Offshore rocket platform fire risk identification method based on AI image identification

Abstract

The invention discloses a sea rocket platform fire risk identification method based on AI image identification. The sea rocket platform fire risk identification method comprises the following steps of performing timestamp labeling and spatial position calibration on a sea rocket platform image sequence; s2, obtaining a preprocessed marine rocket platform image sequence; s3, obtaining a frequency domain image spectrum sequence for frequency domain feature extraction; s4, inputting the preprocessed maritime rocket platform image sequence into a spatial domain Transform branch, and taking the extracted spatial domain image features as spatial domain feature representation; s5, representing the extracted frequency domain image features as frequency domain features; s6, inputting the spatial domain feature representation and the frequency domain feature representation into a mixed feature fusion module to obtain a fused feature representation; and S7, obtaining an offshore rocket platform risk identification result. According to the method, flames and background reflection can still be accurately separated under the condition of complex weather or low contrast, and the fire risk identification effect of the offshore rocket platform is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of offshore rocket platforms, and in particular to a method for identifying fire risks of offshore rocket platforms based on AI image recognition. Background Art

[0002] With the transfer of space launch missions to offshore platforms, the safety risk management issues of offshore rocket platforms in complex environments have become increasingly prominent, especially in the identification of sudden fire risks caused by flames and smoke. The technical difficulty is significantly higher than that of land-based launch platforms. Since offshore platforms are often in dynamic scenes with strong winds, high humidity, and drastic light changes, the traditional fire monitoring methods based on single image processing algorithms or fixed threshold rules have significantly reduced recognition accuracy under the interference of sea conditions.

[0003] Currently, offshore rocket platforms usually deploy multi-angle camera systems to collect environmental images and use traditional image analysis methods such as edge detection, color model discrimination or background difference method to identify flames and smoke. Traditional methods are highly dependent on the environment and often make false detections or missed detections due to changes in lighting, sparse diffusion of smoke or blurred flames. For example, when the sunlight at sea is strong or the equipment is tilted, resulting in imaging distortion, it is difficult for the system to accurately extract feature areas, and it is easy to misjudge sunlight reflections and water surface vapor as flames or smoke; at the same time, when the fire is still in the early and weak stage, the response sensitivity of traditional methods is insufficient, and it is difficult to capture subtle signs of fire in time.

[0004] In addition, existing recognition algorithms generally rely on direct processing of spatial domain images and lack the use of image frequency domain information, resulting in limited modeling capabilities for low-frequency diffusion targets of smoke, unclear recognition boundaries, and inaccurate regional positioning. Some studies have attempted to introduce deep neural networks for discrimination, but they usually use a single modality as input, making it difficult to comprehensively consider the spatial texture and spectral characteristics of the image. The model's generalization ability and interpretability for targets still need to be improved.

[0005] In summary, the existing fire identification technology for offshore rocket platforms still has obvious deficiencies in dynamic environment adaptability, feature extraction integrity, and modal information fusion capabilities. It is urgent to propose a new identification method to solve the technical problems of insufficient multi-source image information fusion, low recognition accuracy, and high sensitivity to environmental interference. Summary of the Invention

[0006] One purpose of the present invention is to propose a method for identifying fire risks of offshore rocket platforms based on AI image recognition. The present invention can accurately separate flames and background reflections under complex weather or low-contrast conditions, thereby improving the effect of identifying fire risks of offshore rocket platforms.

[0007] According to an embodiment of the present invention, a method for identifying fire risks of a marine rocket platform based on AI image recognition includes the following steps:

[0008] S1. Acquire an image sequence of an offshore rocket platform, wherein the image sequence is a continuous frame image acquired under different sea conditions and lighting conditions, and perform time stamping and spatial position calibration on the image sequence;

[0009] S2. Preprocessing the image sequence of the offshore rocket platform to obtain a preprocessed image sequence of the offshore rocket platform;

[0010] S3. Convert the preprocessed marine rocket platform image sequence into the frequency domain based on Fourier transform, generate a frequency domain image spectrum sequence, and perform amplitude normalization to obtain a frequency domain image spectrum sequence for frequency domain feature extraction;

[0011] S4. Construct a hybrid frequency-domain and spatial-domain Transformer structure, input the preprocessed maritime rocket platform image sequence into the spatial-domain Transformer branch, extract spatial-domain image features through the spatial-domain Transformer branch, and use the extracted spatial-domain image features as the spatial-domain feature representation.

[0012] S5. Input the frequency domain image spectrum sequence into the frequency domain Transformer branch, extract the frequency domain image features through the frequency domain Transformer branch, and use the extracted frequency domain image features as the frequency domain feature representation;

[0013] S6. Input the spatial domain feature representation and the frequency domain feature representation into the hybrid feature fusion module, and realize the dynamic interaction and fusion of spatial domain features and frequency domain features through the cross-branch self-attention mechanism to obtain the fused feature representation;

[0014] S7. Input the fused feature representation into the classification and discrimination module built based on the deep classification network, and use the classification and discrimination module to jointly identify the flames and smoke of the offshore rocket platform to obtain the risk identification results of the offshore rocket platform.

[0015] Optionally, the S1 includes the following steps:

[0016] S11. Deploy multiple sets of imaging and acquisition equipment on the offshore rocket platform, set acquisition point locations based on the spatial coordinates of different monitoring areas, form an offshore rocket platform image acquisition network covering key areas of the offshore rocket platform, and uniformly schedule the offshore rocket platform image acquisition time to obtain the original monitoring offshore rocket platform image sequence I. raw ={I t |t∈T}, where I t represents a frame of offshore rocket platform image collected at time point t, and T represents the time series of offshore rocket platform image collection;

[0017] S12. The original monitoring image sequence of the sea rocket platform Iraw Input timestamp annotation module, according to the global synchronous clock, each frame of the sea rocket platform image I t Add a unique time stamp τ t , constructing a time-tagged image sequence for monitoring a sea rocket platform I timestamp ={(I t ,τ t )|t∈T}, where Representing time scalars in the real number domain to maintain the temporal continuity of images of the maritime rocket platform and the consistency of dynamic event analysis;

[0018] S13. Image sequence I for monitoring the offshore rocket platform timestamp Perform spatial position calibration based on the fixed spatial coordinates of each imaging acquisition device Indicates the position of the imaging acquisition device and imaging direction parameters Indicates the orientation of the imaging acquisition device, that is, the rotation angle around the XYZ axis, and the image pose matrix of the offshore rocket platform By the rotation matrix R i With the translation vector T i Composition, where R i =θ i , T i =[x i ,y i ,z i ] T ;

[0019] S14. Based on the image pose matrix P of the offshore rocket platform i Perform perspective geometric correction on the image frames of the offshore rocket platform, reconstruct the distribution relationship of the monitoring offshore rocket platform image sequence in a unified spatial reference system, and generate the spatial calibration offshore rocket platform image sequence I calibrated ={(I t ,τ t ,P i )|t∈T}.

