Railway freeze injury monitoring method and system based on AI visual technology

Abstract

The invention provides a railway freezing damage monitoring method and system based on an AI visual technology, and the method comprises the steps: separating interference signals caused by environment illumination and ice layer thickness through a spectral analysis method according to a preliminary reflection characteristic data set, extracting a pure substance specific reflection peak value from the interference signals, and determining the core feature vectors of a water ice type and a salt ice type; if a reflection peak value in the core feature vector exceeds a preset threshold value, judging that the region is a salt ice type dominant region, and performing clustering analysis on the core feature vector through a hierarchical processing system to obtain a distribution mapping graph after type distinguishing; and carrying out deep learning training on the distribution mapping graph by adopting a machine learning model, fusing reflection characteristics and environmental interference data, judging the severity of the mixed freezing injury area, and obtaining a risk assessment index.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and system for monitoring railway frost damage based on AI vision technology. Background Technology

[0002] In the field of transportation safety, especially in railway operations, the identification and prevention of frost damage is a crucial aspect of ensuring the safe operation of trains.

[0003] Frost damage can not only affect the normal operation of track equipment, but may also increase the risk of train operation. Therefore, accurate identification of the type of frost damage is particularly important.

[0004] Whether it's ice and snow cover or freezing caused by salt erosion, both pose a threat to track safety. Researching how to accurately distinguish between different types of freezing damage has become an urgent issue to be addressed.

[0005] However, current research and practice often struggle to address the challenges posed by the diverse properties of various frost-damaging substances in complex frost-damaging environments.

[0006] Many methods, when analyzing the characteristics of frost damage, lack in-depth exploration of the light reflection properties of different materials, resulting in a significant reduction in the accuracy of identification when faced with a mixture of various frost damages or environmental interference.

[0007] This limitation makes it difficult to quickly determine the specific type and severity of frost damage in practical applications, which in turn affects safety decisions for railway operations.

[0008] The deeper technical challenge lies in the difficulty of effectively capturing and distinguishing the subtle differences in light reflection characteristics among different frost-damaged substances.

[0009] In particular, the reflectivity of substances such as water ice and salt ice is affected by a variety of factors, including ambient light, ice thickness, and surface roughness, which greatly increases the complexity of feature extraction.

[0010] As this problem intensifies, a further challenge lies in how to systematically stratify and organize these complex reflective properties so that they can accurately correspond to specific types of frost damage in different scenarios.

[0011] For example, on an ice-covered track, water ice and salt ice may coexist. However, due to subtle differences in reflectivity and environmental interference, current technology often cannot clearly distinguish the distribution areas and proportions of these two types of ice, thus affecting the assessment of track safety risks.

[0012] Therefore, how to construct a technical system that can systematically stratify and accurately distinguish different frost-damaging substances has become a key issue in ensuring railway operation safety and scientific decision-making. Summary of the Invention

[0013] This invention provides a railway frost damage monitoring method based on AI vision technology, mainly including: The orbital surface image data was acquired by a spectral imaging device. A preliminary scan was performed to identify the differences in reflectance characteristics between water ice and salt ice types. The original spectral reflectance curves, which included environmental interference factors, were obtained to generate a preliminary reflectance characteristic dataset. Based on the preliminary reflection characteristic dataset, the interference signals caused by ambient light and ice thickness are separated by spectral analysis. Specific reflection peaks of pure substances are extracted from these signals to determine the core feature vectors of water ice type and salt ice type. If the reflection peak in the core feature vector exceeds a preset threshold, it is determined to be a salt ice type dominant region. The core feature vector is then subjected to cluster analysis through a hierarchical processing system to obtain a distribution mapping map after type differentiation. A machine learning model is used to train the distribution map using deep learning, and by fusing reflection characteristics and environmental disturbance data, the severity of mixed frost damage areas is determined, and risk assessment indicators are obtained. By comparing the risk assessment indicators with the track safety standards, if the indicators exceed the preset safety threshold, the feature processing module is activated to recalibrate the feature extraction process, determine the optimized hierarchical processing parameters, generate a real-time frost damage identification report based on the optimized hierarchical processing parameters, and integrate relevant processing paths to obtain the final safety decision basis. Based on the aforementioned safety decision-making criteria, a classification algorithm is used to analyze and process subsets of reflective properties of different frost-damaging substances to obtain accurate type proportion distribution data. The track equipment monitoring database is updated using the type proportion distribution data to achieve continuous iterative optimization of frost damage identification and to determine the dynamic trend of potential risk assessment.

[0014] This invention provides a railway frost damage monitoring system based on AI vision technology, mainly comprising: The preliminary scanning module is used to acquire image data of the orbital surface through a spectral imaging device, perform a preliminary scan based on the differences in reflectance characteristics between water ice and salt ice types, obtain the original spectral reflectance curves including environmental interference factors, and obtain a preliminary reflectance characteristic dataset. The feature extraction module is used to separate the interference signals caused by ambient light and ice thickness using spectral analysis methods based on the preliminary reflectance characteristic dataset, extract the specific reflectance peaks of pure substances, and determine the core feature vectors of water ice type and salt ice type. The type determination module is used to determine that if the reflection peak value in the core feature vector exceeds a preset threshold, it is a salt ice type dominant region. The core feature vector is then subjected to cluster analysis through a hierarchical processing system to obtain a distribution mapping map after type differentiation. The risk assessment module is used to train the distribution map using a machine learning model, integrate reflection characteristics and environmental disturbance data, determine the severity of mixed frost damage areas, and obtain risk assessment indicators. The decision generation module is used to compare the risk assessment indicators with the track safety standards. If the indicators exceed the preset safety threshold, the feature processing module is activated to recalibrate the feature extraction process, determine the optimized hierarchical processing parameters, generate a real-time frost damage identification report based on the optimized hierarchical processing parameters, and integrate relevant processing paths to obtain the final safety decision basis. The optimization and update module is used to analyze and process subsets of reflective characteristics of different frost-damaging substances using a classification algorithm based on the safety decision criteria, obtain accurate type proportion distribution data, update the track equipment monitoring database through the type proportion distribution data, realize continuous iterative optimization of frost damage identification, and judge the dynamic change trend of potential risk assessment.

