DAS microseismic event detection method based on segmentation model and connected domain generation algorithm

Abstract

The present invention discloses a DAS microseismic event detection method based on a segmentation model and a connected domain generation algorithm, comprising the steps of: constructing a data set D based on DAS microseismic data containing microseismic events in different scenarios; selecting a semantic segmentation model, and constructing its loss function Loss; using the data set D to minimize the Loss to train the semantic segmentation model, and obtaining a microseismic event segmentation model; generating a prediction frame of a sample, and calculating the confidence score of the prediction frame; calculating the optimal confidence threshold and the optimal area threshold of the sample; and obtaining the DAS microseismic data to be detected for detection. The detection result of the present invention can provide the location information of the DAS microseismic event, facilitate users to locate the microseismic event more quickly and conveniently, and can provide more accurate detection results, reduce the missed detection rate and the false detection rate, improve the detection accuracy of the DAS microseismic event and the robustness of the model, and provide stronger interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a DAS microseismic event detection method based on a segmentation model and a connected domain generation algorithm. Background Art

[0002] Microseismicity is of great significance in revealing geological structures. In addition to natural microseismicity, many engineering activities can also induce microseismicity events. For example, the natural gas collection process of "strong injection and strong production, alternating injection and production" in gas reservoir-type gas storage will cause the geological structure of the gas storage to become unstable and produce microseismicity events. Monitoring these microseismicity events is very important for evaluating the gas tightness and geological stability of the gas storage. At the same time, many geological disasters (landslides, mudslides, etc.) are often accompanied by tiny geological movements in the early stages of their triggering, which will stimulate microseismicity to transmit outward. Using DAS (Distributed acoustic sensing) to monitor microseisms can provide early warnings before landslides and rock pile disasters occur, thereby avoiding losses caused by geological disasters. However, the DAS microseismic detection method has the following problems:

[0003] (1) There are a lot of noisy data in DAS seismic data, and the noise components are complex, such as random noise, traffic noise, optical cable coupling noise, optical noise, etc., which brings great challenges to the detection of DAS microseismic events. Therefore, it is particularly important to quickly and accurately detect microseismic events from DAS data in complex noise scenes.

[0004] (2) Existing DAS microseismic detection methods include the traditional STA / LTA algorithm, CNN-based DAS microseismic detection model, DAS microseismic detection model based on target detection models such as YOLO or Retinanet, DAS microseismic detection model based on template matching technology, and DAS microseismic detection model based on deep residual neural network. Although the above methods can detect DAS microseismic events to a certain extent, classification detection models based on CNN and deep residual network can better classify DAS microseismic events, but they cannot provide detailed location information of DAS microseismic events. Target detection models based on YOLO v3 can provide classification and location information at the same time, but their false detection and missed detection rates are large and their interpretability is poor. The DAS microseismic detection model based on the STA / LTA algorithm has a fast operation speed, but its detection accuracy still needs to be improved.

[0005] IIoU: Improved Intersection of Union, this article refers to the improved intersection of union ratio. TP, FP, TN, and FN are commonly used technical names, respectively: TP: True Positive, the classifier predicts a positive sample, and it is actually a positive sample, that is, the number of positive samples correctly identified. FP: False Positive, the classifier predicts a positive sample, and it is actually a negative sample, that is, the number of false positive negative samples. TN: True Negative, the classifier predicts a negative sample, and it is actually a negative sample, that is, the number of negative samples correctly identified. FN: False Negative, the classifier predicts a negative sample, and it is actually a positive sample, that is, the number of false positive samples. Summary of the invention

[0006] The purpose of the present invention is to provide a DAS microseismic event detection method based on a segmentation model and a connected domain generation algorithm to solve the above problems, which can simultaneously provide positioning information and maintain accuracy, and at the same time have good interpretability of DAS microseismic event detection results.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a DAS microseismic event detection method based on a segmentation model and a connected domain generation algorithm comprises the following steps;

