Submarine bubble plume identification method for deep-sea extreme environment

Abstract

The application provides a seabed bubble plume identification method for deep-sea extreme environment, and relates to the cross field of marine environment detection and computer vision. The probability distribution of bubble shape is learned based on a generative diffusion model, and classification is realized by minimizing noise prediction error to significantly improve data efficiency and anti-interference ability, so that precise and efficient identification operation can be realized in complex deep-sea environment, especially in the exploration area with sample scarcity, serious suspended particulate matter interference and great environmental difference in different sea areas. The identification system with the following architecture is applied: an underwater image acquisition unit, which is carried on a ROV / AUV high-definition industrial camera, is used to acquire a sequence video stream of a deep-sea detection area; an edge computing processing unit, which is provided with a pre-trained and fine-tuned latent diffusion classification model, is used to process image frames in real time or near real time; and a shipborne upper computer unit is used for human-computer interaction, abnormal thermal map visualization and long-time sequence data storage.

Description

Technical Field

[0001] This application proposes a method and system for accurately identifying cold seeps and hydrothermal vent plumes in complex deep-sea environments by using a depth generation model to process underwater optical images, which belongs to the interdisciplinary field of marine environmental detection and computer vision. Background Technology

[0002] Natural gas hydrates, as a new, efficient and environmentally friendly energy source, are widely distributed globally. The total global resource-level hydrate volume is conservatively estimated at tens of trillions of cubic meters, and they are regarded by various countries as the most promising alternative energy source to oil and natural gas in the future.

[0003] Bubble plumes emitted from cold seeps and hydrothermal vents are key indicators of natural gas hydrate reservoirs. On the one hand, bubble detection is crucial for oil and gas and natural gas hydrate resource exploration; for example, many cold seeps discovered globally are associated with large mineral deposits. On the other hand, methane bubble escape directly affects the global carbon cycle; excessive release will exacerbate the greenhouse effect and lead to ocean acidification. Therefore, accurate bubble identification is also of strategic significance for environmental monitoring. Currently, bubble plume detection mainly relies on the following three types of technologies: 1) Acoustic detection: mainly using multibeam echolocation, shallow profiling, and multichannel seismic imaging, suitable for scanning large areas of the ocean, but prone to missed detection due to fluctuations in bubble activity intensity (such as temporary absence of plumes), and difficult to distinguish the acoustic response of bubbles from suspended particles; 2) In-situ optical detection: using ROV / AUV equipped with cameras to acquire images, providing detailed information such as bubble shape and flow velocity, but significantly affected by underwater light attenuation and background blur; 3) Chemical and physical sensing: such as Raman spectrometers can identify bubble composition, but the equipment is expensive and has a slow response speed, failing to meet real-time detection requirements.

[0004] With the development of underwater robotics technology, near-bottom optical photography is being used more and more. However, faced with massive amounts of underwater video data, current technologies still rely on manual interpretation, which is extremely inefficient. Furthermore, existing automated identification methods based on traditional deep learning (such as discriminative models like CNN, YOLO, and ResNet) face the following four core problems in actual deep-sea exploration operations: 1) The "Matching Paradox" Between the Spatiotemporal Multiscale Characteristics of Bubble Streams and Optical Imaging Resolution: Cold seep bubble plumes are dynamic systems with spatiotemporal multiscale characteristics, and their bubble diameters can span three orders of magnitude. Existing optical imaging equipment struggles to balance "large field of view" and "high resolution": large field-of-view cameras lack sufficient resolution to capture the micrometer-sized bubbles that dominate the diffusion process, easily leading to "observation blind spots"; while high-resolution microscopes have extremely small fields of view, failing to reflect the macroscopic morphology of the plume. Furthermore, optical cameras acquire two-dimensional planar projections, and the vertical overlap of bubbles causes severe "bubble stacking" in the image, making it difficult to distinguish bubbles at different depths, thus rendering traditional two-dimensional feature-based recognition algorithms ineffective.

