A station waterlogging intelligent monitoring method and system

Abstract

The application discloses a kind of station waterlogging intelligent monitoring methods, including obtaining degradation monitoring image and auxiliary modal data;Through the auxiliary modal data, the physical attribute graph of waterlogging area is predicted, and the degradation monitoring image is guided to carry out feature enhancement using the physical attribute graph, and enhanced feature is obtained;The enhanced feature is reconstructed using diffusion model, and the physical attribute graph and shallow water equation are used as physical constraint in the reverse denoising process of the diffusion model Iterative correction is carried out, and a reconstructed image is generated;Waterlogging monitoring is carried out according to the reconstructed image;Meanwhile, the application also discloses a kind of station waterlogging intelligent monitoring system, including data acquisition module, physical perception enhancement module, diffusion reconstruction module and waterlogging analysis module.The application solves the problem that it is difficult to robustly and accurately reconstruct waterlogging monitoring visual features from severely degraded monitoring images in extreme rain and fog weather.

Description

Technical Field

[0001] This invention belongs to the field of urban rail transit construction technology, and in particular relates to an intelligent monitoring method and system for station water accumulation. Background Technology

[0002] Station water accumulation monitoring is a crucial aspect of ensuring the safe operation of urban rail transit. Traditional manual inspection methods are inefficient and slow to respond, making it difficult to meet the real-time monitoring needs under extreme weather conditions. In recent years, with the development of computer vision and deep learning technologies, intelligent water accumulation monitoring methods based on surveillance images have gradually been applied.

[0003] Existing technologies include using surveillance cameras to capture images of station areas, employing image segmentation algorithms to identify waterlogged areas, and combining this data with information from water level sensors and rainfall sensors for comprehensive judgment. However, these solutions still face significant technical bottlenecks in engineering applications.

[0004] First, under extreme rain and fog conditions, surveillance images suffer severe degradation, with a significant drop in image quality. Existing methods often treat this weather degradation as "noise" and attempt to eliminate it, employing preprocessing algorithms such as rain removal and fog removal to restore the image. However, these methods neglect the intrinsic relationship between the degradation process and the physical environment (such as raindrop size, fog concentration, and lighting conditions). In complex weather conditions with both rain and fog, the restored images often exhibit texture distortion or loss of detail. Image quality is fundamental to subsequent water accumulation identification and segmentation; distortion of the entry point information will jeopardize the accuracy of subsequent algorithms.

[0005] Secondly, existing solutions for fusing multi-sensor data are relatively simple, typically involving logical judgments or data overlay between image analysis results and sensor readings. This shallow fusion fails to deeply explore the physical consistency and complementary relationships between different modal data, resulting in a lack of effective robustness guarantees for the system when sensor data conflicts occur or a particular modality fails.

[0006] Furthermore, current deep learning models primarily rely on data-driven approaches and lack explicit constraints on physical laws such as hydrodynamics. While these models perform well in scenarios covered by training data, their predictions may defy physical realities in extreme or unseen scenarios, resulting in insufficient generalization ability. Although some research has attempted to use Physical Information Neural Networks (PINNs) for image reconstruction or to introduce physical constraints into diffusion models, no systematic approach has yet been developed for the specific scenario of station flooding monitoring, particularly a technical route combining cross-modal semantic guidance with physical constraint generation, which is not disclosed in existing technologies.

[0007] In summary, the core technical challenge that limits the performance of existing station water accumulation monitoring systems is how to robustly and accurately reconstruct visual features for water accumulation monitoring from severely degraded surveillance images under extreme rain and fog conditions. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a smart monitoring method for station water accumulation, addressing the shortcomings of the prior art.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for intelligent monitoring of water accumulation in railway stations, comprising: Acquire degraded monitoring images and auxiliary modal data; The physical property map of the water accumulation area is predicted by the auxiliary modal data, and the physical property map is used to guide the feature enhancement of the degraded monitoring image to obtain enhanced features; The enhanced features are reconstructed using a diffusion model. In the reverse denoising process of the diffusion model, the physical property map and shallow water equation are used as physical constraints for iterative correction to generate a reconstructed image. Water accumulation monitoring is performed based on the reconstructed images.

[0010] Furthermore, the auxiliary modal data includes millimeter-wave radar point cloud data and / or thermal infrared image data.

[0011] Furthermore, the physical property map of the waterlogged area is predicted using the auxiliary modal data, and the physical property map is used to guide feature enhancement of the degraded monitoring image to obtain enhanced features, specifically including: Multi-scale image features and multi-scale auxiliary modal features are extracted from the degraded monitoring images and the auxiliary modal data, respectively. Predict the physical property map based on the multi-scale auxiliary modal features; Using the physical property map as a query, attention-weighted enhancement is applied to the multi-scale image features to obtain the enhanced features.

[0012] Furthermore, the physical property map includes a probability map of water accumulation areas and / or a relative depth distribution map.

[0013] Furthermore, using the physical property map as a query, attention-weighted enhancement is applied to the multi-scale image features, specifically including: The physical property map is converted into a query vector through an embedding layer, and the multi-scale image features are converted into key vectors and value vectors. Calculate the attention weights between the query vector and the key vector; The enhanced feature is obtained by weighting and summing the value vector with the attention weights.

[0014] Furthermore, before predicting the physical property map of the water accumulation area using the auxiliary modal data, the method further includes: training the physical perception enhancement network using a contrast-reconstruction joint learning objective; The joint contrast-reconstruction learning objective includes cross-modal contrast loss and self-supervised reconstruction loss; The cross-modal contrastive loss is used to force the enhanced visual features to remain semantically close to the physical features extracted from the auxiliary modal data; The self-supervised reconstruction loss is used to utilize the spatial self-similarity of the image itself for reconstruction constraints.