[0020] Optionally, S3 includes the following steps:

[0021] S31. Preprocess the image sequence I of the sea rocket platform pre Input frequency domain transformation module to pre-process the image of the sea rocket platform for each frame I t Applying the two-dimensional discrete Fourier transform operation, the pixel information of the marine rocket platform image in the spatial domain is mapped to the complex spectrum representation in the frequency domain, and the frequency domain image spectrum F is obtained. t (u,v), where F t (u, v) represents the complex amplitude of the image of the sea rocket platform in the horizontal and vertical directions corresponding to frequencies u and v respectively;

[0022] S32. Based on the frequency domain image spectrum F t (u,v) calculates its corresponding amplitude spectrum A t (u, v), the amplitude spectrum is obtained by squaring the real and imaginary parts of the frequency domain image spectrum, summing them up and then taking the square root. The amplitude spectrum A t (u, v) represents the energy of the sea rocket platform image at the frequency position (u, v), reflecting the response intensity of the pre-processed sea rocket platform image at different frequency components;

[0023] S33. Amplitude spectrum A t (u, v) is jointly logarithmically scaled and normalized to its maximum value. By taking the logarithm of each amplitude value and dividing it by the maximum value of all the logarithmized amplitude values in the same frame of the sea rocket platform image, strong responses are suppressed and weak textures are enhanced to obtain the normalized frequency domain amplitude spectrum.

[0024] S34. Normalize the frequency domain amplitude spectrum of each frame The corresponding time scalar τ t and the marine rocket platform image pose matrix P i Combining to form frequency domain image spectrum sequence

[0025] Optionally, the S4 includes the following steps:

[0026] S41. Preprocess the image sequence I of the sea rocket platform pre Input the spatial domain Transformer branch and perform a transformation on each frame of the sea rocket platform image. Perform fixed window division processing and divide it into a set of image blocks of size P×P in is the number of image blocks, represents the kth image block;

[0027] S42. For each image block p t,k After flattening, it is projected into an embedding vector through linear mapping And add the sea rocket platform image pose matrix P i The solved position code Construct an input sequence X with spatial orientation awareness t :

[0028]

[0029] Positional encoding and the pose matrix P of the sea rocket platform image i Related, used to identify the orientation and physical perspective of the image patch in the space of the sea rocket platform;

[0030] S43. Construct a flame-smoke attention guidance function to guide the spatial domain Transformer branch to enhance the attention of key areas in high-light and weak-boundary scenes. The flame-smoke attention guidance function is based on the average brightness of the image block l t,k and edge gradient magnitude g t,k Constructing the image patch attention bias term δ t,k , the image block attention bias term is used as a saliency factor to reflect the saliency of the potential flame or smoke area in the image block:

[0031]

[0032] in g t,k is the mean gradient of the edge of the image block, α, β∈[0,1] are balance parameters;

[0033] S44. Introduce the saliency bias matrix Δ in the multi-head attention calculation of each layer in the spatial domain Transformer branch t , the i,j element of the saliency bias matrix is Δ t (i,j)=γ·(δ t,i +δ t,j ), where γ is the amplification factor, which is used to adjust the focus of the attention map on the flame-smoke salient area, and finally adjust the spatial domain Transformer multi-head attention module:

[0034]

[0035] Among them, Q 空域 is the spatial query matrix, K 空域 is the spatial bond matrix, V 空域 is the space value matrix, is the spatial domain attention score matrix, which represents the attention weight of image block i to image block j. The attention score considers the inner product relationship between the feature vectors of the two image blocks and introduces the significance bias matrix Δ t (i, j) Enhance the characteristic transmission capability of the flame-smoke key area;

[0036] S45. Repeat the L-layer encoding process constructed by combining the spatial attention score matrix and the feedforward network to obtain the output feature sequence of the spatial transformer. Represents the spatial feature expression after saliency guidance;

[0037] S46. Output feature sequence Input flame-smoke feature reconstruction module, based on the physical properties of flames in marine environment with bright high-frequency boundary changes and smoke with fuzzy low-frequency diffusion characteristics, define the flame response weight matrix W fire, smoke response weight matrix W smoke , respectively extract the flame space features and smoke space features:

[0038]

[0039] S47. Flame space characteristics and smoke spatial characteristics Concatenate by channel dimension to construct spatial feature representation:

[0040]

[0041] The spatial feature representation A fusion perception mechanism is used to jointly construct a collaborative recognition model with frequency domain features to achieve separable modeling and saliency enhancement of flame and smoke features under the complex background of offshore rocket platforms.

[0042] Optionally, the S5 includes the following steps:

[0043] S51. The frequency domain image spectrum sequence I freq Input frequency domain Transformer branch, normalize the frequency domain amplitude spectrum A dual-scale sliding window is used to perform local scale division and global scale division respectively to form a local frequency domain sub-block set in, is the local frequency domain sub-block, K1 is the number of local frequency domain sub-blocks;

[0044] S52. Extract frequency domain feature representation Obtain spatial guidance vectors by mean pooling And the spatial guidance vector is passed through the cross-modal guidance mapping matrix W cross Mapped to frequency domain steering vector Used to guide the attention focus of local features in the frequency domain;

[0045] S53. Calculate local frequency domain sub-block set Each local frequency domain sub-block The average amplitude And define the spectral significance factor

[0046]

[0047] Spectral significance factor Constructed as a significant bias matrix The elements of the significance bias matrix are defined as follows:

[0048]

[0049] in, Represents the spectrum energy value of the normalized frequency domain amplitude spectrum at the frequency coordinate (u, v), Indicates the kth local frequency domain sub-block in the tth frame image The average amplitude spectrum value of , P1 represents the side length of the local frequency domain sub-block, μ is the significance scaling coefficient, At time point t, Indicates the maximum value of the average amplitude value in all local frequency domain sub-blocks of the t-th frame image, is the (i, j)th element of the saliency bias matrix constructed in the local frequency domain self-attention module for the t-th frame image, indicating the degree of saliency linkage between the i-th and j-th frequency domain sub-blocks;

[0050] S54. Construct the frequency domain Transformer multi-head attention module through the saliency bias matrix and frequency domain guidance vector:

[0051]

[0052] Among them, Q 频域 is the frequency domain query matrix, K 频域 is the frequency domain key matrix, V 频域 is the frequency domain value matrix;

[0053] S55. Repeat the frequency domain attention score matrix and the feedforward layer combination to build an L-layer local frequency domain Transformer encoding structure and output the frequency domain local feature sequence Constructing frequency domain feature representation

