An AI-based high-speed train image capture and processing system and method

Abstract

The present invention relates to the field of image processing technology, and discloses an AI-based high-speed train image capture and processing system and method. An AI-based high-speed train image capture and processing method includes: processing continuously acquired high-speed train images based on a sliding window algorithm; constructing a dynamic association graph network model and applying a temporal graph attention network for target spatial relationship processing; using a multi-hop anomaly association analysis algorithm to predict the propagation of anomalies between targets; implementing an adaptive resource allocation system to dynamically adjust image acquisition parameters according to a scene importance evaluation model; executing an intelligent cache optimization algorithm to achieve value-based data management and optimized allocation of storage resources; enhancing the anomaly detection ability, and improving the anomaly warning accuracy through dynamic association graph network modeling and multi-hop anomaly association analysis.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically, to an AI-based high-speed train image capture and processing system and method. Background Art

[0002] High-speed train monitoring systems generate massive amounts of image data daily, containing information about equipment, track infrastructure, and environmental conditions along the line. Traditional image capture methods use fixed-interval sampling, which is unable to intelligently identify key scenes, resulting in wasted storage space and missing important information. Furthermore, the high-speed train operating environment is complex, with potential correlations between various monitoring targets. An anomaly in one device can trigger a chain reaction affecting other connected devices, and traditional methods of processing images independently cannot effectively capture these correlations.

[0003] Current image monitoring technology suffers from the following major issues: excessive storage space occupied by redundant images; missed detection of critical scenes; inappropriate resource allocation; difficulty identifying event correlations; and omission of key diagnostic information. These issues lead to inefficient monitoring systems, making them ineffective in supporting the safe operation and maintenance of high-speed trains. Summary of the Invention

[0004] The present invention provides an AI-based high-speed train image capture and processing system and method to solve the technical problems of redundant data storage and omission of key information in traditional image monitoring technology in related technologies.

[0005] The present invention provides an AI-based high-speed train image capture and processing method, comprising:

[0006] Continuously acquired high-speed train images are processed based on a sliding window algorithm to identify key changes in the scene;

[0007] By utilizing key changes in the scene, a dynamic association graph network model is constructed and a temporal graph attention network is applied to process target spatial relationships and generate a target association graph.

[0008] Based on the target association graph, a multi-hop anomaly association analysis algorithm is used to predict the propagation of anomalies between targets, and an anomaly propagation path probability matrix is obtained. The multi-hop anomaly association analysis algorithm is used to predict the propagation of anomalies between targets, including:

[0009] Calculate the adjacency matrix and degree matrix corresponding to the target association graph;

[0010] Construct a normalized adjacency matrix;

[0011] Initialize the node’s abnormal attention vector;

[0012] Use the attention propagation algorithm to iteratively calculate the k-hop anomaly association;

[0013] Build an exception propagation model to predict the diffusion path of exceptions on the target graph;

[0014] Based on the key changes in the scenario and the exception propagation path probability matrix, implement an adaptive resource allocation system, and dynamically adjust the image acquisition parameters according to the scenario importance evaluation model;

[0015] According to the exception propagation path probability matrix and the results of the adaptive resource allocation system, execute an intelligent cache optimization algorithm to achieve value-based data management and optimized allocation of storage resources.

[0016] Furthermore, the processing of continuously acquired high-speed train images based on the sliding window algorithm includes:

[0017] Use the sliding window algorithm to maintain the sequence of the most recent N frames of images;

[0018] Calculate the information entropy difference between adjacent image frames;

[0019] Calculate the structural similarity between adjacent image frames;

[0020] Combine the information entropy difference and the structural similarity to construct a scene change index.

[0021] Furthermore, the construction of the dynamic association graph network model includes:

[0022] Use the object detection algorithm to identify the set of key objects from the image;

[0023] Construct a scene dynamic association graph;

[0024] Calculate the association weights between objects;

[0025] Apply the temporal graph attention network to process the dynamic associations between objects.

[0026] Furthermore, the adaptive resource allocation system includes:

[0027] Based on the scene change index, construct a scene importance evaluation model;

[0028] Dynamically adjust the time resolution and adaptively allocate the sampling frequency according to the scene importance;

[0029] Construct a hierarchical cache structure;

[0030] When an important event is detected, trigger the backtracking analysis algorithm to temporarily increase the retention density of historical frames.