[0005] 2) The challenge of accurate target-background segmentation and anti-interference in complex background environments: The complex background environment of the detection area leads to a significant decrease in the accuracy of traditional image segmentation algorithms, specifically manifested in: Low contrast and blurred boundaries: Under the influence of weak light and light scattering in the deep sea, bubble edges exhibit "abrupt changes in grayscale gradient" or haloing, especially since the grayscale values ​​of microbubbles are close to background noise, causing gradient-based edge detection algorithms to fail. Adhesion of dense bubble clusters: The bubble density in the plume core area is extremely high, exhibiting a "cluster-like" adhesion structure, making it difficult for traditional algorithms to distinguish adhesion boundaries, easily resulting in under-segmentation. Multi-source heterogeneous interference: The deep-sea environment contains multiple interference sources, including resuspended sediments (marine snow) with morphology highly similar to microbubbles, planktonic / benthic organisms that easily form trails, and noise from the equipment itself. Existing algorithms lack universal interference removal models, resulting in extremely high false positive rates.

[0006] 3) Deep learning models face "generalization bottlenecks" and "data traps" in extreme environments; issues include domain adaptation failure: the optical properties of water bodies differ significantly between different sea areas (such as the South China Sea and the Gulf of Mexico), causing models trained on data from specific sea areas to experience severe "domain shift" when facing new environments, resulting in poor generalization ability. Small sample size and long-tailed distribution: acquiring cold seep bubble data is extremely costly, and the sample distribution is highly imbalanced (the number of microbubbles is far greater than that of large bubbles, and samples of abnormal plumes are extremely rare). Existing discriminative models (such as YOLO and U-Net) are prone to "overfitting" or "class bias" under long-tailed data distributions, making it difficult to capture the features of scarce samples.

[0007] 4) In terms of model interpretability, existing methods lack anomaly detection capabilities and interpretability. The "black box" nature of traditional deep learning models makes it impossible for researchers to intuitively understand whether the model makes judgments based on the contour features of bubbles or background noise, lacking the interpretability urgently needed for scientific research.

[0008] In summary, there is an urgent need for a novel method for identifying bubbles in deep-sea cold seeps that can achieve high-precision training even with extremely scarce samples, possesses the ability to fundamentally distinguish bubbles from suspended particles, adapts to different marine environments by learning data distribution, and provides visualized judgment criteria. Therefore, this application is hereby submitted. Summary of the Invention

[0009] The proposed method for identifying seabed bubble plumes in extreme deep-sea environments addresses the problems of existing technologies by learning the probability distribution of bubble morphology based on a generative diffusion model. It aims to achieve classification by minimizing noise prediction errors, thereby significantly improving data efficiency and anti-interference capabilities. This enables accurate and efficient identification operations in complex deep-sea environments, especially in exploration areas with scarce samples, severe interference from suspended particulate matter, and significant differences in environmental conditions across different sea areas.

[0010] To address this, the aforementioned method for identifying seabed bubble plumes in extreme deep-sea environments employs an identification system with the following architecture: an underwater image acquisition unit, mounted on a high-definition industrial camera on an ROV / AUV, for acquiring sequential video streams of the deep-sea exploration area; an edge computing processing unit, equipped with a pre-trained and fine-tuned latent diffusion classification model, for real-time or near-real-time processing of image frames; and a shipborne host computer unit, for human-computer interaction, visualization of anomaly heatmaps, and long-term data storage. Includes the following steps: Step 1: Data Acquisition and Preprocessing; The collected underwater images were dehazed and color corrected to construct a small sample dataset containing categories such as "bubbles", "suspended particles", and "seabed background". Step 2: Construction and fine-tuning of the potential diffusion model; A pre-trained generative pedestal model is used, and the variational autoencoder (VAE) compresses the image into the latent space, while the U-Net network is used as a noise predictor. Step 3: Rapid reasoning and recognition based on successive elimination; During the training phase, the generative pedestal model is trained to predict Gaussian noise added to the latent features of the image by using the input class vector as a condition. During the inference phase, a successive elimination strategy based on "noise prediction fidelity" is adopted. Step 4: Anomaly detection and out-of-distribution screening; Using normality scoring, an image is labeled as an "abnormal / unknown target" when the noise prediction error is higher than a set threshold in all known categories. Step 5: Verification of interpretability based on counterfactual heatmaps; Step 6: Post-processing and geographic information fusion.

[0011] Step 1 involves acquiring continuous video stream data and performing frame extraction at fixed time intervals to establish a small-scale cold spring image dataset. The category labels of the dataset include Bubble plume, Suspended particles / marine snow Underwater background.

