Method for detecting, positioning and classifying underwater gas leakage through multiple beams based on deep learning

Abstract

The invention discloses a method for detecting, positioning and classifying underwater gas leakage through multiple beams based on deep learning, and relates to the field of underwater acoustic detection and deep learning. And carrying out azimuth correction on the multi-beam data with attitude offset to obtain multi-beam data without attitude offset. And through operations of seabed detection, median filtering, morphological expansion and the like, a high-quality multi-beam sonar image is generated. And constructing a data set for training an instance segmentation model and a flux classification model. And performing gas plume detection and segmentation on the preprocessed sonar image, and outputting mask information. And extracting a target contour based on a mask, determining plume coordinates through minimum circumscribed rotating rectangle fitting, and calculating the position of a seabed leakage point in combination with a conversion factor. And selecting complete plume data near the position right below the sonar equipment, and classifying leakage flux levels based on plume morphological characteristics by adopting a flux classification model. Multi-beam sonar and a deep learning algorithm are integrated, and detection, positioning and classification of underwater gas leakage are realized.

Description

Technical Field

[0001] This invention relates to the fields of underwater acoustic detection and deep learning technology, and in particular to a method for detecting, locating, and classifying underwater gas leaks based on deep learning using multi-beam detection. Background Technology

[0002] With the advancement of offshore oil and gas exploration and seabed carbon sequestration projects, the scale of subsea pipeline laying and gas geological storage is expanding, leading to a corresponding increase in the risk of gas leaks. Once a gas leak occurs, it will have a significant negative impact on relevant economic entities and the marine ecological environment.

[0003] Active acoustic technology, especially multi-beam sonar technology, with its high-resolution two-dimensional or three-dimensional imaging capabilities, wide coverage, and robustness to environmental noise, effectively compensates for the image quality degradation problem of optical imaging under turbid or low-light conditions, as well as the shortcomings of passive acoustics in intuitive visualization and its susceptibility to bubble interference, and has become the core technology for current gas leak monitoring.

[0004] Deep learning technology excels in balancing detection accuracy and processing speed, making it suitable for real-time monitoring of underwater gas leaks. Currently, deep learning-based multi-beam gas leak monitoring algorithms can achieve high-precision detection and location of gas leaks. However, further in-depth exploration and practice are needed in the field of real-time classification of leak flux levels. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the current technology of underwater gas leak monitoring by providing a deep learning-based multibeam detection, localization, and classification method for underwater gas leaks, capable of detecting and locating plume targets and classifying gas leak flux levels.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for detecting, locating, and classifying underwater gas leaks based on deep learning using multi-beam echolocation includes the following steps:

[0008] S1: Attitude correction: Using a complementary filtering algorithm, the attitude of the raw data from the attitude sensor is calculated to obtain the attitude angle; based on the attitude angle, the multi-beam data with attitude offset is processed to obtain multi-beam data without attitude offset.

[0009] S2: Data preprocessing: Seabed detection is performed on the attitude-corrected multibeam data, redundant data at the lower end of the seabed line is removed, and median filtering is used to suppress noise; gas plume features are enhanced through morphological dilatation to generate high-quality sonar images;

[0010] S3: Model Training: Construct a dataset based on the generated multibeam images, and train instance segmentation models (such as YOLOv11-seg) and flux classification models (such as YOLOv11-cls).

[0011] S4: Target Detection: Input the preprocessed image into the trained instance segmentation model to detect and segment the gas plume region in real time and output the plume mask information;

[0012] S5: Target localization: Reconstruct the binary image based on mask information, extract the target contour and fit the minimum bounding rectangle, calculate the center coordinates (x, y) of the rectangle. c , y c ), combined with conversion factor k x k y The actual position is calculated using the reference point (x0, y0);

[0013] S6: Flux Classification: Filter complete plume data located directly below or near the sonar equipment, input the trained flux classification model to classify flux levels, and output the flux level evaluation results.

[0014] In step S1, the attitude calculation includes the following steps:

[0015] (1) Calculate the data error of the gyroscope The error formula is as follows:

[0016]

[0017] in, This represents the data from the gyroscope. This is a proportional term used to control the reliability of the sensor; This is the integral term, used to eliminate static error; The larger the accelerometer, the more trust there is in it. The smaller the size, the more trust is placed in the gyroscope; for:

[0018]

[0019] in, , , This represents the theoretical gravitational acceleration component.