[0031] Furthermore, the intelligent cache optimization algorithm includes:

[0032] Optimize the storage resource allocation strategy;

[0033] Construct an intelligent elimination algorithm based on information value to evaluate the retention value of each frame of image;

[0034] Combine storage resource constraints to implement an optimized elimination algorithm;

[0035] Implement an incremental learning model by continuously monitoring system performance metrics;

[0036] Construct a cache access index system.

[0037] Furthermore, the calculation formula for constructing the scene change index is:

[0038] ;

[0039] where represents the scene change index at time , and respectively represent the weight coefficients of information entropy difference and structural similarity in the scene change index, represents the information entropy difference between adjacent image frames, represents the structural similarity index, , respectively represent the images at time , time .

[0040] Furthermore, the calculation formula for applying the temporal graph attention network to process the dynamic association between targets is:

[0041] ;

[0042] where is the scene dynamic association graph, is the hidden state at the current time, containing the feature representations of all nodes; is the observed feature at the current time, that is, the target feature extracted from the image; represents the temporal graph attention network, which combines the graph neural network and the attention algorithm and can capture the association information between targets and maintain temporal coherence; is the updated hidden state, containing temporal and spatial relationship information.

[0043] Furthermore, the calculation formula for constructing the anomaly propagation model to predict the diffusion path of anomalies on the target graph is:

[0044] ;

[0045] where represents that under the condition that an anomaly occurs at node , node Conditional probability of anomaly occurrence; Represents the source node to the destination node Anomaly propagation attention value; and respectively represent the state features of the source node and the destination node ; is a conditional probability calculation function.

[0046] The present invention provides an AI-based high-speed train image capture and processing system for performing the above-mentioned AI-based high-speed train image capture and processing method, including:

[0047] A multi-dimensional information entropy analysis module for processing high-speed train images based on a sliding window algorithm to identify key scene changes;

[0048] A dynamic association graph construction module for constructing a spatial relationship network between targets based on key change frames to generate a target association graph containing temporal information;

[0049] An anomaly propagation prediction module for calculating the anomaly diffusion path between targets through multi-hop anomaly association analysis to form a risk probability matrix;

[0050] A resource adaptive allocation module for dynamically adjusting monitoring parameters according to scene importance and anomaly risk to improve the data quality of key areas;

[0051] An intelligent cache management module for optimizing the storage strategy based on anomaly propagation risk and information value evaluation to achieve value-oriented management of data.

[0052] The beneficial effects of the present invention are as follows: Through multi-dimensional information entropy analysis, dynamic association graph network, multi-hop anomaly association analysis, adaptive resource allocation, and intelligent cache optimization, efficient processing of high-speed train image capture is achieved;

[0053] The storage efficiency and the capture rate of key scenes are improved, and the data storage requirements are reduced; The anomaly detection ability is enhanced, and through dynamic association graph network modeling and multi-hop anomaly association analysis, the anomaly early warning accuracy is improved; The resource utilization efficiency is optimized, the computing resource requirements are reduced, and the system response speed is increased; The integrity of diagnostic information is improved, and the accuracy of fault root cause analysis is increased; The system scalability is enhanced, and a modular design is adopted to support continuous optimization and expansion of system functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flowchart of an AI-based high-speed train image capture and processing method of the present invention;

[0055] Figure 2It is the flowchart of Step 1 in the present invention;

[0056] Figure 3 It is the flowchart of Step 2 in the present invention;

[0057] Figure 4 It is the flowchart of Step 3 in the present invention;

[0058] Figure 5 It is the flowchart of Step 4 in the present invention;

[0059] Figure 6 It is the flowchart of Step 5 in the present invention. Detailed implementation manners

[0060] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the protection scope of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0061] In at least one embodiment of the present invention, a method for processing high - speed train image capture based on AI is disclosed. As Figures 1 to 6 shown, it includes the following steps:

[0062] Step 1: Process the continuously acquired high - speed train images based on the sliding window algorithm to identify key changes in the scene;

[0063] In this step, the most recent N frames of images in the sequence are maintained by constructing a sliding window, and the information entropy difference and structural similarity between adjacent frames are calculated, and then a scene change index is generated for identifying key scenes. The specific implementation is as follows:

[0064] Step 1.1: Use the sliding window algorithm to maintain the sequence of the most recent N frames of images;

[0065] The sequence of the most recent N frames of images is expressed as:

[0066] ;

[0067] where , , represent the image at the current moment t, the earliest image in the window, and the second - earliest image in the window, respectively, and N is the window size;

[0068] Step 1.2: Calculate the information entropy difference between adjacent image frames;

[0069] The calculation formula for the information entropy difference is:

[0070] ;

[0071] where represents the information entropy difference between adjacent image frames, represents the moment of the image, and respectively represent the information entropy of images and ;

[0072] The calculation formula is:

[0073] ;

[0074] where represents the probability that the pixel with the gray value of appears in the image; represents the logarithm with base 2; represents the summation over all gray values from 0 to 255;

[0075] Step 1.3, calculate the structural similarity between adjacent image frames;

[0076] The similarity is expressed as:

[0077] ;

[0078] where represents the structural similarity index, and respectively represent the images at moments and moment of the image, , respectively represent the means of images , ; , respectively represent the standard deviations of images , ;[[ID=8l]] represents the covariance between images and ; and [[ID=9l]]

[0079] Step 1.4, combine the information entropy difference and the structural similarity to construct a scene change index;

[0080] The calculation formula of the scene change index is:

[0081] ​ ;

[0082] wherein represents the scene change index at time , represents the information entropy difference between adjacent image frames, represents the structural similarity index, and respectively represent the images at time and time , and respectively represent the weight coefficients of the information entropy difference and the structural similarity in the scene change index;

[0083] The output result of this step is the scene change index between each pair of adjacent image frames . The larger the index value, the greater the scene change, which can be used for subsequent key frame judgment. In addition, a series of key change frames are identified through threshold judgment:

[0084] ;

[0085] wherein , , respectively represent the 1st, 2nd, th key change frames, is the number of key change frames;

[0086] These key change frames will be used as the core input of the dynamic association graph network model in step 2. At the same time, the scene change index will be used as one of the basic inputs for constructing the scene importance evaluation model in step 4. This intelligent screening mechanism based on information entropy is the basis for the entire system to achieve efficient storage and accurate analysis, laying a data foundation for all subsequent steps.

[0087] Step 2: Utilize the key changes in the scene to construct a dynamic association graph network model and apply a temporal graph attention network to process the target spatial relationship to generate a target association graph;

[0088] This step takes the key change frames identified in step 1 as inputs to construct a dynamic association graph network model to represent the spatial relationship between targets and conduct relevance analysis. The key change frames output by step 1 provide the key time segments that the system focuses on, greatly reducing the amount of data processed and ensuring that no important information is missed. The specific implementation is as follows:

[0089] Step 2.1: Use an object detection algorithm to identify a set of key targets from the image;

[0090] The set of key targets is represented as:

[0091] ;

[0092] wherein 、 、 respectively represent the 1st, 2nd, and th detected targets, represents the number of detected targets;

[0093] Step 2.2, construct a scene dynamic association graph;

[0094] The scene dynamic association graph is represented as:

[0095] ;

[0096] wherein is the scene dynamic association graph:

[0097] ;

[0098] wherein is the node set, 、 、 respectively represent the 1st, 2nd, and th target nodes, represents the number of detected targets;

[0099] ;

[0100] wherein is the edge set, representing the potential association relationship between targets; represents the edge between node and node , and are node indices;

[0101] ;

[0102] wherein is the weight set, represents the strength of the association between node and node ;

[0103] Step 2.3, calculate the association weight between targets;

[0104] The weight calculation considers the following factors:

[0105] Spatial distance:

[0106] ;

[0107] Among them represents the spatial distance between nodes and nodes ; and respectively represent the position coordinates of nodes and nodes ; represents the Euclidean distance between two position coordinates;

[0108] Functional correlation:

[0109] represents the degree of functional association between nodes and nodes ;

[0110] Historical interaction:

[0111] represents the historical interaction frequency between nodes and nodes ;

[0112] The calculation formula for the association weight is:

[0113] ;

[0114] Among them , , respectively represent the weight coefficients of spatial distance, functional correlation, and historical interaction; is the influence range parameter of the spatial distance, controlling the rate of distance attenuation; is the exponential function;

[0115] Step 2.4, Apply the temporal graph attention network to process the dynamic association between targets;

[0116] The calculation formula is:

[0117] ;

[0118] Among them is the scene dynamic association graph, is the hidden state at the current moment, containing the feature representations of all nodes; is the observed feature at the current moment, that is, the target feature extracted from the image; represents the temporal graph attention network, which integrates the graph neural network and the attention algorithm, and can capture the association information between targets and maintain temporal coherence; is the updated hidden state, containing temporal and spatial relationship information.