[0054] Optionally, the S6 includes the following steps:

[0055] S61. Representing spatial features and frequency domain feature representation Input fusion channel conversion module, respectively, through independent linear transformation method to project the spatial domain feature representation and frequency domain feature representation into the unified dimensional fusion feature space, forming the mapped spatial domain feature representation And the mapped frequency domain feature representation The two have the same vector dimension in the fusion feature space;

[0056] S62. Represent the mapped spatial features As the query vector, the mapped frequency domain features are represented as As key vectors and value vectors, the spatial feature output of frequency domain enhancement is constructed through the heterogeneous source attention mechanism. It is used to capture the complementary effect of frequency domain information on spatial domain information and introduce frequency domain energy and pattern characteristics based on spatial structure;

[0057] S63. Represent the mapped frequency domain features As the query vector, the mapped spatial feature representation As key vectors and value vectors, the frequency domain feature output of spatial enhancement is constructed through the heterogeneous source attention mechanism. It is used to model the complementary effect of spatial domain structure on spectral content and introduce guidance information of target structure and texture features based on frequency domain distribution;

[0058] S64. Represent the mapped spatial features Spatial feature output with frequency domain enhancement Splicing by channel dimension to form enhanced spatial feature representation The enhanced spatial feature representation integrates the significant response area information guided by the frequency domain while maintaining the original spatial structure expression;

[0059] S65. Represent the mapped frequency domain features Frequency domain feature output with spatial domain enhancement Splicing by channel dimension to form enhanced frequency domain feature representation The enhanced frequency domain feature representation integrates the modulation effect of the spatial domain structure on the local frequency response while maintaining the original spectrum information expression;

[0060] S66. Enhanced spatial feature representation and enhanced frequency domain feature representation Perform the maximum pooling operation on all image block dimensions to extract the enhanced spatial feature vector and enhanced frequency domain eigenvector The vector is used to compress spatial information and frequency information into a global representation of uniform length to support subsequent fusion classification tasks;

[0061] S67. Enhance the spatial feature vector and enhanced frequency domain eigenvector Splice to form a fusion feature representation

[0062] Optionally, the S7 includes the following steps:

[0063] S71. Represent the fusion features Input classification and discrimination module, which is a discrimination system built on a multi-task deep classification network and has a flame recognition submodule and a smoke recognition submodule, respectively completing the parallel judgment of flame features and smoke features in the image of the offshore rocket platform;

[0064] S72. Represent the fusion features The flame recognition submodule and the smoke recognition submodule are input in sequence, and the flame recognition output results are obtained through the nonlinear activation function and the fully connected layer respectively. Smoke recognition output results in, Indicates the confidence that there is a flame in the t-th frame image, Indicates the confidence that there is smoke in the t-th frame image;

[0065] S73. Output results based on flame recognition Smoke recognition output results Combined with the preset threshold rules for joint judgment, the fire risk identification rules for offshore rocket platforms are constructed, and the flame existence threshold θ is defined. fire , smoke presence threshold θ smoke , and divide the flame and smoke states of the t-th frame offshore rocket platform image into safe state, abnormal state and high-risk state;

[0066] S74. Based on the state classification result, the t-th frame of the offshore rocket platform image is output as the flame recognition result Smoke recognition results and its corresponding confidence level, and output the final risk identification results of the offshore rocket platform

[0067] Optionally, the safety state: meets the flame recognition output result And the smoke recognition output result The abnormal state: meets the flame recognition output result And the smoke recognition output result Or flame recognition output results And the smoke recognition output result The high-risk state: meets the flame recognition output result And the smoke recognition output result

[0068] The beneficial effects of the present invention are:

[0069] (1) The present invention adopts a hybrid Transformer structure that fuses frequency domain and spatial domain, innovatively introduces a frequency domain feature extraction branch, and constructs a cross-modal heterogeneous attention fusion mechanism. Through frequency domain Fourier spectrum energy extraction, the system can capture the high-frequency boundary characteristics of flames and the low-frequency diffusion characteristics of smoke. A saliency guidance mechanism is constructed on the spatial domain Transformer side to highlight the areas with fire risks in the image. At the same time, a spectral saliency bias term is designed on the frequency domain side to guide the attention focus at the frequency level. The bidirectional heterogeneous attention mechanism in the fusion module allows the frequency domain and spatial domain to dynamically interact and jointly optimize at the structural level, effectively making up for the problem that traditional methods are insufficient in responding to fuzzy smoke and weak flames.

[0070] (2) The “flame-smoke saliency guidance mechanism” proposed in this paper constructs an attention bias matrix in the spatial domain Transformer branch by jointly constructing the image block brightness mean and edge gradient amplitude. Combined with the physical characteristics of “dramatic changes in the bright edges of flames and weak diffusion of blurred smoke boundaries” in marine fire images, a regional-level saliency attention guidance function is constructed. The attention weight distribution of potential abnormal areas is strengthened at the multi-head attention level, which can dynamically improve the robustness of the model in strong sunlight reflection, water vapor interference and low-light night scenes, especially in complex weather or low-contrast conditions, it can still accurately separate flames and background reflections.

[0071] (3) The classification and discrimination module of the present invention is based on a dual-branch multi-task deep classification network, which constructs independent recognition sub-modules for flame and smoke features respectively, and makes risk level judgments through a joint confidence judgment mechanism. The whole process mapping from feature extraction to risk judgment is automatically completed through an end-to-end learning method. At the same time, by introducing a risk joint judgment strategy of flame existence threshold and smoke existence threshold, the recognition sensitivity and response speed of cross-anomalies are effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0073] Figure 1 This is a flow chart of a method for identifying fire risks of offshore rocket platforms based on AI image recognition proposed by the present invention. DETAILED DESCRIPTION

[0074] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0075] refer to Figure 1 A method for identifying fire risks of offshore rocket platforms based on AI image recognition includes the following steps:

[0076] S1. Acquire an image sequence of an offshore rocket platform, wherein the image sequence is a continuous frame image acquired under different sea conditions and lighting conditions, and perform time stamping and spatial position calibration on the image sequence;

[0077] S2. Preprocessing the image sequence of the offshore rocket platform to obtain a preprocessed image sequence of the offshore rocket platform;

[0078] S3. Convert the preprocessed marine rocket platform image sequence into the frequency domain based on Fourier transform, generate a frequency domain image spectrum sequence, and perform amplitude normalization to obtain a frequency domain image spectrum sequence for frequency domain feature extraction;

[0079] S4. Construct a hybrid frequency-domain and spatial-domain Transformer structure, input the preprocessed maritime rocket platform image sequence into the spatial-domain Transformer branch, extract spatial-domain image features through the spatial-domain Transformer branch, and use the extracted spatial-domain image features as the spatial-domain feature representation.

