Railway foreign matter invasion real-time monitoring and alarming system based on SAR (Synthetic Aperture Radar)

Abstract

The invention relates to the technical field of rail traffic safety monitoring, and discloses a railway foreign matter invasion real-time monitoring and alarm system based on an SAR radar, and the system comprises an SAR multi-dimensional information depth perception module which extracts polarization, coherence, texture and other multi-dimensional features from SAR data; the multi-scale space-time attention fusion module adaptively focuses the key space-time change through a space and time attention mechanism to generate space-time fusion features; and the railway foreign matter invasion evolution knowledge graph construction and reasoning module integrates the space-time fusion features and multi-source information such as railway infrastructure, environment and historical events to construct a dynamic knowledge graph, and performs risk reasoning and alarming by using a graph neural network. According to the invention, through deep fusion of multi-dimensional perception, intelligent attention and knowledge graph reasoning, the accuracy, robustness and intelligent level of foreign matter invasion monitoring are significantly improved, and deep understanding and adaptive early warning of risks are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of railway traffic safety monitoring, and specifically to a real-time monitoring and alarm system for railway foreign object intrusion based on SAR radar. Background Art

[0002] At present, there are various monitoring means for railway foreign object intrusion, including manual inspection, video monitoring, fiber optic sensing, lidar, etc. Manual inspection has problems such as high labor intensity, low efficiency, poor real-time performance, and being greatly affected by bad weather, and it is difficult to meet the requirements of large-scale and all-weather monitoring. Although the traditional video monitoring system is intuitive, its performance will drop sharply under low visibility conditions such as at night, rain, snow, and fog, and it is easily affected by factors such as vegetation occlusion and viewing angle limitation, and its recognition ability for concealed foreign objects or intrusion events occurring in remote and complex terrain areas is limited. Although fiber optic sensing technology can achieve long-distance distributed monitoring, it is mainly sensitive to physical quantities such as vibration and strain, and it is not very direct in recognizing slowly deformed or static foreign objects, and the installation and maintenance costs are relatively high, and it is easily interfered by environmental noise. Lidar technology can obtain high-precision three-dimensional point cloud data, but it will also be affected by attenuation under weather conditions such as rain, snow, and fog, and its coverage range and data processing efficiency still face challenges.

[0003] Synthetic Aperture Radar (SAR), as an active microwave remote sensing technology, has the ability to observe all day and all weather, can penetrate clouds, rain, snow, and is highly sensitive to surface deformation and changes in target scattering characteristics. Therefore, it shows great potential in fields such as landslide monitoring and infrastructure deformation monitoring. Applying SAR technology to railway foreign object intrusion monitoring can theoretically overcome the limitations of traditional optical and some contact sensing technologies. However, the existing monitoring methods based on SAR technology still face many challenges in practical applications. For example, the inherent speckle noise and complex scattering mechanism in SAR images make it difficult to directly and accurately identify foreign object features from the images. At the same time, the environment along the railway is dynamically changing, and the SAR signal changes caused by human or natural disturbances such as vegetation growth, seasonal changes, and normal construction have a low distinguishability from the signal changes caused by real foreign object intrusion events, which easily leads to a high false alarm rate. In addition, the understanding of foreign object intrusion events often stays at the level of simple change detection or target recognition, lacking in-depth analysis and intelligent risk assessment capabilities for the causes, development trends of foreign objects, and their complex interaction relationships with the surrounding environment. Existing systems also rarely have a closed-loop mechanism for adaptive learning and model optimization according to the actual monitoring effects, and it is difficult to continuously adapt to the changing monitoring scenarios and new types of foreign objects.

[0004] Therefore, how to give full play to the advantages of SAR radar in all-weather and all-time operation, effectively overcome its own limitations, and integrate multi-source information and artificial intelligence technology to achieve precise, robust, and intelligent monitoring and early warning of railway foreign object intrusion events is a technical problem that urgently needs to be solved in the current field of railway traffic safety. Summary of the Invention

[0005] The present invention aims to provide a system and method for more intelligent and robust real-time monitoring and alarm of railway foreign object intrusion, so as to overcome the problems of insufficient robustness in identifying concealed, sudden, and diverse foreign objects and lack of intelligent and adaptive learning ability in the prior art under complex dynamic environments.

[0006] In the first aspect of the present invention, a real-time monitoring and alarm system for railway foreign object intrusion based on SAR radar is provided. The system innovatively integrates multi-dimensional SAR information perception, multi-scale spatio-temporal attention mechanism, and evolutionary knowledge graph reasoning to achieve in-depth understanding and precise early warning of railway foreign object intrusion risks. The system includes:

[0007] A deep perception module for multi-dimensional SAR information, whose core function is to refine and extract multi-dimensional feature information from the original data obtained by the SAR radar, providing a rich data basis for subsequent analysis.

[0008] Specifically, after preprocessing the SAR data, this module is dedicated to extracting:

[0009] Multi-polarization feature F pol : If the SAR system supports, polarization entropy H (such as Cloude-Pottier decomposition or Freeman-Durden decomposition) is obtained through polarization decomposition, pol anisotropy degree A, pol average Alpha angle α, pol and surface scattering power P, s even-order scattering power P, d and volume scattering power P, v etc. These features can reflect the fine scattering mechanism of the ground object.

[0010] Coherence feature F coh : For multi-temporal SAR data, the coherence γ between image pairs is calculated, coh and its calculation method can be: where s 1,i and s 2,i are the complex values at pixel i in two SAR images respectively, and N is the number of pixels in the estimation window. This feature helps to identify the stability changes of the scene.

[0011] Texture feature F tex: For example, the contrast T extracted through the gray-level co-occurrence matrix (GLCM), com energy T, asm etc., are used to characterize the spatial structure characteristics of the ground object surface. These multi-dimensional SAR feature information constitutes a comprehensive description of the physical properties of the monitoring scene.

[0012] The multi-scale spatio-temporal attention fusion module is innovative in simulating the human cognitive process, adaptively learning and focusing on the key spatio-temporal patterns indicating the risk of foreign object intrusion from high-dimensional and complex SAR feature sequences, while suppressing background noise and irrelevant changes.

[0013] This module receives the multi-dimensional SAR feature information X output by the SAR multi-dimensional information depth perception module feat , and proceeds in sequence:

[0014] Multi-scale spatial attention processing: First, capture the features at different spatial scales through a multi-scale feature extraction network (such as a feature pyramid network). Subsequently, introduce the channel attention and spatial attention mechanisms.

[0015] The channel attention map M c is calculated as follows to learn the importance of each feature channel:

[0016] M c (X feat ) = σ(MLP(AvgPool(X feat )) + MLP(MaxPool(X feat )));

[0017] where MLP represents a multi-layer perceptron, AvgPool and MaxPool are average pooling and max pooling respectively, and σ is the Sigmoid activation function.