[0008] S1, construct data set D;

[0009] S11, obtain DAS microseismic data containing microseismic events in N different scenarios, and obtain N normalized microseismic data after normalization, which are marked as D1~D N , where the nth normalized microseismic data is D n , 1≤n≤N, each normalized microseismic data includes C×T data points, C is the number of channels, and T is the number of time points;

[0010] S12, to D n Perform semantic segmentation annotation and generate target boxes;

[0011] D n Inner channel c, data point x at time point t ct The label is y ct , if x ct Located in the microseismic event area, then y ct =1, otherwise y ct = 0, all y ct Constitute the first labeled sample D n1 , 1≤c≤C, 1≤t≤T; then based on D n1 Generate a target frame for each microseismic event and obtain the second annotation sample D containing the target frame n2;

[0012] S13, generate D based on breadth-first search algorithm n The distance weight matrix A n , A n The size is C×T, and the element in the cth row and tth column is a ct ;

[0013] S14, D n1 , D n2 , A n As Tags n , generating D n The corresponding sample ;

[0014] S15, process D1~D in sequence according to S12~S14 N , generating D1~D N The corresponding sample ~ , forming a data set ;

[0015] S2, select a semantic segmentation model and construct a loss function Loss, the semantic segmentation model is used to input samples and output data points x in the samples. ct is the predicted probability of microseismic events p ct ;

[0016] ,

[0017] S3, use the dataset D to minimize the loss to train the semantic segmentation model and obtain the microseismic event segmentation model;

[0018] S4, generated The prediction box is calculated, and the confidence score of the prediction box is calculated, including steps S41-S42;

[0019] S41, according to p ct Generate a segmentation mask map, and use the connected domain generation algorithm to obtain the circumscribed rectangle of each connected domain as the prediction box;

[0020] S42, calculate the confidence score conf of each prediction box;

[0021] S5, calculation sample The optimal confidence threshold and the optimal area threshold of the method include steps S51 to S56;

[0022] S51, preset the number of iterations and the confidence threshold Th for each iteration conf , area threshold Th area ;

[0023] S52, in one iteration, according to the corresponding area threshold Th area , delete the area less than Th in step S42 area The prediction box of

[0024] S53, calculate each remaining prediction box and D n2 The improved intersection-over-union (IIoU) of the corresponding target box is calculated, and the remaining prediction boxes are numbered, where the conf and IIoU of the jth remaining prediction box are marked as conf j and IIoU j ;

[0025] S54, count TP, FP and FN according to IIoU and conf of the remaining prediction boxes;

[0026] For the jth remaining prediction box, if conf j ≥Th conf IIoU j ≥0.5, then TP increases by 1; if conf j ≥Th area IIoU j <0.5, then FP increases by 1; if conf j <Th area And IIoU j ≥0.5, then FN increases by 1; otherwise, the prediction box is skipped;

[0027] S55, calculate the F1 score based on TP, FP and FN;

[0028] S56, repeating steps S52 to S55 until the iteration ends, selecting the confidence threshold and area threshold when the F1 score is the largest as the optimal confidence threshold and optimal area threshold;

[0029] S6, obtaining the DAS microseismic data to be detected for detection;

[0030] S61, obtaining the DAS microseismic data to be detected, normalizing it, and sending it to the microseismic event segmentation model obtained in S3, and outputting the prediction result;

[0031] S62, generate prediction boxes and confidence scores of prediction boxes according to steps S41-S42, and then delete prediction boxes whose areas are smaller than the optimal area threshold according to the optimal area threshold.

[0032] Preferably, the normalization process is to normalize the DAS microseismic data channel by channel, and is calculated according to the following formula:

[0033] ,

[0034] In the formula, is the value of channel c and time point t in DAS microseismic data, Max(x) and Min(x) are The maximum and minimum values ​​on the channel, 1≤c≤C, 1≤t≤T.