[0020] (2) Perform error compensation on the gyroscope data. The compensation formula is as follows:

[0021]

[0022] in, The data is the compensated gyroscope data. The gyroscope data before compensation;

[0023] (3) Calculate quaternions using the following formula:

[0024]

[0025] in, , , , It is a quaternion. , , , For quaternions that are updated in real time, , , To compensate for the angular velocities of each axis;

[0026] (4) Calculate the attitude angles based on quaternions. The calculation formula is as follows:

[0027]

[0028] in, This is the roll angle. The pitch angle, This is the yaw angle.

[0029] In step S1, the processing of multibeam data with attitude offset includes the following steps:

[0030] (1) Determine the area scanned by the sonar in the original sonar data;

[0031] (2) Create a mask for the area scanned by the sonar in the original sonar data;

[0032] (3) Extract the area scanned by the sonar from the original sonar data;

[0033] (4) Create a rotation matrix;

[0034] (5) Rotate the area scanned by the sonar in the original sonar data;

[0035] (6) Rotating mask;

[0036] (7) Fill the sonar scan area in the original sonar data with background data;

[0037] (8) Add rotated region data to the sonar data that has been filled with background data.

[0038] In step S2, the data preprocessing includes the following steps:

[0039] (1) By setting a predefined intensity threshold, the shallowest depth points exceeding the threshold are identified, and these points are smoothed to form a continuous seafloor stratigraphic line. Redundant data with a depth more than a few meters (e.g., 2 meters) below the bottom profile are systematically removed.

[0040] (2) Median filtering uses a sliding window to denoise the sonar image. The filtering function is:

[0041]

[0042] In the formula, and represents the grayscale values ​​of pixel (x, y) before and after filtering, respectively, and S and K represent the width and height of the filtering window, respectively;

[0043] (3) Morphological expansion operations utilize structuring elements to expand the target region, thereby enhancing the contrast between the gas plume and the surrounding acoustic clutter; its simple definition is as follows: Assume there exists a set M, and expand set M using N as the structuring element, the mathematical expression is:

[0044] N Z

[0045] This expression represents Translation Vector after AND The intersection of them is not empty.

[0046] In step S3, the model training includes the following steps:

[0047] (1) Based on the generated multibeam images, construct the datasets required for instance segmentation and flux classification deep learning models respectively; use the segmentation annotation program to complete the construction of the segmentation dataset; filter out the complete plume data located near the multibeam, and divide the multibeam images into several gas flux levels (such as [1, 3, 5, 7, 8, 10, 12] L / min) to complete the construction of the classification dataset;

[0048] (2) Training and validation of deep learning models for instance segmentation based on segmentation datasets. The model training calls the stochastic gradient descent optimizer to update and converge parameters. Average precision mAP is introduced as an evaluation index of the overall performance of the model. The trained model has the highest detection and segmentation accuracy and is used as the model for object detection.

[0049] (3) Training and validation of the throughput classification deep learning model based on the classification dataset.

[0050] In step S4, the target detection includes the following steps:

[0051] (1) Input the real-time data stream image into the trained instance segmentation model, detect and segment the target region in real time, and output the target confidence and mask information;

[0052] (2) Using the trained instance segmentation deep learning model, plume detection is performed on the validation set. The output confidence score is based on the binomial distribution principle. An optimal confidence score threshold that balances high recall and low false alarm rate is set, and targets exceeding the threshold are judged as plumes.

[0053] In step S5, the target localization includes the following steps:

[0054] (1) Based on the target detection results, if the plume target exists, the target mask information is extracted;

[0055] (2) Based on the mask coordinate information, the ray method is used to judge the interior of the mask polygon for each pixel and reconstruct the mask image;

[0056] (3) By extracting the outer contour of the mask region and performing minimum bounding rectangle fitting on each contour, the center coordinates (x) are obtained. c , y c );

[0057] (4) Obtain the spatial location of the plume based on the conversion factor. The conversion factor is calculated based on the image size and the actual spatial distance. The mathematical formula is as follows:

[0058] , in = 0.05997, = -80

[0059] , in = 0.017053033

[0060] Where, the x-axis This reflects the horizontal distance between the plume and the sonar, with the vertical axis corresponding to... It refers to the depth of the plume;

[0061] (5) According to , Based on satellite positioning information, the latitude and longitude coordinates of the plume are output; according to the principle of spherical triangles, the formula for calculating latitude coordinates is:

[0062]

[0063] The formula for calculating longitude coordinates is:

[0064]

[0065] in, The latitude of the plume. For the longitude of the plume, For the Earth's radius, .

