overall structure of clamping the sensing optical fiber is formed.

3. The method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensing grid according to claim 1 or 2, characterized in that: Step S3 specifically includes: S31, in the optical fiber strain field data set strain1 collected by the optical fiber sensor grid based on the longitude and latitude grid structure, the blank area between the longitudinal and transverse optical fibers is the point to be interpolated, denoted as x0, and the strain estimation value corresponding to the point to be interpolated is Z * (x0); There are n sampling data points along the sensing optical fiber around the interpolation point x0, denoted as 1, 2..., n, and the distance from x0 is x i , x j The corresponding strain sampling values ​​of the optical fiber sampling points are Z(x i ), Z(x j ), the mean of the strain sampling values ​​of all sampling points is recorded as m; S32, strain estimation value Z corresponding to the interpolation point * The calculation formula for (x0) is: Among them, λ i is assigned to the i-th strain sampling value Z(x i )’s residual weights; When the variance of the error of the strain estimate is the smallest, the first-order partial derivative of each weight is calculated and set to 0, and we can get: Among them, C(...) represents the covariance; Solving the problem yields the residual weights λ1,λ2,...λ n Substituting the residual weight into equation (1) we can get the strain estimate Z corresponding to the interpolation point: * (x0).

4. The method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensing grid according to claim 1 is characterized in that: Step S4 specifically includes: S41, interpolate the strain data of the optical fiber sensor grid through step S3 to obtain a continuous strain field, make a distribution diagram of the strain field, and use the four farthest corner points of the upper, lower, left, and right of the strain change influence range as boundaries to determine the area D where the strain change occurs on the optical fiber sensor grid due to the cracking of the well wall i (i=1,2...,n), D i The strain field image d i (i=1,2...,n) is made into optical fiber strain field image data set strain0; S42, using a CCD explosion-proof industrial camera, in the area D of the well wall described in step S41 i (i=1,2...,n), take a photo of the cracks in the well wall c i (i=1,2...,n), perform binarization on the collected photos of wellbore cracks, separate the cracks from the wellbore background, mark the cracks in black, and mark the wellbore background in white, and obtain the photos of wellbore cracks after binarization ε i (i=1,2...,n), and establish the crack data set cracks0 to facilitate the subsequent feature extraction and analysis; S43, using SIFT algorithm to pick up feature points of corresponding images in the optical fiber strain field image dataset strain0 and the wellbore crack dataset cracks0; Spatially register and synchronize the optical fiber strain field image in strain0 with the photos of wellbore cracks in cracks0 to ensure the spatial consistency between the two. S44, enhancing the image data sets strain0 and cracks0 by cropping, mirroring, flipping, scaling and translating, etc., to generate the expanded data sets strain1 and cracks1; S45. Label the fracture dataset. In the labelimg environment, create a polygonal labeling format to label the well wall fractures in the well wall fracture dataset cracks1 one by one.

5. The method for identifying cracks in a coal mine shaft in a non-lighting environment based on an optical fiber sensing grid according to claim 1 is characterized in that: Step S5 specifically includes: S51, constructing a wellbore crack recognition model, which includes two parallel branches, responsible for processing the optical fiber strain field data set strain1 and the wellbore crack data set cracks1 respectively; Each branch consists of multiple downsampling convolutional layers, which gradually extract high-level features of the image. The pre-trained CNN model is used as the basic network to improve the feature extraction capability. After the feature extraction of the two branches is completed, their feature maps are spliced ​​in the depth direction to form a new feature map. The splicing operation is performed at a deeper layer of the network to ensure that the fused features are sufficiently abstract. S52, performing feature fusion and decoding of two types of data, including: After feature fusion, the feature map can be decoded by upsampling through a series of convolutional layers, gradually restoring it to the same spatial resolution as the input image; The convolution operation of the decoding layer is designed with a ReLU activation function, and the upsampling operation uses bilinear interpolation; S53, define the semantic segmentation output layer, the specific steps include: In the last layer of the decoding network, a semantic segmentation output layer is defined to output the segmented crack image; After the output layer, a softmax layer is applied to obtain the class probability distribution of each pixel; S54. Define the loss function of the model, the specific steps of which include: In order to improve the sensitivity of the model to the crack area, the weighted cross entropy loss Weighted CE Loss is used as the loss function, and its expression is: Where N is the number of samples, w yi It is judged as crack type y i The weight of the crack area is higher than that of the non-crack area, p yi is the model for crack category y i The predicted probability of S55, training the established model, the specific steps of which include: Use the labeled training data sets strain1 and cracks1 to train the model, ensuring that each sample has a corresponding label, where the crack area is marked as 1 and the non-crack area is marked as 0; During the training process, mini-batch gradient descent or Adam optimization algorithm is used to update parameters, and the performance indicators on the validation set are regularly evaluated. Based on the evaluation results, the model parameters and training strategies are adjusted to improve the accuracy and robustness of target detection.

Images