Mine time shift resistivity monitoring data labeling method and system, storage medium and software

Through the mine time-shift resistivity monitoring data annotation model with deep learning technology, the image semantic segmentation network and multi-objective tracking and matching algorithm are used for automatic annotation, which solves the problem of high time-consuming manual annotation in the existing technology, and improves the data annotation efficiency and quality.

CN119068485APending Publication Date: 2024-12-03XIAN RES INST OF CHINA COAL TECH & ENG GRP CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411148474.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The labeling of time-shift resistivity monitoring data of existing mines relies on manual operations, is time-consuming and costly, and is insufficiently intelligent, making it difficult to effectively process and interpret data.

Method used

The mine time-shift resistivity monitoring data annotation model based on deep learning is adopted, including a preselected target area identification model and a preselected target area classification model. The image semantic segmentation network and multi-objective tracking and matching algorithm are used for automatic annotation to reduce manual intervention.

Benefits of technology

Automatic labeling of time-shift resistivity monitoring data in mines is realized, improving data labeling efficiency and quality, and reducing manual labeling costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0005002894860000061
    Figure BDA0005002894860000061
  • Figure BDA0005002894860000062
    Figure BDA0005002894860000062
  • Figure BDA0005002894860000071
    Figure BDA0005002894860000071
Patent Text Reader

Abstract

The invention discloses a mine time shift resistivity monitoring data labeling method and system, a storage medium and software. According to the disclosed scheme, image segmentation is carried out on mine time-shifting resistivity monitoring data by using an image segmentation network; tracking and matching the potential resistivity anomalies of the plurality of frames by using a multi-target tracking and matching algorithm to obtain the resistivity anomalies simultaneously appearing in the front and back frame ranges as pre-selected anomalies; and according to the resistivity statistical data of the whole area, the anomalies are divided into high resistance, low resistance and background. According to the artificial intelligence-based automatic labeling method for the mine time-shifting resistivity monitoring data, automatic pre-labeling can be realized, only manual rechecking is needed, and the data labeling efficiency and quality are greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of mine water prevention and control, and relates to an automatic labeling method for mine time-lapse resistivity monitoring data based on artificial intelligence. Background Art

[0002] Time-lapse resistivity monitoring in mines is a scientific and effective technical means for coal mine geological transparency and water hazard monitoring and early warning. At present, the time-lapse resistivity monitoring data of mines are displayed in the form of resistivity cloud maps, which mainly rely on manual experience for data screening and anomaly identification. The degree of intelligence in data processing and interpretation is insufficient. Therefore, it is urgent to integrate big data mining, cloud computing, artificial intelligence and other technologies to carry out research on the processing and interpretation methods of mine time-lapse resistivity data based on deep learning.

[0003] Artificial intelligence is applied to the processing and interpretation of mine time-lapse resistivity monitoring data, mainly using deep neural networks to achieve intelligent recognition of resistivity anomalies. Deep learning-based resistivity recognition networks are generally trained using supervised learning methods. To obtain a neural network that meets actual needs for robustness and generalization requires a large amount of labeled data. At present, the annotation of time-lapse resistivity data is generated by manual annotation by scientific researchers. This method requires the personnel involved in the annotation to have rich experience in processing mine electrical monitoring data, and manual annotation is particularly time-consuming, and the data annotation cost is extremely high. Summary of the invention

[0004] In view of the defects or shortcomings of the prior art, the present invention first provides a mine time-lapse resistivity monitoring data annotation model.