[0015] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a method for identifying track surface frost damage. It addresses the challenges of inaccurate identification, untimely risk assessment, and lack of continuous optimization mechanisms caused by the significant differences in reflectance characteristics between water ice and salt ice types of track surface frost damage, complex environmental interference factors, and the resulting operational problems. The solution addresses the interconnected issues of accurately distinguishing between water ice and salt ice frost damage types in complex environments, assessing the severity of mixed areas, and dynamically optimizing safety decisions to ensure the reliability of track equipment monitoring. This invention acquires raw spectral reflectance curves using a spectral imaging device, employs spectral analysis to separate interference signals and extract core feature vectors. If the reflectance peak exceeds a threshold, it is determined to be dominated by salt ice, and cluster analysis is performed to generate a distribution map. Then, a machine learning model is used to fuse data and train risk indicators. After comparison with safety standards, the feature processing module is activated to calibrate parameters and generate real-time reports. Iterative optimization is achieved by analyzing the type proportion distribution through a classification algorithm and updating the database. This solves the aforementioned problems and ultimately automates high-precision identification, dynamic risk assessment, and safety decision-making for track frost damage, improving the monitoring efficiency and reliability of track equipment. Attached Figure Description

[0016] The above and other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which: Figure 1 This is a flowchart of a railway frost damage monitoring method based on AI vision technology according to the present invention.

[0017] Figure 2 This is a schematic diagram of a railway frost damage monitoring method based on AI vision technology according to the present invention.

[0018] Figure 3 This is another schematic diagram of a railway frost damage monitoring method based on AI vision technology according to the present invention.

[0019] Figure 4 This is a schematic diagram of the structure of a railway frost damage monitoring system based on AI vision technology according to the present invention.

[0020] Figure 5 This is a comparison of the spectral reflectance curves of water ice and salt ice according to the present invention.

[0021] Figure 6 A comparison chart showing the detection response time and treatment time for different types of frost damage.

[0022] Figure 7 This is a schematic diagram showing the distribution of frost damage areas on the track surface according to the present invention.

[0023] Figure 8 This is a clustering distribution diagram of the feature vectors of this invention.

[0024] Figure 9 This is a graph comparing the risk assessment errors of different methods under environmental disturbance conditions.

[0025] Figure 10 This is a distribution mapping diagram of frost damage types according to the present invention.

[0026] Figure 11 This is a timing diagram of the processing flow of the railway frost damage monitoring system of the present invention.

[0027] Figure 12 This is a comparison chart of the accuracy of frost damage identification between the present invention and traditional methods.

[0028] Figure 13 This is a schematic diagram of the structure of the spectral imaging device of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown in the figure, a railway frost damage monitoring method and system based on AI vision technology in this embodiment may specifically include: Step S101: Acquire orbital surface image data using a spectral imaging device, perform a preliminary scan to identify the differences in reflectance characteristics between water ice and salt ice types, obtain the original spectral reflectance curves including environmental interference factors, and obtain a preliminary reflectance characteristic dataset.

[0031] Data was collected from the orbital surface using spectral imaging technology. Based on the differences in reflectance between water ice and salt-deposited ice, raw curves containing environmental interference were obtained, forming a preliminary reflectance characteristic dataset. Wavelet denoising was performed on the raw curves in the preliminary reflectance characteristic dataset using Python's SciPy library. Signals were obtained from the raw curves, decomposed using wavelet transform, and noise coefficients were filtered out by setting a threshold to reconstruct the cleaned reflectance curve data. Based on the cleaned reflectance curve data, LIBSVM was used for support vector machine classification, taking into account the reflectance differences between water ice and salt-deposited ice. Feature vectors were extracted from the cleaned reflectance curve data, and the support vector machine model was trained to determine the reflectance characteristic distribution range of the two types of materials. By analyzing the reflectance characteristic distribution range, the differences in reflectance intensity of the two types of materials at different wavebands were obtained, determining their spatial distribution characteristics on the orbital surface. Based on the spatial distribution characteristics, the final distribution ratio of water ice and salt-deposited ice on the orbital surface was obtained, determining the relative content differences between the two types of materials.

[0032] Specifically, when acquiring orbital surface image data using a spectral imaging device, a high-resolution multispectral camera can be used, with a wavelength range of 0.4 to 2.5 micrometers, covering the visible to near-infrared region, to capture the characteristic reflectance peaks of water ice and salt ice. For example... Figure 13 As shown, the spectral imaging device includes a multispectral camera lens, a collimating lens, a beam splitter, a filter wheel, a focusing lens, and a CCD sensor array. Incident light rays are converged by the collimating lens and then enter the beam splitter. The multi-band light rays are then filtered by a rotatable filter wheel, which has six filters for different wavelengths, covering a spectral range of 0.4 to 2.5 micrometers. The light rays are finally focused by the focusing lens onto the CCD sensor array for imaging. The sensor array uses a 4×10 pixel matrix arrangement, enabling simultaneous multi-band acquisition. The device has a field of view of 50°, a focal length of 50mm, and a total length of 220mm. It can be mounted on an orbital probe to acquire spectral image data of the orbital surface at a frequency of 10 frames per second.

[0033] The imaging device, mounted on an orbital probe, acquires data at a frequency of 10 frames per second, achieving a resolution of 0.5 meters per pixel, ensuring coverage of a 100-square-kilometer area on the orbital surface, and generating a raw image dataset. During the initial scan to identify the differences in reflectance characteristics between water ice and salt ice, spectral analysis algorithms are used to extract reflectance data for each pixel, focusing on the absorption bands of water ice at 1.5 and 2.0 micrometers, and the characteristic reflection peak of salt ice at 1.8 micrometers. Significantly different regions are identified by calculating the correlation coefficient between reflectance and wavelength (with a threshold of 0.8), forming a preliminary classification map. When acquiring the raw spectral reflectance curves, which include environmental interference factors, the system automatically records the reflectance data of each pixel across the entire wavelength band. This data is then corrected using an environmental model to account for the solar incidence angle (assumed to be 30 degrees) and atmospheric scattering effects (correction coefficient of 0.95), generating a reflectance curve dataset containing noise. The wavelength-reflectance pairs for each curve are stored in CSV format. Upon obtaining the preliminary reflectance characteristic dataset, Principal Component Analysis (PCA) was used to reduce the dimensionality of the original curves, extracting the first three principal components (with a cumulative contribution rate of 90%) to reduce data redundancy. Simultaneously, K-means clustering (K=3) was used to classify the data into three categories: water ice, salt ice, and background. The classification results and a table of reflectance characteristic parameters were output, with the mean reflectance of water ice approximately 0.7, salt ice approximately 0.5, and background approximately 0.3, forming the final dataset. These steps are automated into a logical chain, from data acquisition to feature extraction and classification analysis, ensuring the continuity of the technical process and the reliability of the results.