[0119] The output of this step is the target association graph and its updated node representation , these outputs are directly used as the basic data structure and computing resources for multi-hop anomaly association analysis in step 3. This association graph not only represents the spatial association relationship between targets, but also captures the dynamic changes of association in the time dimension through the time series graph attention network, providing a complete relationship network model for subsequent anomaly association analysis. In addition, the node representation It contains rich target feature information, which will be further used in steps 3 and 4 to form a progressive relationship of information flow in the system, thereby ensuring the coherence of the entire processing flow and the consistency of information transmission.

[0120] Step 3: Based on the target association graph, use the multi-hop anomaly association analysis algorithm to predict the propagation of anomalies between targets and obtain the anomaly propagation path probability matrix;

[0121] This step makes full use of the target association graph output in step 2 and node representation On this basis, multi-hop anomaly association analysis is performed to predict the anomaly diffusion path on the target graph. The association graph constructed in step 2 provides a complete target relationship network for anomaly propagation analysis, enabling this step to accurately simulate the anomaly propagation process. The specific implementation is as follows:

[0122] Step 3.1, calculate the adjacency matrix and degree matrix corresponding to the target association graph;

[0123] Based on the target association graph constructed in step 2 , calculate its corresponding adjacency matrix Sum degree matrix :

[0124] Adjacency Matrix:

[0125] ;

[0126] in is the adjacency matrix; express OK A real matrix of columns; is the field of real numbers;

[0127] ;

[0128] in Representation node and nodes The edge weight between nodes and nodes If there is no connection between ;

[0129] Degree matrix:

[0130] ;

[0131] Where is the degree matrix;

[0132] ;

[0133] Where represents the degree of node , that is, the sum of the weights of all edges connected to node ; represents the sum of from 1 to ; represents the edge weight between node and node ;

[0134] Step 3.2, construct the normalized adjacency matrix;

[0135] The expression of the normalized adjacency matrix is:

[0136] ;

[0137] Where is the normalized adjacency matrix, is the adjacency matrix; is the identity matrix, with diagonal elements being 1 and the rest being 0;

[0138] Construct the normalized degree matrix:

[0139] ;

[0140] Where is the normalized degree matrix, is the degree matrix;

[0141] Step 3.3, initialize the abnormal attention vector of the nodes:

[0142] The initial abnormal attention vector is expressed as:

[0143] ;

[0144] Where is the initial abnormal attention vector; represents dimensional real vector space;

[0145] Step 3.4, use the attention propagation algorithm to iteratively calculate the k-hop abnormal correlation;

[0146] The calculation formula is:

[0147] ;

[0148] in is an activation function (such as ReLU), which is used to introduce nonlinear transformation; Represents the normalized degree matrix The diagonal elements are raised to the power of negative half. For the The weight matrix of the layer is used to learn the influence transfer pattern between different nodes; Indicates the Abnormal attention vector after iterations; Indicates the Abnormal attention vector after iterations;

[0149] Step 3.5: Build an anomaly propagation model to predict the diffusion path of anomalies on the target graph;

[0150] The calculation formula is:

[0151] ;

[0152] in Indicates that the node Under abnormal conditions, the node The conditional probability of an anomaly occurring; Represents a slave node To Node Abnormal propagation attention value of ; and Represents nodes respectively and nodes Status characteristics; is the conditional probability calculation function, which can be in the following form:

[0153] ;

[0154] in is the state compatibility function, used to measure the node Status and nodes The compatibility of the state in the exception propagation; is an exponential function; Indicates that All except Perform the summation.