[0005] The mine time-lapse resistivity monitoring data annotation model provided by the present invention includes a pre-selected target area recognition model and a pre-selected target area classification model;

[0006] The pre-selected target area recognition model includes an image semantic segmentation network and a first image feature extraction network;

[0007] The method for constructing the classification model of the pre-selected target area includes training an initial network using a data set;

[0008] The data set includes a plurality of high-resistance abnormal images, a plurality of low-resistance abnormal images and a plurality of non-abnormal images, and the data set acquisition method includes:

[0009] Step (1) extracts the resistivity values ​​at different x spatial positions in the mine time-lapse resistivity monitoring data, and normalizes the resistivity values ​​at the same x spatial coordinate position. n All resistivity values ​​at the spatial coordinate position are normalized to obtain x = x n Normalized resistivity sequence R corresponding to the spatial coordinate positionscale,n ; n=1,2,3,…,N,N≥2, and x n With x n+1 The x-direction spatial position of the mine is adjacent; the mine time-lapse resistivity monitoring data includes several data collections of the mine area, and each data collection obtains the resistivity value under the spatial coordinates (x, y, z);

[0010] Step (2), visualize the N normalized resistivity series respectively, and obtain the corresponding atlas Img = {img1, img2, img3, ...img n ,…,img N}, img n For x = x n Resistivity plane contour map corresponding to the spatial coordinate position; n = 1, 2, 3, ..., N;

[0011] Step (3), manually marking and cutting each resistivity plane contour map obtained in step (2), obtaining a high-resistance abnormal image, a low-resistance abnormal image and / or a non-abnormal image in each resistivity plane contour map, and forming a training data set;

[0012] The initial network structure includes a second image feature extraction network, an image feature embedding vector layer, a linear layer and a classification layer connected in sequence; the image feature extraction network is used to extract the image features of each resistivity plane contour map and output the corresponding feature map; the image feature embedding vector layer flattens each feature map and then adds all the flattened feature maps to obtain a feature embedding vector; the linear layer purifies the feature embedding vector, and the classification layer outputs a category according to the purified feature embedding vector;

[0013] The second image feature extraction network has the same structure and network parameters as the first image feature extraction network, and the network parameters of the second image feature extraction network are frozen when the initial network is trained using the data set.

[0014] An optional solution is that the image semantic segmentation network adopts a segmentation model Segment Anything Model.

[0015] An optional solution is that the first image feature extraction network is an Encoder-Decoder model.

[0016] An optional solution is that the normalization of the resistivity values ​​at the same x spatial coordinate position in step (1) includes: first performing mean-variance normalization on multiple resistivity values, and then performing maximum normalization on the mean-variance normalized values; x = x n All the normalized values ​​corresponding to the spatial coordinate positions constitute x=x n Normalized resistivity sequence R corresponding to the spatial coordinate positionscale,n .

[0017] An optional solution is that the step (2) of respectively visualizing the N normalized resistivity sequences corresponding to the m-th monitoring data includes: for any x=x n Normalized resistivity sequence R corresponding to the spatial coordinate position scale,n Gray value conversion, interpolation processing, edge expansion to adjust image size and colorization are performed in sequence.

[0018] The present invention also provides a method for labeling mine time-lapse resistivity monitoring data, wherein the mine time-lapse resistivity monitoring data includes several data collections of a mine area to be monitored, and each data collection obtains a resistivity value at a spatial coordinate (x, y, z); the method comprises: using the method described in steps (1) and (2) above to obtain a resistivity plane contour map atlas of the corresponding monitoring data of the mine area to be monitored; and then labeling the monitoring data using steps 3-6;

[0019] Step 3, using the image semantic segmentation network to segment each frame image in the atlas, and obtain the segmentation mask of each resistivity plane contour map, each segmentation mask contains at least one irregular anomaly; the minimum circumscribed rectangle coordinates, center point coordinates, mask area and external contour points of each irregular anomaly in each frame segmentation mask are obtained, wherein the minimum circumscribed rectangle frame of each irregular anomaly constitutes the detection frame of the corresponding frame segmentation mask;

[0020] Step 4, using the first image feature extraction network to obtain a feature map of each frame image in the atlas;