[0034] Step S102: Based on the preliminary reflectance characteristic dataset, use spectral analysis to separate the interference signals caused by ambient light and ice thickness, extract the specific reflectance peaks of pure substances, and determine the core feature vectors of water ice type and salt ice type.

[0035] For the preliminary reflectance characteristic dataset, spectral analysis is used to decompose the signals in the dataset, separating interference signals caused by ambient light and ice thickness, resulting in pre-cleaned signal data. Based on this pre-cleaned signal data, filtering tools are used to further suppress interference signals, extracting reflectance peak signals related to the unique properties of the substances, and determining the main reflectance peak set. Analysis of the main reflectance peak set is used to compare the reflectance peak signals of water ice and salt ice types, obtaining the unique signal differences between the two types of substances in different wavelength bands, and determining the core characteristics of the two types of substances. Based on the core characteristics of the two types of substances, a support vector machine algorithm is used to classify the feature data, constructing classification boundaries for water ice and salt ice types, and obtaining a set of classified feature vectors. Using the classified feature vector set, signal recombination is performed on the feature vectors of each type of substance to obtain complete signal patterns corresponding to the unique properties of the substances, determining the signal identifiers of the two types of substances. Based on the signal identifiers of the two types of substances, the signals in the original dataset are matched one by one. If the matching degree between the signal pattern and the identifier exceeds a preset threshold, it is classified into the corresponding substance type, obtaining the final substance classification result.

[0036] like Figure 5 As shown, water ice and salt ice exhibit significantly different spectral reflectance characteristics in the 0.4 to 2.5 micrometer wavelength range. The spectral reflectance curve of water ice (solid line) shows two significant absorption valleys at 1.50 micrometers and 2.00 micrometers, with reflectance decreasing to 0.45 and 0.35 respectively, while the overall reflectance remains at a relatively high level of 0.6 to 0.8. In contrast, the spectral reflectance curve of salt ice (dashed line) shows a characteristic reflection peak at 1.80 micrometers, with a peak reflectance reaching 0.68, and an overall reflectance range between 0.4 and 0.6, significantly lower than that of water ice. By extracting the reflection peaks and absorption valleys in these specific wavelength bands, core feature vectors for water ice and salt ice types can be constructed, providing material characteristic basis for subsequent identification of frost damage types.

[0037] like Figure 6As shown, the detection response time varies significantly among different types of frost damage. The total time for water ice is 13.6 seconds, for salt ice it is 16.7 seconds, and for mixed frost damage it is 24.0 seconds. The time distribution for each processing stage is as follows: the data acquisition stage takes 4.5 seconds, 4.8 seconds, and 5.2 seconds for the three types, with little difference; the feature extraction stage takes 3.2 seconds, 4.5 seconds, and 6.8 seconds, with mixed frost damage showing a significant increase in time; the classification and identification stage takes 2.1 seconds, 3.2 seconds, and 5.5 seconds, reflecting an increasing complexity; and the risk assessment stage takes 3.8 seconds, 4.2 seconds, and 6.5 seconds, with mixed frost damage taking the longest time. The horizontally stacked bar chart in the figure uses four different gray levels (0.85, 0.70, 0.55, 0.40) and filling patterns (diagonal lines, backslashes, dots, and intersections) to distinguish the four processing stages. The specific time for each stage is marked in the center, and the total time is highlighted in a white box on the right. Data shows that feature extraction and classification of mixed frost damage are the main performance bottlenecks, providing a clear direction for system optimization.

[0038] Specifically, based on the preliminary reflectance characteristic dataset, the system first processes the data using a spectral decomposition algorithm to separate ambient light interference signals. A linear spectral unmixing model is employed to decompose the reflectance data of each pixel into ambient light components and intrinsic material reflectance components. A reference reflectance of 0.2 is set as the baseline, with the average reflectance of the ambient light spectrum in the 0.4 to 0.7 micrometer range. The ambient light contribution ratio of each pixel is calculated using the least squares method, with the average error controlled within 0.05. Next, to address signal interference caused by ice thickness, the system introduces a deep learning-based thickness estimation model. A trained convolutional neural network is used to analyze the shape changes of the reflectance curve in the 1.0 to 2.2 micrometer band to predict the ice thickness. Assuming a thickness range of 0.1 to 5.0 meters, the prediction accuracy reaches 0.2 meters. The deviation in absorption intensity is then inferred from the thickness value and corrected, reducing the corrected reflectance error to 0.03. Subsequently, the system extracts specific reflectance peaks of pure substances from the corrected data. For water ice, it identifies the characteristic absorption peak at 1.4 micrometers, calculates the ratio of peak depth to baseline reflectance, and sets a threshold of 0.6 to confirm feature significance. For salt ice, it focuses on the reflectance peak at 1.9 micrometers, extracts the contrast between the peak height and surrounding bands, and sets a contrast threshold of 0.4 to ensure the accuracy of feature extraction. Finally, the system constructs feature vectors from the extracted reflectance peak data using a support vector machine algorithm, mapping the core feature vectors of water ice and salt ice to high-dimensional spaces respectively. It calculates the Euclidean distance between feature vectors, sets a distance threshold of 1.5 to distinguish between the two types of substances, generates the final feature vector matrix, and stores it as a standard format file, forming a complete automated analysis chain from interference separation to feature determination.

