Method and system for detecting tearing opening of transformer oil tank in fire scene

Abstract

The invention discloses a method and a system for detecting a tearing opening of a transformer oil tank in a fire scene. The method comprises the following steps: acquiring infrared sequence images of the transformer oil tank under different wave bands; the method comprises the following steps: after preprocessing an infrared sequence image, extracting interested angular points in the infrared sequence image, and respectively obtaining the angular points of the infrared sequence image under different wavebands; obtaining a sparse three-dimensional model from the infrared sequence image by applying a motion recovery structure algorithm according to the angular points of the infrared sequence image under different wavebands, and carrying out meshing on the sparse three-dimensional model to obtain a dense three-dimensional model; establishing an oil tank tearing opening detection model, and taking the preprocessed infrared sequence image and the dense three-dimensional model as input to obtain position and size information of a tearing opening; and training the established oil tank tearing opening detection model, and carrying out online detection by using the trained oil tank tearing opening detection model, so that electrical faults and fire risks caused by the tearing opening can be avoided.

Description

Technical Field

[0001] The invention relates to the technical field of transformer fault detection, and in particular to a method and system for detecting a tear in a transformer oil tank in a fire scene. Background Art

[0002] As we all know, computer 3D reconstruction is an important research field in computer-aided design and computer graphics. There are many ways to build a 3D model, and image-based modeling is a typical method. This method uses the 2D image information of the target object collected by infrared to restore the spatial geometric information of the target object, that is, to study the relationship between the matching points of the 2D image and the corresponding points in the 3D space, calculate and determine the relationship and size of the 3D space points of the target object, so as to achieve 3D reconstruction of the target object. At the same time, the 3D geometric information and texture information are combined through mapping to obtain a model with photo-realistic feeling. The modeling process is automated, the labor intensity is low, and the required equipment is also simple, which greatly reduces the modeling cost. Therefore, in situations where realistic modeling is required, image-based modeling undoubtedly has high practical value and has become a current research hotspot.

[0003] Three-dimensional image reconstruction is an emerging technology that integrates knowledge from computer vision, image processing, computer graphics, etc. It has characteristics that two-dimensional graphics cannot match. Its model can be observed intuitively from multiple different angles, and has realistic effects, and can achieve real-time virtualization and real-time interaction.

[0004] During the operation of the transformer, due to various reasons, such as mechanical stress, material fatigue, corrosion, manufacturing defects, etc., it may cause tears. These tears may threaten the normal operation and safety of the equipment. Therefore, timely and accurate detection of transformer tears is crucial to the maintenance and safety of the equipment. After a fire occurs, due to the difficulty in obtaining information on transformer tears, workers need to approach high-voltage equipment, which poses a great safety hazard. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for detecting a tear in a transformer oil tank in a fire scene based on infrared thermal images.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A method for detecting a tear in a transformer oil tank in a fire scene, comprising:

[0008] Obtain infrared sequence images of transformer oil tank in different bands;

[0009] After preprocessing the infrared sequence images, the corner points of interest in the infrared sequence images are extracted, and the corner points of the infrared sequence images in different bands are obtained respectively;

[0010] According to the corner points of infrared sequence images in different bands, the motion recovery structure algorithm is used to obtain a sparse three-dimensional model from the infrared sequence images, and then gridding is performed to obtain a dense three-dimensional model;

[0011] A fuel tank tear detection model is established, including a feature extraction network and a multimodal data fusion detection network. The preprocessed infrared sequence image is used as the input of the feature extraction network, which outputs a two-dimensional splicing feature. The two-dimensional splicing feature and the dense three-dimensional model are then used as the input of the multimodal data fusion detection network to obtain the location and size information of the tear.

[0012] The established fuel tank tear detection model is trained, and then the trained fuel tank tear detection model is used for online detection.

[0013] In one embodiment of the present invention, infrared sequence images of the transformer oil tank in different bands are obtained by applying the following method:

[0014] The UAV carries a multi-band infrared camera to obtain infrared sequence images, including a first infrared sequence image applicable to a flame band and a second infrared sequence image applicable to a solid metal band; and the infrared lens of the multi-band infrared camera selects a 5-8μm filter lens.