[0018] The features weighted by channel attention [[]] ( represents element-wise multiplication), and then generate a spatial attention map through the spatial attention mechanism:

[0019] M s (X′f eat ) = σ(Conv s ([AvgPool s (X′ feat )); MaxPool s (X′ feat ))));

[0020] where Conv s is the convolutional layer, AvgPool s and MaxPools Pooling is performed along the channel dimension.

[0021] Finally, the spatial attention enhancement feature of the current phase is obtained This mechanism allows the system to focus on significant changes in key areas such as along railway lines.

[0022] Time series attention processing: The spatial attention enhancement feature sequence of T consecutive time phases {A s,1 ,A s,2 ,...,A s,T Input to a time series model such as a Long Short-Term Memory (LSTM) or Transformer Encoder.

[0023] Taking LSTM as an example, it uses the internal gating mechanism (input gate i t 、Forget Gate t , output gate o t ) effectively captures long-term dependencies in sequence data and dynamic evolution trends of scenarios:

[0024] i t =σ(W xi A s,t +W hi h t-1 +b i );

[0025] f t =σ(W xf A s,t +W hf h t-1 +b f );

[0026] o t =σ(W xo A s,t +W ho h t-1 +b o );

[0027] g t =tanh(W xc A s,t +W hc h t-1 +b c );

[0028] c t =f t ⊙c t-1 +i t ⊙g t ;

[0029] h t =ot ⊙tanh(c t );

[0030] Among them, h t is the hidden state at the current moment, representing the context-aware time feature A at this moment t This processing allows the system to understand the normal dynamic baseline of the scene, allowing it to more effectively identify unexpected deviations from the norm.

[0031] Spatiotemporal feature fusion: Enhance the spatial attention feature A of the current phase s and the corresponding context-aware temporal feature A t (For example, the h t ) are effectively fused (such as concatenated and passed through a convolutional layer or attention mechanism) to generate the final spatiotemporal fusion feature A s,t The feature A s,t It includes a deep understanding of the spatiotemporal changes of the scene to identify potential foreign objects intruding.

[0032] The construction and reasoning module of the railway foreign object intrusion evolution knowledge graph is one of the core innovations of this invention. It elevates railway safety monitoring from the traditional data-driven model to a new paradigm that combines "knowledge guidance and data-driven".

[0033] This module is based on the spatiotemporal fusion feature A output by the multi-scale spatiotemporal attention fusion module. s,t , combined with externally input railway infrastructure information (such as track, tunnel, and slope protection location and attributes), environmental factor information (such as weather and vegetation conditions), and historical foreign object event information, the following operations are performed:

[0034] Construct and dynamically evolve the railway foreign object intrusion knowledge graph G kg : Abstract the above multi-source information into entities and relationships in the knowledge graph. Knowledge Graph It can structuredly represent the complex state of railway scenarios and the context of foreign object intrusion events. This map is not static but continuously updated and evolved based on new observations and feedback.

[0035] Learning and reasoning based on graph neural networks (GNN): Apply GNN (such as graph convolutional networks (GCNs) or graph attention networks (GATs)) to the constructed knowledge graph to learn the deep evolutionary patterns of foreign object intrusion and perform intelligent risk reasoning. For example, GCN learns the representation of nodes (entities) by aggregating neighborhood information.

[0036]

[0037] in, is a neighbor of node v, deg(v) is the node degree, and W (l) is a learnable weight matrix.

[0038] GAT assigns different weights to neighbor nodes through an attention mechanism

[0039]

[0040] In this way, the system can learn the complex associations between, for example, "specific weather conditions", "historical slope instability data of a certain type", and "tiny deformation features observed by the current SAR", so as to predict the risk level P of foreign object intrusion risk and generate an alarm message. This reasoning based on the knowledge graph and GNN enables the system to not only detect known patterns, but also understand potential and emerging threats based on relationships and context.

[0041] Optionally, the system provided by the present invention may further include a human-machine collaborative alarm review and model iteration closed-loop module. This module receives the alarm message, supports manual review, and feeds back the review results (such as real foreign objects, false alarm types, new foreign object information) to the multi-scale spatio-temporal attention fusion module and the railway foreign object intrusion evolution knowledge graph construction and reasoning module. This feedback is used to fine-tune model parameters (for example, through the loss function L feedback ) to update the content of the knowledge graph (add new knowledge, adjust relationship confidence), and optimize the GNN model. This forms a closed loop of continuous learning and adaptive evolution, significantly improving the long-term monitoring performance of the system and its adaptability to new situations.

[0042] The second aspect of the present invention provides a method for real-time monitoring and alarming of railway foreign object intrusion based on SAR radar. Corresponding to the system described in the first aspect, this method realizes the intelligent perception and early warning of railway foreign object intrusion by performing the following steps:

[0043] 1. Multi-dimensional SAR feature information extraction step: Through the SAR multi-dimensional information deep perception module, multi-dimensional SAR feature information is extracted from the SAR radar data, such as the polarization feature, coherence feature, and texture feature described above.

[0044] 2. Multi-scale spatiotemporal attention fusion and feature generation steps: The extracted multi-dimensional SAR feature information is processed through the multi-scale spatiotemporal attention fusion module. First, multi-scale spatial attention processing is performed, using a multi-scale feature extraction network and combining channel attention and spatial attention to output spatial attention-enhanced features. Subsequently, time series attention processing is performed on the sequence of spatial attention-enhanced features in continuous phases, for example, using LSTM to capture dynamic evolution patterns to obtain context-aware temporal features. Finally, the spatial attention-enhanced features and context-aware temporal features are fused to generate spatiotemporal fusion features, which can indicate key spatiotemporal changes related to potential foreign object intrusion.

[0045] 3. Risk Reasoning and Alarming Based on the Evolving Knowledge Graph: The Railway Foreign Object Intrusion Evolving Knowledge Graph Construction and Reasoning Module constructs and dynamically evolves a Railway Foreign Object Intrusion Knowledge Graph based on the spatiotemporal fusion features generated in Step 2, combined with information on railway infrastructure, environmental factors, and historical foreign object incidents. Next, a graph neural network is applied to the dynamically evolving Railway Foreign Object Intrusion Knowledge Graph to learn about the evolutionary patterns of foreign object intrusion, perform risk reasoning on the current scenario, assess the risk level, and ultimately generate alarm information.

[0046] Optionally, the method may also include human-machine collaborative alarm review and model iteration steps, and the manual review results are used to optimize the model parameters of the multi-scale spatiotemporal attention fusion module and the knowledge graph content and graph neural network model in the railway foreign object intrusion evolution knowledge graph construction and reasoning module.