[0039] Step S103: If the reflection peak value in the core feature vector exceeds a preset threshold, it is determined to be a salt ice type dominant region. The core feature vector is then subjected to cluster analysis through a hierarchical processing system to obtain a distribution mapping map after type differentiation.

[0040] If the reflection peak value in the core feature vector exceeds a preset threshold, it is determined to be a salt ice type dominant region. Signal intensity analysis is used to initially screen the feature vectors, resulting in a preliminary classification set of regions. Based on this preliminary classification set, a hierarchical processing method is used to group the feature vectors within each region, obtaining a grouped signal cluster set. Using this grouped signal cluster set, the reflection peak value and signal intensity within each signal cluster are compared and analyzed to determine the material type identifier corresponding to each cluster. Based on the material type identifier, clustering techniques are used to further integrate the signal cluster set, resulting in type-differentiated signal groups. Using these type-differentiated signal groups, spatial mapping processing is performed on the regional division data within each group to obtain the corresponding distribution map. Based on the distribution map, signal intensity is checked against the boundaries between the salt ice type dominant region and other regions to determine the final region division result. Using the final region division result, the core features and reflection peak values ​​within each region are recorded and stored, resulting in a complete type distribution archive.

[0041] like Figure 7 As shown, the track surface exhibits distribution characteristics of various types of frost damage. Within a 24-meter-long track section, water ice-covered areas (diagonally filled areas), salt ice-covered areas (dot-like filled areas), and mixed frost damage areas (cross-filled areas) are distributed between and around the double rails. Water ice-covered areas A and C are mainly distributed on the left and upper right sides of the track, exhibiting irregular shapes; salt ice-covered area B is located in the middle of the track, approximately 3.0 meters long; the mixed frost damage area is located on the right side of the track, approximately 0.65 meters long. This area, due to the overlap of water ice and salt ice, forms a mixed frost damage zone and is classified as a severely frost damage area. By using a hierarchical processing system to perform cluster analysis on the core feature vectors of different areas, the spatial distribution and coverage of each type of frost damage can be accurately identified, providing spatial positioning basis for subsequent risk assessment.

[0042] like Figure 8As shown, the extracted core feature vectors are mapped to a two-dimensional feature space after dimensionality reduction using Principal Component Analysis (PCA). The horizontal axis represents Principal Component 1 (PC1), and the vertical axis represents Principal Component 2 (PC2), with coordinates ranging from -3 to 3. In the feature space, the three types of substances exhibit obvious clustering distribution characteristics: water ice type data points (circular markers) are mainly concentrated in the upper left region, forming a tight cluster; salt ice type data points (triangular markers) are mainly distributed in the lower right region, clearly separated from water ice type; background or other type data points (square markers) are scattered in the middle region. Black cross markers indicate the center point of each cluster, and dashed ellipses delineate the boundary range of each cluster. This clear inter-class separation and intra-class aggregation characteristic verifies the effectiveness of the feature extraction method and provides a theoretical basis for accurately distinguishing different types of frost-damaged substances.

[0043] Specifically, the system first performs threshold detection on the reflection peaks in the core feature vectors, setting the threshold for determining the salt ice type to 2.3. It scans the reflection peak data in the 1.8-micron band; if the detected peak intensity exceeds this threshold, the region is automatically marked as a salt ice-dominant region. The data processing module stores the feature vector of this region separately in a temporary database and generates a preliminary classification label file. Subsequently, the system initiates a hierarchical processing system, employing a density-based spatial clustering algorithm to group and analyze the labeled core feature vectors. The cluster radius is set to 0.8, and the neighborhood point threshold is 5. By calculating the density distribution of each feature vector in high-dimensional space, different clusters are automatically identified. For the salt ice-dominant region, the system further extracts its secondary feature peaks in the 1.6 to 2.0-micron band, calculates their intensity ratio to the primary peak, and sets the ratio range to 0.3 to 0.7 as an auxiliary classification criterion to improve clustering accuracy. Next, the system maps the clustering results to a two-dimensional space, generating a type distribution mapping map. An interpolation algorithm is used to smooth the distribution boundaries, with the interpolation grid spacing set to 0.1 meters to ensure a natural boundary transition. At the same time, different types of regions are distinguished by color coding: salt ice regions are marked in dark blue, and water ice regions are marked in light blue. The generated visualization layer is automatically saved as a high-resolution image file for subsequent analysis modules to use. The entire process is automated through scripts to realize data flow and result integration, forming a complete processing chain from threshold detection to distribution mapping.

[0044] Step S104: A machine learning model is used to train the distribution map using deep learning, and the reflection characteristics and environmental interference data are fused to determine the severity of the mixed frost damage area and obtain risk assessment indicators.

[0045] The distribution map is processed using a machine learning model, and deep learning methods are employed to analyze the reflectance data and environmental interference information to obtain preliminary frost damage area division results. Based on the preliminary frost damage area division results, the reflectance data within each area is processed, and the severity level of the area is determined by combining the impact of environmental interference. Using the severity level within each area, data integration technology is used to comprehensively analyze the reflectance data and environmental interference information of each area to obtain a median risk assessment value. If the median risk value of a certain area exceeds a preset threshold, its reflectance data undergoes secondary information processing to determine the frost damage risk level of that area. Based on the frost damage risk level, environmental interference information in high-risk areas is analyzed in detail to obtain a corresponding set of assessment indicators. Based on the set of assessment indicators, the severity and risk level of each area are correlated and matched to determine the final regional risk distribution map. Using the final regional risk distribution map, the data integration results for high-risk areas are stored and processed to obtain a complete frost damage analysis archive.

[0046] like Figure 9 As shown, within the environmental interference intensity range of 0 to 10, the risk assessment error of the traditional threshold method increases sharply from 5% to 35%, a 30 percentage point increase; the error of the single machine learning method increases from 4% to 22%, an 18 percentage point increase; while the error of the method of this invention increases only slowly from 2% to 8%, a 6 percentage point increase. In the figure, the horizontal axis represents the environmental interference intensity level, and the vertical axis represents the percentage of risk assessment error. The three curves are distinguished by solid lines with square markers, dashed lines with triangle markers, and dotted lines with circles markers, respectively. Comparative data shows that the method of this invention, by integrating spectral analysis and environmental interference data correction technology, reduces the assessment error by 27 percentage points compared to the traditional method and by 14 percentage points compared to the single machine learning method under strong interference (level 10) conditions, verifying the stability and anti-interference capability of this invention under complex environmental conditions.