[0021] Step 5: Based on the position information of each detection frame in the corresponding segmentation mask and the corresponding feature map, a multi-target tracking matching algorithm is used to match the detection frames in the segmentation masks of the previous and next frames. The positions of the detection frames that match each other in the previous and next frames belong to the same pre-selected target area; wherein the position of the detection frame in the segmentation mask of the first frame is used as the initial pre-selected target area, and the unmatched detection frame that appears after the second frame is used as a newly added pre-selected target area; then, the pre-selected target area in each resistivity plane contour map is obtained; the pre-selected target area in each resistivity plane contour map is cropped, and the cropped area belonging to the same pre-selected target area is used as the cropped atlas corresponding to the corresponding pre-selected target area;

[0022] Step 6, inputting the cropped atlas belonging to the same pre-selected target area into the pre-selected target area classification model of claim 1, outputting the anomaly category of the corresponding pre-selected target area, and labeling the corresponding area in the corresponding resistivity plane contour map according to the category.

[0023] An optional solution is that in step 3, a connected domain obtaining algorithm and a contour line obtaining algorithm are used to obtain the minimum circumscribed rectangle coordinates, center point coordinates, mask area and external contour points of each irregular anomaly in each frame segmentation mask.

[0024] An optional solution is that step 5 includes:

[0025] 5.1 According to the position of the detection frame in each frame segmentation mask obtained in step 3, predict the position of each detection frame in each frame segmentation mask in the next frame segmentation mask, and obtain the predicted frame in each segmentation mask after the second frame;

[0026] 5.2 According to the detection box position in each frame segmentation mask obtained in step 3, the real feature representation f in the corresponding frame feature map is obtained through position mapping t ; According to the prediction box in each frame segmentation mask obtained in step 5.1, the prediction feature representation f in the corresponding frame feature map is obtained by position mapping p ;

[0027] 5.3 For the segmentation masks and feature maps of each frame after the second frame: calculate the distance and intersection-over-union ratio between each prediction box and each detection box in the same segmentation mask, and calculate the real feature representation f in the same feature map t With each prediction feature representation f p similarity;

[0028] 5.4 For the segmentation masks of each frame after the second frame, select the detection frames that match each prediction frame in the same segmentation mask according to the distance, intersection-over-union ratio and similarity calculated in step 5.3. The detection frame positions corresponding to the same prediction frame in the previous and next frames belong to the same pre-selected target area. The detection frame position in the segmentation mask of the first frame is used as the initial pre-selected target area, and the unmatched detection frames that appear after the second frame are used as a new pre-selected target area.

[0029] An optional solution is to use the Kalman filter algorithm in step 5.1 to predict the position of each detection box in each frame segmentation mask in the next frame segmentation mask.

[0030] An optional solution is to set the weight ratio of distance, intersection-over-union ratio and similarity in step 5.4, then calculate the product of the weight ratio of the three and the corresponding parameter value, and filter the detection boxes that match each prediction box in the same segmentation mask according to the sum of the three product values.

[0031] The present invention also provides a corresponding mine time-lapse resistivity monitoring data annotation system, which includes a data visualization module, a pre-selected target area annotation model and a pre-selected target area classification module; the data visualization module is used for the above steps (1) and (2); the pre-selected target area annotation module is used to execute the above steps 3-5; the data visualization module is used to execute the above step 6.

[0032] The present invention also provides a corresponding storage medium on which a computer program / instruction is stored, characterized in that the steps of the above method are implemented when the computer program / instruction is executed by a processor.

[0033] The present invention also provides a corresponding software product, including a computer program / instruction, which implements the steps of the above method when executed by a processor.

[0034] The present invention uses an image segmentation network to perform image segmentation on the mine time-lapse resistivity monitoring data to obtain the outline of each potential resistivity anomaly; uses a multi-target tracking method to track and match the potential resistivity anomalies of multiple frames, and obtains the resistivity anomalies that appear simultaneously in the range of several frames before and after as pre-selected anomalies; and divides the anomalies into three categories: high resistance, low resistance, and background (non-abnormal) according to the resistivity statistical data of the entire area. Further, the pre-labeled results can be manually reviewed by scientific researchers.