[0035] As a preferred method, the DAS microseismic data is normalized channel by channel before performing data enhancement operation.

[0036] As a preferred embodiment: S13 specifically includes steps Sa1 to Sa4;

[0037] Sa1, preset an initial distance weight matrix A 1n , the size is C×T, if y ct =1, then A 1n The element a in row c and column t ct =0, if y ct =0,a ct =1500;

[0038] Sa2, with a ct = 0 as the starting point, traverse A using the breadth-first search algorithm 1n For each element in , get the distance value between each element and the starting point, and use the distance value to update the value of the element. If an element is traversed by multiple starting points and each gets a distance value, then use the minimum distance value to update the value of the element;

[0039] Sa3, A 1n Each element in is reduced by 1000 times to obtain the intermediate distance weight matrix A 2n ;

[0040] Sa4, modified A 2n The value of the element in A 2n Middle ct =0, change it to A 2n The maximum value of the elements in D n The distance weight matrix A n .

[0041] Preferably, the semantic segmentation model is UNet, FCN, Deeplabv3, Deeplabv3+ and PSPNet.

[0042] As a preferred embodiment: In S41, according to p ct Generate a segmentation mask map, specifically generate a mask according to the following formula: ct , as the value of the cth row and tth column in the segmentation mask image;

[0043] .

[0044] Preferably, in S42, the confidence score of the prediction box is calculated according to the formula conf=sum1 / sum2, wherein sum1 is the predicted probability sum of the elements in the connected domain where the prediction box is located, and sum2 is the mask sum of the elements in the connected domain where the prediction box is located.

[0045] Preferably, the F1 score is obtained according to the following formula:

[0046] ,

[0047] ,

[0048] ,

[0049] In the formula, precision is the accuracy and recall is the recall rate.

[0050] Preferably, in S53, the improved intersection-over-union ratio IIoU is calculated according to the following formula:

[0051] ,

[0052] Where Area P Area is the predicted box area. T Area P The corresponding target box area, Inter is the intersection area between the predicted box and the real box.

[0053] Compared with the prior art, the advantages of the present invention are:

[0054] Compared with the DAS microseismic event detection method based on classification detection, the present invention can provide the location information of DAS microseismic events, which is convenient for users to locate microseismic events more quickly and conveniently. This is because the data points marked in the present invention are points in the microseismic event area, so the position of the prediction box of the present invention in the DAS map is the position of the microseismic event. Compared with the DAS microseismic event detection model based on the target detection model such as YOLOv3, the present invention can use more accurate semantic segmentation annotation information to provide more accurate detection results, greatly reducing the missed detection rate and false detection rate. At the same time, its detection results are directly derived from the segmentation results, and are more interpretable; compared with traditional algorithms such as STA / LTA, the present invention can provide higher detection accuracy.

[0055] The DAS microseismic event detection method based on the segmentation model and the connected domain generation algorithm provided by the present invention greatly improves the DAS microseismic event detection accuracy and model robustness, reduces the missed detection rate and the false detection rate, and at the same time provides stronger interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of the present invention;

[0057] Figure 2 This is a comparison chart of the DAS microseismic detection effects of the present invention and the RetinaNet model;

[0058] Figure 3 This is a visualization diagram of a sample in Example 2;

[0059] Figure 4 for Figure 3 Visualization of the distance weight matrix corresponding to the samples;

[0060] Figure 5 This is a detection image of a sample in Example 3 using the RetinaNet model;

[0061] Figure 6 for Figure 5 A visualization of the test results of the corresponding sample using the present invention;

[0062] Figure 7 for Figure 6 The segmentation mask image generated when the corresponding sample is detected using the present invention;

[0063] Figure 8 This is a detection image of a sample in Example 4 using the RetinaNet model;

[0064] Fig. 9 for Figure 8 Visualization diagram of the detection results of the corresponding samples using the present invention;

[0065] Fig.10 for Fig. 9 The segmentation mask map generated when the present invention is used for the corresponding sample. DETAILED DESCRIPTION

[0066] The present invention will be further described below in conjunction with the accompanying drawings.