[0012] Step 2 includes: First, construct a "multi-scale feature fusion" encoder. encoder It includes multiple parallel branches, which extract features from local slices focusing on microbubble texture and panoramic images focusing on the overall morphology of the plume, respectively. The multi-scale features are then cascaded and fused with attention in the latent space to obtain a latent feature tensor containing rich scale information. Define seabed optical observation images as The observed images are processed using a sliding window to obtain several To preserve the high-frequency texture details of tiny bubbles, the observed image was downsampled using bilinear interpolation. To preserve the overall morphological structure, the encoder of the variational autoencoder (VAE) is then frozen. Will and Mapping to the compressed underwater feature latent space, we obtain the latent feature tensor as expressed in the following expression. : Then, in the conditional injection phase, sets of geological target categories corresponding to bubble plumes, suspended sediments, and deep-sea background features are defined as follows: ,in For bubble plumes, For sediments, As background, category labels It is converted to one-hot encoding and passed through a learnable projection matrix. Mapped to condition vector This is then injected into the cross-attention mechanism of the denoising network; The training process is modeled as a reverse thermodynamic diffusion process within the latent feature space; specifically, the discrete diffusion time step is defined as... From the standard Gaussian distribution Medium-sampled random perturbation tensor, i.e., noise In time step Towards latent features Adding perturbations yields noise features ; At the same time, class balancing weights with the following expression are introduced into the training objective function. : in, This represents the number of samples in that category; thus, it forces the model to focus more on scarce samples during training. Training the denoising network The parameters are To minimize the following weighted variational objective function : in, For the sampled random Gaussian perturbation tensor; For the network with given noisy features Time step and geological conditions Denoising perturbations in the prediction; This represents the square of the Frobenius norm, used to quantify the differences between tensors; These are time-step weighting coefficients used to balance the learning weights under different levels of noise. These are category weighting coefficients used to balance the learning weights of samples with different distribution numbers; The training model learns to target specific seabed geological features. Given the prior conditions, the ordered latent feature distribution corresponding to the target can be recovered from the chaotic random field.

[0013] In step 3, during the inference phase, a real-time screenshot of a ROV to be identified is taken. First, it is encoded as Define category The "cumulative noise reconstruction residual" The implementation process, which forms the basis for the judgment, includes: right Perform multiple rounds of random noise addition to generate Group sample pairs ; Category Collection Assuming the images belong to bubble plumes Sediments or background Input the model to calculate the predicted disturbance; No. The formula for calculating the cumulative noise reconstruction residual for each category is as follows: in, This is the sampling weighting factor; A dynamic hypothesis testing strategy is adopted: during the cumulative calculation process, a statistic of the residual difference between the two competing categories is constructed. : in, For the completed [projects] under the assumption of bubble plume categories The sample variance of the single noise prediction residual relative to its mean for each random noise-added sampling; For the completed work under the assumption of suspended sediment categories The sample variance of the single noise prediction residual relative to its mean for each random noise-added sampling; This represents the current cumulative number of random, noisy samples. like Exceeding the set significance threshold indicates that... The noise prediction under the assumption deviates significantly from the true distribution, meaning the assumption is invalid, thus terminating the calculation for this category prematurely; Finally, the output is the following category that minimizes the residual. :

[0014] Step 4, the recognition stage based on noise prediction error, for the image to be detected Encode it and add random noise to obtain Assuming the image to be detected belongs to either "bubble" or "non-bubble", the corresponding category conditions are applied. Input the model and calculate the L2 norm error between the model's predicted noise and the actual added noise; the category with the smallest noise prediction error is the prediction result; if the denoising error under the "bubble" condition is significantly lower than the error under the "suspended particle" condition, then the target is determined to be a bubble; Define image Geological anomaly index The expected value of the minimum reconstruction residual under all known prior categories is expressed as follows: GAI distribution of normal samples constructed based on the training set And set an abnormal threshold. ; During the detection process, if the following conditions are met: If the target does not conform to any known cold seep geological features, it is marked as an "out-of-distribution anomalous target" and a higher-priority storage instruction is triggered for storage.

[0015] Step 5 includes: First, a "fluid dynamics morphology residual map" is generated, and it is assumed that the model will identify the image as a bubble plume. ; Calculate target adversarial category With prediction category Perturbation prediction difference tensor in the latent space The expression is as follows: Then, the difference tensor is projected back into pixel space by the decoder to obtain the following visualized heatmap. : Heat map The high-response region on the surface is the key morphological feature by which the model distinguishes bubbles from sediments.