[0047] In summary, the present invention significantly improves the intelligence level and early warning capability of SAR radar in railway foreign object intrusion monitoring through the introduction of an innovative multi-scale spatiotemporal attention mechanism and an evolving knowledge graph. It can more effectively respond to diverse threats in complex scenarios and has the potential to continuously learn and adapt to new situations.

[0048] The present invention provides a railway foreign body intrusion real-time monitoring and alarm system based on SAR radar.

[0049] It has the following beneficial effects:

[0050] 1. This invention uses the SAR multi-dimensional information depth perception module to extract rich features of SAR data, such as multipolarization, coherence, and texture, providing more comprehensive physical attribute information for target recognition. Combined with the multi-scale spatiotemporal attention fusion module, the system can adaptively learn and focus on key spatiotemporal regions and change patterns that indicate foreign object intrusion risk, effectively suppressing interference from human or natural environmental dynamics such as seasonal changes in vegetation and normal construction, thereby significantly reducing false alarm rates and improving the robustness and accuracy of identifying true foreign object intrusion events in complex dynamic environments.

[0051] 2. The present invention innovatively introduces a knowledge graph construction and reasoning module for railway foreign object intrusion evolution, which structurally associates SAR perception features with multi-source information such as railway infrastructure, environmental factors, and historical foreign object events. By learning and reasoning on the knowledge graph through graph neural networks, the system is no longer limited to identifying objects in isolation, but can understand the context, potential associated factors, and possible evolution trends of foreign object intrusion events, thereby achieving a deep assessment of risks and more intelligent early warning decisions, such as distinguishing occasional small falling rocks from continuous small falling rocks that may indicate a larger-scale instability.

[0052] 3. The present invention endows the system with the ability of continuous learning and self-optimization through the dynamic evolution mechanism of the knowledge graph construction and reasoning module for railway foreign object intrusion evolution, and an optional human-machine collaborative alarm review and model iteration closed-loop module. New observation data, environmental change information, and feedback results of manual review can all be used to update the content of the knowledge graph, optimize the parameters of the attention model, and the reasoning logic of the graph neural network. This closed-loop adaptive learning mechanism enables the system to continuously adapt to new foreign object types, changing environmental conditions, and accumulated monitoring experience, thereby maintaining and improving its monitoring and early warning performance during long-term operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the system architecture of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0055] Please refer to the attached Figure 1 , the present invention provides a real-time monitoring and alarm system for railway foreign object intrusion based on SAR radar, which realizes efficient and accurate monitoring and early warning of railway foreign object intrusion events through the deep integration of multi-dimensional information perception, intelligent attention mechanism, and knowledge graph reasoning.

[0056] As Figure 1 shown, referring to an exemplary system architecture, the system mainly consists of a SAR multi-dimensional information deep perception module, a multi-scale spatio-temporal attention fusion module, a knowledge graph construction and reasoning module for railway foreign object intrusion evolution, and a human-machine collaborative alarm review and model iteration closed-loop module.

[0057] The SAR multi-dimensional depth perception module, serving as the system's perception front end, receives raw observation data from the SAR radar. This module performs necessary preprocessing on the raw data and extracts multi-dimensional SAR feature information that characterizes the scattering and geometric characteristics of ground objects. This information forms the basis for subsequent analysis.

[0058] The Multiscale Spatiotemporal Attention Fusion Module is connected to the output of the SAR Multidimensional Information Depth Perception Module. Its core function is to perform in-depth analysis of the input multidimensional SAR feature information. By incorporating multiscale spatial attention and time series attention mechanisms, it automatically learns and focuses on spatiotemporal variations that are highly relevant to potential foreign object intrusion events. The module's processed output is a spatiotemporal fusion feature that effectively indicates abnormal signals while suppressing background noise and regular dynamic changes.

[0059] The railway foreign object intrusion evolution knowledge graph construction and reasoning module is the core of the system's intelligent decision-making. It is connected to the output of the multi-scale spatiotemporal attention fusion module. This module not only utilizes the spatiotemporal fusion features output by the former, but also integrates external inputs such as railway infrastructure information, real-time and historical environmental factors, and historical foreign object incident records. Based on this multi-source heterogeneous information, the module constructs a dynamically evolving railway foreign object intrusion knowledge graph. Furthermore, AI technologies such as graph neural networks are used to learn and reason about this knowledge graph, uncovering the complex evolutionary patterns of foreign object intrusion, assessing the risks in the current monitoring scenario, and ultimately generating alarm information.

[0060] Optionally, a human-machine collaborative alarm review and model iteration closed-loop module is connected to the alarm output of the railway foreign object intrusion evolution knowledge graph construction and reasoning module and the manual operation interface of the monitoring center. This module presents the alarm information generated by the system to the operator for review. The operator's review results, such as confirmed real alarms, false alarms and their causes, or newly discovered foreign object types, will be fed back to the system. This feedback information is used to guide the fine-tuning and optimization of the model parameters of the multi-scale spatiotemporal attention fusion module on the one hand, and to update the knowledge content and graph neural network model in the railway foreign object intrusion evolution knowledge graph construction and reasoning module on the other hand, thereby forming a closed-loop iterative learning mechanism to continuously improve the performance of the system.

[0061] In terms of data flow, SAR radar data is first input into the SAR multi-dimensional information deep perception module; the multi-dimensional SAR feature information output after processing by this module flows to the multi-scale spatio-temporal attention fusion module; the spatio-temporal fusion features generated by the multi-scale spatio-temporal attention fusion module, together with other external information, are jointly input into the railway foreign object intrusion evolution knowledge graph construction and reasoning module; this module finally outputs alarm information to the user or the alarm review and model iteration closed-loop module of human-machine collaboration; if the latter is included, its feedback information will act on the aforementioned attention module and knowledge graph module in reverse. Through such module division and data flow design, this system can achieve comprehensive, intelligent and adaptive monitoring and alarm of railway foreign object intrusion.

[0062] In a specific embodiment of the present invention, the system can adopt an edge-cloud collaborative computing architecture. The monitoring front end (edge end) deployed along the railway is responsible for real-time data collection and preliminary processing, while the cloud server is responsible for complex model training, knowledge graph management and global decision-making. Data can be efficiently transmitted between the edge end and the cloud through wireless communication networks such as 5G.

[0063] In addition, in order to enhance the ability to confirm and analyze alarm events, this system can further integrate optical or infrared imaging devices (such as high-definition intelligent spherical cameras). When the SAR multi-dimensional information deep perception module initially identifies potential abnormal signals, the system can automatically guide the imaging device to aim at the target area for optical imaging, capture and video recording. After the SAR data and optical image data are aligned through a spatio-temporal registration algorithm, they are jointly used as multi-modal information to be input into subsequent modules for analysis, effectively breaking through the monitoring limitations of single sensing means in complex environments such as thick fog, heavy precipitation and insufficient night light.