[0015] In one embodiment of the present invention, corner points of infrared sequence images in different bands are obtained respectively, including: firstly extracting edge contour lines of target objects from the infrared sequence images through a Canny edge detection operator, and then calculating the sharpness of the edge contour lines using an image sharpness algorithm to obtain key corner points in the infrared images.

[0016] In one embodiment of the present invention, obtaining a dense three-dimensional model includes: applying a Crust algorithm to expand a sparse three-dimensional model obtained from an infrared sequence image using a structure-from-motion algorithm to obtain a dense three-dimensional model.

[0017] In one embodiment of the present invention, the location and size information of the tear is obtained by the following formula:

[0018] R tear =Mask R-CNN(F fusion , M 3D );

[0019] in,

[0020] F fusion =Concatenate(ConvNet(I λ1 ), ConvNet(I λ2 ));

[0021] M3D =Crust(P cloud );

[0022] P cloud =SfM(C λ1 , C λ2 );

[0023] C λi =Sharpness(E λi ); i = 1, 2;

[0024] E λi =Canny(I λi );

[0025] In the formula, R tear It represents the location and size of the tear, Mask R-CNN represents the mask region convolutional neural network, and F fusion It is represented as a two-dimensional concatenation feature, Concatenate is represented as a concatenation function, ConvNet is represented as a convolutional neural network for feature extraction, ConvNet (I λ1 ) represents the feature extraction network to extract the features of the first infrared sequence image, ConvNet(I λ2 ) represents the feature extraction network extracting the second infrared sequence image, I λ1 Represented as the first infrared sequence image, I λ2 Represented as the second infrared sequence image, M 3D Represented as a dense three-dimensional model, P cloud It is represented as a sparse 3D model obtained from infrared image sequences using the structure from motion algorithm SfM. cloud ) represents the expansion of the 3D point cloud obtained from the infrared sequence images using the Crust algorithm. λi Represented as using the Canny edge detector on image I λi The result of edge detection, C λi Expressed as using the image sharpness algorithm Sharpness from the edge image E λi The set of corner points extracted from .

[0026] The present invention provides a system using the above-mentioned method for detecting a tear in a transformer oil tank in a fire scene, comprising:

[0027] Sequence image module, used to obtain infrared sequence images of transformer oil tank in different bands;

[0028] The corner point module is used to extract the corner points of interest in the infrared sequence images after preprocessing the infrared sequence images, and obtain the corner points of the infrared sequence images in different bands respectively;

[0029] The 3D model module is used to obtain a sparse 3D model from the infrared sequence images according to the corner points of the infrared sequence images in different bands by applying the motion recovery structure algorithm, and to grid the sparse 3D model to obtain a dense 3D model.

[0030] A modeling module is used to establish a fuel tank tear detection model, including a feature extraction network and a multimodal data fusion detection network; the preprocessed infrared sequence image is used as the input of the feature extraction network, which outputs a two-dimensional splicing feature; the two-dimensional splicing feature and the dense three-dimensional model are then used as the input of the multimodal data fusion detection network to obtain the location and size information of the tear;

[0031] The training and detection module is used to train the established fuel tank tear detection model, and then use the trained fuel tank tear detection model for online detection.

[0032] In one embodiment of the present invention, infrared sequence images of the transformer oil tank in different bands are obtained by applying the following method:

[0033] The UAV carries a multi-band infrared camera to obtain infrared sequence images, including a first infrared sequence image applicable to a flame band and a second infrared sequence image applicable to a solid metal band; and the infrared lens of the multi-band infrared camera selects a 5-8μm filter lens.

[0034] In one embodiment of the present invention, corner points of infrared sequence images in different bands are obtained respectively, including: firstly extracting edge contour lines of target objects from the infrared sequence images through a Canny edge detection operator, and then calculating the sharpness of the edge contour lines using an image sharpness algorithm to obtain key corner points in the infrared images.

[0035] In one embodiment of the present invention, obtaining a dense three-dimensional model includes: applying a Crust algorithm to expand a sparse three-dimensional model obtained from an infrared sequence image using a structure-from-motion algorithm to obtain a dense three-dimensional model.

