Road surface icing temperature special point monitoring and collecting method and system based on thermal spectrum map and XGBoost

Abstract

The invention discloses a method and a system for monitoring and collecting icing-prone temperature special points of a road surface based on a thermal spectrum map and XGBoost. At present, position data of easy freezing of a road surface cannot be fed back in time according to temperature changes, special jump points and hysteresis exist, and an existing monitoring method cannot be matched with dynamic changes of the freezing state of the road surface. According to the road surface icing temperature special point monitoring and collecting method, a three-dimensional data set combining weather, a road surface and a position is obtained to serve as multi-source data collection, and an XGBoost temperature prediction model combining multi-source data and quantitative weather factors is established; based on an XGBoost temperature prediction model, an improved Z-score algorithm is combined with thermal spectrum map spatial characteristics to determine accurate positioning, monitoring and real-time early warning processes of road surface freezing-prone temperature special points; according to the road surface icing temperature special point monitoring and collecting method based on the thermal spectrum map and the XGBoost, the quantitative accurate process capable of standardizing the system to determine the road surface icing temperature special point is provided.

Description

Technical Field

[0001] Specifically, this invention relates to a method and system for monitoring and collecting data on special temperature points on roads prone to icing, based on thermal spectral maps and XGBoost. Background Technology

[0002] Special points on highway surfaces prone to icing in winter are important factors causing traffic accidents and road surface damage. For example, special sections of highways such as bridge-roadbed transition sections, tunnel entrances and exits, and sharp bends in shaded areas are more likely to have their surface temperatures drop below freezing point due to structural or environmental factors, making them prone to icing. The regular changes in road surface temperature at specific locations caused by these geographical conditions are a crucial indicator of whether icing will occur at that location. When the road surface temperature is below 0°C and there is snow accumulation, the snow will gradually freeze from a loose state into a hard ice layer. At this time, the tire friction coefficient will drop from 0.7-0.8 on a normal dry road surface to 0.3-0.4. Furthermore, for low-temperature points, the lower the temperature compared to 0°C, the more severe the icing and the lower the friction coefficient. If the temperature drops further below -5°C, semi-melted ice will completely harden into a dense ice layer, with a friction coefficient of only 0.1-0.2. This increases vehicle braking distance by 3-5 times, and the traffic accident rate soars to 14.3 times that of non-snowy days, while also exacerbating maintenance costs. Therefore, detecting specific road surface temperatures prone to icing can indirectly analyze the impact of the unique road surface structure on the probability of icing in winter at that location. More efficient and accurate early identification of low-temperature road surface points can ensure early prediction of road icing conditions, assisting road maintenance managers in developing appropriate maintenance strategies.

[0003] The unique temperature characteristics that make existing road surfaces prone to icing are mainly found in various road structures, particularly bridge structures. Bridge decks, in particular, are suspended in the air, with their top, bottom, and sides directly exposed to the air. This means that the surface area of ​​a bridge that loses heat through convection is much larger than that of the road surface in contact with the ground. During winter nights, when ambient temperatures drop, bridges can more quickly and effectively transfer heat to the surrounding cold air, causing the bridge deck temperature to drop rapidly. In contrast, ordinary road surfaces are in close contact with the soil or roadbed below, and the soil, with its heat capacity, acts as insulation, slowing the rate of temperature drop. Furthermore, mountain roads may be affected by the shadow of mountains, resulting in insufficient sunlight, especially in winter when short daylight hours reduce heat absorption, making them cooler more easily at night. Additionally, valley topography can lead to the accumulation of cold air, forming localized low-temperature zones where the road surface temperature is lower than that of surrounding unaffected roads. Furthermore, if there are poorly drained, damp areas, underground cavities, or structures such as ramps or box girder structures of elevated roads that are not in direct contact with the ground beneath the road surface, it can obstruct heat conduction between the road surface and the foundation, or increase convective heat loss beneath the road surface, potentially leading to lower local road surface temperatures. Currently, the selection of snow removal points is mainly based on historical experience and the subjective judgment of road maintenance departments, making it difficult to accurately define the distribution characteristics of specific icing-prone points on particular road sections through effective means.

[0004] In summary, the current methods for controlling highway icing in winter have significant technical shortcomings, mainly in the following two aspects: First, macro-level icing warnings rely excessively on data from regional meteorological stations for forecasting. However, meteorological stations are generally spaced 5-10km apart, making it impossible to accurately capture the micro-environmental differences at specific points on the road surface prone to icing, such as bridge expansion joints and shaded bends. Furthermore, they can only reflect the overall atmospheric condition and cannot distinguish local temperature fluctuations caused by factors such as road structure and terrain undulations. This leads to misjudgments in warnings, such as macro-level absence of ice but localized presence of ice, which fails to meet the needs of refined management. Second: The core technical deficiency of current methods for controlling highway icing in winter lies in the fact that the location of specific icing-prone temperature points on the road surface still relies excessively on manual inspection and experience-based judgment. Maintenance personnel need to drive patrol vehicles to inspect each section of the route, which is not only inefficient but also poses safety risks to patrol personnel in severe weather such as blizzards and freezing fog. Furthermore, different personnel's judgment standards for critical states such as thin ice and frost freezing are influenced by experience and can lead to deviations. At the same time, there is a 3 to 4-hour lag in data recording and summarization, which cannot match the dynamic changes in the road surface icing state. This often causes de-icing operations to miss the best time, resulting in a waste of maintenance resources and difficulty in ensuring road traffic safety. There is a lack of standardized, accurate, dynamic, and timely data collection methods. Summary of the Invention

[0005] This invention provides a method and system for monitoring and collecting temperature-specific points on roads prone to icing based on thermal spectral maps and XGBoost, in order to solve the above-mentioned problems.