[0064] The design goal of the system is to achieve a high-timeliness response. For example, a process of radar scanning, data processing and platform alarm can be completed within seconds (such as 5 seconds), and relevant supporting materials including optical images can be presented to the monitoring personnel within a short time (such as 20 seconds).

[0065] In this embodiment, the SAR multi-dimensional information deep perception module is used to extract rich and discriminative feature information from the original observation data obtained from the SAR radar system. These feature information form the basis for subsequent intelligent analysis and decision-making.

[0066] First, the input original SAR data I rawA series of standard preprocessing operations are performed to generate a SAR image for analysis. These preprocessing steps include, but are not limited to, radiometric calibration, which is used to convert the pixel values of the SAR image into physically meaningful backscatter coefficients; geometric correction, which is used to eliminate geometric distortions caused by terrain undulations and sensor attitudes, ensuring that the image is aligned with the geographic reference system; and speckle noise suppression, using algorithms such as Lee filtering, Frost filtering, or GammaMAP filtering, to reduce the inherent speckle noise in the SAR image while preserving the edges and detail information of the image as much as possible. The SAR image obtained after preprocessing is denoted as I raw .

[0067] After obtaining the standardized SAR image I raw the module further extracts multi-dimensional features.

[0068] If the adopted SAR radar system has multi-polarization imaging capabilities (for example, it can acquire data of four polarization channels: HH, HV, VH, VV), then polarization feature extraction is performed.

[0069] A commonly used method is to adopt polarization decomposition techniques. For example, the Cloude-Pottier decomposition can be applied. This method is based on the eigenvalue decomposition of the coherence matrix or covariance matrix, and three key parameters are obtained:

[0070] Polarization entropy H pol , which represents the randomness or disorder degree of the scattering process;

[0071] Average Alpha angle α pol , which indicates the type of the main scattering mechanism (such as surface scattering, double-bounce scattering, volume scattering);

[0072] Degree of anisotropy A pol , which describes the relative intensity of the secondary scattering mechanism.

[0073] Another method is the Freeman-Durden three-component decomposition, which decomposes the total backscatter power into the contributions of three basic scattering mechanisms: surface scattering power P s , double-bounce scattering power P d (usually from dihedral structures), and volume scattering power P v (usually from complex structures such as vegetation). These extracted multi-polarization features together constitute the feature set F pol ={H pol ,A pol ,α pol ,P s ,P d ,P v,...}, which can provide fine information about the material, structure, and orientation of ground objects, and are crucial for differentiating different types of ground objects and foreign objects.

[0074] When the system monitors in the Interferometric SAR (InSAR) mode or uses multi-temporal SAR data for change detection, the extraction of coherence features becomes particularly important. For a pair of precisely registered SAR complex images s1 and s2 acquired at different times, the coherence γ within a local window coh can be calculated by the following formula:

[0075]

[0076] where s 1,k and s 2,k respectively represent the complex values of the k-th pixel in the master and slave SAR images, is the complex conjugate of s 2,k , and N is the total number of pixels contained in the estimation window used to calculate coherence. The coherence value γ coh ranges from 0 to 1. High coherence indicates stable scattering characteristics of the ground object during the two imaging periods, while low coherence indicates changes or the presence of decoherence factors. The coherence feature is denoted as F coh .

[0077] To characterize the spatial structure and arrangement law of ground objects in SAR images, this module also extracts texture features. A widely used method is based on the Gray Level Co-occurrence Matrix (GLCM). First, for the SAR image intensity or its derived feature map (such as the intensity map of a certain polarization channel), calculate the GLCM in different directions and distances. Then, calculate a series of texture descriptions based on the GLCM, such as contrast T com (measuring the amplitude of local gray level changes), energy or angular second moment T asm (measuring the uniformity of image texture), entropy T ent (measuring the randomness or complexity of image texture), and correlation T cor (measuring the linear dependence degree of image gray levels in space), etc. These texture features together constitute the feature set F tex = {T con , T asm , T ent , T cor ,...}, which helps to distinguish regions with different surface roughnesses or internal structures.

[0078] In addition, according to the specific performance and application requirements of the SAR radar, other types of features can also be extracted. For example, if the SAR system has a high PRF (Pulse Repetition Frequency) and advanced signal processing capabilities, the micro-Doppler feature F md, which has the potential to detect and identify foreign objects with small vibrating or rotating components (such as swaying obstacles).

[0079] In summary, through the above series of processing steps, the SAR multi-dimensional information depth perception module transforms the original SAR observation data into a set of multi-dimensional SAR feature information sets containing intensity information and rich polarization features F pol , coherence features F coh , texture features F tex and other potential features F other (such as F md ).

[0080] This feature set provides high-quality and multi-perspective input data for the subsequent multi-scale spatio-temporal attention fusion module, which is an important prerequisite for achieving accurate foreign object intrusion recognition and risk assessment.

[0081] In a preferred embodiment of the present invention, in order to improve the accuracy of change detection, the SAR multi-dimensional information depth perception module further includes a dynamic background update mechanism. The system first establishes an initial reference background model of the monitoring scene by continuously collecting SAR observation data for a period of time, and this model constitutes the benchmark for subsequent intrusion detection. Subsequently, during the operation of the system, the reference background model is continuously iteratively optimized by periodically fusing new observation data and learning environmental features. This dynamic update mechanism can effectively adapt to the gradual changes of the scene caused by seasonal changes, vegetation growth, etc., reduce its interference with the performance of foreign object detection, and thus improve the long-term stability and reliability of the system.

[0082] In addition, this module can use differential interferometric synthetic aperture radar (D-InSAR) technology to extract high-precision surface deformation features. By performing differential interferometric processing on multi-temporal SAR complex images, the small deformation fields of key infrastructure along the railway (such as slopes, subgrades, and tunnel entrances) can be inverted, and the measurement accuracy can reach the millimeter level. These deformation data, as an important feature (which can be classified into F coh or as an independent deformation feature F def ), are input into the subsequent module for early identification and trend warning of slow disasters such as landslides and settlements.

[0083] In this embodiment, the multi-scale spatio-temporal attention fusion module receives the multi-dimensional SAR feature information output by the SAR multi-dimensional information depth perception module, and its core task is to perform in-depth spatio-temporal analysis on these feature information to adaptively identify and enhance the key signals related to railway foreign object intrusion events, while suppressing the interference of irrelevant backgrounds and normal dynamics.

[0084] The processing flow of this module mainly includes multi-scale spatial attention processing, time series attention processing, and finally spatio-temporal feature fusion.