[0155] The output of this step is the abnormal propagation path probability matrix, where Indicates an abnormal slave node Propagate to nodes Probability. This probability matrix is the core reference data for resource allocation and cache optimization in subsequent steps. It directly affects the scenario importance assessment in Step 4, as scenarios predicted to be high-risk areas for abnormal propagation will be assigned higher importance weights; at the same time, it is also the key input for intelligent cache optimization in Step 5, providing a risk-based priority division basis for storage resource allocation. In this way, the system can preferentially allocate limited computing and storage resources to areas where anomalies are most likely to occur, thereby achieving efficient capture of critical anomalies.

[0156] Step 4. Based on the key changes in the scenario and the probability matrix of abnormal propagation paths, implement an adaptive resource allocation system to dynamically adjust the image acquisition parameters according to the scenario importance assessment model;

[0157] This step is based on the output results of the previous steps, especially the scenario change metrics in Step 1 and the probability matrix of abnormal propagation paths in Step 3 , to construct an adaptive resource allocation system to dynamically optimize the time resolution and cache strategy of the monitoring system. This design that depends on the previous steps ensures that resource allocation decisions fully consider scenario changes and abnormal propagation risks, making the system resource utilization more efficient. The specific implementation is as follows:

[0158] Step 4.1. Based on the scenario change metrics, construct a scenario importance assessment model;

[0159] The calculation formula for scenario importance is:

[0160] ;

[0161] where represents the scenario importance at time , represents the scenario change metrics at time , , , respectively represent the weight coefficients of the scenario change metrics, the degree of abnormality, and the context importance; represents the degree of abnormality at time , which can be extracted from the probability matrix of abnormal propagation paths in Step 3; represents the context importance at time , considering the time relationship between the time point and the key event.

[0162] Step 4.2. Dynamically adjust the time resolution and adaptively allocate the sampling frequency according to the scenario importance;

[0163] The calculation formula is:

[0164] ;

[0165] wherein represents the sampling frequency at time , and represents the result of scene importance evaluation; the function is a piecewise function, expressed as:

[0166] ;

[0167] wherein , , respectively represent high, medium, and low time resolutions, corresponding to different sampling frequencies; and[[ID=**28**]] are the high and low threshold parameters respectively, used to divide the importance levels.

[0168] Step **4.3**, construct a hierarchical cache structure;

[0169] ... [[ID=3**5**]]The hierarchical cache structure includes a high-resolution short-term cache and a low-resolution long-term cache , and the relationship between the two is:

[0170] : Save high-resolution image data within the most recent time for detailed analysis and anomaly diagnosis;

[0171] : Save low-resolution or key-frame image data within the most recent time for long-term trend analysis;

[0172] ;

[0173] wherein and respectively represent the time window sizes of the short-term and long-term caches, represents much greater than;

[0174] Step **4.4**, when an important event is detected, trigger the backtracking analysis algorithm to temporarily increase the historical frame retention density;

[0175] The calculation formula is:

[0176] ;

[0177] wherein and respectively represent the historical frame retention densities at the adjusted and pre-adjusted time points , that is, the number of image frames retained per unit time, is a time - weighted function used to adjust the retention density according to the correlation between the time point and the event , and can be defined as:

[0178] ;

[0179] where is the time when the event occurs; is the density boost coefficient, which controls the maximum boost amplitude; is the time - influence range parameter, which controls the decay rate of the time influence; is the exponential function.

[0180] The output result of this step is the adaptive time resolution and the dynamically adjusted caching strategy. These outputs will directly affect the data acquisition and storage behavior of the system. By dynamically adjusting the sampling frequency according to the scene importance, the system can obtain higher - quality data in critical scenes while saving resources in non - critical scenes. In addition, the design of the hierarchical caching structure and the backtracking analysis algorithm ensures that the system can provide complete historical context information when abnormal events occur. These optimization strategies and caching decisions provide an operation framework and execution environment for the intelligent caching optimization in Step 5, enabling the final data management to achieve more refined value - oriented optimization on this basis.