[0080] S5. Input the frequency domain image spectrum sequence into the frequency domain Transformer branch, extract the frequency domain image features through the frequency domain Transformer branch, and use the extracted frequency domain image features as the frequency domain feature representation;

[0081] S6. Input the spatial domain feature representation and the frequency domain feature representation into the hybrid feature fusion module, and realize the dynamic interaction and fusion of spatial domain features and frequency domain features through the cross-branch self-attention mechanism to obtain the fused feature representation;

[0082] S7. Input the fused feature representation into the classification and discrimination module built based on the deep classification network, and use the classification and discrimination module to jointly identify the flames and smoke of the offshore rocket platform to obtain the risk identification results of the offshore rocket platform.

[0083] In this embodiment, S1 includes the following steps:

[0084] S11. Deploy multiple sets of imaging and acquisition equipment on the offshore rocket platform, set acquisition point locations based on the spatial coordinates of different monitoring areas, form an offshore rocket platform image acquisition network covering key areas of the offshore rocket platform, and uniformly schedule the offshore rocket platform image acquisition time to obtain the original monitoring offshore rocket platform image sequence I. raw ={I t |t∈T}, where I t represents a frame of offshore rocket platform image collected at time point t, and T represents the time series of offshore rocket platform image collection;

[0085] S12. The original monitoring image sequence of the sea rocket platform I raw Input timestamp annotation module, according to the global synchronous clock, each frame of the sea rocket platform image I t Add a unique time stamp τ t , constructing a time-tagged image sequence for monitoring a sea rocket platform I timestamp ={(I t ,τ t )|t∈T}, where Representing time scalars in the real number domain to maintain the temporal continuity of images of the maritime rocket platform and the consistency of dynamic event analysis;

[0086] S13. Image sequence I for monitoring the offshore rocket platform timestamp Perform spatial position calibration based on the fixed spatial coordinates of each imaging acquisition device Indicates the position of the imaging acquisition device and imaging direction parameters Indicates the orientation of the imaging acquisition device, that is, the rotation angle around the XYZ axis, and the image pose matrix of the offshore rocket platform By the rotation matrix R i With the translation vector T i Composition, where R i =θ i , T i =[x i ,y i ,z i ] T ;

[0087] S14. Based on the image pose matrix P of the offshore rocket platform i Perform perspective geometric correction on the image frames of the offshore rocket platform, reconstruct the distribution relationship of the monitoring offshore rocket platform image sequence in a unified spatial reference system, and generate the spatial calibration offshore rocket platform image sequence I calibrated ={(I t ,τ t ,P i )|t∈T}.

[0088] In this embodiment, S3 includes the following steps:

[0089] S31. Preprocess the image sequence I of the sea rocket platform pre Input frequency domain transformation module to pre-process the image of the sea rocket platform for each frame I t Applying the two-dimensional discrete Fourier transform operation, the pixel information of the marine rocket platform image in the spatial domain is mapped to the complex spectrum representation in the frequency domain, and the frequency domain image spectrum F is obtained. t (u,v), where F t (u, v) represents the complex amplitude of the image of the sea rocket platform in the horizontal and vertical directions corresponding to frequencies u and v respectively;

[0090] S32. Based on the frequency domain image spectrum F t (u,v) calculates its corresponding amplitude spectrum A t (u, v), the amplitude spectrum is obtained by squaring the real and imaginary parts of the frequency domain image spectrum, summing them up and then taking the square root. The amplitude spectrum A t(u, v) represents the energy of the sea rocket platform image at the frequency position (u, v), reflecting the response intensity of the pre-processed sea rocket platform image at different frequency components;

[0091] S33. Amplitude spectrum A t (u, v) is jointly logarithmically scaled and normalized to its maximum value. By taking the logarithm of each amplitude value and dividing it by the maximum value of all the logarithmized amplitude values in the same frame of the sea rocket platform image, strong responses are suppressed and weak textures are enhanced to obtain the normalized frequency domain amplitude spectrum.

[0092] S34. Normalize the frequency domain amplitude spectrum of each frame The corresponding time scalar τ t and the marine rocket platform image pose matrix P i Combining to form frequency domain image spectrum sequence

[0093] In this embodiment, S4 includes the following steps:

[0094] S41. Preprocess the image sequence I of the sea rocket platform pre Input the spatial domain Transformer branch and perform a transformation on each frame of the sea rocket platform image. Perform fixed window division processing and divide it into a set of image blocks of size P×P in is the number of image blocks, represents the kth image block;

[0095] S42. For each image block p t,k After flattening, it is projected into an embedding vector through linear mapping And add the sea rocket platform image pose matrix P i The solved position code Construct an input sequence X with spatial orientation awareness t :

[0096]

[0097] Positional encoding and the image pose matrix P of the offshore rocket platform i Related, used to identify the orientation and physical perspective of the image patch in the space of the sea rocket platform;

[0098] S43. Construct a flame-smoke attention guidance function to guide the spatial domain Transformer branch to enhance the attention of key areas in high-light and weak-boundary scenes. The flame-smoke attention guidance function is based on the average brightness of the image block l t,k and edge gradient magnitude g t,kConstructing the image patch attention bias term δ t,k , the image block attention bias term is used as a saliency factor to reflect the saliency of the potential flame or smoke area in the image block:

[0099]

[0100] in g t,k is the mean gradient of the edge of the image block, α, β∈[0,1] are balance parameters;

[0101] S44. Introduce the saliency bias matrix Δ in the multi-head attention calculation of each layer in the spatial domain Transformer branch t , the i,j element of the saliency bias matrix is Δ t (i,j)=γ·(δ t,i +δ t,j ), where γ is the amplification factor, which is used to adjust the focus of the attention map on the flame-smoke salient area, and finally adjust the spatial domain Transformer multi-head attention module:

[0102]

[0103] Among them, Q 空域 is the spatial query matrix, K 空域 is the spatial bond matrix, V 空域 is the space value matrix, is the spatial domain attention score matrix, which represents the attention weight of image block i to image block j. The attention score considers the inner product relationship between the feature vectors of the two image blocks and introduces the significance bias matrix Δ t (i, j) Enhance the characteristic transmission capability of the flame-smoke key area;

[0104] S45. Repeat the L-layer encoding process constructed by combining the spatial attention score matrix and the feedforward network to obtain the output feature sequence of the spatial transformer. Represents the spatial feature expression after saliency guidance;

[0105] S46. Output feature sequence Input flame-smoke feature reconstruction module, based on the physical properties of flames in marine environment with bright high-frequency boundary changes and smoke with fuzzy low-frequency diffusion characteristics, define the flame response weight matrix W fire , smoke response weight matrix W smoke , respectively extract the flame space features and smoke space features:

[0106]

[0107] S47. Flame space characteristics and smoke spatial characteristics Concatenate by channel dimension to construct spatial feature representation:

[0108]

[0109] Spatial feature representation A fusion perception mechanism is used to jointly construct a collaborative recognition model with frequency domain features to achieve separable modeling and saliency enhancement of flame and smoke features under the complex background of offshore rocket platforms.