[0015] Furthermore, the reverse denoising process of the diffusion model uses the physical property map and shallow water equation as physical constraints for iterative correction, specifically including: In each denoising iteration, the currently generated intermediate image is obtained; Extract the morphological and boundary features of the water accumulation area from the intermediate image; The morphological features and the boundary features are compared with the spatial constraints given by the physical property map and the motion law constraints given by the shallow water equation. When the intermediate image does not meet the physical constraints, the generated result will be fine-tuned to conform to the physical constraints.

[0016] Furthermore, the motion constraints given by the shallow water equation include constraints on the continuity of morphological changes in the water accumulation area between adjacent time frames.

[0017] Furthermore, water accumulation monitoring is performed based on the reconstructed image, specifically including: The reconstructed image is input into an adaptive prompt generator, which automatically generates segmentation prompts based on the image features of the reconstructed image and the segmentation results of the previous time step. The segmentation prompts are input into the visual model, which then segments the water accumulation region in the reconstructed image based on the segmentation prompts and outputs a water accumulation region segmentation mask. Water depth is estimated based on the segmentation mask of the water accumulation area.

[0018] Based on the same inventive concept, this invention also proposes an intelligent monitoring system for station water accumulation, comprising: The data acquisition module is used to acquire degraded monitoring images and auxiliary modal data; The physical perception enhancement module is used to predict the physical property map of the water accumulation area through the auxiliary modal data, and use the physical property map to guide the feature enhancement of the degraded monitoring image to obtain enhanced features; The diffusion reconstruction module is used to reconstruct the enhanced features using a diffusion model. In the reverse denoising process of the diffusion model, the physical property map and shallow water equation are used as physical constraints for iterative correction to generate a reconstructed image. A water accumulation analysis module is used to monitor water accumulation based on the reconstructed image.

[0019] Compared with the prior art, the present invention has the following advantages: 1. This invention introduces millimeter-wave radar and / or thermal infrared imaging, which are insensitive to weather conditions, as auxiliary modal data. It actively predicts the physical property map of the water accumulation area from the auxiliary modal data and uses the physical property map as semantic guidance information to perform cross-modal attention-weighted enhancement on degraded monitoring images. This achieves a shift from passive denoising to active perception, enabling accurate location and enhancement of water accumulation-related features in extreme weather conditions where rain and fog coexist. It effectively avoids the texture distortion and detail loss problems caused by traditional rain and fog removal algorithms, and significantly improves the targeting and effectiveness of degraded image enhancement.

[0020] 2. This invention uses physical property maps as spatial constraints and shallow water equations as motion constraints in the reverse denoising process of the diffusion model. It performs real-time correction on the intermediate images generated in each denoising iteration based on the physical model and fine-tunes the generated results that do not meet the physical constraints to conform to the physical laws. This achieves a deep integration of physical laws and data-driven approaches, ensuring that the water accumulation pattern in the final reconstructed image is accurate in space and coherent and smooth in time. It significantly improves the model's generalization ability and prediction reliability in extreme or unseen scenarios.

[0021] 3. This invention uses a cross-modal attention mechanism to enhance visual features by using the physical attribute map of the auxiliary modality prediction as the query. It also uses cross-modal contrast loss to force the enhanced visual features to remain close to the physical features extracted by the auxiliary modality in the semantic space. This achieves deep semantic fusion of multi-source heterogeneous data at the feature level, effectively overcoming the shallow fusion defects of simple data superposition or logical judgment in the prior art. It significantly improves the synergy of multi-source data fusion and the robustness of the system when a single modality fails.

[0022] 4. This invention trains the physical perception enhancement network by employing a contrast-reconstruction joint learning objective that combines cross-modal contrast loss and self-supervised reconstruction loss. It utilizes cross-modal contrast loss to maintain cross-modal semantic consistency and utilizes self-supervised reconstruction loss to maintain the visual naturalness of the image. This enables model training under conditions of clear-degraded image data that do not rely on pairing, significantly reducing data acquisition costs and annotation difficulties, and significantly improving the feasibility and scalability of the solution in practical engineering.

[0023] 5. This invention uses an adaptive prompt generator to automatically generate segmentation prompts based on the image features of the reconstructed image and the segmentation results of the previous time step. The visual big data model then uses these prompts to segment the water accumulation area, achieving automation and high precision in water accumulation area segmentation. This effectively solves the application bottleneck of the visual big data model in specific tasks such as water accumulation monitoring, where manual prompts are required, and fully leverages the powerful segmentation capabilities of the visual big data model.

[0024] 6. This invention links the water accumulation monitoring results with the station's drainage equipment and / or floodgates for control, enabling the automatic activation of the drainage equipment and / or the automatic closure of the floodgates when the water accumulation exceeds a preset threshold. This constructs a complete intelligent flood control closed loop from perception, enhancement, analysis to decision-making and execution, significantly improving the response speed and automation level of emergency response to water accumulation at stations.

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] Figure 1 The present invention provides an overall flowchart of a method for intelligent monitoring of water accumulation in railway stations.

[0027] Figure 2 This is a structural schematic diagram of a smart monitoring system for station water accumulation provided by the present invention. Detailed Implementation

[0028] This invention provides a method for intelligent monitoring of water accumulation in railway stations, such as... Figure 1 As shown, it includes: Step S1: Acquire degradation monitoring images and auxiliary modal data; Step S2: Predict the physical property map of the water accumulation area using the auxiliary modal data, and use the physical property map to guide feature enhancement of the degraded monitoring image to obtain enhanced features; Step S3: Reconstruct the enhanced features using a diffusion model. In the reverse denoising process of the diffusion model, the physical property map and shallow water equation are used as physical constraints for iterative correction to generate a reconstructed image. Step S4: Monitor water accumulation based on the reconstructed image.