[0181] Step 5: According to the abnormal propagation path probability matrix and the results of the adaptive resource allocation system, execute the intelligent caching optimization algorithm to achieve value - based data management and optimal allocation of storage resources;

[0182] This step is based on all the previous steps, especially integrating the abnormal propagation path probability matrix in Step 3 and the adaptive resource allocation strategy in Step 4 to achieve intelligent caching optimization, allocate more storage resources to the potentially affected targets, and retain the most valuable information. This step is the final execution layer of the entire system, transforming the analysis results of all the previous steps into specific data management strategies. The specific implementation is as follows:

[0183] Step 5.1: Optimize the storage resource allocation strategy;

[0184] Based on the abnormal propagation path probability matrix predicted in Step 3 , optimize the storage resource allocation strategy:

[0185] ;

[0186] where represents the amount of storage resources allocated to the node ; Indicates the probability that an anomaly propagates from a slave node to a node , from the output of step 3; Indicates the importance of node , which can be evaluated based on factors such as target type, function, etc.; The function can be defined as:

[0187] ;

[0188] where represents the storage resource allocation function, is the basic storage resource allocation amount, the minimum storage resource amount allocated to all targets; and respectively represent the influence coefficients of the anomaly propagation probability and target importance on resource allocation.

[0189] Step 5.2, construct an intelligent elimination algorithm based on information value to evaluate the retention value of each frame of image;

[0190] The calculation formula is:

[0191] ;

[0192] where represents the retention value score of image , used to determine the retention priority of the image when storage resources are limited; represents the information amount of image , which can be measured by indicators such as information entropy or the number of key targets; represents the storage duration of image , that is, the difference between the current time and the image acquisition time; represents the correlation between image and the current concerned event , which can be calculated by the matching degree between the image content and the event characteristics;

[0193] The function can be defined as:

[0194] ;

[0195] where represents the image value scoring function after comprehensively considering the information amount, storage duration, and event correlation; , , respectively represent the information amount, storage duration, and correlation with the event of the image; , , respectively represent the weight coefficients of the information amount, storage duration, and event correlation; is the time decay factor, which controls the rate at which the image value decays over time; is an exponential function.

[0196] Step 5.3, combined with storage resource constraints, implement an optimized elimination algorithm;

[0197] When the cache space is insufficient, preferentially eliminate images with lower value scores and retain high-value images.

[0198] Step 5.4, implement an incremental learning model by continuously monitoring system performance metrics;

[0199] System performance metrics such as key event capture rate, storage utilization, etc., dynamically adjust the weight parameters , , , , , etc., and continuously optimize the cache policy.

[0200] Step 5.5, construct a cache access index system;

[0201] Support fast data retrieval based on multiple dimensions such as time, space, target type, and event association, and improve the access efficiency of historical data.

[0202] The output result of this step is a complete intelligent cache optimization system, which not only guides the data storage behavior of the system, but also realizes self-optimization through incremental learning. In this way, the system can continuously improve the efficiency and accuracy of data management during long-term operation. At the same time, this step also forms a closed loop of the entire processing flow, because the optimized cache system will affect the acquisition and processing of future image data, and these data will become the input of Step 1, thus forming a continuously optimized closed-loop system. Through the close connection and data circulation of the five steps, this embodiment realizes a high-speed train image capture processing system with high efficiency, accuracy, and resource conservation.

[0203] An AI-based high-speed train image capture processing system for performing the above-mentioned AI-based high-speed train image capture processing method, including:

[0204] A multi-dimensional information entropy analysis module for processing high-speed train images based on the sliding window algorithm to identify key scene changes;

[0205] A dynamic association graph construction module for constructing a spatial relationship network between targets based on key change frames and generating a target association graph containing temporal information;

[0206] An abnormal propagation prediction module for calculating the abnormal diffusion path between targets through multi-hop abnormal association analysis to form a risk probability matrix;

[0207] A resource adaptive allocation module, which is used to dynamically adjust monitoring parameters according to the scene importance and abnormal risk, so as to improve the data quality of key areas;

[0208] An intelligent cache management module, which is used to optimize the storage strategy based on the abnormal propagation risk and information value assessment, so as to realize the value-oriented management of data.

[0209] Here, the present invention provides an implementation example:

[0210] An actual application test was carried out in a high-speed rail line monitoring system. The length of this line is about 1318 kilometers, and about 680 monitoring points are distributed along the line. Each point generates about 8640 images per day on average (one image is collected every 10 seconds). The application scenario, implementation process and technical effect verification are introduced in detail below.

[0211] In this application, the catenary, track and bridge monitoring systems of a certain section of a high-speed rail were selected as the experimental platform. The length of this section is about 150 kilometers, and 72 image acquisition monitoring points are set, including 28 catenary monitoring points, 32 track monitoring points and 12 bridge structure monitoring points. The duration of the experiment was 30 days, and about 187 million original images were collected, with a total data volume of about 28TB.