[0110] In this embodiment, S5 includes the following steps:

[0111] S51. The frequency domain image spectrum sequence I freq Input frequency domain Transformer branch, normalize the frequency domain amplitude spectrum A dual-scale sliding window is used to perform local scale division and global scale division respectively to form a local frequency domain sub-block set in, is the local frequency domain sub-block, K1 is the number of local frequency domain sub-blocks;

[0112] S52. Extract frequency domain feature representation Obtain spatial guidance vectors by mean pooling And the spatial guidance vector is passed through the cross-modal guidance mapping matrix W cross Mapped to frequency domain steering vector Used to guide the attention focus of local features in the frequency domain;

[0113] S53. Calculate local frequency domain sub-block set Each local frequency domain sub-block The average amplitude And define the spectral significance factor

[0114]

[0115] Spectral significance factor Constructed as a significant bias matrix The elements of the significance bias matrix are defined as follows:

[0116]

[0117] in, Represents the spectrum energy value of the normalized frequency domain amplitude spectrum at the frequency coordinate (u, v), Indicates the kth local frequency domain sub-block in the tth frame image The average amplitude spectrum value of , P1 represents the side length of the local frequency domain sub-block, μ is the significance scaling coefficient, At time point t, Indicates the maximum value of the average amplitude value in all local frequency domain sub-blocks of the t-th frame image, is the (i, j)th element of the saliency bias matrix constructed in the local frequency domain self-attention module for the t-th frame image, indicating the degree of saliency linkage between the i-th and j-th frequency domain sub-blocks;

[0118] S54. Construct the frequency domain Transformer multi-head attention module through the saliency bias matrix and frequency domain guidance vector:

[0119]

[0120] Among them, Q 频域 is the frequency domain query matrix, K 频域 is the frequency domain key matrix, V 频域 is the frequency domain value matrix;

[0121] S55. Repeat the frequency domain attention score matrix and the feedforward layer combination to build an L-layer local frequency domain Transformer encoding structure and output the frequency domain local feature sequence Constructing frequency domain feature representation

[0122] In this embodiment, S6 includes the following steps:

[0123] S61. Representing spatial features and frequency domain feature representation Input fusion channel conversion module, respectively, through independent linear transformation method to project the spatial domain feature representation and frequency domain feature representation into the unified dimensional fusion feature space, forming the mapped spatial domain feature representation And the mapped frequency domain feature representation The two have the same vector dimension in the fusion feature space;

[0124] S62. Represent the mapped spatial features As the query vector, the mapped frequency domain features are represented as As key vectors and value vectors, the spatial feature output of frequency domain enhancement is constructed through the heterogeneous source attention mechanism. It is used to capture the complementary effect of frequency domain information on spatial domain information and introduce frequency domain energy and pattern characteristics based on spatial structure;

[0125] S63. Represent the mapped frequency domain features As the query vector, the mapped spatial feature representation As key vectors and value vectors, the frequency domain feature output of spatial enhancement is constructed through the heterogeneous source attention mechanism. It is used to model the complementary effect of spatial domain structure on spectral content and introduce guidance information of target structure and texture features based on frequency domain distribution;

[0126] S64. Represent the mapped spatial features Spatial feature output with frequency domain enhancement Splicing by channel dimension to form enhanced spatial feature representation The enhanced spatial feature representation integrates the significant response area information guided by the frequency domain while maintaining the original spatial structure expression;

[0127] S65. Represent the mapped frequency domain features Frequency domain feature output with spatial domain enhancement Splicing by channel dimension to form enhanced frequency domain feature representation The enhanced frequency domain feature representation integrates the modulation effect of the spatial domain structure on the local frequency response while maintaining the original spectrum information expression;

[0128] S66. Enhanced spatial feature representation and enhanced frequency domain feature representation Perform the maximum pooling operation on all image block dimensions to extract the enhanced spatial feature vector and enhanced frequency domain eigenvector The vector is used to compress spatial information and frequency information into a global representation of uniform length to support subsequent fusion classification tasks;

[0129] S67. Enhance the spatial feature vector and enhanced frequency domain eigenvector Splice to form a fusion feature representation

[0130] In this embodiment, S7 includes the following steps:

[0131] S71. Represent the fusion features Input classification and discrimination module, which is a discrimination system built on a multi-task deep classification network. It has a flame recognition submodule and a smoke recognition submodule, which respectively complete the parallel judgment of flame features and smoke features in the image of the offshore rocket platform;

[0132] S72. Represent the fusion features The flame recognition submodule and the smoke recognition submodule are input in sequence, and the flame recognition output results are obtained through the nonlinear activation function and the fully connected layer respectively. Smoke recognition output results in, Indicates the confidence that there is a flame in the t-th frame image, Indicates the confidence that there is smoke in the t-th frame image;

[0133] S73. Output results based on flame recognition Smoke recognition output results Combined with the preset threshold rules for joint judgment, the fire risk identification rules for offshore rocket platforms are constructed, and the flame existence threshold θ is defined. fire , smoke presence threshold θ smoke , and divide the flame and smoke states of the t-th frame offshore rocket platform image into safe state, abnormal state and high-risk state;

[0134] S74. Based on the state classification result, the t-th frame of the offshore rocket platform image is output as the flame recognition result Smoke recognition results and its corresponding confidence level, and output the final risk identification results of the offshore rocket platform

[0135] In this embodiment, the safe state: the flame recognition output result is satisfied. And the smoke recognition output result Abnormal state: meet the flame recognition output result And the smoke recognition output result Or flame recognition output results And the smoke recognition output result High-risk status: meet the flame recognition output results And the smoke recognition output result

[0136] Example 1:

[0137] In June 2024, during a "Comprehensive Safety Drill for Offshore Long-Range Solid-Propellant Carrier Rocket Platform Launches" jointly held by the Satellite Launch Center of City A and the Wenchang Space Launch Site in Hainan, the project team deployed the fire risk identification system based on AI image recognition proposed in this invention. The platform's operating area is 18.5 degrees north latitude and 110.4 degrees east longitude, a sea area with high temperature and humidity, variable wind speeds and drastic light contrasts. Traditional video surveillance systems in this area frequently give false alarms due to interference from reflections, water vapor and smoke, posing a risk of failure to provide timely warnings of fires and inaccurate identification of smoke.