[0029] In step S1, degraded monitoring images of the station's monitoring area and synchronously acquired auxiliary modal data are acquired. Degraded monitoring images refer to visible light images severely degraded under extreme rain and fog conditions, where rain lines, fog, and other interference cause blurring, reduced contrast, and loss of detail. Auxiliary modal data includes millimeter-wave radar point cloud data and / or thermal infrared image data. Millimeter-wave radar uses millimeter-wave electromagnetic waves to penetrate rain and fog for detection, while thermal infrared imaging senses temperature distribution by detecting the infrared radiation emitted by objects themselves. Both are unaffected by rain and fog and can stably provide spatial location information for water accumulation areas.

[0030] In step S2, a physical attribute map of the waterlogged area is predicted using auxiliary modal data. This physical attribute map includes a probability map of the waterlogged area and / or a relative depth distribution map, which are spatial distribution information with clear physical meaning, used to characterize the possible spatial location and relative depth distribution of the waterlogged area. Then, using this physical attribute map as semantic guidance information, a cross-modal attention mechanism is employed to enhance the features of the degraded monitoring image, resulting in enhanced features.

[0031] In step S3, the enhanced features are reconstructed using a diffusion model. During the reverse denoising process of the diffusion model, iterative correction is performed using the physical property map and the shallow water equation as physical constraints to generate the reconstructed image. The shallow water equation provides the mass conservation law for water movement, including the constraint on the continuity of morphological changes in the water accumulation area between adjacent time frames. That is, water movement satisfies the law of mass conservation, and the change in water accumulation morphology over time is continuous and smooth, without abrupt breaks or jumps.

[0032] In step S4, water accumulation monitoring is performed based on the reconstructed image. The reconstructed image is input into an adaptive prompt generator, which automatically generates segmentation prompts based on the image features of the reconstructed image and the segmentation results of the previous time step. Then, the segmentation prompts are input into a large visual model to segment the water accumulation area, outputting a water accumulation area segmentation mask. Water depth is then estimated based on the segmentation mask, and finally, the water accumulation monitoring results are output.

[0033] The aforementioned technical solution introduces millimeter-wave radar and / or thermal infrared imaging as auxiliary modal data. It actively predicts the physical property map of the water accumulation area from this auxiliary modal data and uses this physical property map as semantic guidance information to perform cross-modal attention-weighted enhancement on degraded images. This effectively solves the problem of difficulty in locating and extracting water accumulation features caused by severe degradation of monitoring images under extreme rain and fog conditions. Since millimeter-wave radar and thermal infrared imaging are unaffected by rain and fog, even when visible light images are almost unavailable, the auxiliary modal data can still provide reliable prior knowledge of the spatial location of water accumulation, thereby guiding the visual model to achieve targeted enhancement and avoiding the texture distortion and detail loss problems caused by blind denoising in traditional rain and fog removal algorithms. Based on this, through iterative correction using the physical property map as a spatial constraint and the shallow water equation as a motion law constraint in the inverse denoising process of the diffusion model, it ensures that the water accumulation morphology in the final reconstructed image is accurate in space and coherent and smooth in time. This achieves a deep fusion of physical laws and data-driven approaches, significantly improving the model's generalization ability and prediction reliability in extreme or unseen scenarios. Furthermore, the adaptive prompt generator automatically generates segmentation prompts, which are then used by a large visual model to segment waterlogged areas. This automates and increases the accuracy of the segmentation process, effectively addressing the bottleneck of requiring manual prompts in specific tasks. The entire solution constructs a complete intelligent flood control closed loop encompassing "perception-enhancement-analysis-decision-execution," achieving a comprehensive leap from passive noise reduction to proactive perception, from purely data-driven to a dual physical-data-driven approach, and from manual intervention to automated processing.

[0034] In step S1 of the present invention, the auxiliary modal data includes millimeter-wave radar point cloud data and / or thermal infrared image data.

[0035] In practical engineering deployments, millimeter-wave radar point cloud data and / or thermal infrared image data are selected as auxiliary modal data. Millimeter-wave radar works by utilizing electromagnetic waves in the millimeter-wave band for detection. These waves have relatively long wavelengths, effectively penetrating raindrops and fog particles. Therefore, radar point cloud data is largely unaffected by rain or fog, and can stably detect the approximate outline of surface structures and waterlogged areas. Thermal infrared imaging works by detecting the infrared radiation emitted by objects themselves, without relying on external light sources. Water has a high heat capacity (specific heat capacity approximately 4.2 × 10³ J / (kg·℃)), while concrete pavement has a relatively low heat capacity. A significant temperature difference exists between water and pavement under day and night conditions. Therefore, water forms a distinct thermal characteristic in thermal infrared images compared to the surrounding pavement, similarly unaffected by rain, fog, or lighting conditions. Both auxiliary modal data can be used individually or in combination, offering flexible configuration options.

[0036] This invention utilizes millimeter-wave radar and / or thermal infrared imaging as auxiliary modal data, taking advantage of the physical characteristics of these two sensing technologies being insensitive to rain and fog, thus providing a reliable data source for subsequent cross-modal semantic guidance. Compared with traditional solutions that rely solely on single visible light images or simple water level sensor data, the auxiliary modal data introduced in this invention can independently and stably provide prior spatial location data of waterlogged areas under adverse weather conditions, fundamentally solving the problem of data ingress distortion in extreme weather conditions. Whether using a single auxiliary modality or fusing dual modalities, the robustness and accuracy of waterlogging monitoring under rain and fog are effectively improved. For cost-sensitive stations or stations with limited installation space, millimeter-wave radar or thermal imagers can be deployed separately; for stations with higher monitoring accuracy requirements, both sensors can be deployed simultaneously to achieve complementary advantages. This flexible configuration allows this invention to adapt to the diverse deployment needs of different stations, possessing broad engineering applicability.