[0066] In step S6, the flux classification includes the following steps:

[0067] (1) Based on the x output by the positioning module real Distance information is used to filter complete plume data located directly below or near the sonar equipment to classify the gas leakage flux level;

[0068] (2) Using a trained flux classification deep learning model, plume feature changes under different flux categories are identified to classify different gas flux levels.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] This invention provides a deep learning-based multi-beam detection, localization, and classification method for underwater gas leaks. This method integrates modules such as attitude correction, data preprocessing, detection, localization, and flux classification, constructing a complete underwater gas leak monitoring process. It effectively solves the problem of fragmented processes in traditional methods, providing a macro-framework and process approach for the development of leak monitoring algorithms.

[0071] This invention employs mask reconstruction and minimum bounding rectangle fitting techniques, combined with the ray casting method, transformation factor calculation, and the principle of spherical triangles, to achieve high-precision positioning of plume targets from image mask processing to outputting actual latitude and longitude. Compared with traditional positioning methods, it has smaller errors and more accurate position calculations.

[0072] The instance segmentation and throughput classification models involved can be flexibly replaced with current advanced algorithm models, adapting to different hardware environments and monitoring needs, facilitating continuous optimization, improving leakage monitoring effectiveness, and adapting to technological iterations.

[0073] The method proposed in this invention has strong versatility. The attitude correction stage uses a complementary filtering algorithm, which can adapt to complex underwater attitude changes; the instance segmentation and classification models involved can be flexibly replaced with the most advanced algorithm models currently available, thereby improving the leakage monitoring effect. Attached Figure Description

[0074] Figure 1 The overall flowchart of the underwater gas leak detection, localization, and classification method based on deep learning of the present invention is shown.

[0075] Figure 2The performance graphs for object detection and instance segmentation of different versions of YOLOv11 are presented. Among them, (a) shows the object detection performance of different versions of YOLOv11 models; (b) shows the instance segmentation performance of different versions of YOLOv11 models.

[0076] Figure 3 The image shows the detection results of YOLOv11x-seg based on image data acquired by multibeam sonar (the detection and segmentation results of plume targets at different lateral distances).

[0077] Figure 4 The localization process is as follows: (original multibeam image → mask reconstruction → rectangle fitting → center point mapping).

[0078] Figure 5 The results of flux classification (seven-class flux confusion matrix). Detailed Implementation

[0079] The following embodiments will provide a more detailed description of the present invention in conjunction with the accompanying drawings.

[0080] See Figure 1 The present invention relates to a method for detecting, locating, and classifying underwater gas leaks based on deep learning using multi-beam detection, includes the following steps:

[0081] S1: Attitude Correction: This covers attitude calculation and processing of images with attitude shifts.

[0082] The attitude calculation includes the following steps:

[0083] (1) Calculate the data error of the gyroscope The error formula is as follows:

[0084]

[0085] in, This represents the data from the gyroscope. This is a proportional term used to control the reliability of the sensor; This is the integral term, used to eliminate static error; The larger the accelerometer, the more reliable it is. The smaller the size, the more trust is placed in the gyroscope; for:

[0086]

[0087] in, , , For theoretical gravitational acceleration in , and The directional component.

[0088] (2) Error compensation is applied to the gyroscope data. The compensation formula is as follows:

[0089]

[0090] in, The data is the compensated gyroscope data. The data is the gyroscope data before compensation.

[0091] (3) Perform quaternion calculations, the formula for which is as follows:

[0092]

[0093] in, , , , It is a quaternion. , , , For quaternions that are updated in real time, , , To compensate for the angular velocities of each axis.

[0094] (4) Calculate the attitude angles based on quaternions. The calculation formula is as follows:

[0095]

[0096] in, This is the roll angle. The pitch angle, This is the yaw angle.