[0212] In this application example, we performed multi-dimensional information entropy analysis on the image sequences collected at the same monitoring point at different time periods, and set the sliding window size N = 20, the weight coefficient = 0.6, = 0.4.

[0213] The corresponding relationship between the calculated scene change index and the actual abnormal events under different weather and lighting conditions is shown in Table 1:

[0214] Table 1: Corresponding relationship between scene change index and abnormal events under different conditions

[0215]

[0216] The multi-dimensional information entropy analysis algorithm can effectively identify abnormal scenes under various complex weather and lighting conditions, with an average accuracy rate of 88.3% and an average recall rate of 85.6%, showing good robustness.

[0217] In the catenary monitoring example, we applied the dynamic association graph network model to model key components such as poles, positioners, suspension strings, and contact wires. The comparison of the association accuracy of this model at different sampling frequencies is shown in Table 2:

[0218] Table 2: Comparison of the association accuracy of the dynamic association graph network model at different sampling frequencies

[0219]

[0220] It can be found through comparison that the dynamic association graph network model of this embodiment can improve the target association accuracy while maintaining low computational resource consumption under the condition of adaptive sampling frequency.

[0221] We conducted multi-hop anomaly association analysis on components such as sleepers, fasteners, and rails in the track monitoring system. The comparison of the accuracy of anomaly propagation prediction at different hop counts is shown in Table 3:

[0222] Table 3: Anomaly Propagation Prediction Results under Different Hop Count Conditions

[0223]

[0224] It can be seen that the multi-hop anomaly association analysis algorithm of this embodiment can effectively predict the propagation path of anomalies. Especially within the range of 1 hop and 2 hops, the prediction accuracy rate exceeds 85%, and it can give early warnings of anomaly propagation events 8 - 24 minutes in advance, which wins valuable intervention time for maintenance personnel.

[0225] During the entire experiment period, we compared the storage efficiency of the traditional fixed sampling method and the adaptive resource allocation method of this embodiment. The comparison results of the storage efficiency are shown in Table 4:

[0226] Table 4: Comparison of Storage Efficiency of Different Sampling Strategies

[0227]

[0228] The data shows that the adaptive resource allocation algorithm of this embodiment achieves a storage compression rate of 65%, while increasing the detection rate of key events to 94.6%, far superior to the traditional fixed sampling and simple threshold methods.

[0229] Through 30 consecutive days of application in a certain high-speed rail monitoring system, we comprehensively evaluated the long-term technical effects of this embodiment. The improvement effects of key performance indicators are shown in Table 5:

[0230] Table 5: Evaluation of the Long-term Operating Effects of this Embodiment

[0231]

[0232] In practical applications, we recorded multiple cases of anomaly propagation prediction. The three most representative cases are shown in Table 6:

[0233] Table 6: Analysis of Practical Application Cases of Anomaly Propagation Prediction

[0234]

[0235] The comparison of the monitoring system deployment costs under different embodiments is shown in Table 7 as follows:

[0236] Table 7: Comparison of the monitoring system deployment costs under different embodiments

[0237]

[0238] This embodiment not only has better technical indicators than the traditional method, but also shows advantages in terms of economic benefits. By reducing storage and maintenance costs and improving the fault prevention ability at the same time, the overall return on investment of the system has increased by 45.9 percentage points.

[0239] In practical applications, this embodiment improves the efficiency and reliability of the high-speed train image monitoring system, reduces the waste of storage resources, enhances the prediction and diagnosis capabilities of abnormal events, and provides strong technical support for the safe operation of high-speed railways.