[0035] The present invention can realize automatic labeling of mine time-lapse resistivity monitoring data, and only requires manual review, which greatly improves the efficiency and quality of data labeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A SegNet network structure diagram for image feature extraction applicable to the present invention;

[0037] Figure 2 Schematic diagram of a visualized image set in an embodiment of the present invention;

[0038] Figure 3 for Figure 2 Segmentation results of the image set;

[0039] Figure 4 for Figure 3 Multi-target tracking results in ;

[0040] Figure 5 for Figure 4 The classification and labeling results of the pre-selected target areas ID1-3 are shown in FIG. DETAILED DESCRIPTION

[0041] Unless otherwise specified, the science and technology in this article are based on common knowledge in the relevant fields.

[0042] For the data coordinates selected by the present invention, in physical space, x generally represents the excavation direction of the mine tunnel, y coordinate represents the width direction of the tunnel, and z coordinate represents the height direction of the tunnel.

[0043] The present invention normalizes the resistivity values ​​at the same x spatial coordinate position. In a specific solution, the existing normalization method can be used to normalize the original resistivity value to a value of [0, 1]. In a specific example, the original data can be firstly normalized by mean variance, and then the mean variance normalized value can be normalized by the maximum value. The specific calculation process is as follows:

[0044] Step 1.1: Extract x=x from the mth monitoring data n All resistivity values ​​at spatial locations; n = 1, 2, 3, ..., N;

[0045] Step 1.2, find x=x n The mean value of all resistivity values ​​at a spatial position is R mean,n and variance R σ,n , n = 1, 2, 3, ..., N;

[0046] Step 1.3: According to formula (1), x = x n Each resistivity value at a spatial position is normalized by mean variance; n = 1, 2, 3, ..., N;

[0047]

[0048] In the formula, r xn For x = x n Any resistivity value at a spatial position, R xn For r xn The mean-variance normalized value of ;

[0049] Step 1.4: Count the x=x obtained in step 1.3 n The maximum value R among the normalized values ​​of the resistivity values ​​at the spatial position xn,max and the minimum value R xn,min , according to formula (2), the mean variance normalized values ​​obtained in step 1.3 are normalized to the maximum value;

[0050]

[0051] R xn,scale For R xn The maximum normalized value of

[0052] x=x n All the normalized values ​​corresponding to the spatial positions constitute x=x n Normalized resistivity series R corresponding to the spatial position scale,n .

[0053] The image semantic segmentation network of the present invention is used to perform resistivity anomaly segmentation on the resistivity plane contour map to obtain a segmentation mask containing irregular anomalies. The existing or improved related functional networks based on the existing technology can be selected for implementation. An example of an image semantic segmentation network uses the Segment Anything Model (SAM) pre-trained with open source data as the segmentation network. The Image Encoder of the network uses VIT-H, removes the original promote branch of the SAM network, and only retains the three modules of Image Encoder, Image embedding, and Mask Encoder. When working, the resistivity plane contour map is input into the SAM segmentation network frame by frame, and each frame of the image is segmented according to the Segment everything mode to obtain a segmentation mask for each frame.

[0054] The first image feature extraction network and the second image feature extraction network of the present invention are both used to extract the feature map of the resistivity plane contour map, and the structures and network parameters of the two are the same. The existing or improved related functional networks based on the existing technology can be selected for implementation. A specific example is the SegNet neural network pre-trained with the Cityscapes data set. More specifically, as shown in the attached Figure 3 The symmetrical SegNet network shown in the figure adopts the Encoder-Decoder model: the left side is the encoder, which uses a two-dimensional convolution layer conv Layer with a convolution kernel of (3*3) and a maximum pooling layer Maxpool to downsample the input four times to achieve feature extraction; the right side is the decoder Decoder, which uses deconvolution to upsample the encoded feature map and concatenates it with the feature map on the left through skip connections to achieve feature fusion of different levels; the last layer is the classification layer, which uses a convolution with a convolution kernel size of 1*1 to map the feature representation to the category space, and obtains the probability of each pixel belonging to each category through the normalized exponential function shown in formula (3).