[0097] Furthermore, processing images with pose shifts involves the following steps:

[0098] (1) Determine the area covered by the sonar scan in the original sonar data;

[0099] (2) Create a mask for the sonar scanning area in the original sonar data;

[0100] (3) Extract the sonar scan area from the raw sonar data;

[0101] (4) Create a rotation matrix;

[0102] (5) Rotate the area scanned by the sonar in the original sonar data;

[0103] (6) Rotating mask;

[0104] (7) Fill the sonar scan area in the original sonar data with background data;

[0105] (8) Add rotated region data to the sonar data that has been filled with background data.

[0106] S2: Data Preprocessing

[0107] First, by setting a predefined intensity threshold, the shallowest depth points exceeding that threshold are identified. These points are then smoothed to form continuous seafloor stratigraphic lines, while redundant data at the lower depths of the bottom profile are systematically removed.

[0108] Secondly, median filtering uses a sliding window to denoise the sonar image; the filtering function is:

[0109]

[0110] In the formula, and These represent the pixel points before and after filtering, respectively. The grayscale value is given by S and K, where S and K represent the width and height of the filter window, respectively.

[0111] Finally, the morphological expansion operation uses a structuring element to expand the region of interest, thereby enhancing the contrast between the gas plume and the surrounding acoustic clutter. Its simple definition is as follows: Assuming there exists a set M, and M is expanded using N as the structuring element, the mathematical expression is:

[0112] N Z

[0113] This expression represents Translation After the vector, its relation to The intersection of them is not empty.

[0114] S3: Model Training

[0115] First, based on the generated multibeam images, the datasets required for instance segmentation and flux classification deep learning models are constructed respectively. The segmentation dataset is constructed using a segmentation annotation program. Near-range data where the plume is directly below the multibeams are selected, and the multibeam images are divided into several gas flux levels (e.g., [1, 3, 5, 7, 8, 10, 12] L / min) to complete the construction of the classification dataset.

[0116] Secondly, the training and validation of the instance segmentation deep learning model were completed based on the segmentation dataset, taking YOLOv11-seg as an example. Five versions of YOLOv11-seg (YOLOv11n-seg, YOLOv11s-seg, YOLOv11m-seg, YOLOv11l-seg, and YOLOv11x-seg) were trained and evaluated. The model training used the stochastic gradient descent (SGD) optimizer to update and converge parameters; details of the parameter settings during training are shown in Table 1. Mean accuracy (mAP) was introduced as an evaluation metric for the overall model performance. The results showed that YOLOv11x-seg had the highest detection and segmentation accuracy, and it was chosen as the model for object detection.

[0117] Figure 2 This section showcases the object detection and instance segmentation performance of different versions of YOLOv11. (a) shows the object detection performance of different versions of the YOLOv11 model; (b) shows the instance segmentation performance of different versions of the YOLOv11 model. Figure 2 As can be seen, all versions of the YOLOv11-seg model exhibit good performance in gas leak detection and segmentation tasks. In terms of target detection performance, the YOLOv11x-seg model achieves 0.995 on the mAP0.5 metric, almost reaching 100%; even at the more challenging mAP0.5-0.95 metric, it still reaches 0.895, demonstrating excellent generalization ability and robustness. Instance segmentation performance shows a similar trend, with the YOLOv11x-seg model achieving 0.995 on mAP0.5 and 0.895 on mAP0.5-0.95. Experiments show that more complex models, by increasing network depth and width, can more fully learn the acoustic feature patterns of plumes, especially exhibiting stronger feature extraction capabilities when distinguishing low-contrast plume edges from seabed clutter.

[0118] Finally, the throughput classification deep learning model was trained and validated based on the classification dataset, using YOLOv11-cls as an example. The throughput classification model was trained according to the training parameters in Table 1, and the results show that YOLOv11x-cls has the highest classification accuracy.

[0119] Table 1 Model training parameters

[0120]

[0121] S4: Target Detection

[0122] First, the real-time acquired data stream image needs to be input into the already trained instance segmentation model. This model will perform real-time detection and accurate segmentation of the target regions in the image, and then output the confidence score of each target and the corresponding mask information for subsequent analysis and processing.