[0085] First, multi-scale spatial attention processing is carried out. The input single-temporal multi-dimensional SAR feature map (where H and W are the height and width of the feature map respectively, and C in is the number of input feature channels) will pass through a multi-scale feature extraction network. This network aims to capture information at different spatial scales and receptive fields. For example, a Feature Pyramid Network (FPN) structure can be adopted, or dilated convolutional layers with different dilation rates can be used in parallel.

[0086] After obtaining the multi-scale spatial features, channel attention and spatial attention mechanisms are introduced to further refine the features.

[0087] The channel attention mechanism is used to adaptively learn the importance of different feature channels. A specific implementation method is that first, the input feature X feat is respectively subjected to global average pooling AvgPool(X feat ) and global max pooling MaxPool(X feat ) operations to obtain two 1×1×C in channel descriptors. These two descriptors are then respectively passed through a shared multi-layer perceptron (MLP). The outputs are added element-wise and passed through the Sigmoid activation function σ to generate the channel attention weight map Its calculation can be expressed as:

[0088] M c (X feat ) = σ(MLP(AvgPool(X feat )) + MLP(MaxPool(x feat )));

[0089] Among them, MLP usually contains two fully connected layers. The first fully connected layer compresses the dimension from C in to a smaller value, and the second fully connected layer restores it to C in . By multiplying M c with the original input feature X feat element-wise along the channel dimension, the channel attention weighted feature is obtained, where represents element-wise multiplication.

[0090] Next, based on the channel attention weighted feature X' feat , the spatial attention mechanism is applied to learn the importance of spatial regions. First, along the channel dimension of X' featPerform average pooling AvgPool separately s (X' feat ) and max pooling MaxPool s (X' feat ), obtaining two spatial descriptors of H×W×1. Concatenate these two descriptors along the channel dimension, and then process them through a convolutional layer Conv s (for example, using a 7x7 convolutional kernel), and pass through the Sigmoid activation function σ to generate the spatial attention weight map Its calculation can be expressed as:

[0091]

[0092] where, [·;·] represents the concatenation operation along the channel dimension. Finally, multiply the spatial attention weight map M s element-wise with its input feature X' feat to obtain the spatially attention-enhanced feature at the current time phase This feature A s highlights the key spatial regions related to potential foreign objects.

[0093] Subsequently, perform temporal attention processing. The module takes the sequence of spatially attention-enhanced features {A s,1 , A s,2 ,..., A s,T} obtained after the above multi-scale spatial attention processing for T consecutive time phases (time step t = 1, 2,..., T) as input to capture the dynamic evolution law and temporal dependence relationship of the scene.

[0094] One implementation is to use a long short-term memory network (LSTM). At each time step t, the LSTM cell updates the current cell state c s,t (usually first flatten it or reduce its dimension to a one-dimensional vector through a convolutional layer) and the previous hidden state h t-1 and cell state c t-1 and the current cell state c t and hidden state h t . The core units of LSTM include the input gate i t , the forget gate f t , the output gate o t and the candidate cell state (sometimes also denoted as g t ). Its state update equation is usually expressed as:

[0095] i t = σ(W xi A s,t + W hi h t-1+b i );

[0096] f t = σ(W xf A s,t + W hf h t-1 +b f );

[0097] o t = σ(W xo A s,t + W ho h t-1 +b o );

[0098] g t = tanh(W xc A s,t + W hc h t-1 +b c );

[0099] c t = f t ⊙ c t-1 + i t ⊙ g t ;

[0100] h t = o t ⊙ tanh(c t );

[0101] Among them, W x* , W h* are the corresponding weight matrices, b * is the bias term, σ g is usually the Sigmoid function, σ c and σ h are usually the Tanh function, and ⊙ represents element-wise multiplication. After processing the entire sequence through the LSTM, the obtained hidden state sequence (or the hidden state h T ) at the last time step is the context-aware temporal feature A t that contains the scene temporal context information.

[0102] Another way to implement temporal sequence attention processing is to adopt the Transformer Encoder structure. It calculates the mutual dependence degree between the features at different time steps in the sequence through the self-attention mechanism (Self-Attention), and its core calculation is:

[0103]

[0104] Among them, the Q, K, V matrices are obtained by performing different linear transformations on the input sequence {A s,1 , A s,2 ,..., A s,T}, and d k is the dimension of the key vector K. The Transformer Encoder, through components such as multi-head attention, feed-forward neural network, residual connection, and layer normalization, can effectively capture long-range dependencies and output the sequence feature A t enhanced by temporal attention.

[0105] Finally, spatio-temporal feature fusion is performed. This step aims to combine the spatial key information at the current moment with the dynamic evolution patterns extracted from the historical time series. A specific approach is to fuse the spatially attention-enhanced feature A s,t at the current phase t with the context-aware temporal feature A t (such as the output h t of LSTM or the output of Transformer at the corresponding position). The fusion method can be a simple concatenation operation, and then further feature learning and dimensionality reduction are performed through one or more fully connected layers or convolutional layers to obtain the final spatio-temporal fusion feature A s,t . More complex fusion strategies can also be adopted, such as attention-based fusion, to dynamically adjust the contribution weights of spatial and temporal features.

[0106] The generated spatio-temporal fusion feature A s,t synthesizes the significant regional information in space and the dynamic evolution information in time of the scene, can more accurately indicate potential foreign object intrusion events, and provides high-quality input for the subsequent knowledge graph reasoning module.

[0107] In this embodiment, the railway foreign object intrusion evolution knowledge graph construction and reasoning module is the intelligent analysis and decision-making center of the system. It receives the spatio-temporal fusion feature A s,t output by the multi-scale spatio-temporal attention fusion module, and integrates multi-source heterogeneous information. By constructing, reasoning, and evolving the knowledge graph, it realizes in-depth understanding and intelligent prediction of railway foreign object intrusion risks.

[0108] The core functions of this module include the construction of the knowledge graph, the learning and reasoning of abnormal evolution patterns based on graph neural networks, and the dynamic evolution of the knowledge graph.

[0109] First, the railway foreign object intrusion knowledge graph is constructed. This includes ontology definition, knowledge extraction, and knowledge storage.

[0110] In the ontology definition stage, it is necessary to pre-define the schema layer of the knowledge graph, and clarify the included entity types and relation types.

[0111] Entity types can include, for example: railway sections Tunnel E tunnel , Bridge E bridge , Slope protection E slope , Monitoring points SAR image snapshots SAR perception feature sets Potential foreign objects Historical foreign object events Weather conditions Geographical environment units etc.

[0112] Relation types can include, for example: located in Contains R contains , Adjacent to Has features Occurs at time Triggers Affects R affects , Has attributes etc. These definitions provide a framework for the subsequent structured organization of knowledge.