[0037] In step S2 of the present invention, the physical property map of the waterlogged area is predicted by the auxiliary modal data, and the physical property map is used to guide the feature enhancement of the degraded monitoring image to obtain enhanced features. Specifically, this includes: extracting multi-scale image features and multi-scale auxiliary modal features from the degraded monitoring image and the auxiliary modal data, respectively; predicting the physical property map based on the multi-scale auxiliary modal features; and using the physical property map as a query to perform attention-weighted enhancement on the multi-scale image features to obtain the enhanced features.

[0038] This step is implemented using a cross-modal semantically guided physical perception enhancement network (CSP-Net), which includes a multimodal feature extraction module, a physical prior prediction head, and a cross-modal attention enhancement module. First, the multimodal feature extraction module employs two encoder branches: a visual encoder extracts multi-scale image features from the degraded monitoring image, and an auxiliary encoder extracts multi-scale auxiliary modal features from auxiliary modal data. Both encoders are implemented using multi-layer convolutional neural networks, capturing structural, textural, and physical information of their respective modalities at different scales. Second, the physical prior prediction head is a lightweight prediction network composed of convolutional layers. It predicts a physical property map based on the multi-scale auxiliary modal features. The physical property map includes a probability map of water accumulation areas and / or a relative depth distribution map, providing prior guidance in terms of spatial location and scale for subsequent attention enhancement. Finally, the cross-modal attention enhancement module uses the physical property map as a query to perform attention-weighted enhancement on the multi-scale image features, enabling the model to accurately focus on water-related areas in the degraded image, enhancing key features and suppressing irrelevant noise, resulting in enhanced image features.

[0039] The core design idea of ​​the above technical solution is to use weather-independent auxiliary modal data to "guide" the visual model in finding and enhancing features related to water accumulation in severely degraded surveillance images. The principle lies in the fact that the auxiliary modal data provides independent and reliable spatial location priors about water accumulation areas. These priors, acting as "semantic landmarks," guide the visual model to accurately focus on degraded images through a cross-modal attention mechanism. Specifically, multi-scale feature extraction ensures that the model can simultaneously capture global contextual information and local detail information of the image. The physical prior prediction head explicitly extracts the physical information contained in the auxiliary modal data into attribute maps, and the cross-modal attention mechanism uses these physical attribute maps as guidance for feature weighting. When the physical prior indicates that "this area is likely to be flooded," the attention mechanism causes the model to focus on searching for texture and edge features in that area in the degraded image and assign higher weights to these features; conversely, for areas where the physical prior indicates a low probability of water accumulation, the attention mechanism suppresses the feature responses of that area. In this way, the model can accurately "focus" on key water accumulation areas in blurred, degraded images, preserving key details and suppressing irrelevant noise during the enhancement process. Compared with traditional single-modal image enhancement methods, this scheme achieves deep semantic fusion of multi-source data at the feature level, effectively overcoming the shallow fusion defects of simple data superposition or logical judgment in existing technologies, and significantly improving the synergy of multi-source data fusion and the robustness of the system when a single modality fails.

[0040] The physical property map includes a probability map of water accumulation areas and / or a relative depth distribution map.

[0041] The physical property maps include a water accumulation area probability map and / or a relative depth distribution map. Each pixel value in the water accumulation area probability map represents the probability of water accumulation at that spatial location, ranging from 0 to 1. Values ​​closer to 1 indicate a higher probability of water accumulation. This map is obtained by performing convolutional mapping and sigmoid activation on auxiliary modal features. Each pixel value in the relative depth distribution map represents the relative water depth at that location, reflecting the relative depth relationships between different locations within the water accumulation area. This map is a continuous value distribution map obtained by performing regression mapping on auxiliary modal features. These two maps are spatial distribution information with clear physical meaning predicted from auxiliary modal data. They are not the final monitoring results but serve as "semantic soft labels" and physical constraints for subsequent enhancement and diffusion reconstruction processes. The water accumulation area probability map primarily serves the cross-modal attention enhancement module, providing spatial location guidance for feature enhancement; the relative depth distribution map primarily serves the physical constraint correction of the diffusion model, providing spatial constraints for the depth distribution of the reconstructed image.

[0042] This invention explicitly extracts the physical information contained in auxiliary modal data by predicting probability maps and relative depth distribution maps of water accumulation areas from the auxiliary modal data. This information is then presented as a visualized attribute map, allowing the physical knowledge within the auxiliary modal data to participate in subsequent deep learning processes in an interpretable manner. The water accumulation area probability map provides spatial guidance for the cross-modal attention enhancement module, enabling the model to focus its attention on areas most likely to have water accumulation, thus effectively avoiding wasting computational resources in non-water accumulation areas. The relative depth distribution map provides spatial depth constraints for the inverse denoising process of the diffusion model, ensuring that the depth distribution of water accumulation areas in the reconstructed image is consistent with the physical prior. The two physical attribute maps complement each other, providing comprehensive physical prior information from the two dimensions of "whether water accumulation exists" and "how deep the water is." Compared to existing technologies that only use a single water level value for judgment, the physical attribute maps of this invention provide richer and more detailed spatial distribution information of water accumulation, providing sufficient spatial constraint guidance for subsequent image enhancement and diffusion reconstruction, significantly improving the spatial resolution and precision of water accumulation monitoring.

[0043] Specifically, the attention-weighted enhancement of the multi-scale image features using the physical property map as the query includes: converting the physical property map into a query vector through an embedding layer; converting the multi-scale image features into key vectors and value vectors; calculating the attention weights of the query vector and the key vector; and performing a weighted summation of the value vectors using the attention weights to obtain the enhanced features.