[0047] like Figure 10As shown, this distribution map displays the spatial distribution characteristics of frost damage risk within a 100-meter-long and 5-meter-wide track area using grayscale contour lines. The horizontal axis represents the track length (0-100 meters), and the vertical axis represents the track width (0-5 meters). Risk levels are distinguished by grayscale shades: dark black areas represent high-risk areas (risk index > 7), medium gray areas represent medium-risk areas (risk index 4-7), and light gray areas represent low-risk areas (risk index < 4). Four key contour lines are marked with values ​​of 3, 5, 7, and 9, clearly defining the boundaries of different risk zones. Black pentagrams mark the locations of three typical high-risk points: (25m, 2.5m), (60m, 3.8m), and (85m, 1.2m). The grayscale scale on the right provides a continuous mapping relationship of risk indices, facilitating rapid assessment of the risk level in any area. This mapping map integrates deep learning training results of reflection characteristics and environmental interference data, providing an intuitive spatial reference for track equipment maintenance decisions and inspection route planning.

[0048] Specifically, the system first preprocesses the distribution map, automatically extracting the features of frost-damaged areas. For each pixel, it calculates the reflection intensity value in the 2.2-micron band, sets an initial filtering threshold of 1.5, and marks pixels below this threshold as low-risk areas. The data is then input into a pre-trained convolutional neural network model using the ResNet-50 architecture with a learning rate of 0.001. By performing convolution operations on the feature layers of the map, deep spatial features are extracted. The system then fuses the extracted features with environmental interference data, such as temperature fluctuation values ​​within the range of -5 to 5 degrees Celsius. Principal component analysis is used to reduce the dimensionality of the fused data, retaining the first three principal components to reduce computational complexity. Next, the system initiates a deep learning training module, employing a gradient descent optimization algorithm with 1000 iterations and a batch size of 32. This module performs classification training on the fused feature data, automatically generating severity labels for mixed frost damage areas. These labels are categorized into mild, moderate, and severe. Severe areas are determined by a combined reflectance intensity and environmental disturbance index exceeding 3.2. Subsequently, the system calculates risk assessment indicators for each area using a weighted average method, setting the reflectance feature weight at 0.6 and the environmental disturbance weight at 0.4. This generates a final risk index value, ranging from 0 to 10. Areas with an index higher than 7 are marked as high-risk. The data processing module automatically stores the risk assessment results in a cloud database and generates a corresponding risk distribution heatmap with a grid resolution of 0.2 meters. An automated script then pushes the results to the relevant analysis system, forming a complete data processing flow from feature extraction to risk assessment.

[0049] Step S105: The risk assessment index is compared with the track safety standard. If the index exceeds the preset safety threshold, the feature processing module is activated to recalibrate the feature extraction process, determine the optimized hierarchical processing parameters, generate a real-time frost damage identification report based on the optimized hierarchical processing parameters, and integrate relevant processing paths to obtain the final safety decision basis.

[0050] See Figure 2 The risk assessment indicators are compared with track safety standards. If the assessment result exceeds a preset safety threshold, the feature calibration module is activated to perform preliminary screening of the raw data and obtain an adjusted feature set. Based on the adjusted feature set, a feature extraction process is used to perform hierarchical processing of the data. Combined with the optimization and adjustment of hierarchical parameters, the hierarchical data groups are determined. Through the hierarchical data groups, in-depth analysis is performed on the information within each group. If the data fluctuation within a group exceeds a preset range, it is calibrated a second time to obtain calibrated data units. Based on the calibrated data units and combined with the business logic of frost damage identification, the information within each unit is classified and labeled to obtain a classified frost damage identifier set. Through the classified frost damage identifier set, a support vector machine algorithm is used to perform pattern matching on the identifiers to determine the distribution of potential high-risk areas. Based on the distribution of high-risk areas and combined with the processing path, the data is integrated and mapped to generate real-time archives and determine the basis for the final safety decision.

[0051] like Figure 11 As shown, the complete processing flow of the system of this invention includes 7 consecutive stages, with a total processing time of 28 seconds. The data acquisition stage (0-5 seconds) acquires track surface image data using a spectral imaging device; the preprocessing stage (5-8 seconds) performs noise reduction and interference signal separation; the feature extraction stage (8-12 seconds) extracts core feature vectors; the clustering analysis stage (12-16 seconds) performs hierarchical processing and type differentiation; the risk assessment stage (16-22 seconds) uses a machine learning model for deep learning training; the decision generation stage (22-25 seconds) generates a safety decision report; and the database update stage (25-28 seconds) achieves continuous iterative optimization. In the figure, horizontal bars use gradient grayscale (0.9 to 0.0) and different fill patterns to represent the process progression, and key processing nodes are marked with vertical dashed lines to ensure that the time sequence of each stage is clearly distinguishable. This time sequence diagram shows that the system can complete the entire process from data acquisition to decision output within 30 seconds, meeting the timeliness requirements of real-time monitoring of railway frost damage.