[0067] Example 1: See Figure 1 ,A DAS microseismic event detection method based on segmentation model and connected domain generation algorithm, comprising the following steps;

[0068] S1, construct data set D;

[0069] S11, obtain DAS microseismic data containing microseismic events in N different scenarios, and obtain N normalized microseismic data after normalization, which are marked as D1~D N , where the nth normalized microseismic data is D n , 1≤n≤N, each normalized microseismic data includes C×T data points, C is the number of channels, and T is the number of time points;

[0070] S12, to D n Perform semantic segmentation annotation and generate target boxes;

[0071] D n Inner channel c, data point x at time point t ct The label is y ct , if x ct Located in the microseismic event area, then y ct =1, otherwise y ct = 0, all y ct Constitute the first labeled sample D n1 , 1≤c≤C, 1≤t≤T; then based on D n1 Generate a target frame for each microseismic event and obtain the second annotation sample D containing the target frame n2 ;

[0072] S13, generate D based on breadth-first search algorithm n The distance weight matrix A n , A n The size is C×T, and the element in the cth row and tth column is a ct ;

[0073] S14, D n1 , D n2 , A n As Tags n , generating D n The corresponding sample ;

[0074] S15, process D1~D in sequence according to S12~S14 N , generating D1~D N The corresponding sample ~ , forming a data set ;

[0075] S2, select a semantic segmentation model and construct a loss function Loss, the semantic segmentation model is used to input samples and output data points x in the samples. ct is the predicted probability of microseismic events p ct ;

[0076] ,

[0077] S3, use the dataset D to train the semantic segmentation model with minimal loss to obtain the microseismic event segmentation model;

[0078] S4, generated The prediction box is calculated, and the confidence score of the prediction box is calculated, including steps S41-S42;

[0079] S41, according to p ct Generate a segmentation mask map, and use the connected domain generation algorithm to obtain the circumscribed rectangle of each connected domain as the prediction box;

[0080] S42, calculate the confidence score conf of each prediction box;

[0081] S5, calculation sample The optimal confidence threshold and the optimal area threshold of the method include steps S51 to S56;

[0082] S51, preset the number of iterations and the confidence threshold Th for each iteration conf , area threshold Th area ;

[0083] S52, in one iteration, according to the corresponding area threshold Th area , delete the area less than Th in step S42 area The prediction box of

[0084] S53, calculate each remaining prediction box and D n2 The improved intersection-over-union (IIoU) of the corresponding target box is calculated, and the remaining prediction boxes are numbered, where the conf and IIoU of the jth remaining prediction box are marked as conf j and IIoU j ;

[0085] S54, count TP, FP and FN according to IIoU and conf of the remaining prediction boxes;

[0086] For the jth remaining prediction box, if conf j ≥Th conf And IIoU j ≥0.5, then TP increases by 1; if conf j ≥Th area And IIoU j <0.5, then FP increases by 1; if conf j <Th area And IIoU j ≥0.5, then FN increases by 1; otherwise, the prediction box is skipped;

[0087] S55, calculate the F1 score based on TP, FP and FN;

[0088] S56, repeating steps S52 to S55 until the iteration ends, selecting the confidence threshold and area threshold when the F1 score is the largest as the optimal confidence threshold and optimal area threshold;

[0089] S6, obtaining the DAS microseismic data to be detected for detection;

[0090] S61, obtaining the DAS microseismic data to be detected, normalizing it, and sending it to the microseismic event segmentation model obtained in S3, and outputting the prediction result;

[0091] S62, generate prediction boxes and confidence scores of prediction boxes according to steps S41-S42, and then delete prediction boxes whose areas are smaller than the optimal area threshold according to the optimal area threshold.