[0044] The specific implementation of the cross-modal attention enhancement module is as follows: First, the physical property map (probability map of water accumulation area and relative depth distribution map) is converted into a query vector Q through an embedding layer. The embedding layer consists of convolutional layers and fully connected layers, used to encode the two-dimensional spatial distribution of the physical property map into a sequence of feature vectors. Multi-scale image features are converted into key vectors K and value vectors V, respectively. The key vectors are used for relevance matching with the query vector, and the value vectors carry the actual content information of the image features. Then, the dot product of the query vector Q and the key vector K is calculated, scaled by a scaling factor d (d is the vector dimension) to prevent the gradient from entering the saturation region, and then normalized by softmax to obtain the attention weight matrix A = softmax(Q·K^T / d). Each element in the attention weight A reflects the correlation between the high-probability water accumulation area indicated by the physical property map and various spatial locations in the image features. Finally, the attention weight A is used to weight and sum the value vector V to obtain the enhanced image feature F_e = A·V. The above process is the mathematical implementation of the cross-modal attention mechanism. Through this mechanism, prior information in the physical property map is applied to image features as attention weights, thereby achieving feature enhancement guided by physical priors.

[0045] The aforementioned technical solution incorporates physical priors into the forward propagation process of a neural network in a differentiable manner through the mathematical design of a cross-modal attention mechanism. The underlying principle is that the calculation of attention weights essentially involves matching the spatial location information in the physical property map with the visual information in the image features. Regions of interest based on the physical prior (high-probability water accumulation areas) receive higher weights in the attention calculation, thus enhancing the corresponding image features after weighted summation. Conversely, image features corresponding to regions not of interest based on the physical prior (non-water accumulation areas) are suppressed. Compared to traditional channel attention or spatial attention mechanisms (which rely solely on image-specific information to calculate attention weights), this invention's cross-modal attention mechanism introduces external physical priors as a "query" for attention calculation, making the attention distribution more accurate and targeted. This is because the information in the degraded image itself is unreliable due to rain and fog interference, and attention weights calculated solely based on the image itself may drift. In contrast, the physical property map comes from auxiliary modal data unaffected by weather, providing reliable prior information. By using reliable physical priors as queries to calculate attention weights, the attention drift problem caused by simply relying on the features of the degraded image itself is effectively avoided, making the feature enhancement results more accurate and robust.

[0046] In this invention, before predicting the physical property map of the water accumulation area using the auxiliary modal data in step S2, the method further includes: training the physical perception enhancement network using a contrast-reconstruction joint learning objective; the contrast-reconstruction joint learning objective includes cross-modal contrast loss and self-supervised reconstruction loss; the cross-modal contrast loss is used to force the enhanced visual features to remain close to the physical features extracted from the auxiliary modal data in the semantic space; the self-supervised reconstruction loss is used to utilize the spatial self-similarity of the image itself for reconstruction constraints.

[0047] CSP-Net training employs a joint contrast-reconstruction learning objective, which includes cross-modal contrastive loss and self-supervised reconstruction loss. The cross-modal contrastive loss uses the InfoNCE loss function. For the same training sample, the degraded surveillance image and the auxiliary modal data describe the same scene, and their semantic content should be consistent. Therefore, the visual features enhanced by CSP-Net and the physical features extracted from the auxiliary modal data should be close to each other in the semantic space. The InfoNCE loss shortens the distance between paired positive sample pairs (enhanced visual features and their corresponding physical features) in the semantic space, while simultaneously widening the distance between unpaired negative sample pairs (enhanced visual features and the physical features of other samples), forcing the enhanced visual features and the physical features extracted from the auxiliary modality to remain close in the semantic space, ensuring cross-modal semantic consistency. The self-supervised reconstruction loss adopts the implementation of a masked autoencoder, which uses the spatial self-similarity of the image itself to perform reconstruction constraints. It randomly masks a part of the image features (such as 50% of the spatial location in the image), and then trains the model to predict the masked part based on the unmasked part. By minimizing the difference between the prediction result and the original features, the model learns the intrinsic structural prior of the image, ensuring the visual naturalness of the enhanced image.

[0048] The training strategy described in this invention addresses the data bottleneck problem of obtaining paired clear-degraded images in practical engineering. The principle is that acquiring paired multimodal data (i.e., data pairs containing both clear and degraded images of the same scene) is extremely difficult in real-world scenarios, as it's impossible to simultaneously capture a "clear" version of the same scene under rainy or foggy conditions. To solve this problem, this invention employs a self-supervised learning strategy, completing model training without relying on paired clear images. The principle of cross-modal contrastive loss is to utilize the semantic consistency of different modalities describing the same scene as a supervisory signal, using the pairing relationships of multimodal data as natural data labels. The principle of self-supervised reconstruction loss is to utilize the spatial self-similarity of the image itself (i.e., similar texture regions in the image have similar feature representations) to construct a supervisory signal, ensuring that the image's structural prior remains even when affected by rain or fog. Through joint optimization of these two losses, CSP-Net can learn effective feature enhancement capabilities without labeled data. Compared to existing supervised training methods that require a large number of paired clear-degraded images, the training strategy of this invention significantly reduces data acquisition costs and annotation difficulties, and significantly improves the feasibility and scalability of the solution in practical engineering, making large-scale deployment possible.

[0049] In step S3 of the present invention, the reverse denoising process of the diffusion model uses the physical property map and the shallow water equation as physical constraints for iterative correction, specifically including: in each denoising iteration, obtaining the currently generated intermediate image; extracting the morphological features and boundary features of the water accumulation area from the intermediate image; comparing the morphological features and the boundary features with the spatial constraints given by the physical property map and the motion law constraints given by the shallow water equation; when the intermediate image does not meet the physical constraints, fine-tuning the generation result to conform to the direction of the physical constraints.