[0123] Secondly, based on the initial detection, a pre-trained instance segmentation deep learning model is used to perform detailed plume detection on a specific validation set, outputting the confidence value for each detection result. To ensure the accuracy and reliability of the detection results, an optimal confidence threshold is set based on the statistical principle of binomial distribution. This threshold simultaneously addresses the dual requirements of high recall and low false positive rate. Only when the confidence of a detected target exceeds this preset threshold is it classified as a plume, thereby effectively improving detection accuracy and the overall performance of the system.

[0124] S5: Target Positioning:

[0125] First, based on the results of the target detection algorithm, if a plume target is indeed detected, the mask information of that plume target is then extracted. This step provides the basic data for subsequent image analysis and processing.

[0126] Secondly, after successfully obtaining the mask information, the efficient geometric algorithm of ray casting is used based on the coordinates of these masks to examine each pixel in the image one by one to determine whether it is located inside the mask polygon. Through this point-by-point examination, accurate reconstruction of the mask image can be achieved, ensuring that the reconstructed image is highly consistent with the original mask information.

[0127] Then, based on the reconstructed mask image, the outer contour of the mask region is further extracted. For each extracted contour, the minimum bounding rectangle of the rotation is fitted. This step allows for the accurate acquisition of the center coordinates of each contour, denoted as (x...). c , y c With these center coordinates, the specific spatial location of the plume can be determined by combining them with a transformation factor. It's important to note that this transformation factor is not arbitrarily set, but rather derived through rigorous mathematical derivation based on the conversion relationship between image size and actual spatial distance. The specific mathematical formula is shown below:

[0128] , in = 0.05997, = -80

[0129] , in = 0.017053033

[0130] Where the x-coordinate is real The vertical axis represents the horizontal distance between the plume and the sonar equipment. This distance is obtained through precise measurement and is used to describe the specific position of the plume in the horizontal direction; while the vertical axis y... real This corresponds to the depth of the plume, that is, the specific distribution of the plume in the vertical direction. This data is also obtained through high-precision detection methods.

[0131] Finally, after a comprehensive analysis of x real y real Based on these two key parameters and the precise location information obtained through the satellite positioning system, the system will output the latitude and longitude coordinates of the plume. To ensure the accuracy of these coordinates, based on the geometric principles of spherical triangles, and through rigorous mathematical derivation and calculation, the specific formula for determining the latitude coordinates can be derived as follows:

[0132]

[0133] The formula for calculating longitude coordinates is:

[0134]

[0135] in, The latitude of the plume. For the longitude of the plume, For the Earth's radius, .

[0136] S6: Flux Classification:

[0137] First, based on the x output by the positioning module real Distance information is used to filter complete plume data located directly below the sonar equipment to classify the gas leakage flux level. In this embodiment, complete plume data within ±10% of the detection radius directly below the sonar equipment are filtered to ensure that the plume pattern is not truncated by the edges.

[0138] Secondly, a trained flux classification deep learning model is used to identify changes in plume characteristics under different flux categories, thereby achieving classification of different gas flux levels.

[0139] from Figure 3 As can be seen, the trained YOLOv11x-seg model can accurately identify plume targets at different lateral distances. This shows that the model can maintain stable and efficient detection performance even with varying distances, fully demonstrating its reliability in complex underwater environments.

[0140] Figure 4 The demonstration shows the localization process, specifically: original multibeam image → mask reconstruction → rectangle fitting → center point mapping, to achieve the localization of the leaky plume. Details are as follows:

[0141] (1) The original multibeam image is a visualization result (colored triangle) of the original acoustic data acquired by the multibeam sonar, which includes the underwater scene and acoustic reflection information of the gas plume.

[0142] (2) Mask reconstruction is to accurately extract the plume region from the original image, and use a deep learning model (such as YOLOv11-seg) to identify and segment it, so as to separate the plume from the complex background and generate a mask image that retains only the plume pixels.

[0143] (3) Rectangle fitting is a method for fitting a rectangular frame to the plume region after mask reconstruction, using the minimum bounding rectangle algorithm to fit the actual shape of the plume. Compared with ordinary rectangles, the rotated rectangle can more tightly wrap the plume (adapting to the tilted and irregular shape of the plume), thereby reducing positioning errors and accurately obtaining the spatial occupancy information of the plume.

[0144] (4) Center point mapping is to calculate the center coordinates of the fitted rectangle, and then combine the sonar equipment parameters and conversion factors to map the image coordinates to the actual underwater spatial coordinate system, calculate the horizontal distance and depth of the plume, and finally combine satellite positioning to obtain latitude and longitude, thus realizing the leap from image pixels to the real location.