[0138] In order to compare the performance differences between the method of the present invention and the traditional method, this drill set up a fire simulation of "local short circuit of high-temperature cables, igniting trace liquid fuel residues". The fire occurred in the cooling cabin area in the middle section of the platform, with an area of about 8 square meters. The camera angles covered the main deck of the platform, the fuel pump room, the cooling pipeline channel and the outside of the tail propulsion engine. Twelve 4K high-definition AI cameras were distributed and deployed, combined with a three-axis stabilized bracket and a GPS+IMU time synchronization module to realize real-time image sequence acquisition of the platform.

[0139] In the data preparation phase, the present invention uses a total of 25,000 frames of historical platform operation image data for pre-training, and collects real platform images in different time periods (including sunrise, noon, dusk, and night) and different sea conditions (calm, choppy, rainy, and sunny) to model the feature coupling relationship between the air domain and the frequency domain. The training data sample consists of:

[0140] Flame sample image frames: 8200 (open flame area, micro-fire area, edge spark);

[0141] Smoke sample image frames: 9400 (white smoke, black smoke, transparent oil and gas, hot air disturbance);

[0142] Image frames without abnormalities: 7400 (normal environmental background, including high light reflection, water mist and steam).

[0143] The control group used a traditional fire detection system based on YOLOv4 and edge enhancement filtering algorithm. Its training data was a subset of the above samples, totaling 12,000 images. The flame area was identified by using spatial image grayscale change and brightness threshold segmentation, and the smoke was identified by using mean background modeling and color feature filtering.

[0144] At the beginning of the drill, the platform slowly sailed in a sailing state. The lighting environment was direct strong afternoon sunlight, wind force 4, temperature 32°C, and humidity 75%. After receiving multi-angle video input, the system first completed the image timestamp and spatial position calibration based on the method of the present invention, and automatically corrected some tilted images through the IMU. The system then entered the preprocessing and frequency domain transformation process. The spectrum diagram obtained after Fourier transformation showed that the simulated fire source area had an amplitude surge at the frequency point (20, 22) and was automatically normalized to suppress high-contrast noise.

[0145] At this time, the spatial domain Transformer divides the platform image into local blocks of 64×64 windows and introduces a brightness + edge composite saliency scoring mechanism, assigning a saliency weight δ = 0.87 to the bottom area of the cooling cabin. It identifies abnormal polygon edge continuity in this area, indicating suspected spark aggregation. At the same time, the frequency domain Transformer guides attention to the area with concentrated frequency energy, identifying a 93.4% similarity between the diffuse spectrum and typical black smoke texture features.

[0146] Through the cross-modal fusion module, the system achieves dynamic cross-guidance of spatial structure and spectral energy. The final fused feature vector is synchronously input into the flame discrimination sub-network and the smoke discrimination sub-network in the deep discriminant network. The discrimination results are as follows:

[0147] At the 73rd second image frame, the flame confidence level is 0.82 and the smoke confidence level is 0.65, indicating that the system is in an abnormal state.

[0148] At the 78th second image frame, the flame confidence level rose to 0.91 and the smoke confidence level reached 0.87, and the system upgraded its risk level to a high-risk state.

[0149] The system's average recognition delay is 0.56 seconds, and the overall recognition lead time is 3.4 times that of traditional systems.

[0150] Compared with the traditional YOLO+ filtering method, the detailed comparative data of the method of the present invention in this exercise scenario are as follows:

[0151] Indicator Category Traditional methods Method of the present invention Improvement Flame recognition recall rate (R@95) 81.3％ 94.7％ ↑13.4％ Smoke recognition accuracy 68.9％ 88.5％ ↑19.6％ Average lead time in seconds for abnormal state recognition 3.1s 6.7s ↑116％ Nighttime recognition effectiveness Very poor (misjudgment of reflection) Stable recognition ↑ Usability enhancements Low flame / little smoke false detection rate 21.2％ 7.5％ ↓13.7% decrease Model processing speed (fps) 21.4fps 26.8fps ↑Increase by 25.2%

[0152] In addition, under the simulation of low-light and heavy fog environments at night, the recognition model of the present invention can accurately identify the high-frequency jumping characteristics of open flames in the fog, while the traditional model has halo misjudgments in more than 80% of the test images, especially in the background of the high-temperature tail flame of the platform's tail thruster, with a false alarm rate of 32.7%. The present invention uses frequency domain features to mark these low-intensity flame spectrum energy concentration areas, significantly reducing false detections.

[0153] To further verify the system's stability under actual long-term deployment conditions, the project team conducted multiple monitoring sessions over the next two months, including a sunny day (June 23), a rainy day (July 8), and a night shift (July 20). A total of more than 420,000 frames of image data were collected. The measured flame accuracy rate remained between 93.2% and 95.6%, and the smoke accuracy rate remained above 87.1%. The system's false alarm rate was less than 3.4%, far better than the existing system's average false alarm rate of 12.7%.

[0154] Through the above simulation and data testing, the AI image recognition system of the present invention not only overcomes the problems of misjudgment and missed reporting caused by environmental factors in traditional fire monitoring, but also significantly improves the accuracy of risk identification, response speed and environmental adaptability through the triple improvement mechanism of frequency-space domain interaction, adaptive significance guidance, and deep multi-task classification, providing a practical and feasible intelligent solution for fire early warning of offshore rocket platforms.

[0155] The present invention adopts a hybrid Transformer structure that fuses frequency domain and spatial domain, innovatively introduces a frequency domain feature extraction branch, and constructs a cross-modal heterogeneous attention fusion mechanism. Through frequency domain Fourier spectrum energy extraction, the system can capture the high-frequency boundary characteristics of flames and the low-frequency diffusion characteristics of smoke. A saliency guidance mechanism is constructed on the spatial domain Transformer side to highlight areas with fire risks in the image. At the same time, a spectral significance bias item is designed on the frequency domain side to guide the attention focus at the frequency level. The bidirectional heterogeneous attention mechanism in the fusion module allows the frequency domain and spatial domain to dynamically interact and jointly optimize at the structural level, effectively making up for the problem that traditional methods have insufficient response to fuzzy smoke and weak flames.

[0156] The "flame-smoke saliency guidance mechanism" proposed in the present invention jointly constructs an attention bias matrix through the image block brightness mean and edge gradient amplitude in the spatial domain Transformer branch. Combined with the physical characteristics of "dramatic changes in the bright edges of flames and weak diffusion of blurred boundaries of smoke" in marine fire images, a regional-level saliency attention guidance function is constructed, which strengthens the attention weight distribution of potential abnormal areas at the multi-head attention level. It can dynamically improve the robustness of the model in strong sunlight reflection, water vapor interference and low-light night scenes, and can still accurately separate flames and background reflections under complex weather or low-contrast conditions, thereby improving the fire risk identification effect of marine rocket platforms.