[0092] In this embodiment: the normalization process is to normalize the DAS microseismic data channel by channel, and is calculated according to the following formula:

[0093] ,

[0094] In the formula, is the value of channel c and time point t in DAS microseismic data, Max(x) and Min(x) are The maximum and minimum values ​​on the channel, 1≤c≤C, 1≤t≤T.

[0095] After the DAS microseismic data are normalized channel by channel, data enhancement is performed.

[0096] S13 specifically includes steps Sa1 to Sa4;

[0097] Sa1, preset an initial distance weight matrix A 1n , the size is C×T, if y ct =1, then A 1n The element a in row c and column t ct =0, if y ct =0,a ct =1500;

[0098] Sa2, with a ct = 0 as the starting point, traverse A using the breadth-first search algorithm 1n For each element in , get the distance value between each element and the starting point, and use the distance value to update the value of the element. If an element is traversed by multiple starting points and each gets a distance value, then use the minimum distance value to update the value of the element;

[0099] Sa3, A 1n Each element in is reduced by 1000 times to obtain the intermediate distance weight matrix A 2n ;

[0100] Sa4, modified A 2n The value of the element in A 2n Middle ct =0, change it to A 2n The maximum value of the elements in D n The distance weight matrix A n .

[0101] The semantic segmentation models are UNet, FCN, Deeplabv3, Deeplabv3+ and PSPNet.

[0102] According to p in S41 ct Generate a segmentation mask map, specifically generate a mask according to the following formula: ct , as the value of the cth row and tth column in the segmentation mask image;

[0103] .

[0104] In S42, the confidence score of the prediction box is calculated according to the formula conf=sum1 / sum2, where sum1 is the predicted probability sum of the elements in the connected domain where the prediction box is located, and sum2 is the mask sum of the elements in the connected domain where the prediction box is located.

[0105] The F1 score is obtained according to the following formula:

[0106] ,

[0107] ,

[0108] ,

[0109] In the formula, precision is the accuracy and recall is the recall rate.

[0110] In S53, the improved intersection-over-union ratio IIoU is calculated according to the following formula:

[0111] ,

[0112] Where Area P Area is the predicted box area. T Area P The corresponding target box area, Inter is the intersection area between the predicted box and the real box.

[0113] Example 2: See Figures 1 to 4 Based on Example 1, the dataset D constructed by S1 is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. A comparative experiment is performed on the test set using the method of the present invention and the existing RetinaNet model. Figure 2 shown. Figure 2In the data, when two methods were used to detect DAS microseismic events, the precision, recall and F1 scores were 0.951, 0.937 and 0.944 respectively. The precision, recall and F1 scores were much higher than those of RetinaNet, and were 17.8%, 21.5% and 19.6% higher than those of the RetinaNet model respectively. This shows that the present invention can significantly reduce the false detection rate and missed detection rate compared to directly using the target detection model for DAS microseismic event detection, and can provide detection performance far exceeding that of conventional target detection methods. In addition, in order to more vividly display the distance weight matrix, we selected a sample from the test set as Figure 3 As shown, the corresponding distance weight matrix visualization is generated as follows Figure 4 As shown in Figure 2. Since the distance weight matrix and sample size are both C×T, their elements correspond one to one, that is, a ct With x ct One to one correspondence, Figure 4 The elements in the white area correspond to the data points in the sample located in the microseismic event area. Figure 4 The set of white areas in the middle corresponds to a microseismic event in a sample.