[0016] In step 6, the recognition result output by the processor unit is bound to a timestamp, and the ultra-short baseline USBL positioning data and depth sensor data carried by the ROV are read simultaneously. When the identification result is "bubble plume", extract the latitude and longitude of the ROV at the current moment. and depth By combining the camera's attitude angles (Pitch, Roll, Yaw), the absolute geographic coordinates of the bubble nozzle on the seabed are calculated. These are the longitude and latitude values ​​of the ROV's reference center in the world geodetic coordinate system. These refer to the pitch angle, yaw angle, and heading angle of the camera coordinate system relative to the geographic coordinate system, respectively. Finally, the distribution map of the seabed cold seep vents was drawn in real time in the form of a 3D point cloud in the host computer software, which intuitively shows the potential location of mineral resources in the exploration area.

[0017] In summary, this application has the following advantages and beneficial effects compared to the prior art: 1. This application addresses the problem of blind spots in the recognition caused by the large particle size range and "two-dimensional stacking" of existing bubble plumes. It innovatively proposes a "multi-scale feature fusion coding" solution. By constructing a multi-resolution image pyramid, it extracts the texture details of microbubbles and the macroscopic morphological features of plumes in parallel. In addition, attention fusion is performed in the latent space. This can effectively solve the contradiction that traditional optical detection cannot take into account both micron-sized bubbles and macroscopic plume morphology, as well as the counting and recognition failure caused by the vertical overlap (stacking) of bubbles. It achieves high-fidelity perception of full-scale bubble plumes. 2. This application addresses the high false positive rate interference caused by complex deep-sea backgrounds (such as sea snow and biomass trails) by innovatively proposing a technical means for fluid morphology fidelity inference. On the training side, the model is forced to learn the essential differences in the generation and distribution of sediment particles, biomass trails, and bubble fluids. On the inference side, "noise prediction residual" is used as the criterion for fluid morphology consistency, thus effectively solving the problem that traditional algorithms misclassify suspended particles as bubbles due to edge blurring and grayscale confusion in extreme environments with high turbidity, thereby reducing the false positive rate in complex backgrounds. 3. This application addresses the generalization bottleneck caused by the scarcity of samples in extreme deep-sea environments and "domain drift" across sea areas. It innovatively adopts a generative classification architecture based on probability density distribution, utilizes the powerful modeling capability of diffusion models for data distribution (rather than simple decision boundaries), and introduces balanced weights to force the model to focus on scarce anomalous plume samples. This successfully solves the pain point of discriminative models being prone to overfitting on small sample and long-tailed data, endowing the system with extremely strong cross-sea area operation capabilities, and maintaining high-precision identification without large-scale re-labeling for new sea areas. Attached Figure Description

[0018] The following figures further illustrate the design concept and features of this application.

[0019] Figure 1 This is a logical architecture diagram of a submarine cold seep bubble identification system based on a diffusion model; Figure 2 It is a flowchart of reasoning and classification; Figure 3 This is a schematic diagram of noise prediction using a diffusion model; Figure 4 This is a schematic diagram comparing the noise prediction errors of generated bubbles and suspended particulate matter. Figure 5 This is a comparative illustration of the effects of counterfactual heatmaps; the left image is the original murky image, and the right image is the heatmap that highlights the bubble outlines and ignores the surrounding floating noise. Figure 6 This is a schematic diagram illustrating the steps of the diffusion model. Detailed Implementation

[0020] The present application will now be further described in conjunction with the accompanying drawings and the following preferred embodiments.

[0021] Example 1, as Figure 1 As shown, this application proposes an identification system with the following architecture to address the problem of accurate identification of seabed bubble plumes in complex deep-sea environments, especially under conditions of scarce samples, severe interference from suspended particulate matter, and significant differences in environmental characteristics across different sea areas: The underwater image acquisition unit is equipped with a high-definition industrial camera on an ROV / AUV and is used to acquire sequential video streams of the deep-sea exploration area. Edge computing processing unit, which is equipped with a pre-trained and fine-tuned latent diffusion classification model for real-time or near-real-time processing of image frames; The shipborne host computer unit is used for human-computer interaction, visualization of abnormal heat maps, and long-term data storage.

[0022] Based on the construction and application of the aforementioned identification system, the seabed bubble plume identification method for deep-sea extreme environments described in this application abandons the discriminative classification approach (directly predicting labels) adopted by existing technologies. Instead, it innovatively adopts a generative approach, that is, using a diffusion model to model the conditional probability density. .