[0052] Specifically, the system first uses an automated comparison module to match the calculated risk assessment indicators with track safety standards. Assuming a safety threshold of 6.5, if an indicator value exceeds this threshold, for example, reaching 7.2, the system automatically triggers an alarm signal and records a deviation of 0.7. This data is then transmitted to the feature processing module for further processing. Upon receiving the alarm, the feature processing module initiates a calibration procedure, recalculating the feature extraction weights using an adaptive adjustment algorithm. The initial weight is set to 0.5, and based on historical data analysis, an optimized weight value of 0.55 is obtained. Simultaneously, combined with spatial correlation analysis, the grid resolution for feature extraction is adjusted from 0.3 meters to 0.25 meters to improve accuracy. Next, based on the optimized hierarchical processing parameters, such as adjusting the temperature sensitivity parameter for frost damage identification from 2.0 to 1.8, the system automatically generates a real-time frost damage identification report. The report includes the probability value of frost damage distribution, setting areas with a probability higher than 0.75 as key monitoring targets. A data fusion algorithm integrates the processing path with historical frost damage trend data to calculate a comprehensive impact coefficient. Assuming a coefficient value of 0.68, areas below 0.7 are considered controllable. Finally, based on the above analysis results, the system automatically generates safety decision criteria, which include a priority ranking of frost damage repair. The priority calculation adopts a weighted scoring method, with the influence of temperature set to 0.45, the influence of humidity set to 0.35, and other factors set to 0.2. Areas with a comprehensive score higher than 8.0 will be marked as emergency treatment objects. All data are transmitted to the decision execution system through an automated interface, forming a complete closed-loop process from risk comparison to decision generation.

[0053] Step S106: Based on the safety decision criteria, a classification algorithm is used to analyze and process subsets of reflective characteristics of different frost-damaging substances to obtain accurate type proportion distribution data. The track equipment monitoring database is updated through the type proportion distribution data to achieve continuous iterative optimization of frost damage identification and to determine the dynamic trend of potential risk assessment.

[0054] like Figure 12As shown, the method of this invention achieves identification accuracies of 92%, 89%, and 85% for water ice, salt ice, and mixed frost damage areas, respectively. Compared with the traditional threshold method (65%, 58%, 45%) and the single support vector machine method (72%, 68%, 62%), the identification accuracy is improved by 27, 31, and 40 percentage points, respectively. The improvement in identification accuracy for mixed frost damage areas is most significant, increasing from 45% to 85% (a 40 percentage point increase), verifying the effectiveness of this invention's fusion of spectral analysis and machine learning techniques for complex frost damage scenarios. The three sets of bar charts in the figure correspond to water ice type (diagonal line filling), salt ice type (cross-grid filling), and mixed region (dot filling), respectively. The 85% baseline marks the high-performance identification threshold. The method of this invention reaches or exceeds this baseline for all frost damage types, demonstrating the stability and reliability of the system.

[0055] See Figure 3 To address the reflective properties of substances susceptible to freezing damage, a classification algorithm is used to initially analyze the data, obtaining the distribution information of different substance types and determining preliminary classification results. Based on the preliminary classification results, and in conjunction with the monitoring database of track equipment, the distribution information of different substance types is compared and analyzed. If the proportion of a certain substance type exceeds a preset threshold, its relevant data is deeply labeled, resulting in a labeled dataset. Using this labeled dataset, to meet the needs of potential risk assessment, the data is stratified and filtered to obtain the distribution information of high-risk substances and determine their impact range on track safety. Based on the distribution information of high-risk substances, combined with dynamically changing monitoring data, the status of equipment within the impact range is updated in real time, obtaining updated equipment status records. Using these updated equipment status records, to meet the needs of trend judgment, equipment operating parameters are continuously tracked. If parameter fluctuations exceed a preset range, they are prioritized to determine the monitoring focus of critical equipment. Based on the monitoring focus of critical equipment, and combined with iterative optimization processing logic, historical data in the monitoring database is retrospectively compared to obtain the distribution characteristics of abnormal patterns and determine the evolution trend of potential risks.

[0056] Specifically, the system first uses a support vector machine (SVM) classification algorithm to automatically analyze a subset of reflectance characteristics of frost-damaged materials, based on safety decision-making criteria. Assuming the reflectance data includes three types: ice, snow, and mixed frozen materials, the initial dataset contains 1000 sample points. The reflectance range for ice is set to 0.6 to 0.8, for snow 0.3 to 0.5, and for mixed frozen materials 0.4 to 0.7. After algorithm training, the classification accuracy reaches 0.85, yielding a type distribution of 40% ice, 35% snow, and 25% mixed frozen materials. Subsequently, the system automatically imports this distribution data into the track equipment monitoring database and compares it with historical frost damage records. Assuming the frost damage type distribution in the database over the past 30 days is 45% ice, 30% snow, and 25% mixed frozen materials, the Euclidean distance between the current data and historical data is calculated to be 0.05, indicating minimal change in distribution, but the decreasing trend in the proportion of ice requires attention. Next, based on the updated database, the system uses a time series analysis algorithm to predict the dynamic trend of potential risks. The prediction period is set to the next 7 days. The calculated ice layer freezing damage risk index rises from the current 0.42 to 0.48, exceeding the safe range of 0.45. The system automatically marks this trend as requiring key monitoring and generates a risk warning log. The log records that the risk rise rate is approximately 0.01 / day. At the same time, it correlates with the aging data of the track equipment. Assuming that the average service life of the equipment is 5 years and the current aging coefficient is 0.3, the comprehensive analysis shows that the equipment's freeze resistance has decreased by 10%, further deducing that the probability of risk aggravation is 0.6. Finally, a complete dynamic risk assessment report is generated for subsequent business modules to use.

[0057] like Figure 4 As shown, the present invention also provides a railway frost damage monitoring system based on AI vision technology, which mainly includes: The preliminary scanning module is used to acquire image data of the orbital surface through a spectral imaging device, perform a preliminary scan based on the differences in reflectance characteristics between water ice and salt ice types, obtain the original spectral reflectance curves including environmental interference factors, and obtain a preliminary reflectance characteristic dataset. The feature extraction module is used to separate the interference signals caused by ambient light and ice thickness using spectral analysis methods based on the preliminary reflectance characteristic dataset, extract the specific reflectance peaks of pure substances, and determine the core feature vectors of water ice type and salt ice type. The type determination module is used to determine that if the reflection peak value in the core feature vector exceeds a preset threshold, it is a salt ice type dominant region. The core feature vector is then subjected to cluster analysis through a hierarchical processing system to obtain a distribution mapping map after type differentiation. The risk assessment module is used to train the distribution map using a machine learning model, integrate reflection characteristics and environmental disturbance data, determine the severity of mixed frost damage areas, and obtain risk assessment indicators. The decision generation module is used to compare the risk assessment indicators with the track safety standards. If the indicators exceed the preset safety threshold, the feature processing module is activated to recalibrate the feature extraction process, determine the optimized hierarchical processing parameters, generate a real-time frost damage identification report based on the optimized hierarchical processing parameters, and integrate relevant processing paths to obtain the final safety decision basis. The optimization and update module is used to analyze and process subsets of reflective characteristics of different frost-damaging substances using a classification algorithm based on the safety decision criteria, obtain accurate type proportion distribution data, update the track equipment monitoring database through the type proportion distribution data, realize continuous iterative optimization of frost damage identification, and judge the dynamic change trend of potential risk assessment.