[0114] Example 3: See Figure 5 , Figure 6 and Figure 7 , select a sample corresponding to a weak DAS microseismic event from the test set, and use the RetinaNet model and the present invention to detect it. Figure 5 , Figure 6 shown. Figure 6 The box in the middle is the predicted box, and 0.857 is the confidence score of the predicted box. Figure 5 The segmentation mask image generated when the corresponding sample is detected using the present invention is visualized as follows: Figure 7 As shown, Figure 7 The white area in the middle is the area with a mask value of 1, corresponding to the microseismic events in the sample. Figure 5 and Figure 6 It can be seen that for some samples with large noise (i.e., weak DAS microseismic events), the method provided by the present invention can correctly detect the DAS microseismic events and provide a large confidence score, while the RetinaNet model cannot detect the weak DAS microseismic events and misses the detection. Figure 7 It can be seen that the method provided by the present invention can correctly segment the microseismic event, and at the same time proves that the detection results of the method provided by the present invention are highly interpretable, and the model correctly learns the characteristics and location information of the DAS microseismic event.

[0115] Example 4: See Figure 8 , Fig. 9 and Fig.10, select a sample corresponding to a DAS microseismic event with obvious amplitude from the test set, and use the RetinaNet model and the present invention to detect it. Figure 8 , Fig. 9 As shown, Fig. 9 The box in the middle is the prediction box, and 0.811 is the confidence score of the prediction box. Fig. 9 The segmentation mask image generated when the corresponding sample is detected using the present invention is visualized as follows: Fig.10 shown. Fig.10 The white area in the middle is the area where the mask is 1, corresponding to the microseismic events in the sample. Figure 8 and Fig. 9 It can be seen that RetinaNet cannot correctly detect some DAS microseismic events with obvious amplitudes in the test set, but the method provided by the present invention can correctly detect them. This proves that the method provided by the present invention is more robust than RetinaNet in DAS microseismic detection, and further proves that the method provided by the present invention can provide higher detection accuracy and strong interpretability.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A DAS microseismic event detection method based on a segmentation model and a connected domain generation algorithm, characterized in that: The steps include: S1, construct data set D; S11, obtain DAS microseismic data containing microseismic events in N different scenarios, and obtain N normalized microseismic data after normalization, which are marked as D1~D N , where the nth normalized microseismic data is D n , 1≤n≤N, each normalized microseismic data includes C×T data points, C is the number of channels, and T is the number of time points; S12, to D n Perform semantic segmentation annotation and generate target boxes; D n Inner channel c, data point x at time point t ct The label is y ct , if x ct Located in the microseismic event area, then y ct =1, otherwise y ct = 0, all y ct Constitute the first labeled sample D n1 , 1≤c≤C, 1≤t≤T; then based on D n1 Generate a target frame for each microseismic event and obtain the second annotation sample D containing the target frame n2 ; S13, generate D based on breadth-first search algorithm n The distance weight matrix A n , A n The size is C×T, and the element in the cth row and tth column is a ct ; S14, D n1 , D n2 , A n As Tags n , generating D n The corresponding sample ; S15, process D1~D in sequence according to S12~S14 N , generating D1~D N The corresponding sample ~ , forming a data set ; S2, select a semantic segmentation model and construct a loss function Loss, the semantic segmentation model is used to input samples and output data points x in the samples. ct is the predicted probability of microseismic events p ct ; ， S3, use the dataset D to train the semantic segmentation model with minimal loss to obtain the microseismic event segmentation model; S4, generated The prediction box is calculated, and the confidence score of the prediction box is calculated, including steps S41-S42; S41, according to p ct Generate a segmentation mask map, and use the connected domain generation algorithm to obtain the circumscribed rectangle of each connected domain as the prediction box; S42, calculate the confidence score conf of each prediction box; S5, calculation sample The optimal confidence threshold and the optimal area threshold of the method include steps S51 to S56; S51, preset the number of iterations and the confidence threshold Th for each iteration conf , area threshold Th area ; S52, in one iteration, according to the corresponding area threshold Th area , delete the area less than Th in step S42 area The prediction box of S53, calculate each remaining prediction box and D n2 The improved intersection-over-union (IIoU) of the corresponding target box is calculated, and the remaining prediction boxes are numbered, where the conf and IIoU of the jth remaining prediction box are marked as conf j and IIoU j ; S54, count TP, FP and FN according to IIoU and conf of the remaining prediction boxes; For the jth remaining prediction box, if conf j ≥Th conf IIoU j ≥0.5, then TP increases by 1; if conf j ≥Th area IIoU j <0.5, then FP increases by 1; if conf j <Th area IIoU j ≥0.5, then FN increases by 1; otherwise, the prediction box is skipped; S55, calculate the F1 score based on TP, FP and FN; S56, repeating steps S52 to S55 until the iteration ends, selecting the confidence threshold and area threshold when the F1 score is the largest as the optimal confidence threshold and optimal area threshold; S6, obtaining the DAS microseismic data to be detected for detection; S61, obtaining the DAS microseismic data to be detected, normalizing it, and sending it to the microseismic event segmentation model obtained in S3, and outputting the prediction result; S62, generating prediction boxes and confidence scores of the prediction boxes according to steps S41-S42, and then deleting prediction boxes whose areas are smaller than the optimal area threshold according to the optimal area threshold; In S53, the improved intersection-over-union ratio IIoU is calculated according to the following formula: ， Where Area P Area is the predicted box area. T Area P The corresponding target box area, Inter is the intersection area between the predicted box and the real box.