[0157] The classification and discrimination module of the present invention is based on a dual-branch multi-task deep classification network. Independent recognition sub-modules are constructed for flame and smoke features respectively, and risk level judgment is performed through a joint confidence judgment mechanism. The whole process mapping from feature extraction to risk judgment is automatically completed through an end-to-end learning method. At the same time, by introducing a risk joint judgment strategy of flame existence threshold and smoke existence threshold, the recognition sensitivity and response speed of cross-anomalies are effectively improved.

[0158] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying fire risks of offshore rocket platforms based on AI image recognition, characterized in that: The steps include: S1. Acquire an image sequence of an offshore rocket platform, wherein the image sequence is a continuous frame image acquired under different sea conditions and lighting conditions, and perform time stamping and spatial position calibration on the image sequence; S2. Preprocessing the image sequence of the offshore rocket platform to obtain a preprocessed image sequence of the offshore rocket platform; S3. Convert the preprocessed marine rocket platform image sequence into the frequency domain based on Fourier transform, generate a frequency domain image spectrum sequence, and perform amplitude normalization to obtain a frequency domain image spectrum sequence for frequency domain feature extraction; S4. Construct a hybrid frequency-domain and spatial-domain Transformer structure, input the preprocessed maritime rocket platform image sequence into the spatial-domain Transformer branch, extract spatial-domain image features through the spatial-domain Transformer branch, and use the extracted spatial-domain image features as the spatial-domain feature representation. S5. Input the frequency domain image spectrum sequence into the frequency domain Transformer branch, extract the frequency domain image features through the frequency domain Transformer branch, and use the extracted frequency domain image features as the frequency domain feature representation; S6. Input the spatial domain feature representation and the frequency domain feature representation into the hybrid feature fusion module, and realize the dynamic interaction and fusion of spatial domain features and frequency domain features through the cross-branch self-attention mechanism to obtain the fused feature representation; S7. Input the fused feature representation into the classification and discrimination module built based on the deep classification network, and use the classification and discrimination module to jointly identify the flames and smoke of the offshore rocket platform to obtain the risk identification results of the offshore rocket platform.

2. The method for identifying fire risks of offshore rocket platforms based on AI image recognition according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Deploy multiple imaging and acquisition devices on the offshore rocket platform, set acquisition point locations based on the spatial coordinates of different monitoring areas, and form an offshore rocket platform image acquisition network covering key areas of the offshore rocket platform to obtain the original monitoring image sequence I of the offshore rocket platform. raw ={I t |t∈T}, where I t represents a frame of offshore rocket platform image collected at time point t, and T represents the time series of offshore rocket platform image collection; S12. The original monitoring image sequence of the sea rocket platform I raw Input timestamp annotation module, according to the global synchronous clock, each frame of the sea rocket platform image I t Add a unique time stamp τ t , constructing a time-tagged image sequence for monitoring a sea rocket platform I timestamp ={(I t ,τ t )|t∈T}, where τ t Represents a time scalar in the real number domain; S13. Image sequence I for monitoring the offshore rocket platform timestamp Perform spatial position calibration based on the fixed spatial coordinates (x i ,y i ,z i ) represents the position of the imaging acquisition device, and the imaging direction parameter θ i It represents the orientation of the imaging acquisition device, that is, the rotation angle around the XYZ axis, and the image pose matrix P of the offshore rocket platform i By the rotation matrix R i With the translation vector T i Composition, where R i =θ i , T i =[x i ,y i ,z i ] T ; S14. Based on the image pose matrix P of the offshore rocket platform i Perform perspective geometric correction on the image frames of the offshore rocket platform, reconstruct the distribution relationship of the monitoring offshore rocket platform image sequence in a unified spatial reference system, and generate the spatial calibration offshore rocket platform image sequence I calibrated ={(I t ,τ t ,P i )|t∈T}.