[0113] In the knowledge extraction stage, information is extracted from different sources and transformed into instances and relationships in the knowledge graph. From the spatio-temporal fusion feature A output by the multi-scale spatio-temporal attention fusion module s,t , the areas indicating anomalies and their characteristics can be extracted as or instances of entities and their attributes. For example, the significantly changed areas in A s,t can be instantiated as a potential foreign object entity, whose attributes include location, size, and the feature vector encoded by A s,t . Geographic Information System (GIS) data is used to instantiate railway infrastructure entities (such as E tunnel ) and their spatial topological relationships (such as ). The records in the historical foreign object intrusion event database are extracted as entities, and information such as the time, location, type, cause, and consequence of their occurrence is associated. Real-time or historical meteorological data is used to instantiate entities.

[0114] In the knowledge storage stage, all the extracted knowledge is in the form of triples (e h , r, e t) are organized and stored in the form of, where e h is the head entity, e t is the tail entity, and r is the relationship between the two. For example, (Potential Foreign Object A, Railway Section X), (Railway Section X, Steep Slope). These triples together constitute the railway foreign object intrusion knowledge graph where is the set of all entity nodes, and ε kg is the set of all relationship edges. This knowledge graph can be stored in a dedicated graph database, such as Neo4j or Apache Jena.

[0115] Next, a graph neural network (GNN) is used to learn the constructed knowledge graph to mine the complex associations and evolution patterns of foreign object intrusion events and perform real-time risk reasoning.

[0116] First is node representation learning. GNN models, such as the aforementioned graph convolutional network (GCN) or graph attention network (GAT), are used to learn a low-dimensional, dense vector representation (also known as an embedding) for each entity node in the knowledge graph.

[0117] For GCN, at the l-th layer, the feature vector of node v is updated to and is calculated as:

[0118]

[0119] where, is the set of first-order neighbor nodes of node v in graph G kg , deg(v) is the degree of node v (the number of connected edges), W (l) is the trainable weight matrix at the l-th layer, is the feature vector of neighbor node u at the l-th layer, and σ act is a non-linear activation function, such as ReLU. By stacking multiple layers of GCN, nodes can aggregate information from their multi-order neighborhoods.

[0120] For GAT, the aggregation weights of node features are not fixed but are dynamically learned through an attention mechanism. The feature vector of node v is updated as:

[0121]

[0122] where the attention coefficient represents the importance of the information of node u to node v and is calculated as:

[0123]

[0124] Here, || represents the vector concatenation operation, is a trainable attention parameter vector, and LeakyReLU is a non-linear activation function. GAT can assign different importance to different neighbors, thus capturing complex relationships more flexibly.

[0125] The node embedding H learned through GNN v encodes the semantic information of each entity in its graph context. Based on these embeddings, various inference tasks can be performed. For example, link prediction can be carried out, that is, predicting whether there is a certain specific relationship between two entities in the knowledge graph, or whether a newly emerging SAR perception feature entity is associated with a high-risk pattern. Node classification or regression can also be performed. For example, for a certain monitoring area or an identified entity, predicting the risk probability P risk of foreign object intrusion or the risk level.

[0126] To capture the evolutionary characteristics of foreign object intrusion, time information can be incorporated into the GNN model to form a temporal graph network (Temporal GNN). For example, when aggregating neighborhood information in each layer of the GNN, not only the current state of neighbor nodes but also their historical state sequences are considered. Or, a sequence model (such as LSTM or Transformer) can be applied to the node representations output by the GNN to learn the typical precursor feature sequences before the occurrence of specific types of foreign object events and the evolutionary paths of related environmental factors.

[0127] Finally, this module is also responsible for the dynamic evolution of the knowledge graph. With the continuous input of new SAR observation data, environmental information, and manual review feedback, the knowledge graph needs to be updated accordingly. For example, when a new foreign object intrusion event is confirmed, the relevant entities and relationships will be added to G kg ; when there are significant changes in the environmental factors (such as continuous rainfall) in a certain area, the attributes of the corresponding entities will be updated. This dynamic evolution enables the knowledge graph to reflect the latest situation along the railway in real time. At the same time, the GNN model also needs to be retrained or incrementally learned regularly or after significant changes in the knowledge graph to adapt to the new data distribution and knowledge patterns, ensuring the accuracy and timeliness of inference.

[0128] Through the above construction, inference, and evolution mechanisms, the railway foreign object intrusion evolution knowledge graph construction and inference module can extract deep associated knowledge from multi-source data and perform intelligent risk assessment and early warning based on this knowledge, significantly improving the intelligence level of the monitoring system.

[0129] In a specific application of the present invention, in order to effectively distinguish normal train passing events from abnormal foreign object intrusion events and prevent false alarms triggered by regular operation activities, dedicated recognition rules can be constructed in the knowledge graph or a GNN model can be trained to learn such patterns. A specific implementation method is to utilize the correlation features of the time-series SAR signals. For example, for the representation of railway track entities in the knowledge graph, a rule can be defined: when it is detected that the time correlation coefficient of the radar signals in its corresponding area significantly drops from a stable high value (such as a first threshold greater than 0.99) to a lower value (such as a second threshold lower than 0.85) within a short period of time and then returns to the high value, this event is inferred as a "train passing" event. Such events will be recorded by the system but do not trigger an alarm output.

[0130] Similarly, for other known non-intrusive targets or regular operation activities, corresponding knowledge rules can be constructed or models can be trained for recognition, so as to achieve precise detection, positioning, and classification recognition of foreign object intrusion. For example, the system can identify rolling stones, construction objects larger than a specific specification (such as 15 cm x 15 cm x 15 cm), engineering vehicles, personnel entering the monitoring perimeter, etc., and distinguish them from regular events. These specific knowledge and rules greatly enrich the connotation of the knowledge graph and improve the intelligent level of system reasoning.

[0131] In this embodiment, the human-machine collaborative alarm review and model iteration closed-loop module is a key link to ensure the long-term stable operation of the system, continuously improve performance, and adapt to new situations. This module aims to effectively combine the experience knowledge of human operators with the analysis capabilities of artificial intelligence systems to form a dynamic learning and optimization loop.

[0132] This module receives the alarm information generated by the railway foreign object intrusion evolution knowledge graph construction and reasoning module. These alarm information usually includes the alarm location, the speculated foreign object type (if the knowledge graph can infer), the risk level or confidence level P risk , as well as the summary of relevant SAR sensing features and timestamps. The alarm information is presented to professional operators in the monitoring center through the user interface.

[0133] The operator reviews the alarms issued by the system based on their professional knowledge, experience, and available auxiliary information (such as on-site cameras, inspection records, etc.). The main results of the review include:

[0134] 1. Confirmed as a real foreign object intrusion: The operator confirms that the alarm is true and can further mark detailed information such as the accurate type, actual range, and harm degree of the foreign object.