2. The DAS microseismic event detection method based on the segmentation model and the connected domain generation algorithm according to claim 1 is characterized in that: The normalization process is to normalize the DAS microseismic data channel by channel, and is calculated according to the following formula: ， In the formula, is the value of channel c and time point t in DAS microseismic data, Max(x) and Min(x) are The maximum and minimum values ​​on the channel, 1≤c≤C, 1≤t≤T.

3. The DAS microseismic event detection method based on the segmentation model and the connected domain generation algorithm according to claim 2 is characterized in that: After the DAS microseismic data are normalized channel by channel, data enhancement is performed.

4. The DAS microseismic event detection method based on the segmentation model and the connected domain generation algorithm according to claim 1 is characterized in that: S13 specifically includes steps Sa1 to Sa4; Sa1, preset an initial distance weight matrix A 1n , the size is C×T, if y ct =1, then A 1n The element a in row c and column t ct =0, if y ct =0,a ct =1500; Sa2, with a ct = 0 as the starting point, traverse A using the breadth-first search algorithm 1n For each element in , get the distance value between each element and the starting point, and use the distance value to update the value of the element. If an element is traversed by multiple starting points and each gets a distance value, then use the minimum distance value to update the value of the element; Sa3, A 1n Each element in is reduced by 1000 times to obtain the intermediate distance weight matrix A 2n ; Sa4, modified A 2n The value of the element in A 2n Middle ct =0, change it to A 2n The maximum value of the elements in D n The distance weight matrix A n .

5. The DAS microseismic event detection method based on segmentation model and connected domain generation algorithm according to claim 1, characterized in that: The semantic segmentation models are UNet, FCN, Deeplabv3, Deeplabv3+ and PSPNet.

6. The DAS microseismic event detection method based on segmentation model and connected domain generation algorithm according to claim 1, characterized in that: According to p in S41 ct Generate a segmentation mask map, specifically generate a mask according to the following formula: ct , as the value of the cth row and tth column in the segmentation mask image; 。 7. The DAS microseismic event detection method based on segmentation model and connected domain generation algorithm according to claim 1 is characterized in that: In S42, the confidence score of the prediction box is calculated according to the formula conf=sum1 / sum2, where sum1 is the predicted probability sum of the elements in the connected domain where the prediction box is located, and sum2 is the mask sum of the elements in the connected domain where the prediction box is located.

8. The DAS microseismic event detection method based on segmentation model and connected domain generation algorithm according to claim 1, characterized in that: The F1 score is obtained according to the following formula: ， ， ， In the formula, precision is the accuracy and recall is the recall rate.

Images