[0023] The method specifically includes the following steps: Step 1: Data Acquisition and Preprocessing; The collected underwater images were dehazed and color corrected to construct a small sample dataset containing categories such as "bubbles", "suspended particles", and "seabed background". Specifically, an underwater remotely operated vehicle (ROV) equipped with a high-definition industrial camera was deployed to the target area of ​​cold seep hydrothermal activity on the seabed. Based on changes in the seabed topography, the ROV was kept at a height of 1 to 5 meters above the seabed interface for close-range observation. Continuous video stream data was collected, and frames were extracted at fixed time intervals to create a small-scale cold seep image dataset.

[0024] The category labels of the dataset include: (bubble plume) (Suspended particles / Marine snow) (Underwater background).

[0025] Considering the blue-green color cast caused by light attenuation in the deep-sea environment, firstly, the dark channel prior algorithm is used to restore the color and dehaze the original image to enhance the contrast of the bubble edges; then, the image size is resampled to 512*512 pixels, and the pixel intensity value is normalized from [0, 255] to the range of [-1, 1] to adapt to the input requirements of the subsequent diffusion model. The key to this step is that, thanks to the characteristics of the subsequent generative model, only 10 to 50 training samples are needed for each category to meet the requirements of high-precision training, which greatly reduces the cost of deep-sea data annotation.

[0026] Step 2: Construction and fine-tuning of the potential diffusion model; A variational autoencoder (VAE) is used to compress the image into the latent space, and a U-Net network is used as a noise predictor. Specifically, a pre-trained generative pedestal model is used, with the Stable Diffusion architecture preferred in this embodiment to make targeted improvements and solve scale matching and background interference problems.

[0027] First, a "multi-scale feature fusion" encoder is constructed to address the scale mismatch and stacking issues; the encoder It includes multiple parallel branches, which extract features from local slices focusing on microbubble texture (high resolution) and panoramic images focusing on the overall morphology of plumes (low resolution), respectively. The multi-scale features are then cascaded and fused with attention in the latent space to obtain a latent feature tensor containing rich scale information. ; Specifically, seabed optical observation images are defined as The encoder of the variational autoencoder (VAE) by freezing parameters , to pixel space and Mapping to the compressed underwater feature latent space, we obtain the latent feature tensor as expressed in the following expression. : This enables the model to recognize both tiny bubble textures and macroscopic plume structures, and to correctly identify bubbles even when they are stacked, by using texture statistics. Then, in the conditional injection stage, instead of a general text encoder, a new dedicated "submarine geological entity embedding layer" is proposed; specifically, sets of geological target categories corresponding to bubble plumes, suspended sediments, and deep-sea background features are defined as follows: , category labels It is converted to one-hot encoding and passed through a learnable projection matrix. Mapped to condition vector This is then injected into the cross-attention mechanism of the denoising network; The training process is modeled as a reverse thermodynamic diffusion process within the latent feature space; specifically, the discrete diffusion time step is defined as... From the standard Gaussian distribution Medium-sampled random perturbation tensor (i.e., noise) At time step Towards latent features Adding perturbations yields noise features ; Meanwhile, to address the issues of "small sample size and imbalanced data," the following class balancing weights are introduced into the training objective function. : in, This represents the number of samples in that category; thus, the model is forced to focus more on scarce samples (such as anomalous plumes) during training. Training the denoising network (Parameters are) To minimize the following weighted variational objective function : in, It is a sampled stochastic perturbation tensor. For the network with given noisy features Time step and geological conditions Denoising perturbations in the prediction; This represents the square of the Frobenius norm, used to quantify the differences between tensors; These are time-step weighting coefficients used to balance the learning weights under different levels of noise. These are category weighting coefficients used to balance the learning weights of samples with different distribution numbers; The physical meaning of the above formula is: training the model to learn about specific seabed geological targets. Given the prior conditions, the ordered latent feature distribution corresponding to the target can be recovered from the chaotic random field.