[0145] Table 2 Target detection results based on simulation data

[0146]

[0147] Table 3 Target localization results based on simulation data

[0148]

[0149] Table 2 shows that the YOLOv11-based detection model achieved a detection rate of 99.5% and a false alarm rate of 1%. The 99.5% detection rate means that the model can identify almost all leak events, demonstrating extremely high sensitivity and reliability in gas leak detection. Meanwhile, the 1% false alarm rate indicates that the model has an extremely low false alarm rate, meaning it rarely misclassifies non-leak situations as leaks, greatly reducing interference from misjudgments. These results collectively demonstrate that the model possesses both high accuracy and robustness, enabling efficient and accurate gas leak detection in complex and variable environments.

[0150] As shown in Table 3, the model's average horizontal positioning error in a 30-meter water depth scenario is approximately 0.04 meters, with an error occurrence probability of 8%, indicating that the model exhibits minimal deviation in positioning tasks. This superior performance strongly validates the effectiveness of using the minimum circumscribed rotating rectangle for plume contour fitting.

[0151] Figure 5 The results show that the flux classification model exhibits strong robustness and generalization performance in terms of classification accuracy. However, the model is also prone to confusion when dealing with high flux levels and small differences in flux levels. This limitation is mainly due to the high similarity in morphological features of high-flux plumes, which poses a significant challenge to the model's classification.

[0152] From a computational performance perspective, experiments show that the model achieves an average image processing speed of approximately 20 milliseconds per frame on a server equipped with an NVIDIA GeForce RTX 3060 graphics card, highlighting its deployment potential in real-time monitoring environments.

[0153] The above embodiments are merely preferred embodiments of the present invention and should not be considered as limiting the scope of the present invention. All equivalent variations and improvements made in accordance with the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A method for detecting, locating, and classifying underwater gas leaks based on deep learning using multi-beam echo detection, characterized in that... Includes the following steps: S1: Attitude correction: Using a complementary filtering algorithm, the attitude of the raw data from the attitude sensor is calculated to obtain the attitude angle; based on the attitude angle, the multi-beam data with attitude offset is processed to obtain multi-beam data without attitude offset. S2: Data preprocessing: Seabed detection is performed on the attitude-corrected multibeam data, redundant data at the lower end of the seabed line is removed, and median filtering is used to suppress noise; gas plume features are enhanced through morphological dilatation to generate high-quality sonar images; S3: Model Training: Construct a dataset based on the generated multibeam images, and train an instance segmentation model and a flux classification model. S4: Target Detection: Input the preprocessed image into the trained instance segmentation model to detect and segment the gas plume region in real time and output the plume mask information; S5: Target localization: Reconstruct the binary image based on mask information, extract the target contour and fit the minimum bounding rectangle, calculate the center coordinates (x, y) of the rectangle. c , y c ), combined with conversion factor k x k y The actual position is calculated using the reference point (x0, y0); S6: Flux Classification: Filter complete plume data located directly below or near the sonar equipment, input the trained flux classification model to classify flux levels, and output the flux level evaluation results.

2. The method for detecting, locating, and classifying underwater gas leaks based on deep learning as described in claim 1, characterized in that... In step S1, the attitude calculation includes the following steps: (1) Calculate the data error of the gyroscope The error formula is as follows: in, This represents the data from the gyroscope. This is a proportional term used to control the reliability of the sensor; This is the integral term, used to eliminate static error; The larger the accelerometer, the more trust there is in it. The smaller the size, the more trust is placed in the gyroscope; for: in, , , This represents the theoretical gravitational acceleration component. (2) Perform error compensation on the gyroscope data. The compensation formula is as follows: in, The data is the compensated gyroscope data. The gyroscope data before compensation; (3) Calculate quaternions using the following formula: in, , , , It is a quaternion. , , , For quaternions that are updated in real time, , , To compensate for the angular velocities of each axis; (4) Calculate the attitude angles based on quaternions. The calculation formula is as follows: in, This is the roll angle. The pitch angle, This is the yaw angle.