[0240] The above describes the embodiments of the present invention. However, these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. An AI-based method for capturing and processing high-speed train images, characterized in that, Including: Processing continuously acquired high-speed train images based on a sliding window algorithm to identify key changes in the scene; Utilizing the key changes in the scene to construct a dynamic association graph network model and applying a temporal graph attention network for target spatial relationship processing to generate a target association graph; Based on the target association graph, using a multi-hop anomaly association analysis algorithm to predict the propagation of anomalies between targets and obtaining an anomaly propagation path probability matrix. The use of the multi-hop anomaly association analysis algorithm to predict the propagation of anomalies between targets includes: Calculating the adjacency matrix and degree matrix corresponding to the target association graph; Constructing a normalized adjacency matrix; Initializing the anomaly attention vector of the nodes; Iteratively calculating the k-hop anomaly association using the attention propagation algorithm; Constructing an anomaly propagation model to predict the diffusion path of anomalies on the target graph; Based on the key changes in the scene and the anomaly propagation path probability matrix, implementing an adaptive resource allocation system to dynamically adjust image acquisition parameters according to the scene importance evaluation model; According to the anomaly propagation path probability matrix and the results of the adaptive resource allocation system, executing an intelligent cache optimization algorithm to achieve value-based data management and optimized allocation of storage resources.

2. The method for capturing and processing high-speed train images based on AI according to claim 1, wherein The processing of continuously acquired high-speed train images based on the sliding window algorithm includes: Maintaining the most recent N-frame image sequence using the sliding window algorithm; Calculating the information entropy difference between adjacent image frames; Calculating the structural similarity between adjacent image frames; Combining the information entropy difference and the structural similarity to construct a scene change index.

3. The method for capturing and processing high-speed train images based on AI according to claim 1 is characterized in that, The construction of the dynamic association graph network model includes: Identifying a set of key targets from the image using an object detection algorithm; Constructing a scene dynamic association graph; Calculating the association weight between targets; Applying a temporal graph attention network to process the dynamic association between targets.

4. The method for capturing and processing high-speed train images based on AI according to claim 1, wherein The adaptive resource allocation system includes: Constructing a scene importance evaluation model based on the scene change index; Dynamically adjusting the time resolution and adaptively allocating the sampling frequency according to the scene importance; Constructing a hierarchical cache structure; When an important event is detected, triggering a backtracking analysis algorithm to temporarily increase the retention density of historical frames.

5. A method for capturing and processing high-speed train images based on AI according to claim 1, characterized in that, The intelligent cache optimization algorithm includes: Optimizing the storage resource allocation strategy; Constructing an intelligent elimination algorithm based on information value to evaluate the retention value of each frame of image; Combining the storage resource constraints to implement the optimized elimination algorithm; Implementing an incremental learning model by continuously monitoring the system performance metrics; Constructing a cache access index system.

6. The method for capturing and processing high-speed train images based on AI according to claim 2, wherein The calculation formula for constructing the scene change index is: ； Among them represents the scene change index at a moment , and respectively represent the weight coefficients of information entropy difference and structural similarity in the scene change index represents the information entropy difference between adjacent image frames represents the structural similarity index , respectively represent the images at the moment and the moment .

7. A method for processing high-speed train image capture based on AI according to claim 3, characterized in that, The calculation formula for applying the temporal graph attention network to process the dynamic association between targets is: ； Among them is the scene dynamic association graph, is the hidden state at the current moment, including the feature representations of all nodes; is the observation feature at the current moment, that is, the target feature extracted from the image; represents the temporal graph attention network, which integrates the graph neural network and the attention algorithm, and can capture the correlation information between targets and maintain temporal coherence; is the updated hidden state, including temporal and spatial relationship information.

8. A method for processing high-speed train image capture based on AI according to claim 1, characterized in that, The calculation formula for constructing the anomaly propagation model to predict the diffusion path of anomalies on the target graph is: ； Among them represents the conditional probability of an abnormality occurring at node under the condition that an abnormality occurs at node ; represents the abnormal propagation attention value from node to node ; and respectively represent the state characteristics of node and node ; is a conditional probability calculation function.

9. An AI-based high-speed train image capture and processing system, characterized in that, A method for processing high-speed train image capture based on AI according to any one of claims 1-8, including: A multi-dimensional information entropy analysis module for processing high-speed train images based on a sliding window algorithm to identify key scene changes; A dynamic association graph construction module for constructing a spatial relationship network between targets based on key change frames to generate a target association graph containing temporal information; An anomaly propagation prediction module for calculating the anomaly diffusion path between targets through multi-hop anomaly association analysis to form a risk probability matrix; A resource adaptive allocation module, which is used to dynamically adjust monitoring parameters according to scene importance and anomaly risk, and improve the data quality of key areas; An intelligent cache management module, which is used to optimize the storage strategy based on anomaly propagation risk and information value assessment, and realize value-oriented management of data.

Images