[0058] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for monitoring railway frost damage based on AI vision technology, characterized in that, The method includes: The orbital surface image data was acquired by a spectral imaging device. A preliminary scan was performed to identify the differences in reflectance characteristics between water ice and salt ice types. The original spectral reflectance curves, which included environmental interference factors, were obtained to generate a preliminary reflectance characteristic dataset. Based on the preliminary reflection characteristic dataset, the interference signals caused by ambient light and ice thickness are separated by spectral analysis. Specific reflection peaks of pure substances are extracted from these signals to determine the core feature vectors of water ice type and salt ice type. If the reflection peak in the core feature vector exceeds a preset threshold, it is determined to be a salt ice type dominant region. The core feature vector is then subjected to cluster analysis through a hierarchical processing system to obtain a distribution mapping map after type differentiation. A machine learning model is used to train the distribution map using deep learning, and by fusing reflection characteristics and environmental disturbance data, the severity of mixed frost damage areas is determined, and risk assessment indicators are obtained. By comparing the risk assessment indicators with the track safety standards, if the indicators exceed the preset safety threshold, the feature processing module is activated to recalibrate the feature extraction process, determine the optimized hierarchical processing parameters, generate a real-time frost damage identification report based on the optimized hierarchical processing parameters, and integrate relevant processing paths to obtain the final safety decision basis. Based on the aforementioned safety decision-making criteria, a classification algorithm is used to analyze and process subsets of reflective properties of different frost-damaging substances to obtain accurate type proportion distribution data. The track equipment monitoring database is updated using the type proportion distribution data to achieve continuous iterative optimization of frost damage identification and to determine the dynamic trend of potential risk assessment.

2. The railway frost damage monitoring method based on AI vision technology according to claim 1, characterized in that, The process involves acquiring orbital surface image data using a spectral imaging device, performing a preliminary scan to assess the differences in reflectance characteristics between water ice and salt ice types, obtaining raw spectral reflectance curves that include environmental interference factors, and generating a preliminary reflectance characteristic dataset, including: Data was collected from the orbital surface using spectral imaging technology. Based on the differences in reflectance between water ice and salt-deposited ice, the original curves containing environmental interference were obtained, forming a preliminary set of reflectance characteristic data. The SciPy library in Python is used to perform wavelet denoising on the original curves in the preliminary reflection characteristic dataset. The signal is obtained from the original curve, the signal is decomposed by applying wavelet transform, the noise coefficient is filtered out by setting a threshold, and the signal is reconstructed to obtain the cleaned reflection curve data. Based on the cleaned reflection curve data, and considering the differences in reflection between water ice and salt-deposited ice, the LIBSVM tool was used for support vector machine classification. Feature vectors were extracted from the cleaned reflection curve data, and the support vector machine model was trained to determine the distribution range of the reflection characteristics of the two types of materials. By measuring the reflection intensity distribution range of the two types of materials in different wavebands, the spatial distribution characteristics on the orbital surface can be determined. Based on spatial distribution characteristics, the final distribution ratio of water ice and salt-deposited ice on the orbital surface is obtained, and the relative content difference between the two types of substances is determined.

3. The railway frost damage monitoring method based on AI vision technology according to claim 1, characterized in that, Based on the preliminary reflectance characteristic dataset, the interference signals caused by ambient light and ice thickness are separated using spectral analysis methods. Specific reflectance peaks of pure substances are extracted from these signals to determine the core feature vectors of water ice and salt ice types, including: For the preliminary reflection characteristic dataset, the signals in the dataset are decomposed and processed using spectral analysis technology to separate the interference signals caused by ambient light and ice thickness, and the preliminary cleaned signal data is obtained. Based on the signal data after preliminary cleaning, filtering tools are used to further suppress interference signals, extract reflection peak signals related to the unique properties of the material, and determine the main set of reflection peaks; By analyzing the main set of reflection peaks, comparing the reflection peak signals of water ice and salt ice, obtaining the unique signal differences of the two types of substances in different bands, and determining the core characteristics of the two types of substances; Based on the core characteristics of the two types of substances, the support vector machine algorithm is used to classify the feature data, construct the classification boundary between water ice type and salt ice type, and obtain the set of classified feature vectors. By using the classified feature vector set, signal recombination is performed on the feature vector of each type of substance to obtain the complete signal pattern corresponding to the unique properties of the substance and determine the signal identifier of the two types of substances. Based on the signal identifiers of the two types of substances, the signals in the original dataset are matched one by one. If the matching degree between the signal pattern and the identifier exceeds the preset threshold, it is classified into the corresponding substance type, and the final substance classification result is obtained.

4. The railway frost damage monitoring method based on AI vision technology according to claim 1, characterized in that, If the reflection peak value in the core feature vector exceeds a preset threshold, it is determined to be a salt ice type dominant region. A hierarchical processing system is then used to perform cluster analysis on the core feature vector to obtain a distribution map after type differentiation, including: If the reflection peak in the core feature vector exceeds a preset threshold, it is determined to be a salt ice type dominant region. The feature vector is preliminarily screened through signal intensity analysis to obtain a preliminary set of regions. Based on the initially classified set of regions, a hierarchical processing method is used to group the feature vectors within each region to obtain a set of grouped signal clusters; By comparing and analyzing the reflection peak and signal intensity within each signal cluster after grouping, the material type identifier corresponding to each cluster is determined. Based on the material type identifier, clustering techniques are used to further integrate the signal cluster set to obtain signal groups after type differentiation; After grouping the signals by type, spatial mapping processing is performed on the regional data within each group to obtain the corresponding distribution mapping map. Based on the distribution map, signal strength is checked against the boundary between the salt ice type dominant region and other regions to determine the final regional division result. Based on the final regional division results, the core features and reflection peaks of each region are recorded and stored to obtain a complete type distribution archive.