[0050] PGDR uses the Denoising Diffusion Probability Model (DDPM) as its basic architecture and embeds a physical correction mechanism in the reverse denoising process. Let the total number of denoising steps in the diffusion model be N. In the t-th denoising iteration (t = N, N-1, ..., 1), the model predicts x_{t-1} based on the current noisy intermediate image x_t. In standard DDPM, the prediction of x_{t-1} is based solely on the learned data distribution. However, in the PGDR of this invention, a physical correction step is added after obtaining the initial prediction: First, morphological and boundary features of the water accumulation area are extracted from the intermediate image of the initial prediction. Morphological features include geometric measures such as the area, perimeter, and compactness of the water accumulation area, while boundary features include topological properties such as the smoothness, continuity, and closure of the boundary. Then, these features are compared with the spatial constraints given by the physical property map and the motion constraints given by the shallow water equation. Spatial constraints mean that the water accumulation area should be located within the high-probability region indicated by the water accumulation area probability map in the physical property map, and the depth distribution should be consistent with the relative distribution indicated by the depth relative distribution map. Motion constraint means that the morphological changes of the water accumulation area between adjacent time frames should satisfy the water movement laws described by the shallow water equation, i.e., the mass conservation constraint. When the water accumulation morphology in the intermediate image does not meet the above physical constraints, x_{t-1} is fine-tuned to conform to the physical constraints using gradient descent. Specifically, a physical constraint loss function L_physics is constructed to measure the deviation between the water accumulation morphology in the intermediate image and the physical constraints. Then, the gradient is calculated through backpropagation, and x_{t-1} is updated along the gradient descent direction. After N iterations, the final reconstructed image is output.

[0051] The above scheme achieves a deep fusion of physical laws and data-driven approaches by embedding a physical correction mechanism in the reverse denoising process of the diffusion model. The principle is that the standard diffusion model only focuses on the consistency between the generated image and the training data distribution during reverse denoising, without considering whether the generated result conforms to physical laws. This may lead to physically unreasonable water accumulation patterns (such as impossible water boundary shapes or depth distributions that severely contradict physical principles). This invention adds a physical model-based correction step in each denoising iteration, checking in real time whether the currently generated intermediate image conforms to physical constraints, and fine-tuning the generated results that do not meet physical constraints back to conforming to physical laws. The shallow water equation expresses the basic physical principle of water mass conservation; the morphological changes of the water accumulation area between adjacent frames are continuous and smooth, and there should be no sudden boundary breaks, area jumps, or other phenomena that violate mass conservation. The physical property map provides a priori spatial distribution of the water accumulation area within a single frame image. By combining the two, dual physical constraints in the spatial domain (physical property map) and the temporal domain (shallow water equation) are achieved, ensuring that the final reconstructed water accumulation pattern is accurate in space and coherent and smooth in time. Compared to existing data-driven deep learning models, the solution of this invention makes the prediction results more consistent with physical reality, has stronger interpretability and reliability, and has stronger generalization ability in extreme or unseen scenarios.

[0052] The motion constraints given by the shallow water equation include the constraint on the continuity of morphological changes in the water accumulation area between adjacent time frames.

[0053] Shallow water equations are a set of partial differential equations describing the motion of water flow on a free surface. The basic form of their continuity equation is: h / t + (hu) / x + (hv) / y = 0, where h is the water depth, and u and v are the water velocities in the x and y directions, respectively. This equation expresses the fundamental physical principle of water mass conservation: within a unit area, the rate of change of water depth over time is equal to the difference between the amount of water flowing into and out of that unit area. Based on this physical principle, the morphological changes of a water accumulation area between adjacent time frames (including area changes, boundary displacements, depth distribution changes, etc.) should be continuous and smooth, without sudden breaks, jumps, or other unreasonable phenomena that violate mass conservation. The morphological change continuity constraint requires that, when reconstructing images, the morphological changes of water accumulation between adjacent frames should satisfy the above physical continuity. Specifically, this means that the boundaries of the water accumulation area move continuously over time, the rate of change of the water accumulation area is within the physically permissible range, and the depth distribution evolves smoothly over time.

[0054] This invention incorporates the motion constraints of shallow water equations into the inverse denoising process of a diffusion model. This ensures that the model is constrained by physical laws while learning image generation, guaranteeing the physical consistency of the generated results over time. When abrupt changes in the water accumulation morphology between adjacent frames occur in the intermediate images generated by the model (e.g., a sudden break in the water boundary or a sudden expansion of the area in a frame, which is physically impossible), the physical correction step can identify such violations of physical laws and correct the generated results back to a direction consistent with physical laws through gradient fine-tuning. This constraint ensures the physical consistency and visual coherence of the reconstructed image sequence over time, avoiding the jitter and inconsistency problems commonly found in time-series prediction using purely data-driven methods. Compared to schemes that only use spatial constraints (such as physical property maps), combining the motion constraints of shallow water equations with the introduction of temporal physical priors significantly improves the temporal coherence of the reconstruction results. This provides a more stable and reliable input for subsequent water accumulation region segmentation and water depth estimation, especially in water accumulation segmentation tasks with consecutive frames, effectively reducing inter-frame jumps in the segmentation results and improving the temporal consistency of monitoring results.

[0055] In step S4 of the present invention, water accumulation monitoring is performed based on the reconstructed image, specifically including: inputting the reconstructed image into an adaptive prompt generator, wherein the adaptive prompt generator automatically generates segmentation prompts based on the image features of the reconstructed image and the segmentation results of the previous time step; inputting the segmentation prompts into a large visual model, wherein the large visual model performs water accumulation region segmentation on the reconstructed image based on the segmentation prompts and outputs a water accumulation region segmentation mask; and estimating the water depth based on the water accumulation region segmentation mask.