[0036] In one embodiment of the present invention, the location and size information of the tear is obtained by the following formula:

[0037] R tear =Mask R-CNN(F fusion , M 3D );

[0038] in,

[0039] F fusion =Concatenate(ConvNet(I λ1 ), ConvNet(I λ2 ));

[0040] M 3D =Crust(P cloud );

[0041] P cloud =SfM(C λ1 , C λ2 0;

[0042] C λi =Sharpness(E λi ); i = 1, 2;

[0043] E λi =Canny(I λi );

[0044] In the formula, R tear It represents the location and size of the tear, Mask R-CNN represents the mask region convolutional neural network, and F fusion It is represented as a two-dimensional concatenation feature, Concatenate is represented as a concatenation function, ConvNet is represented as a convolutional neural network for feature extraction, ConvNet (I λ1 ) represents the feature extraction network to extract the features of the first infrared sequence image, ConvNet(I λ2 ) represents the feature extraction network extracting the second infrared sequence image, I λ1 Represented as the first infrared sequence image, I λ2 Represented as the second infrared sequence image, M 3D Represented as a dense three-dimensional model, P cloud It is represented as a sparse 3D model obtained from infrared image sequences using the structure from motion algorithm SfM. cloud ) represents the expansion of the 3D point cloud obtained from the infrared sequence images using the Crust algorithm. λi Represented as using the Canny edge detector on image I λi The result of edge detection, C λi Expressed as using the image sharpness algorithm Sharpness from the edge image E λi The set of corner points extracted from .

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] Using a multi-band infrared camera to obtain infrared images of different frequency bands, covering different temperature ranges, can more accurately distinguish between flames and metal surfaces. Infrared images of different frequency bands provide rich information, enhance the contrast between flames and metal surfaces, and reduce the interference of flames on metal surface temperature measurement. By extracting features from infrared images of different frequency bands through independent convolutional neural networks, and fusing the output feature maps of each convolutional neural network together by splicing, a multimodal feature representation is formed. This fusion method can capture the features of the tear from different angles, significantly improving the accuracy of tear detection. The use of a 5-8μm filter lens can effectively reduce the impact of flames, so that the tear can still be accurately detected in complex environments (such as fire scenes). This choice improves the robustness of the system and enables the system to work stably under various conditions.

[0047] A corner point matching algorithm based on image sharpness is used to achieve feature point matching. The Canny edge detection operator is used to extract clear edge contours at the single pixel level. Then, a simple and effective image sharpness algorithm is used to extract corner points, thereby improving the accuracy of 3D reconstruction.

[0048] The method of recovering structure from motion is used to obtain sparse 3D point clouds, and dense point cloud data is obtained by expanding the sparse point cloud. This method can effectively reduce the sparsity of point cloud data and improve the accuracy of 3D models. The Crust algorithm is used to grid point cloud data. After obtaining dense point cloud data, the holes are filled with grid point cloud data to simplify the model and obtain the final 3D model. This method can more finely display the 3D model of the target object, thereby improving the accuracy of tear detection.

[0049] Using MaskR-CNN for tear detection and segmentation can not only detect the location of the tear, but also provide pixel-level segmentation masks, which improves the robustness of tear detection and can accurately distinguish tear from other objects even in complex backgrounds.

[0050] The UAV is equipped with an infrared camera to obtain the infrared thermal image of the transformer, and a three-dimensional model of the transformer is established based on the image. The tear information is obtained by combining the infrared thermal image of the transformer with the three-dimensional model. When the crack is still developing or the crack already exists, the infrared can detect the crack information in time, thereby avoiding the occurrence of fire. This real-time monitoring capability significantly improves the response speed of the system and can discover safety hazards at the first time. By comprehensively using multi-band infrared images, deep learning models, feature point matching, three-dimensional reconstruction and gridding and other technical means, comprehensive detection of transformer tears is achieved, and detection efficiency and reliability are improved. This method can cover a variety of different application scenarios and ensure that it can work effectively in various situations. Real-time detection by using an infrared camera equipped with a UAV reduces the cost of manual detection and improves work efficiency. The UAV can cover a larger area, reduce manpower investment, and increase the detection frequency to ensure that potential problems are discovered in time. By timely detecting the tear, electrical failures and fire risks caused by the tear can be avoided. Not only does it improve the accuracy of detection, but it can also discover potential problems at an early stage, so that timely measures can be taken to repair them and prevent accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The present invention is a flow chart of a method for detecting a tear in a transformer oil tank in a fire scene according to an embodiment of the present invention.