[0055]

[0056] Where C is the number of categories, and the value of C is 3; z is the output before the classification layer (D is the maximum value of z), z i is the output of the i-th neuron in the previous layer of the classification layer; z c is the output corresponding to category c (c=1, 2, 3);

[0057] On the Cityscapes dataset Figure 3The network shown in the figure is pre-trained to obtain an image feature extraction network suitable for the present invention (where the output before the classification layer is a feature map). During pre-training, the number of output categories is 19; the pre-training learning rate is 0.01, and the learning rate is reduced by 10 times at 20, 40, and 50 epochs, and the final learning rate is 10 -5 , and trained for a total of 60 epochs.

[0058] The pre-selected target area classification model of the present invention is constructed on the basis of the first image feature extraction network, forming a network structure including the second image feature extraction network, the image feature embedding vector layer, the linear layer and the classification layer. Figure 3 Based on the network shown in the figure, the feature map output before the last classification layer of the SegNet network is flattened into a vector, and multiple frames of cropped images of the pre-selected target area are respectively input into the backbone of the anomaly classification network to obtain the feature map of each frame image, which are then summed up to obtain a feature embedding vector; then a linear layer with a 64-dimensional output and a linear layer with a 3-dimensional output are connected in sequence; finally, the normalized exponential function shown in formula (4) is used to obtain the probability that each group of anomalies belongs to each category.

[0059] The high-resistance abnormal image, low-resistance abnormal image and non-abnormal image described in the present invention are annotated by the relative high and low resistivity statistics in the image, and in principle, they are annotated according to the high, medium and low statistical values. The manual annotation can also be combined with the knowledge and experience of the technicians.

[0060] The present invention is further explained in detail below with reference to the embodiments, but the present invention is not limited thereto.

[0061] Example:

[0062] This embodiment automatically pre-marks the inversion results of data obtained from a certain monitoring by an underground time-lapse resistivity monitoring system of a working face in a mining area.

[0063] This embodiment assumes that multiple data collection is performed each time monitoring, and the interval between adjacent data collections is 100 meters in the x direction, 20 meters in the y direction, and 10 meters in the z direction; there are 7 sampling points in the x direction (equivalent to 7 YZ slices), that is, N=7; the number of sampling points in the y direction is 31, and the number of sampling points in the z direction is 11;

[0064] The method of the present invention is used to automatically pre-label the monitoring data:

[0065] According to steps 1 and 2 of the present invention, the monitoring data is subjected to the above normalization method and visualization to obtain a visualized atlas corresponding to N=7 YZ slices, as shown below: Figure 2 As shown; the visualization method of this embodiment is as follows:

[0066] (1.1) Map the normalized resistivity data from the interval [0,1] to the grayscale value interval [0,255], and round down the data to obtain the corresponding gray value result;

[0067] (1.2) Perform bicubic spline interpolation on the gray value result, and the interpolation target size (w, h) is:

[0068]

[0069] where y n 、z n are the number of sampling points in the y and z directions respectively, and Δy and Δz are the sampling intervals in the y and z directions respectively.

[0070] (1.3) The interpolated data is expanded to become a square, the size of the square is dsize = max(w, h), the size of the left and right sides that need to be padded is left = (dsize-w) / / 2, the size of the top and bottom sides that need to be padded is top = (dsize-h) / / 2, and the constant 0 is used to fill the padded area;

[0071] (1.4) Perform bicubic spline interpolation again on the square data after edge filling, and the interpolation target size is (224, 224);

[0072] (1.5) OpenCV is used to visualize the slice data with an interpolated size of (224, 224) and a value in the interval [0, 255] to obtain an 8-bit grayscale image of each slice; the grayscale image is then mapped to an RGB three-channel color image to obtain the corresponding resistivity plane contour map.