5. The railway frost damage monitoring method based on AI vision technology according to claim 1, characterized in that, The distribution map is trained using a machine learning model with deep learning, and reflectivity and environmental disturbance data are fused to determine the severity of mixed frost damage areas, resulting in risk assessment indicators, including: The distribution map is processed by a machine learning model, and the reflection data and environmental interference information are analyzed by deep training methods to obtain preliminary results of frost damage area division. Based on the preliminary frost damage area delineation results, information processing was performed on the reflectance data of each area, and the severity level of each area was determined by combining the impact of environmental disturbances. By classifying the severity within the region, and using data integration technology to comprehensively analyze the reflection data and environmental interference information of each region, an intermediate value for risk assessment is obtained. Based on the median value of the risk assessment, if the median value of a certain area exceeds the preset threshold, the reflected data will be processed for secondary information processing to determine the frost damage risk level of that area. Based on the frost damage risk level, we will focus on analyzing environmental disturbance information in high-risk areas to obtain the corresponding set of assessment indicators. Based on the set of assessment indicators, the severity and risk level of each region are correlated and matched to determine the final regional risk distribution map; By using the final regional risk distribution map, data from high-risk areas are combined and stored to obtain a complete frost damage analysis archive.

6. The railway frost damage monitoring method based on AI vision technology according to claim 1, characterized in that, The risk assessment indicators are compared with track safety standards. If the indicators exceed a preset safety threshold, the feature processing module is activated to recalibrate the feature extraction process, determine optimized hierarchical processing parameters, generate a real-time frost damage identification report based on the optimized hierarchical processing parameters, and integrate relevant processing paths to obtain the final safety decision basis, including: By comparing the risk assessment indicators with the track safety standards, if the assessment result exceeds the preset safety threshold, the feature calibration module is activated to perform preliminary screening of the original data and obtain the adjusted feature set. Based on the adjusted feature set, the data is processed in layers using a feature extraction process. Combined with the optimization and adjustment of the layering parameters, the data groups after layering are determined. After the data is grouped into layers, in-depth analysis is performed on the information within each group. If the data fluctuation in a certain group exceeds the preset range, it is calibrated a second time to obtain calibrated data units. Based on the calibrated data units and combined with the business logic of frost damage identification, the information in each unit is classified and labeled to obtain a set of classified frost damage labels. Using the classified set of frost damage markers, a support vector machine algorithm is used to perform pattern matching on the markers to determine the distribution of potential high-risk areas. Based on the distribution of high-risk areas, the data is integrated and mapped according to the processing path to generate real-time profiles and determine the basis for the final security decision.

7. The railway frost damage monitoring method based on AI vision technology according to claim 1, characterized in that, Regarding the aforementioned safety decision-making criteria, a classification algorithm is used to analyze and process subsets of reflectivity characteristics of different frost-damaging substances to obtain accurate type proportion distribution data. This type proportion distribution data is then used to update the track equipment monitoring database, enabling continuous iterative optimization of frost damage identification and determining the dynamic trends of potential risk assessments. This includes: Based on the reflective properties of substances susceptible to frost damage, a classification algorithm is used to perform preliminary data analysis, obtain information on the distribution of different types of substances, and determine the preliminary classification results. Based on the preliminary classification results, and combined with the monitoring database of track equipment, the distribution information of the proportion of types is compared and analyzed. If the proportion of a certain type of material is found to exceed the preset threshold, its relevant data is deeply labeled to obtain the labeled data set. By using the labeled dataset, the data is stratified and filtered to meet the needs of potential risk assessment, obtain the distribution information of high-risk substances, and determine the scope of their impact on orbital safety. Based on the distribution information of high-risk substances and combined with dynamically changing monitoring data, the status of equipment within the affected area is updated in real time, and the updated equipment status records are obtained. By updating the equipment status records, and in response to the need for trend judgment, the equipment operating parameters are continuously tracked. If the parameter fluctuations exceed the preset range, they are prioritized to determine the key monitoring focus of critical equipment. Based on the monitoring priorities of key equipment and combined with iterative optimization processing logic, historical data in the monitoring database is backtracked and compared to obtain the distribution characteristics of abnormal patterns and determine the evolution trend of potential risks.

8. A railway frost damage monitoring system based on AI vision technology, characterized in that, The system includes: The preliminary scanning module is used to acquire image data of the orbital surface through a spectral imaging device, perform a preliminary scan based on the differences in reflectance characteristics between water ice and salt ice types, obtain the original spectral reflectance curves including environmental interference factors, and obtain a preliminary reflectance characteristic dataset. The feature extraction module is used to separate the interference signals caused by ambient light and ice thickness using spectral analysis methods based on the preliminary reflectance characteristic dataset, extract the specific reflectance peaks of pure substances, and determine the core feature vectors of water ice type and salt ice type. The type determination module is used to determine that if the reflection peak value in the core feature vector exceeds a preset threshold, it is a salt ice type dominant region. The core feature vector is then subjected to cluster analysis through a hierarchical processing system to obtain a distribution mapping map after type differentiation. The risk assessment module is used to train the distribution map using a machine learning model, integrate reflection characteristics and environmental disturbance data, determine the severity of mixed frost damage areas, and obtain risk assessment indicators. The decision generation module is used to compare the risk assessment indicators with the track safety standards. If the indicators exceed the preset safety threshold, the feature processing module is activated to recalibrate the feature extraction process, determine the optimized hierarchical processing parameters, generate a real-time frost damage identification report based on the optimized hierarchical processing parameters, and integrate relevant processing paths to obtain the final safety decision basis. The optimization and update module is used to analyze and process subsets of reflective characteristics of different frost-damaging substances using a classification algorithm based on the safety decision criteria, obtain accurate type proportion distribution data, update the track equipment monitoring database through the type proportion distribution data, realize continuous iterative optimization of frost damage identification, and judge the dynamic change trend of potential risk assessment.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images