[0056] The water accumulation monitoring process comprises three sub-steps. First, the reconstructed image is input into an Adaptive Cue Generator (APG), which includes a feature encoding submodule, a temporal modeling submodule, and a cue generation submodule. The feature encoding submodule uses a lightweight convolutional neural network (such as MobileNetV3) to extract multi-scale visual features from the reconstructed image, including color histogram features, Local Binary Pattern (LBP) texture features, Canny edge features, and deep semantic features. The temporal modeling submodule uses a Long Short-Term Memory (LSTM) network to perform temporal modeling on the water accumulation area segmentation results from historical moments. The LSTM contains memory units and gating mechanisms, effectively capturing long-term dependencies in the time series. By modeling historical segmentation results using the LSTM, the temporal modeling submodule can predict the movement trend and morphological changes of the water accumulation area at the current moment. The cue generation submodule integrates the visual features of the current frame and temporal priors, fusing them through an attention fusion mechanism. A coarse heatmap of the water accumulation area is then output via a convolutional layer. Finally, a non-maximum suppression (NMS) algorithm is used to select several high-confidence locations from the heatmap as point cues. Boundary cues are generated based on the bounding rectangles of the connected components of the high-value regions in the heatmap. The point and bounding cues generated by APG are then input into a large visual model (such as the Segment Anything Model, SAM). SAM uses these cues to segment the water accumulation area in the reconstructed image, outputting a water accumulation area segmentation mask (a binary image where a pixel value of 1 represents water accumulation and 0 represents non-water accumulation). Finally, based on the water accumulation area segmentation mask, visual depth information in the reconstructed image, and distance information for the corresponding area provided by millimeter-wave radar, a lightweight water depth regression network is used to estimate the water depth, outputting a water depth distribution map of the entire water accumulation area.

[0057] The above solution automatically generates segmentation cues for the large visual model through an adaptive cue generator, solving the application bottleneck of requiring manual cues in specific tasks such as water accumulation monitoring. This achieves automated and high-precision water accumulation region segmentation. The principle behind this solution is that APG (Adaptive Prompt Generator) comprehensively utilizes the visual features of the current frame (color, texture, edges) and temporal priors (historical segmentation results) to accurately determine the possible locations of water accumulation regions. It then expresses this location information in a cue format (dot cues and box cues) that the large visual model can understand. Dot cues indicate to the large visual model that "these locations belong to water accumulation regions," allowing it to use these locations as positive samples during segmentation. Box cues indicate that "the water accumulation region is roughly located within this rectangular area," limiting the search space of the large visual model. These two methods complement each other, effectively guiding the large visual model to perform accurate segmentation. Compared to the traditional method that relies on manual cues, the cues automatically generated by APG are more timely, stable, and adaptable to dynamic changes in the scene, enabling the large visual model to operate automatically around the clock without manual intervention. By leveraging the powerful zero-shot generalization segmentation capabilities of the large visual model SAM, high accuracy and high generalization of water accumulation area segmentation were achieved. Finally, water depth estimation was performed using distance information from millimeter-wave radar, fully utilizing the advantage that radar data is unaffected by weather conditions, providing a reliable depth reference for water depth estimation. Compared to methods that rely solely on visual information for water depth estimation, the water depth estimation results of this invention are more accurate and stable.

[0058] Based on the same inventive concept, this invention also provides a smart monitoring system for station water accumulation, such as... Figure 2 As shown, it includes: The data acquisition module is used to acquire degraded monitoring images and auxiliary modal data; The physical perception enhancement module is used to predict the physical property map of the water accumulation area through the auxiliary modal data, and use the physical property map to guide the feature enhancement of the degraded monitoring image to obtain enhanced features; The diffusion reconstruction module is used to reconstruct the enhanced features using a diffusion model. In the reverse denoising process of the diffusion model, the physical property map and shallow water equation are used as physical constraints for iterative correction to generate a reconstructed image. A water accumulation analysis module is used to monitor water accumulation based on the reconstructed image.

[0059] The system comprises a data acquisition module, a physical sensing enhancement module, a diffusion reconstruction module, a water accumulation analysis module, and an edge computing unit. The data acquisition module includes binocular monitoring cameras, millimeter-wave radar, and thermal imagers, deployed at key water accumulation risk points such as station entrances / exits, platforms, and / or depots. All sensors are synchronized via a time synchronization module and acquire data synchronously through a data acquisition card. It also includes a data preprocessing submodule for performing preprocessing operations on the acquired multi-source data, such as size normalization, filtering and noise reduction, rasterization, temperature calibration, and geometric correction. (Refer to...) Figure 1 The physical perception enhancement module executes step S2 of the method. It integrates a CSP-Net model, including a multimodal feature extraction submodule, a physical prior prediction head submodule, and a cross-modal attention enhancement submodule. The input is preprocessed multimodal data pairs, and the output is enhanced image features. The diffusion reconstruction module executes step S3 of the method. It integrates a PGDR model, including a diffusion model inverse denoising submodule and a physical constraint correction submodule. The input is enhanced image features, and the output is a reconstructed image. The water accumulation analysis module executes step S4 of the method. It integrates APG and a large visual model SAM. The input is the reconstructed image, and the output is water accumulation region segmentation masks and water depth distribution maps, among other water accumulation monitoring results. All three modules—physical perception enhancement, diffusion reconstruction, and water accumulation analysis—are deployed in software on an edge computing unit. The edge computing unit uses a GPU-accelerated embedded AI computing platform to achieve real-time on-site data processing. The system also includes a communication module for data interaction with the station's flood control platform and a control module for coordinated control of drainage equipment and floodgates based on water accumulation monitoring results.