[0052] Figure 2 The present invention is a block diagram of a system for detecting a tear in a transformer oil tank in a fire scene according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to facilitate those skilled in the art to understand the technical solution of the present invention, the technical solution of the present invention is further described in conjunction with the accompanying drawings of the specification.

[0054] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0055] See also Figure 1 As shown, the present invention provides a method for detecting a tear in a transformer oil tank in a fire scene, comprising:

[0056] S10, obtaining infrared sequence images of the transformer oil tank in different bands.

[0057] In one embodiment of the present invention, the drone carries a multi-band infrared camera to obtain infrared sequence images, including a first infrared sequence image applicable to a flame band and a second infrared sequence image applicable to a solid metal band, and the infrared lens of the multi-band infrared camera selects a 5-8μm filter lens.

[0058] In this embodiment, all objects will radiate heat energy to the surrounding environment. This heat energy exists in the form of infrared rays. Different objects have different temperatures and radiate different infrared energy. Different infrared lenses are selected according to the different infrared bands emitted by flames and solid metals. Here, a 5-8μm filter lens is selected to reduce the impact of flames. Multi-band infrared cameras are used to obtain infrared images of different frequency bands. These images cover different temperature ranges, allowing the system to distinguish between flames and metal surfaces. By comparing the changes in infrared intensity of the same area in different frequency bands, the contrast between the flame and the metal surface can be enhanced, thereby reducing the interference of flames on the temperature measurement of the metal surface.

[0059] S20, after preprocessing the infrared sequence images, extract the corner points of interest in the infrared sequence images, and obtain the corner points of the infrared sequence images in different bands respectively.

[0060] In one embodiment of the present invention, the infrared sequence images are preprocessed for clarity, including: first extracting the edge contour lines of the target objects from the infrared sequence images through the Canny edge detection operator, and then calculating the sharpness of the edge contour lines using an image sharpness algorithm to obtain key corner points in the infrared images.

[0061] In this embodiment, for each infrared image, edge detection is first performed and then corner point extraction is performed. This can be expressed as follows:

[0062] E λi =Canny(I λi );

[0063] C λi =Sharpness(E λi ); i = 1, 2;

[0064] Among them, E λi Represented as using the Canny edge detector on image I λi The result of edge detection, C λi Expressed as using the image sharpness algorithm Sharpness from the edge image E λi The set of corner points extracted from I λ1 Represented as the first infrared sequence image, I λ2 Represented as the second infrared sequence image.

[0065] In this embodiment, the Canny edge detection operator extracts a relatively clear edge contour line at a single pixel level, and then a simple and effective image sharpness algorithm is used to extract corner points.

[0066] S30, according to the corner points of the infrared sequence images in different bands, a motion recovery structure algorithm is applied to obtain a sparse three-dimensional model from the infrared sequence images, and the sparse three-dimensional model is gridded to obtain a dense three-dimensional model.

[0067] In one embodiment of the present invention, the structure is recovered from two images of different perspectives by using the SfM technology to obtain a three-dimensional point cloud:

[0068] P cloud =SfM(C λ1 , C λ2 );

[0069] In the formula, p cloud It is represented as a sparse 3D model obtained from infrared image sequences using the structure from motion algorithm SfM. λ1 It is represented as the set of corner points corresponding to the first infrared sequence image, C λ2 Represented as the set of corner points corresponding to the second infrared sequence image.

[0070] In this embodiment, the Crust algorithm is used to expand the sparse 3D model obtained from the infrared sequence image using the motion recovery structure algorithm to obtain a dense 3D model. The specific formula is as follows:

[0071] M 3D =Crust(P cloud );

[0072] Where M 3D Represented as a dense three-dimensional model, P cloud It is represented as a sparse 3D model obtained from infrared image sequences by the structure from motion algorithm SfM. cliud ) represents the application of the Crust algorithm to expand the sparse three-dimensional model obtained from the infrared sequence image using the motion recovery structure algorithm.