[0073] This embodiment uses the above-mentioned Segment Anything Model (SAM) as the image semantic segmentation network and adopts Figure 1 The network shown is used as an image feature extraction network and a pre-selected target region classification network built on its basis;

[0074] Using Segment Anything Model Figure 2 The segmentation result is as follows Figure 3 As shown; this embodiment adopts a connected domain obtaining algorithm and an isoline obtaining algorithm (specifically, the connected domain obtaining algorithm is used to obtain the mask of each connected domain on each frame segmentation mask; and then the isoline obtaining algorithm is used to obtain the coordinates of the external contour points of the mask of each connected domain) to obtain the minimum circumscribed rectangle coordinates, center point coordinates, mask area and external contour points of each irregular anomaly in each frame segmentation mask;

[0075] right Figure 3 The multiple segmentation results of the seven consecutive frames shown in FIG. 1 are used for multi-target tracking, and the following is obtained: Figure 4The four potential pre-selected target areas ID1-ID4 are shown; the multi-target tracking algorithm is:

[0076] 5.1 According to the position of the detection frame in each frame segmentation mask obtained in step 3, the Kalman filter algorithm is used to predict the position of each detection frame in each frame segmentation mask in the next frame segmentation mask, and the predicted frame in each segmentation mask after the second frame is obtained;

[0077] 5.2 According to the detection box position in each frame segmentation mask obtained in step 3, the real feature representation f in the corresponding frame feature map is obtained through position mapping t ; According to the prediction box in each frame segmentation mask obtained in step 5.1, the prediction feature representation f in the corresponding frame feature map is obtained through position mapping p ;

[0078] 5.3 For the segmentation masks and feature maps of each frame after the second frame: calculate the distance and intersection-over-union ratio between each prediction box and each detection box in the same segmentation mask, and calculate the real feature representation f in the same feature map t With each prediction feature representation f p similarity;

[0079] 5.4 For the segmentation masks of each frame after the second frame, the detection frames that match each prediction frame are selected in the same segmentation mask according to the distance, intersection-over-union ratio and similarity calculated in step 5.3, and the detection frame positions corresponding to the same prediction frame in the previous and next frames belong to the same pre-selected target area; the detection frame position in the segmentation mask of the first frame is used as the initial pre-selected target area, and the unmatched detection frame that appears after the second frame is used as a new pre-selected target area; this step sets the weight ratios of distance, intersection-over-union ratio and similarity to 0.3, 0.3 and 0.4 respectively, and then obtains the product value of the weight ratio of the three and the corresponding parameter value, and selects the detection frame that matches each prediction frame in the same segmentation mask according to the sum of the three product values ​​(specifically, the detection frame with the largest sum of product values ​​is selected as the detection frame that matches the corresponding prediction frame); in this embodiment, the Mahalanobis distance is used to evaluate the distance between the prediction frame and the detection frame; the cosine distance is used as the true feature value f t and the predicted feature value f p similarity;

[0080] Afterwards Figure 3 The four potential pre-selected targets ID1-ID4 are classified and the following is obtained: Figure 5 As shown in the results, after classification, only two groups of anomalies that appeared simultaneously on 7 consecutive frames of slices were retained from the original four pre-selected targets. Based on the abnormal resistivity values, ID1 was determined as a low-resistance anomaly, ID3 was determined as a high-resistance anomaly, and ID2 and ID4 were non-abnormal. This is consistent with the results of manual segmentation and classification.