3. The method for identifying fire risks of offshore rocket platforms based on AI image recognition according to claim 2, characterized in that: The S3 includes the following steps: S31. Preprocess the image sequence I of the sea rocket platform pre Input frequency domain transformation module to pre-process the image of the sea rocket platform for each frame I t Applying the two-dimensional discrete Fourier transform operation, the pixel information of the marine rocket platform image in the spatial domain is mapped to the complex spectrum representation in the frequency domain, and the frequency domain image spectrum F is obtained. t (u,v); S32. Based on frequency domain image spectrum F t (u,v) calculates its corresponding amplitude spectrum A t (u, v), the amplitude spectrum is obtained by squaring the real and imaginary parts of the frequency domain image spectrum, summing them up and then taking the square root. The amplitude spectrum A t (u,v) represents the energy of the image of the rocket platform on the sea at the frequency position (u,v); S33. Amplitude spectrum A t (u, v) is jointly logarithmically scaled and normalized to its maximum value. By taking the logarithm of each amplitude value and dividing it by the maximum value of all the logarithmized amplitude values in the same frame of the sea rocket platform image, strong responses are suppressed and weak textures are enhanced to obtain the normalized frequency domain amplitude spectrum. S34. Normalize the frequency domain amplitude spectrum of each frame The corresponding time scalar τ t and the marine rocket platform image pose matrix P i Combining to form frequency domain image spectrum sequence 4. The method for identifying fire risks of offshore rocket platforms based on AI image recognition according to claim 3, characterized in that: The S4 comprises the following steps: S41. Preprocess the image sequence I of the sea rocket platform pre Input the spatial domain Transformer branch and perform a computation on each frame of the offshore rocket platform image. Perform fixed window division processing and divide it into a set of image blocks of size P×P Where K is the number of image blocks, p t,k represents the kth image block; S42. For each image block p t,k After flattening, it is projected into the embedding vector z through linear mapping t,k , and add the sea rocket platform image pose matrix P i The solved position code Construct an input sequence X with spatial orientation awareness t , position encoding and the image pose matrix P of the offshore rocket platform i Related, used to identify the orientation and physical perspective of the image patch in the space of the sea rocket platform; S43. Construct a flame-smoke attention guidance function to guide the spatial domain Transformer branch to enhance the attention of key areas in high-light and weak-boundary scenes. The flame-smoke attention guidance function is based on the average brightness of the image block l t,k and edge gradient magnitude g t,k Constructing the image patch attention bias term δ t,k , the image block attention bias term is used as a saliency factor to reflect the saliency of the potential flame or smoke area in the image block: in, α,β∈[0,1] are balance parameters; S44. Introduce the saliency bias matrix Δ in the multi-head attention calculation of each layer in the spatial domain Transformer branch t , the i,j element of the saliency bias matrix is Δ t (i,j)=γ·(δ t,i +δ t,j ), where γ is the amplification factor, which is used to adjust the focus of the attention map on the flame-smoke salient area, and finally adjust the spatial domain Transformer multi-head attention module: Among them, Q 空域 is the spatial query matrix, K 空域 is the spatial bond matrix, V 空域 is the space value matrix, is the spatial domain attention score matrix, which represents the attention weight of image block i to image block j. The attention score considers the inner product relationship between the feature vectors of the two image blocks and introduces the significance bias matrix Δ t (i, j) Enhance the characteristic transmission capability of the flame-smoke key area; S45. Repeat the L-layer encoding process constructed by combining the spatial attention score matrix and the feedforward network to obtain the output feature sequence of the spatial transformer. Represents the spatial feature expression after saliency guidance; S46. Output feature sequence Input the flame-smoke feature reconstruction module, and define the flame response weight matrix W based on the physical properties of the flame in the marine environment, which has high-frequency boundary changes and the smoke has fuzzy low-frequency diffusion characteristics. fire , smoke response weight matrix W smoke , respectively extract the flame space features and smoke spatial characteristics S47. Flame space characteristics and smoke spatial characteristics Splicing by channel dimension to construct spatial feature representation 5. The method for identifying fire risks of offshore rocket platforms based on AI image recognition according to claim 4, characterized in that: The S5 comprises the following steps: S51. The frequency domain image spectrum sequence I freq Input frequency domain Transformer branch, normalize the frequency domain amplitude spectrum A dual-scale sliding window is used to perform local scale division and global scale division respectively to form a local frequency domain sub-block set in, is the local frequency domain sub-block, K1 is the number of local frequency domain sub-blocks; S52. Extract frequency domain feature representation Obtain spatial guidance vectors by mean pooling And the spatial guidance vector is passed through the cross-modal guidance mapping matrix W cross Mapped to frequency domain steering vector Used to guide the attention focus of local features in the frequency domain; S53. Calculate local frequency domain sub-block set Each local frequency domain sub-block The average amplitude And define the spectral significance factor Spectral significance factor Constructed as a significant bias matrix The elements of the significance bias matrix are defined as follows: in, Represents the spectrum energy value of the normalized frequency domain amplitude spectrum at the frequency coordinate (u, v), Indicates the kth local frequency domain sub-block in the tth frame image The average amplitude spectrum value of , P1 represents the side length of the local frequency domain sub-block, μ is the significance scaling coefficient, At time point t, Indicates the maximum value of the average amplitude value in all local frequency domain sub-blocks of the t-th frame image, is the (i, j)th element of the saliency bias matrix constructed in the local frequency domain self-attention module for the t-th frame image, indicating the degree of saliency linkage between the i-th and j-th frequency domain sub-blocks; S54. Construct the frequency domain Transformer multi-head attention module through the saliency bias matrix and frequency domain guidance vector: Among them, Q 频域 is the frequency domain query matrix, K 频域 is the frequency domain key matrix, V 频域 is the frequency domain value matrix; S55. Repeat the frequency domain attention score matrix and the feedforward layer combination to build an L-layer local frequency domain Transformer encoding structure and output the frequency domain local feature sequence Constructing frequency domain feature representation 6. The method for identifying fire risks of offshore rocket platforms based on AI image recognition according to claim 5, characterized in that: The S6 comprises the following steps: S61. Representing spatial features and frequency domain feature representation Input fusion channel conversion module, respectively, through independent linear transformation method to project the spatial domain feature representation and frequency domain feature representation into the unified dimensional fusion feature space, forming the mapped spatial domain feature representation And the mapped frequency domain feature representation The two have the same vector dimension in the fusion feature space; S62. Represent the mapped spatial features As the query vector, the mapped frequency domain features are represented as As key vectors and value vectors, the spatial feature output of frequency domain enhancement is constructed through the heterogeneous source attention mechanism. It is used to capture the complementary effect of frequency domain information on spatial domain information and introduce frequency domain energy and pattern characteristics based on spatial structure; S63. Represent the mapped frequency domain features As the query vector, the mapped spatial features are represented as As key vectors and value vectors, the frequency domain feature output of spatial enhancement is constructed through the heterogeneous source attention mechanism. It is used to model the complementary effect of spatial domain structure on spectral content and introduce guidance information of target structure and texture features based on frequency domain distribution; S64. Represent the mapped spatial features Spatial feature output with frequency domain enhancement Splicing by channel dimension to form enhanced spatial feature representation The enhanced spatial feature representation integrates the significant response area information guided by the frequency domain while maintaining the original spatial structure expression; S65. Represent the mapped frequency domain features Frequency domain feature output with spatial domain enhancement Splicing by channel dimension to form enhanced frequency domain feature representation The enhanced frequency domain feature representation integrates the modulation effect of the spatial domain structure on the local frequency response while maintaining the original spectrum information expression; S66. Enhanced spatial feature representation and enhanced frequency domain feature representation Perform the maximum pooling operation on all image block dimensions to extract the enhanced spatial feature vector and enhanced frequency domain eigenvector The vector is used to compress spatial information and frequency information into a global representation of uniform length to support subsequent fusion classification tasks; S67. Enhance the spatial feature vector and enhanced frequency domain eigenvector Splice to form a fusion feature representation 7. The method for identifying fire risks of offshore rocket platforms based on AI image recognition according to claim 6, characterized in that: The S7 comprises the following steps: S71. Represent the fusion features Input classification and discrimination module, which is a discrimination system built on a multi-task deep classification network and has a flame recognition submodule and a smoke recognition submodule, respectively completing the parallel judgment of flame features and smoke features in the image of the offshore rocket platform; S72. Represent the fusion features The flame recognition submodule and the smoke recognition submodule are input in sequence, and the flame recognition output results are obtained through the nonlinear activation function and the fully connected layer respectively. Smoke recognition output results in, Indicates the confidence that there is a flame in the t-th frame image, Indicates the confidence that there is smoke in the t-th frame image; S73. Output results based on flame recognition Smoke recognition output results Combined with the preset threshold rules for joint judgment, the fire risk identification rules for offshore rocket platforms are constructed, and the flame existence threshold θ is defined. fire , smoke presence threshold θ smoke , and divide the flame and smoke states of the t-th frame offshore rocket platform image into safe state, abnormal state and high-risk state; S74. Based on the state classification result, the t-th frame of the offshore rocket platform image is output as the flame recognition result Smoke recognition results and its corresponding confidence level, and output the final risk identification results of the offshore rocket platform 8. The method for identifying fire risks of offshore rocket platforms based on AI image recognition according to claim 7, characterized in that: The safety state: meets the flame recognition output result And the smoke recognition output result The abnormal state: meets the flame recognition output result And the smoke recognition output result Or flame recognition output results And the smoke recognition output result The high-risk state: meets the flame recognition output result And the smoke recognition output result

Images