[0060] The system of this invention, through the coordinated operation of the four modules mentioned above, fully realizes intelligent monitoring of the entire process from data acquisition, feature enhancement, image reconstruction to water accumulation analysis. The data acquisition module utilizes multi-source sensors to simultaneously acquire visible light images and auxiliary modal data, providing a reliable data foundation for subsequent analysis even under extreme weather conditions. The physical perception enhancement module ensures accurate extraction of water accumulation features from degraded images through physical prior-guided feature enhancement. The diffusion reconstruction module ensures the quality and physical reliability of the reconstructed images through physically constrained diffusion reconstruction. The water accumulation analysis module achieves automated and high-precision segmentation by combining an adaptive cue generator with a large visual model, and then uses radar data for water depth estimation. Each module is deployed on an edge computing unit to process data locally, avoiding network bandwidth pressure and transmission delays caused by transmitting large amounts of video data back to the cloud or central server, ensuring the real-time nature of water accumulation monitoring. The control module realizes a complete closed loop from perception to execution, automatically triggering the activation of drainage equipment and / or the closure of floodgates when the water accumulation monitoring results exceed a preset threshold. The entire system has a clear structure, distinct layers, and a complete technical route, demonstrating good engineering feasibility and promising prospects for widespread application.

[0061] In summary, the intelligent monitoring method and system for station flooding provided by this invention effectively solves the technical problem of existing technologies struggling to robustly and accurately reconstruct the visual features of flooding monitoring from severely degraded surveillance images under extreme rain and fog conditions. This is achieved through cross-modal semantic-guided physical perception enhancement and generative diffusion reconstruction constrained by physical information. The method represents a comprehensive leap from passive denoising to active perception, from purely data-driven to a dual physical-data driven approach, and from manual intervention to automated processing. It provides a new generation of intelligent solutions with high robustness, high precision, and high reliability for station flood prevention monitoring. This solution can be widely applied to intelligent monitoring and early warning of flooding in urban rail transit stations, underground passages, sunken plazas, and other scenarios, demonstrating significant practical value and economic significance.

[0062] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent monitoring of water accumulation in railway stations, characterized in that, include: Acquire degraded monitoring images and auxiliary modal data; The physical property map of the water accumulation area is predicted by the auxiliary modal data, and the physical property map is used to guide the feature enhancement of the degraded monitoring image to obtain enhanced features; The enhanced features are reconstructed using a diffusion model. In the reverse denoising process of the diffusion model, the physical property map and shallow water equation are used as physical constraints for iterative correction to generate a reconstructed image. Water accumulation monitoring is performed based on the reconstructed images.

2. The intelligent monitoring method for station water accumulation according to claim 1, characterized in that, The auxiliary modal data includes millimeter-wave radar point cloud data and / or thermal infrared image data.

3. The intelligent monitoring method for station water accumulation according to claim 1, characterized in that, The physical property map of the waterlogged area is predicted using the auxiliary modal data. This physical property map is then used to guide feature enhancement of the degraded monitoring image, resulting in enhanced features, specifically including: Multi-scale image features and multi-scale auxiliary modal features are extracted from the degraded monitoring images and the auxiliary modal data, respectively. Predict the physical property map based on the multi-scale auxiliary modal features; Using the physical property map as a query, attention-weighted enhancement is applied to the multi-scale image features to obtain the enhanced features.

4. The intelligent monitoring method for station water accumulation according to claim 3, characterized in that, The physical property map includes a probability map of water accumulation areas and / or a relative depth distribution map.

5. The intelligent monitoring method for station water accumulation according to claim 3, characterized in that, Using the physical property map as a query, attention-weighted enhancement is applied to the multi-scale image features, specifically including: The physical property map is converted into a query vector through an embedding layer, and the multi-scale image features are converted into key vectors and value vectors. Calculate the attention weights between the query vector and the key vector; The enhanced feature is obtained by weighting and summing the value vector with the attention weights.

6. The intelligent monitoring method for station water accumulation according to claim 1, characterized in that, Before predicting the physical property map of the waterlogged area using the auxiliary modal data, the method further includes: training the physical perception enhancement network using a contrast-reconstruction joint learning objective; The joint contrast-reconstruction learning objective includes cross-modal contrast loss and self-supervised reconstruction loss; The cross-modal contrastive loss is used to force the enhanced visual features to remain semantically close to the physical features extracted from the auxiliary modal data; The self-supervised reconstruction loss is used to utilize the spatial self-similarity of the image itself for reconstruction constraints.

7. The intelligent monitoring method for station water accumulation according to claim 1, characterized in that, The reverse denoising process of the diffusion model uses the physical property map and shallow water equation as physical constraints for iterative correction, specifically including: In each denoising iteration, the currently generated intermediate image is obtained; Extract the morphological and boundary features of the water accumulation area from the intermediate image; The morphological features and the boundary features are compared with the spatial constraints given by the physical property map and the motion law constraints given by the shallow water equation. When the intermediate image does not meet the physical constraints, the generated result will be fine-tuned to conform to the physical constraints.

8. The intelligent monitoring method for station water accumulation according to claim 7, characterized in that, The motion constraints given by the shallow water equation include: the continuity constraint of the morphological changes of the water accumulation area between adjacent time frames.

9. The intelligent monitoring method for station water accumulation according to claim 1, characterized in that, Water accumulation monitoring based on the reconstructed image specifically includes: The reconstructed image is input into an adaptive prompt generator, which automatically generates segmentation prompts based on the image features of the reconstructed image and the segmentation results of the previous time step. The segmentation prompts are input into the visual model, which then segments the water accumulation region in the reconstructed image based on the segmentation prompts and outputs a water accumulation region segmentation mask. Water depth is estimated based on the segmentation mask of the water accumulation area.

10. A smart monitoring system for station water accumulation, characterized in that, include: The data acquisition module is used to acquire degraded monitoring images and auxiliary modal data; The physical perception enhancement module is used to predict the physical property map of the water accumulation area through the auxiliary modal data, and use the physical property map to guide the feature enhancement of the degraded monitoring image to obtain enhanced features; The diffusion reconstruction module is used to reconstruct the enhanced features using a diffusion model. In the reverse denoising process of the diffusion model, the physical property map and shallow water equation are used as physical constraints for iterative correction to generate a reconstructed image. A water accumulation analysis module is used to monitor water accumulation based on the reconstructed image.

Images