Claims

1. A mine time-lapse resistivity monitoring data annotation model, characterized in that: The mine time-lapse resistivity monitoring data annotation model includes a pre-selected target area recognition model and a pre-selected target area classification model; The pre-selected target area recognition model includes an image semantic segmentation network and a first image feature extraction network; The method for constructing the classification model of the pre-selected target area includes training an initial network using a data set; The data set includes a plurality of high-resistance abnormal images, a plurality of low-resistance abnormal images and a plurality of non-abnormal images, and the data set acquisition method includes: Step (1) extracts the resistivity values ​​at different x spatial positions in the mine time-lapse resistivity monitoring data, and normalizes the resistivity values ​​at the same x spatial coordinate position. n All resistivity values ​​at the spatial coordinate position are normalized to obtain x = x n Normalized resistivity series R corresponding to the spatial coordinate position scale,n ; n=1,2,3,…,N,N≥2, and x n With x n+1 The x-direction spatial position of the mine is adjacent; the mine time-lapse resistivity monitoring data includes several data collections of the mine area, and each data collection obtains the resistivity value under the spatial coordinates (x, y, z); Step (2), visualize the N normalized resistivity series respectively, and obtain the corresponding atlas Img = {img1, img2, img3, ...img n ,…,img N }, img n For x = x n Resistivity plane contour map corresponding to the spatial coordinate position; n = 1, 2, 3, ..., N; Step (3), manually marking and cutting each resistivity plane contour map obtained in step (2), obtaining a high-resistance abnormal image, a low-resistance abnormal image and / or a non-abnormal image in each resistivity plane contour map, and forming a training data set; The initial network structure includes a second image feature extraction network, an image feature embedding vector layer, a linear layer and a classification layer connected in sequence; the image feature extraction network is used to extract the image features of each resistivity plane contour map and output the corresponding feature map; the image feature embedding vector layer flattens each feature map and then adds all the flattened feature maps to obtain a feature embedding vector; the linear layer purifies the feature embedding vector, and the classification layer outputs a category according to the purified feature embedding vector; The second image feature extraction network has the same structure and network parameters as the first image feature extraction network, and the network parameters of the second image feature extraction network are frozen when the initial network is trained using the data set.

2. The mine time-lapse resistivity monitoring data annotation model according to claim 1 is characterized in that: The image semantic segmentation network adopts the segmentation model Segment Anything Model.

3. The mine time-lapse resistivity monitoring data annotation model according to claim 1 is characterized in that: The first image feature extraction network Encoder-Decoder model.

4. The mine time-lapse resistivity monitoring data annotation model according to claim 1 is characterized in that: Step (1) of normalizing the resistivity values ​​at the same x spatial coordinate position includes: first performing mean-variance normalization on multiple resistivity values, and then performing maximum normalization on the mean-variance normalized values; x = x n All the normalized values ​​corresponding to the spatial coordinate positions constitute x=x n Normalized resistivity series R corresponding to the spatial coordinate position scale,n .

5. The mine time-lapse resistivity monitoring data annotation model according to claim 1 is characterized in that: Step (2) of respectively visualizing the N normalized resistivity sequences corresponding to the m-th monitoring data includes: for any x=x n Normalized resistivity series R corresponding to the spatial coordinate position scale,n Gray value conversion, interpolation processing, edge expansion to adjust image size and colorization are performed in sequence.