[0073] S40, establishing a fuel tank tear detection model, including a feature extraction network and a multimodal data fusion detection network; using the preprocessed infrared sequence image as the input of the feature extraction network, which outputs a two-dimensional splicing feature; then using the two-dimensional splicing feature and the dense three-dimensional model as the input of the multimodal data fusion detection network to obtain the position and size information of the tear.

[0074] In one embodiment of the present invention, there are now two infrared sequence images I in different wavebands. λ1 and I λ2, and dense 3D model, this information is input into the fuel tank tear detection model, which performs tear detection through multimodal feature fusion.

[0075] In this embodiment, a feature extraction network is first applied to obtain two-dimensional splicing features, which are obtained by the following formula:

[0076] F fusion =Concatenate(ConvNet(I λ1 ), ConvNet(I λ2 ));

[0077] In the formula, F fusion It is represented as a two-dimensional concatenation feature, Concatenate is represented as a concatenation function, ConvNet is represented as a convolutional neural network for feature extraction, ConvNet (I λ1 ) represents the feature extraction network to extract the features of the first infrared sequence image, ConvNet(I λ2 ) represents the feature extraction network extracting the second infrared sequence image.

[0078] Then use Mask R-CNN to detect and segment the tear. Mask R-CNN includes Region Proposal Network and Detection Head. Region Proposal Network generates candidate regions for the tear, and Detection Head classifies and regresses the candidate regions and generates segmentation masks for the tear. Specifically, the location and size information of the tear is obtained through the following formula:

[0079] R tear =MaskR-CNN(F fusion , M 3D );

[0080] In the formula, R tear It represents the location and size information of the tear, and Mask R-CNN represents the masked region convolutional neural network.

[0081] S50, training the established fuel tank tear detection model, and then using the trained fuel tank tear detection model to perform online detection. In this embodiment, the training method is the existing technology.

[0082] See also Figure 2 As shown, the present invention also provides a system for detecting a tear in a transformer oil tank in a fire scene, which uses the above-mentioned method for detecting a tear in a transformer oil tank in a fire scene, and comprises:

[0083] The sequence image module 10 is used to obtain infrared sequence images of the transformer oil tank in different bands.

[0084] The corner point module 20 is used to extract the corner points of interest in the infrared sequence images after preprocessing the infrared sequence images, and obtain the corner points of the infrared sequence images in different bands.

[0085] The three-dimensional model module 30 is used to obtain a sparse three-dimensional model from the infrared sequence images according to the corner points of the infrared sequence images in different bands by applying the motion recovery structure algorithm, and grid it to obtain a dense three-dimensional model.

[0086] The modeling module 40 is used to establish a fuel tank tear detection model, including a feature extraction network and a multimodal data fusion detection network; the preprocessed infrared sequence image is used as the input of the feature extraction network, which outputs a two-dimensional splicing feature; the two-dimensional splicing feature and the dense three-dimensional model are then used as the input of the multimodal data fusion detection network to obtain the location and size information of the tear.

[0087] The training and detection module 50 is used to train the established fuel tank tear detection model, and then use the trained fuel tank tear detection model to perform online detection.

[0088] In this embodiment, the infrared sequence images of the transformer oil tank in different bands are obtained by applying the following method:

[0089] The UAV carries a multi-band infrared camera to obtain infrared sequence images, including a first infrared sequence image applicable to a flame band and a second infrared sequence image applicable to a solid metal band; and the infrared lens of the multi-band infrared camera selects a 5-8μm filter lens.

[0090] In this embodiment, the corner points of infrared sequence images in different bands are obtained respectively, including: firstly extracting the edge contour line of the target object from the infrared sequence image through the Canny edge detection operator, and then calculating the sharpness of the edge contour line using the image sharpness algorithm to obtain the key corner points in the infrared image.

[0091] In this embodiment, obtaining a dense three-dimensional model includes: applying a Crust algorithm to expand a sparse three-dimensional model obtained from an infrared sequence image using a structure-from-motion algorithm to obtain a dense three-dimensional model.