[0028] Step 3: Rapid reasoning and recognition based on successive elimination; During the training phase, unlike traditional classifiers that search for classification boundaries, generative pedestal models learn "how to generate a bubble image from pure noise". That is, by using the input class vector (such as one-hot encoding) as a condition, the model is trained to predict Gaussian noise to be added to the latent features of the image. During the inference phase, a successive elimination strategy based on "noise prediction fidelity" is adopted. Specifically, for a real-time screenshot of a ROV to be identified... First, it is encoded as The identification process does not directly predict labels, but rather seeks out the geological categories that best match the observed data to the generated potential distribution. Define category The "cumulative noise reconstruction residual" The implementation process, which forms the basis for the judgment, includes: right Perform multiple rounds of random noise addition to generate Group sample pairs ; Category Collection Includes target class, interference class, and background: Let's assume the images belong to bubble plumes. Sediments are The background is Input the model to calculate the predicted disturbance; No. The formula for calculating the cumulative noise reconstruction residual for each category is as follows: in, This is the sampling weighting factor; Employing a dynamic hypothesis testing strategy: during the cumulative calculation process (e.g., every calculation) After the second sampling), construct the current two competing categories (e.g. and The residual difference statistic : in, For the completed [projects] under the assumption of bubble plume categories The sample variance of the single noise prediction residual relative to its mean for each random noise-added sampling; For the completed work under the assumption of suspended sediment categories The sample variance of the single noise prediction residual relative to its mean for each random noise-added sampling; This represents the current cumulative number of random, noisy samples. and These two sample variances are used to measure the volatility and dispersion of the noise prediction residual sequence under the corresponding assumptions. The variances of these two samples are standardized in the form of standard errors to account for the difference between the means of the two classes of residuals. The dimensionless meaning of statistical significance is that, when the difference between the means of the residuals of the two categories is the same, the smaller the fluctuation of the residuals, the more significant the difference and the more reliable the successive elimination decision. like Exceeding the set significance threshold (e.g., corresponding to P < 0.002) indicates that in The noise prediction under the assumption deviates significantly from the true distribution, meaning the assumption is invalid, thus terminating the calculation for this category prematurely; Finally, the output is the following category that minimizes the residual. : Unlike traditional segmentation algorithms that rely on sharp boundaries, this embodiment relies on "noise prediction residuals." Even if the bubble edges are blurred or adhere together, as long as their texture and morphological distribution conform to hydrodynamic characteristics (such as the spherical / ellipsoidal tendency caused by surface tension), the model can still perform segmentation. The prediction residuals under these conditions will be very small. Conversely, for aggregated bubble clusters, if we assume... (Sediments are typically irregular aggregates), and the model attempts to reconstruct them into irregular shapes, resulting in significant structural errors. Therefore, by comparison... This application can penetrate the fuzzy boundaries and the appearance of adhesion, and make accurate identification based on the "statistical nature of fluid morphology", thereby effectively solving the failure problem of traditional algorithms caused by unclear boundaries.

[0029] Step 4: Anomaly detection and out-of-distribution screening; Using normality scoring, images are marked as “abnormal / unknown targets” when the noise prediction error is higher than a set threshold in all known categories, which helps to discover new unrecorded cold seep hydrothermal vent morphologies. Specifically, in the recognition stage based on noise prediction error, for the image to be detected... Encode it and add random noise to obtain Assuming the image to be detected belongs to either "bubbles" or "non-bubbles (particles / background)," the corresponding category conditions are applied. Input the model and calculate the L2 norm error (MSE) between the model's predicted noise and the actual added noise; the category with the smallest noise prediction error is the prediction result; if the denoising error under the "bubble" condition is significantly lower than the error under the "suspended particle" condition, then the target is determined to be a bubble; In order to identify unknown deep-sea geological phenomena (such as new biological communities or unknown vent morphology), this application introduces the following "geological anomaly index": Define image Geological anomaly index The expected value of the minimum reconstruction residual under all known prior categories is expressed as follows: GAI distribution of normal samples constructed based on the training set And set an abnormal threshold. (For example, take) (99th percentile) During the detection process, if the following conditions are met: If the target does not conform to any known cold seep geological features, it is marked as an "out-of-distribution anomalous target" and a higher-priority storage instruction is triggered for storage. Step 5: Verification of interpretability based on counterfactual heatmaps; To verify the physical reliability of the generative pedestal model's recognition, firstly, a "fluid dynamics morphology residual map" was generated, and it was assumed that the model would identify the image as a bubble plume. ; To explain "why it was identified as bubbles rather than suspended particles," the target adversarial category was calculated. (i.e., suspended sediments) and predicted categories Perturbation prediction difference tensor in the latent space The expression is as follows: The formula above calculates the prediction error when the assumption is "suspended particles," minus the prediction error when the assumption is "bubbles." Positive values ​​indicate that the region more closely matches bubble characteristics and completely contradicts particle characteristics.