[0135] 2. Confirm as false alarm: The operator determines that the alarm is a false alarm and records the reason for the false alarm as much as possible, such as normal railway maintenance operations, non-intrusive surface changes caused by extreme weather, or other benign interference sources.

[0136] 3. Detect system missed alarms: Although this module mainly processes the alarms issued by the system, in actual applications, if the operator discovers foreign object intrusion events that the system fails to detect through other means (such as manual inspections), this information should also be recorded and incorporated into the feedback system.

[0137] 4. Identify new or unknown situations: The operator may encounter new types of foreign objects, new disaster-causing patterns, or complex environmental interactions that are not covered by the system model or knowledge graph.

[0138] This information that has been manually reviewed and annotated will be used as feedback data to drive the iterative optimization of the artificial intelligence model and the evolution of the knowledge graph in the system.

[0139] For the optimization of the multi-scale spatio-temporal attention fusion module, the TP and FP samples (i.e., SAR feature data with correct labels) annotated by the operator can be used to fine-tune the model parameters of this module. For example, these newly annotated samples can be added to the original training dataset or form a dedicated feedback dataset. In the model fine-tuning stage, a feedback-based loss function L feedback . If the output layer of this module is designed to classify the presence of foreign objects, then L feedback can be the standard cross-entropy loss function or the Focal Loss function to penalize the prediction errors of the model on these manually verified samples. Through the backpropagation algorithm, the weights inside the attention network are adjusted to enable it to better learn the features for distinguishing real foreign object signals from false alarm signals.

[0140] For the optimization of the railway foreign object intrusion evolution knowledge graph construction and reasoning module, the feedback information drives its evolution at two levels:

[0141] 1. Update and enrich the content of the knowledge graph:

[0142] For the confirmed TP events, their detailed information (such as the exact foreign object type, occurrence time, geographical location, associated environmental factors, and the corresponding SAR sensing feature A s,t ) will be structured into new entities and relationships and added to the knowledge graph G kg . For example, if a "small landslide" triggered by "heavy rainfall" is confirmed, new event entities will be created, and relationships will be established between it and related weather entities, geographical location entities, and SAR feature entities, such as (small landslide A, Heavy rainfall X), (small landslide A, SAR feature set Y).

[0143] For FP events, analyzing their causes helps to correct inaccurate knowledge in the knowledge graph. For example, if a certain type of SAR feature A s,t is frequently misreported as a foreign object, but is confirmed as a normal scene change manually, then the rules related to this feature can be adjusted or the confidence or weight of certain paths pointing to "high risk" in graph reasoning can be reduced.

[0144] For new types of foreign objects or unknown situations, the annotations of operators may prompt the expansion of the knowledge graph ontology, that is, define new entity types or relationship types for subsequent accumulation and learning of such new knowledge.

[0145] 2. Retraining or incremental learning of the graph neural network (GNN) model:

[0146] The update of the knowledge graph content (especially the addition of a large number of high-quality, manually confirmed event nodes and relationships) provides richer and more accurate training data for the GNN model. After a certain amount of feedback is accumulated or regularly, the updated knowledge graph G' kg can be used to retrain the GNN model (such as GCN or GAT).

[0147] Manually confirmed risk level P risk or event classification labels can be used as supervision signals for GNN model training. For example, in the node classification task, the manually confirmed foreign object event nodes can be used as positive samples, and the confirmed non-foreign object areas can be used as negative samples to optimize the parameters of the GNN.

[0148] For the incremental learning of the model, some techniques can be adopted to absorb new knowledge and data without completely retraining the entire model to adapt to the dynamic changes of the knowledge graph.

[0149] Through the above-mentioned human-machine collaborative alarm review and model iteration closed-loop mechanism, the system can continuously integrate the domain expert knowledge of operators into the process of model learning and knowledge accumulation. This not only helps to correct the current errors of the system and improve the accuracy of alarms, but more importantly, it enables the entire monitoring and warning system to have the ability of adaptive evolution, be able to continuously learn new foreign object intrusion patterns, adapt to environmental changes, and thus maintain a high level of monitoring performance and robustness in long-term operation. This closed-loop feedback is the key to achieving true intelligent monitoring.

[0150] In this embodiment, a specific application scenario will be used to describe the complete working process of the railway foreign object intrusion real-time monitoring and alarm system provided by the present invention to illustrate how each module works together.

[0151] The workflow starts with the SAR radar system collecting radar data for a target railway section regularly or in key areas according to a predetermined plan or specific trigger conditions (such as severe weather warnings).

[0152] The acquired original SAR data stream I raw is first sent to the SAR multi-dimensional information depth perception module. This module performs preprocessing on I raw including radiometric calibration, geometric correction, and speckle noise suppression, to obtain the corrected SAR image I sar .

[0153] Subsequently, the module extracts multi-dimensional features from I sar For example, if multi-polarization SAR data is used, polarization features F pol are extracted; if multi-temporal data is used, coherence features γ coh are calculated; and texture features F tex are also extracted. These features together constitute the multi-dimensional SAR feature information Fsar of the current observation phase feat .

[0154] Next, the multi-dimensional SAR feature information Fsar feat is passed to the multi-scale spatio-temporal attention fusion module. This module first performs multi-scale spatial attention processing on the input Fsar feat such as by calculating the channel attention map M c [[ID=3३]]and the spatial attention map M s to obtain the spatially attention-enhanced feature A s of the current phase. This A s highlights the areas in the SAR image where significant changes may occur.

[0155] Then, the module inputs the sequence of spatially attention-enhanced features {A s,t-T+1 ,...,A s,t} of the current and the previous T - 1 phases into a time series attention processing unit, such as an LSTM network, to capture the temporal dynamic evolution pattern of the scene and generate the context-aware temporal feature A t . Finally, the module fuses the spatially attention-enhanced feature A s,t of the current phase with its corresponding context-aware temporal feature A t,t to output the spatio-temporal fusion feature A st . This A st contains a deep representation of the spatio-temporal changes related to potential foreign object intrusion in the railway scene.

[0156] Subsequently, the spatio-temporal fusion feature A stis sent to the railway foreign object intrusion evolution knowledge graph construction and reasoning module. This module first instantiates the potential change regions and their features indicated in A st as entities or attributes in the knowledge graph G kg . Meanwhile, the module obtains railway infrastructure information, current and recent weather conditions, and historical foreign object event information in this area from an external database. This information, together with the SAR perception features from A st , updates or expands the knowledge graph G kg .

[0157] Then, the graph neural network model performs reasoning on the updated knowledge graph G kg . For example, the GNN identifies that the change in the SAR features in the slope protection st area indicated by A is similar to the SAR feature pattern before historical small rockfall events, and there is a current strong rainfall as an inducing factor, and the slope protection itself has a slippery property. Combining these associated information, the GNN model outputs a relatively high probability P risk of foreign object intrusion risk or an alarm message with a high risk level, indicating that a rockfall may occur in the slope protection area.