[0092] In this embodiment, the position and size information of the tear is obtained by the following formula:

[0093] R tear =Mask R-CNN(F fusion , M 3D );

[0094] in,

[0095] F fusion =Concatenate(ConvNet(I λ1 ), ConvNet(I λ2 ));

[0096] M 3D =Crust(P cloud );

[0097] P cloud =SfM(C λ1 , C λ2 0;

[0098] C λi =Sharpness(E λi ); i = 1, 2;

[0099] E λi =Canny(I λi );

[0100] In the formula, R tear It represents the location and size of the tear, Mask R-CNN represents the mask region convolutional neural network, and F fusion It is represented as a two-dimensional concatenation feature, Concatenate is represented as a concatenation function, ConvNet is represented as a convolutional neural network for feature extraction, ConvNet (I λ1 ) represents the feature extraction network to extract the features of the first infrared sequence image, ConvNet(I λ2 ) represents the feature extraction network extracting the second infrared sequence image, I λ1 Represented as the first infrared sequence image, I λ2 Represented as the second infrared sequence image, M 3D Represented as a dense three-dimensional model, P cloud It is represented as a sparse 3D model obtained from infrared image sequences using the structure from motion algorithm SfM. cloud ) represents the expansion of the 3D point cloud obtained from the infrared sequence images using the Crust algorithm. λi Represented as using the Canny edge detector on image I λi The result of edge detection, C λi Expressed as using the image sharpness algorithm Sharpness from the edge image E λi The set of corner points extracted from .

[0101] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0102] The above-described embodiments merely represent implementation methods of the invention. The protection scope of the present invention is not limited to the above-described embodiments. For those skilled in the art, several modifications and improvements may be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for detecting a tear in a transformer oil tank in a fire scene, characterized in that: include: Obtain infrared sequence images of transformer oil tank in different bands; After preprocessing the infrared sequence images, the corner points of interest in the infrared sequence images are extracted, and the corner points of the infrared sequence images in different bands are obtained respectively; According to the corner points of infrared sequence images in different bands, the motion recovery structure algorithm is used to obtain a sparse three-dimensional model from the infrared sequence images, and then gridding is performed to obtain a dense three-dimensional model; A fuel tank tear detection model is established, including a feature extraction network and a multimodal data fusion detection network. The preprocessed infrared sequence image is used as the input of the feature extraction network, which outputs a two-dimensional splicing feature. Then, the 2D splicing features and the dense 3D model are used as the input of the multimodal data fusion detection network to obtain the location and size information of the tear. The established fuel tank tear detection model is trained, and then the trained fuel tank tear detection model is used for online detection.

2. The method for detecting a tear in a transformer oil tank in a fire scene according to claim 1, characterized in that: To obtain infrared sequence images of the transformer tank in different bands, the following methods are applied: The UAV carries a multi-band infrared camera to obtain infrared sequence images, including a first infrared sequence image applicable to a flame band and a second infrared sequence image applicable to a solid metal band; and the infrared lens of the multi-band infrared camera selects a 5-8μm filter lens.

3. The method for detecting a tear in a transformer oil tank in a fire scene according to claim 1, characterized in that: The corner points of infrared sequence images in different bands are obtained respectively, including: firstly extracting the edge contour line of the target object from the infrared sequence images through the Canny edge detection operator, and then calculating the sharpness of the edge contour line by using the image sharpness algorithm to obtain the key corner points in the infrared image.

4. The method for detecting a tear in a transformer oil tank in a fire scene according to claim 1, characterized in that: Acquiring a dense three-dimensional model includes: applying a Crust algorithm to expand a sparse three-dimensional model obtained from an infrared sequence image by applying a motion recovery structure algorithm to acquire a dense three-dimensional model.

5. The method for detecting a tear in a transformer oil tank in a fire scene according to claim 1, characterized in that: Get the location and size information of the tear through the following formula: R tear =MaskR-CNN(F fusion ,M 3D ); in, F fusion =Concatenate(ConvNet(I λ1 ),ConvNet(I λ2 )); M 3D =Crust(P cloud ); P cloud =SfM(C λ1 ,C λ2 0; C λi =Sharpness(E λi );i=1,2; E λi =Canny(I λi ); In the formula, R tear It represents the location and size information of the tear, MaskR-CNN represents the mask region convolutional neural network, and F fusion It is represented as a two-dimensional concatenation feature, Concatenate is represented as a concatenation function, ConvNet is represented as a convolutional neural network for feature extraction, ConvNet (I λ1 ) represents the feature extraction network to extract the features of the first infrared sequence image, ConvNet(I λ2 ) represents the feature extraction network extracting the second infrared sequence image, I λ1 Represented as the first infrared sequence image, I λ2 Represented as the second infrared sequence image, M 3D Represented as a dense three-dimensional model, P cloud It is represented as a sparse 3D model obtained from infrared image sequences using the structure from motion algorithm SfM. cloud ) represents the expansion of the 3D point cloud obtained from the infrared sequence images using the Crust algorithm. λi Represented as using the Canny edge detector on image I λi The result of edge detection, C λi Expressed as using the image sharpness algorithm Sharpness from the edge image E λi The set of corner points extracted from .