3. The method for detecting, locating, and classifying underwater gas leaks based on deep learning, as described in claim 1, is characterized in that... In step S1, the processing of multibeam data with attitude offset includes the following steps: (1) Determine the area scanned by the sonar in the original sonar data; (2) Create a mask for the area scanned by the sonar in the original sonar data; (3) Extract the area scanned by the sonar from the original sonar data; (4) Create a rotation matrix; (5) Rotate the area scanned by the sonar in the original sonar data; (6) Rotating mask; (7) Fill the sonar scan area in the original sonar data with background data; (8) Add rotated region data to the sonar data that has been filled with background data.

4. The method for detecting, locating, and classifying underwater gas leaks based on deep learning as described in claim 1, characterized in that... In step S2, the data preprocessing includes the following steps: (1) By setting a predefined intensity threshold, the shallowest depth points exceeding the threshold are identified, and these points are smoothed to form a continuous seabed stratigraphic line. Redundant data with depths exceeding several meters below the bottom profile are systematically removed. (2) Median filtering uses a sliding window to denoise the sonar image. The filtering function is: In the formula, and represents the grayscale values ​​of pixel (x, y) before and after filtering, respectively, and S and K represent the width and height of the filtering window, respectively; (3) Morphological expansion operations utilize structuring elements to expand the target region, thereby enhancing the contrast between the gas plume and the surrounding acoustic clutter; its simple definition is as follows: Assume there exists a set M, and expand set M using N as the structuring element, the mathematical expression is: N Z This expression represents Translation Vector followed by set The intersection of them is not empty.

5. The method for detecting, locating, and classifying underwater gas leaks based on deep learning as described in claim 1, characterized in that... In step S3, the model training includes the following steps: (1) Based on the generated multibeam images, construct the datasets required for instance segmentation and flux classification deep learning models respectively; use the segmentation annotation program to complete the construction of the segmentation dataset; filter out the complete plume data located near the multibeam, and divide the multibeam images into several gas flux levels to complete the construction of the classification dataset; (2) Training and validation of deep learning models for instance segmentation based on segmentation datasets. The model training calls the stochastic gradient descent optimizer to update and converge parameters. Average precision mAP is introduced as an evaluation index of the overall performance of the model. The trained model has the highest detection and segmentation accuracy and is used as the model for object detection. (3) Training and validation of the throughput classification deep learning model based on the classification dataset.

6. The method for detecting, locating, and classifying underwater gas leaks based on deep learning as described in claim 1, characterized in that... In step S4, the target detection includes the following steps: (1) Input the real-time data stream image into the trained instance segmentation model, detect and segment the target region in real time, and output the target confidence and mask information; (2) Using the trained instance segmentation deep learning model, plume detection is performed on the validation set. The output confidence score is based on the binomial distribution principle. An optimal confidence score threshold that balances high recall and low false alarm rate is set, and targets exceeding the threshold are judged as plumes.

7. The method for detecting, locating, and classifying underwater gas leaks based on deep learning as described in claim 1, characterized in that... In step S5, the target localization includes the following steps: (1) Based on the target detection results, if the plume target exists, the target mask information is extracted; (2) Based on the mask coordinate information, the ray method is used to judge the interior of the mask polygon for each pixel and reconstruct the mask image; (3) By extracting the outer contour of the mask region and performing minimum bounding rectangle fitting on each contour, the center coordinates (x) are obtained. c , y c ); (4) Obtain the spatial location of the plume based on the conversion factor. The conversion factor is calculated based on the image size and the actual spatial distance. The mathematical formula is as follows: , in = 0.05997, = -80 , in = 0.017053033 Where, the x-axis This reflects the horizontal distance between the plume and the sonar, with the vertical axis corresponding to... It refers to the depth of the plume; (5) According to , Based on satellite positioning information, the latitude and longitude coordinates of the plume are output; according to the principle of spherical triangles, the formula for calculating latitude coordinates is: The formula for calculating longitude coordinates is: in, The latitude of the plume. For the longitude of the plume, For the Earth's radius, .

8. The method for detecting, locating, and classifying underwater gas leaks based on deep learning as described in claim 1, characterized in that... In step S6, the flux classification includes the following steps: (1) Based on the output of the positioning module Distance information is used to filter complete plume data located directly below or near the sonar equipment to classify the gas leakage flux level; (2) Using a trained flux classification deep learning model, plume feature changes under different flux categories are identified to classify different gas flux levels.

Images