[0158] The alarm message (including location, risk level P risk , relevant evidence chains such as associated SAR features, weather conditions, historical events, etc.) is sent to the human-machine collaborative alarm review and model iteration closed-loop module. The operator at the railway safety monitoring center receives this alarm. The operator retrieves the real-time monitoring video (if available) of this area, consults the daily inspection records, or deploys a drone for rapid investigation. Assuming that the operator confirms it as a small rockfall intrusion into the restricted area, this alarm is marked as a true alarm (TP), and detailed information such as the actual range, volume, and impact on the line of the rockfall is entered into the system. These feedback information will be used for:

[0159] 1. Update the knowledge graph G kg : Add this confirmed event as a new instance, and establish associations between it and entities such as relevant SAR features, weather conditions, geographical locations, etc.

[0160] 2. Optimize the model: Add the SAR features A st of this event and its label (real foreign object) to the training data of the multi-scale spatio-temporal attention fusion module, and perform fine-tuning through the loss function L feadback to improve its recognition ability for similar events. Meanwhile, the updated knowledge graph G kg will also be used for the next round of training or incremental learning of the GNN model to improve the accuracy of its reasoning.

[0161] If the operator determines it to be a false positive (FP), for example, the SAR feature change is caused by a planned slope reinforcement construction, then the operator will label it as a false positive and record the reason. This information will also be fed back:

[0162] 1. Update the knowledge graph G kg : It may add knowledge of scenario patterns related to "normal construction", or adjust the weights of certain features or rules that led to this false alarm.

[0163] 2. Optimize the model: Use the SAR feature A of this false alarm st and its label (non - foreign object) for negative sample training of the attention model to help the model better distinguish real threats from benign interference.

[0164] Through the above work process, the system proposed by the present invention can not only monitor and alarm potential railway foreign object intrusion events in real time, but also continuously learn and evolve through the feedback closed - loop of human - machine collaboration, constantly improving the accuracy and intelligence level of monitoring, so as to provide strong support for ensuring railway operation safety.

[0165] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A railway foreign object intrusion real-time monitoring and alarm system based on SAR radar, characterized by: include: SAR multi-dimensional information depth perception module, used to extract multi-dimensional SAR feature information from SAR radar data; A multi-scale spatiotemporal attention fusion module is connected to the SAR multi-dimensional information depth perception module, and is used to perform multi-scale spatial attention processing and time series attention processing on the multi-dimensional SAR feature information, and fuse them to generate spatiotemporal fusion features to identify spatiotemporal changes related to potential foreign object intrusion; The railway foreign object intrusion evolution knowledge graph construction and reasoning module is connected to the multi-scale spatiotemporal attention fusion module, and is used to construct and dynamically evolve the railway foreign object intrusion knowledge graph based on the spatiotemporal fusion features, railway infrastructure information, environmental factor information and historical foreign object event information, and learn the knowledge graph through the graph neural network to explore the evolution pattern of foreign object intrusion and perform risk reasoning to generate alarm information.

2. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 1 is characterized in that: The multi-dimensional SAR feature information extracted by the SAR multi-dimensional information depth perception module includes at least one feature selected from the following group: multi-polarization feature, coherence feature, and texture feature.

3. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 1 is characterized in that: When the multi-scale spatiotemporal attention fusion module performs multi-scale spatial attention processing, a multi-scale feature extraction network is used to extract spatial features under different receptive fields, and channel attention and spatial attention mechanisms are combined to automatically learn and focus on features related to key sensitive areas of the railway, and output spatial attention enhanced features.

4. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 3 is characterized in that: When the multi-scale spatiotemporal attention fusion module performs time series attention processing, it uses a long short-term memory network or a Transformer Encoder to capture the temporal dependency and the dynamic evolution pattern of the scene for the feature sequence composed of the spatial attention enhancement features of multiple consecutive time phases, and outputs context-aware time features.

5. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 4 is characterized in that: When the multi-scale spatiotemporal attention fusion module fuses and generates spatiotemporal fusion features, the spatial attention enhancement features of the current phase are fused with the context-aware time features of the corresponding time step.

6. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 1 is characterized in that: When the railway foreign object intrusion evolution knowledge graph construction and reasoning module constructs the railway foreign object intrusion knowledge graph, the railway infrastructure, environmental factors, historical foreign object events and SAR perception features represented by the spatiotemporal fusion features are abstracted as entities in the knowledge graph, and the relationship between the entities is defined based on their internal connections.

7. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 6 is characterized in that: When the railway foreign object intrusion evolution knowledge graph construction and reasoning module learns the knowledge graph through a graph neural network, it includes using the graph neural network to learn the low-dimensional vector representation of entity nodes, and based on the learned representation, it mines the typical evolution paths and association rules of foreign object intrusion events, and performs risk level assessment on potential foreign object intrusion events.

8. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 7 is characterized in that: When the railway foreign object intrusion limit evolution knowledge graph construction and reasoning module dynamically evolves the railway foreign object intrusion limit knowledge graph and graph neural network model, it updates the entities, relationships or attributes in the knowledge graph based on newly acquired SAR observation data, environmental information and manual review feedback results, and iteratively optimizes the graph neural network model.

9. The railway foreign object intrusion real-time monitoring and alarm system based on SAR radar according to claim 1 is characterized in that: It also includes a human-machine collaborative alarm review and model iteration closed-loop module, which is used to receive the alarm information generated by the railway foreign object intrusion evolution knowledge graph construction and reasoning module, support manual review of the alarm information, and feed back the review results to the multi-scale spatiotemporal attention fusion module to fine-tune its model parameters, and at the same time feed back to the railway foreign object intrusion evolution knowledge graph construction and reasoning module to update its knowledge content and optimize the graph neural network model.

10. A method for real-time monitoring and alarming of foreign object intrusion on railways based on SAR radar, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: Through the SAR multi-dimensional information depth perception module, multi-dimensional SAR feature information is extracted from SAR radar data; Through the multi-scale spatiotemporal attention fusion module, the multi-dimensional SAR feature information is subjected to multi-scale spatial attention processing and time series attention processing, and the spatiotemporal fusion features are generated by fusion to identify the spatiotemporal changes related to potential foreign object intrusion; Through the railway foreign object intrusion evolution knowledge graph construction and reasoning module, based on the spatiotemporal fusion characteristics, railway infrastructure information, environmental factor information and historical foreign object event information, the railway foreign object intrusion knowledge graph is constructed and dynamically evolved, and the knowledge graph is learned through the graph neural network to explore the evolution pattern of foreign object intrusion and perform risk reasoning to generate alarm information.

Images