6. A system using the method for detecting a tear in a transformer oil tank in a fire scene according to any one of claims 1 to 5, characterized in that: include: Sequence image module, used to obtain infrared sequence images of transformer oil tank in different bands; The corner point module is used to extract the corner points of interest in the infrared sequence images after preprocessing the infrared sequence images, and obtain the corner points of the infrared sequence images in different bands respectively; The 3D model module is used to obtain a sparse 3D model from the infrared sequence images according to the corner points of the infrared sequence images in different bands by applying the motion recovery structure algorithm, and to grid the sparse 3D model to obtain a dense 3D model. A modeling module is used to establish a fuel tank tear detection model, including a feature extraction network and a multimodal data fusion detection network; the preprocessed infrared sequence image is used as the input of the feature extraction network, which outputs a two-dimensional splicing feature; Then, the 2D splicing features and the dense 3D model are used as the input of the multimodal data fusion detection network to obtain the location and size information of the tear. The training and detection module is used to train the established fuel tank tear detection model, and then use the trained fuel tank tear detection model for online detection.

7. The transformer oil tank tear detection system in a fire scene according to claim 6, characterized in that: To obtain infrared sequence images of the transformer tank in different bands, the following methods are applied: The UAV carries a multi-band infrared camera to obtain infrared sequence images, including a first infrared sequence image applicable to a flame band and a second infrared sequence image applicable to a solid metal band; and the infrared lens of the multi-band infrared camera selects a 5-8μm filter lens.

8. The transformer oil tank tear detection system in a fire scene according to claim 6, characterized in that: The corner points of infrared sequence images in different bands are obtained respectively, including: firstly extracting the edge contour line of the target object from the infrared sequence images through the Canny edge detection operator, and then calculating the sharpness of the edge contour line by using the image sharpness algorithm to obtain the key corner points in the infrared image.

9. The transformer oil tank tear detection system in a fire scene according to claim 6, characterized in that: Acquiring a dense three-dimensional model includes: applying a Crust algorithm to expand a sparse three-dimensional model obtained from an infrared sequence image by applying a motion recovery structure algorithm to acquire a dense three-dimensional model.

10. The transformer oil tank tear detection system in a fire scene according to claim 6, characterized in that: Get the location and size information of the tear through the following formula: R tear =MaskR-CNN(F fusion ,M 3D ); in, F fusion =Concatenate(ConvNet(I λ1 ),ConvNet(I λ2 )); M 3D =Crust(P cloud ); P cloud =SfM(C λ1 ,C λ2 ); C λi =Sharpness(E λi );i=1,2; E λi =Canny(I λi ); In the formula, R tear It represents the location and size information of the tear, MaskR-CNN represents the mask region convolutional neural network, and F fusion It is represented as a two-dimensional concatenation feature, Concatenate is represented as a concatenation function, ConvNet is represented as a convolutional neural network for feature extraction, ConvNet (I λ1 ) represents the feature extraction network to extract the features of the first infrared sequence image, ConvNet(I λ2 ) represents the feature extraction network extracting the second infrared sequence image, I λ1 Represented as the first infrared sequence image, I λ1 Represented as the second infrared sequence image, M 3D Represented as a dense three-dimensional model, P cloud It is represented as a sparse 3D model obtained from infrared image sequences using the structure from motion algorithm SfM. cloud ) represents the expansion of the 3D point cloud obtained from the infrared sequence images using the Crust algorithm. λi Represented as using the Canny edge detector on image I λi The result of edge detection, C λi Expressed as using the image sharpness algorithm Sharpness from the edge image E λi The set of corner points extracted from .

Images