[0030] Then, the difference tensor is projected back into pixel space by the decoder to obtain the following visualized heatmap. : Heat map The high-response region on the surface is the key morphological feature (such as the continuity of the plume, edge curvature, etc.) that the model uses to distinguish between bubbles and sediments.

[0031] Step 6: Post-processing and geographic information fusion; The recognition results (category and confidence level) output by the processor unit are bound to the timestamp, and the ultra-short baseline (USBL) positioning data and depth sensor data on the ROV are read simultaneously. When the identification result is "bubble plume", extract the latitude and longitude of the ROV at the current moment. and depth By combining the camera's attitude angles (Pitch, Roll, Yaw), the absolute geographical coordinates of the bubble nozzle on the seabed can be calculated.

[0032] Finally, the distribution map of the seabed cold seep vents was drawn in real time in the form of a 3D point cloud in the host computer software, which intuitively shows the potential location of mineral resources in the exploration area.

[0033] As described above, similar technical solutions can be derived from the solutions presented in conjunction with the accompanying drawings and description. Any solutions that do not depart from the structural content of this invention still fall within the scope of the claims of this invention.

Claims

1. A method for identifying a bubble plume on the seabed in a deep-sea extreme environment, characterized in that: The application has a recognition system with the following architecture. The underwater image acquisition unit is equipped with a high-definition industrial camera on an ROV / AUV and is used to acquire sequential video streams of the deep-sea exploration area. Edge computing processing unit, which is equipped with a pre-trained and fine-tuned latent diffusion classification model for real-time or near-real-time processing of image frames; The shipborne host computer unit is used for human-computer interaction, visualization of anomaly heatmaps, and long-term data storage. Includes the following steps, Step 1: Data Acquisition and Preprocessing; The collected underwater images were dehazed and color corrected to construct a small sample dataset containing categories such as "bubbles", "suspended particles" and "seabed background". Step 2: Construction and fine-tuning of the potential diffusion model; A pre-trained generative pedestal model is used, and the variational autoencoder (VAE) compresses the image into the latent space, while the U-Net network is used as a noise predictor. Step 3: Rapid reasoning and recognition based on successive elimination; During the training phase, the generative pedestal model is trained to predict Gaussian noise added to the latent features of the image by using the input class vector as a condition. During the inference phase, a successive elimination strategy based on "noise prediction fidelity" is adopted. Step 4: Anomaly detection and out-of-distribution screening; Using normality scoring, an image is labeled as an "abnormal / unknown target" when the noise prediction error is higher than a set threshold in all known categories; Step 5: Verification of interpretability based on counterfactual heatmaps; Step 6: Post-processing and geographic information fusion.

2. The method for identifying seabed bubble plumes in extreme deep-sea environments according to claim 1, characterized in that: Step 1 involves acquiring continuous video stream data and performing frame extraction at fixed time intervals to establish a small-scale cold spring image dataset. The category labels of the dataset include Bubble plume, Suspended particles / marine snow Underwater background.

3. The method for identifying seabed bubble plumes in extreme deep-sea environments according to claim 1, characterized in that: Step 2 includes, First, construct a "multi-scale feature fusion" encoder. encoder It includes multiple parallel branches, which extract features from local slices focusing on microbubble texture and panoramic images focusing on the overall morphology of the plume, respectively. The multi-scale features are then cascaded and fused with attention in the latent space to obtain a latent feature tensor containing rich scale information. Define seabed optical observation images as ; Several images were obtained by performing sliding window processing on the observed images. To preserve the high-frequency texture details of tiny bubbles; The observed image was downsampled using bilinear interpolation. To preserve the overall morphological structure; then, the encoder of the variational autoencoder (VAE) with frozen parameters is... Will and Mapping to the compressed underwater feature latent space, we obtain the latent feature tensor as expressed in the following expression. : Then, in the conditional injection phase, the sets of geological target categories corresponding to bubble plumes, suspended sediments, and deep-sea background features are defined as follows: ,in For bubble plumes, For sediments, As background, category labels It is converted to one-hot encoding and passed through a learnable projection matrix. Mapped to condition vector This is injected into the cross-attention mechanism of the denoising network; The training process is modeled as a reverse thermodynamic diffusion process within the latent feature space; specifically, the discrete diffusion time step is defined as... From the standard Gaussian distribution Medium-sampled random perturbation tensor, i.e., noise In time step Towards latent features Adding perturbations yields noise features ; At the same time, class balancing weights are introduced into the training objective function with the following expression. : in, This represents the number of samples in that category; thus, it forces the model to focus more on scarce samples during training. Training the denoising network The parameters are To minimize the following weighted variational objective function : in, For the sampled random Gaussian perturbation tensor; For the network with given noisy features Time step and geological conditions Denoising perturbations in the prediction; This represents the square of the Frobenius norm, used to quantify the differences between tensors; These are time-step weighting coefficients used to balance the learning weights under different levels of noise. These are category weighting coefficients used to balance the learning weights of samples with different distribution numbers; The training model learns to target specific seabed geological features. Given the prior conditions, the ordered latent feature distribution corresponding to the target can be recovered from the chaotic random field.