6. A method for labeling mine time-lapse resistivity monitoring data, wherein the mine time-lapse resistivity monitoring data includes several data collections of the mine area to be monitored, and each data collection obtains the resistivity value under the spatial coordinates (x, y, z); characterized in that: The method comprises: using the method described in steps (1) and (2) of claim 1 to obtain a resistivity plane contour map atlas of monitoring data of the corresponding mine area to be monitored; and then using steps 3-6 to annotate the monitoring data; Step 3, using the image semantic segmentation network to segment each frame image in the atlas, and obtain the segmentation mask of each resistivity plane contour map, each segmentation mask contains at least one irregular anomaly; the minimum circumscribed rectangle coordinates, center point coordinates, mask area and external contour points of each irregular anomaly in each frame segmentation mask are obtained, wherein the minimum circumscribed rectangle frame of each irregular anomaly constitutes the detection frame of the corresponding frame segmentation mask; Step 4, using the first image feature extraction network to obtain a feature map of each frame image in the atlas; Step 5: Based on the position information of each detection frame in the corresponding segmentation mask and the corresponding feature map, a multi-target tracking matching algorithm is used to match the detection frames in the segmentation masks of the previous and next frames. The positions of the detection frames that match each other in the previous and next frames belong to the same pre-selected target area; wherein the position of the detection frame in the segmentation mask of the first frame is used as the initial pre-selected target area, and the unmatched detection frame that appears after the second frame is used as a newly added pre-selected target area; then, the pre-selected target area in each resistivity plane contour map is obtained; the pre-selected target area in each resistivity plane contour map is cropped, and the cropped area belonging to the same pre-selected target area is used as the cropped atlas corresponding to the corresponding pre-selected target area; Step 6, inputting the cropped atlas belonging to the same pre-selected target area into the pre-selected target area classification model of claim 1, outputting the anomaly category of the corresponding pre-selected target area, and labeling the corresponding area in the corresponding resistivity plane contour map according to the category.

7. The mine time-lapse resistivity monitoring data labeling method according to claim 6, characterized in that: Step 3 uses the connected domain finding algorithm and the contour finding algorithm to find the minimum circumscribed rectangle coordinates, center point coordinates, mask area and external contour points of each irregular anomaly in each frame segmentation mask.

8. The mine time-lapse resistivity monitoring data labeling method according to claim 6, characterized in that: The step 5 comprises: 5.1 According to the position of the detection frame in each frame segmentation mask obtained in step 3, predict the position of each detection frame in each frame segmentation mask in the next frame segmentation mask, and obtain the predicted frame in each segmentation mask after the second frame; 5.2 According to the detection box position in each frame segmentation mask obtained in step 3, the real feature representation f in the corresponding frame feature map is obtained through position mapping t ; According to the prediction box in each frame segmentation mask obtained in step 5.1, the prediction feature representation f in the corresponding frame feature map is obtained through position mapping p ; 5.3 For the segmentation masks and feature maps of each frame after the second frame: calculate the distance and intersection-over-union ratio between each prediction box and each detection box in the same segmentation mask, and calculate the real feature representation f in the same feature map t With each prediction feature representation f p similarity; 5.4 For the segmentation masks of each frame after the second frame, select the detection frames that match each prediction frame in the same segmentation mask according to the distance, intersection-over-union ratio and similarity calculated in step 5.

3. The detection frame positions corresponding to the same prediction frame in the previous and next frames belong to the same pre-selected target area. The detection frame position in the segmentation mask of the first frame is used as the initial pre-selected target area, and the unmatched detection frames that appear after the second frame are used as a new pre-selected target area.

9. The mine time-lapse resistivity monitoring data labeling method according to claim 8, characterized in that: In step 5.1, the Kalman filter algorithm is used to predict the position of each detection box in each frame segmentation mask in the next frame segmentation mask.

10. The mine time-lapse resistivity monitoring data labeling method according to claim 8, characterized in that: In step 5.4, by setting the weight ratio of distance, intersection-over-union ratio and similarity, and then calculating the product of the weight ratio of the three and the corresponding parameter value, the detection box that matches each prediction box is screened in the same segmentation mask according to the sum of the three product values.

11. A mine time-lapse resistivity monitoring data annotation system, characterized in that: It includes a data visualization module, a pre-selected target area labeling model and a pre-selected target area classification module; the data visualization module is used to execute steps (1) and (2) described in claim 1; the pre-selected target area labeling module is used to execute steps 3-5 described in claim 8; the data visualization module is used to execute step 6 described in claim 8.

12. A storage medium, characterized in that: A computer program / instruction is stored thereon, wherein the computer program / instruction, when executed by a processor, implements the steps of the method described in any one of claims 6 to 10.

13. A software product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method described in any one of claims 6 to 10 are implemented.