4. The method for identifying seabed bubble plumes in extreme deep-sea environments according to claim 1, characterized in that: In step 3, during the inference phase, a real-time screenshot of a ROV to be identified is taken. First, it is encoded as Define category The "cumulative noise reconstruction residual" The implementation process, which forms the basis for the judgment, includes: right Perform multiple rounds of random noise addition to generate Group sample pairs ; Category Collection Assuming the images belong to bubble plumes Sediments or background Input the model to calculate the predicted disturbance; No. The formula for calculating the cumulative noise reconstruction residual for each category is as follows: in, This is the sampling weighting factor; A dynamic hypothesis testing strategy is adopted: during the cumulative calculation process, a statistic of the residual difference between the two competing categories is constructed. : in, For the completed [projects] under the assumption of bubble plume categories The sample variance of the single noise prediction residual relative to its mean for each random noise-added sampling; For the completed work under the assumption of suspended sediment categories The sample variance of the single noise prediction residual relative to its mean for each random noise-added sampling; This represents the current cumulative number of random, noisy samples. like Exceeding the set significance threshold indicates that... The noise prediction under the assumption deviates significantly from the true distribution, meaning the assumption is invalid, thus terminating the calculation for this category prematurely; Finally, the output is the following category that minimizes the residual. : 。 5. The method for identifying seabed bubble plumes in extreme deep-sea environments according to claim 1, characterized in that: Step 4, the recognition stage based on noise prediction error, for the image to be detected Encode it and add random noise to obtain ; Assuming the image to be detected belongs to either "bubble" or "non-bubble", the corresponding category conditions are applied. Input the model and calculate the L2 norm error between the model's predicted noise and the actual added noise; the category with the smallest noise prediction error is the prediction result; if the denoising error under the "bubble" condition is significantly lower than the error under the "suspended particle" condition, then the target is determined to be a bubble; Define image Geological anomaly index The expected value of the minimum reconstruction residual under all known prior categories is expressed as follows: GAI distribution of normal samples constructed based on the training set And set an abnormal threshold. ; During the detection process, if the following conditions are met: If the target does not conform to any known cold seep geological features, it is marked as an "out-of-distribution anomalous target" and a higher-priority storage instruction is triggered for storage.

6. The method for identifying seabed bubble plumes in extreme deep-sea environments according to claim 1, characterized in that: Step 5 includes, First, a "fluid dynamics morphology residual map" is generated, and it is assumed that the model will identify the image as a bubble plume. ; Calculate target adversarial category With prediction category Perturbation prediction difference tensor in the latent space The expression is as follows: Then, the difference tensor is projected back into pixel space by the decoder to obtain the following visualized heatmap. : Heat map The high-response region on the surface is the key morphological feature by which the model distinguishes bubbles from sediments.

7. The method for identifying seabed bubble plumes in extreme deep-sea environments according to claim 1, characterized in that: In step 6, the recognition result output by the processor unit is bound to a timestamp, and the ultra-short baseline USBL positioning data and depth sensor data carried by the ROV are read simultaneously. When the identification result is "bubble plume", extract the latitude and longitude of the ROV at the current moment. and depth Combined with the camera's attitude angle To calculate the absolute geographic coordinates of the bubble vent on the seabed; among which, These are the longitude and latitude values ​​of the ROV's reference center in the world geodetic coordinate system. These refer to the pitch angle, yaw angle, and heading angle of the camera coordinate system relative to the geographic coordinate system, respectively. Finally, the distribution map of the seabed cold seep vents was drawn in real time in the form of a 3D point cloud in the host computer software, which intuitively shows the potential location of mineral resources in the exploration area.

Images