[0006] A method for monitoring and collecting temperature-specific points on roads prone to icing based on thermal spectral maps and XGBoost is characterized by: acquiring a three-dimensional dataset combining meteorological, road surface, and location data as multi-source data collection; establishing an XGBoost temperature prediction model that combines multi-source data with quantified meteorological factors; and using an improved Z-score algorithm combined with spatial features of the thermal spectral map to determine the precise location, monitoring, and real-time early warning process of temperature-specific points on roads prone to icing.

[0007] As a preferred approach, the multi-source data acquisition process involves deploying fixed weather stations in multiple areas prone to temperature anomalies. These stations collect air temperature, relative humidity, and road surface temperature data as the three primary sources of raw data. Air temperature data is measured in the range of -40 to +60℃ with an accuracy of ±0.1℃; relative humidity data is measured in the range of 0 to 100%RH with an accuracy of ±2%; and road surface temperature data is measured in the range of -40 to +60℃ with an accuracy of ±0.1℃. The collected raw data from these three sources is then preprocessed to construct a three-dimensional dataset that combines meteorological, road surface, and location data.

[0008] As a preferred approach, the preprocessing of the three sources of data includes outlier removal, missing value imputation, reliability verification, and continuous processing of data alignment and structuring. Outlier removal is adopted The criterion processes raw data from three sources. When the deviation of a parameter value from the mean of the data for the time period containing that parameter value exceeds [a certain threshold], [the criterion applies]. At that time, the median of five consecutive adjacent samples is used to replace the error caused by sensor drift, thereby completing the outlier removal process. When the air temperature data σ=3.2℃, the air temperature data values ​​less than -20℃ or greater than 4℃ are judged as outlier data and removed. Missing value imputation uses signal interruption data with a missing rate of less than or equal to 5% as missing data. The missing data is classified and processed as follows: data with a duration of less than 30 seconds is classified as short-term missing data and is imputed using linear interpolation; data with a duration of more than or equal to 30 seconds is classified as long-term missing data and is imputed using the KNN algorithm to ensure the integrity of the three-dimensional dataset. Reliability verification involves comparing air temperature, relative humidity, and road surface temperature data collected by mobile composite sensors with data from fixed weather stations at the same time and location. Data with average errors of less than or equal to 0.7℃ for air temperature, less than or equal to 6% for relative humidity, and less than or equal to 0.5℃ for road surface temperature are used as the data to be used. Data alignment and structuring: Based on timestamps and latitude and longitude coordinates, the data to be used is classified into mobile thermal spectrum data and fixed meteorological data. The mobile thermal spectrum data and fixed meteorological data are correlated and aligned to form a structured dataset with seven input features, including air temperature, relative humidity, road surface temperature, water and snow thickness, latitude and longitude, sampling time, and road segment type.

[0009] As a preferred approach: the structured dataset, formed through continuous processing of outlier removal, missing value imputation, reliability verification, and data alignment and structuring, is sequentially divided, optimized, and verified to form the XGBoost model. The process of establishing the XGBoost temperature prediction model is as follows: The process of dividing the structured dataset involves using air temperature and relative humidity collected by a fixed weather station as input features and road surface temperature as output features to construct a dataset with a total sample size of 100,000 or more. The dataset is then divided into a training set and a test set in an 8:2 ratio. The split training and test sets are subjected to hyperparameter optimization to configure a parameter grid with 120-180 boosting trees, a maximum tree depth of 2-4, and a learning rate of 0.05-0.15. The XGBoost temperature prediction model is formed by grid search with the goal of minimizing the mean absolute error (MAE). The optimized XGBoost temperature prediction model was validated using five-fold cross-validation to evaluate its performance. The training set determination coefficient R² was greater than or equal to 0.97, and the test set MAE was less than or equal to 0.8℃, ensuring that the XGBoost temperature prediction model has the ability to accurately quantify the impact of meteorological factors on road surface temperature.

[0010] As a preferred approach, the calculation steps for determining specific points on the road surface prone to icing using the XGBoost temperature prediction model and the improved Z-score algorithm combined with spatial features of thermal spectral maps are as follows: Step 1: For each data point in the heat map P i A neighborhood is constructed with a radius of 0.1°. N i As the target neighborhood, count the number of valid data points within the target neighborhood. M Ensure that the value of M is within the range of 5 to 100; Step 2: Calculate the average temperature within the target neighborhood. The calculation formula is as follows: ; In the above formula, The average temperature within the target neighborhood. The number of data collection points in the neighborhood. Let i be the set of all data collection points within a radius of 1.5 km around point i. The road surface temperature at each sampling point within the neighborhood; Step 3: Input the air temperature and relative humidity data collected by the mobile thermal spectrometer into the XGBoost temperature prediction model, and output the predicted road surface temperature. T i ; Step 4: Use the difference between the actual temperature and the predicted temperature at each data collection point as a characteristic to determine the special points of road surface prone to icing. The formula for calculating the characteristics of special points of road surface prone to icing is as follows: ; In the above formula, Let i be the neighborhood temperature difference at the point to be measured. The predicted pavement temperature at point i using the XGBoost temperature prediction model. The average temperature within the target neighborhood; Step 5: Calculate the average temperature difference at specific points on the road surface prone to icing. The calculation formula is as follows: ; In the above formula, This represents the average temperature deviation within the neighborhood. The neighborhood temperature deviation at the point i to be measured; Step Six: Calculate the standard deviation of the point temperature difference at specific points on the road surface prone to icing. The calculation formula is: ; In the above formula, Let i be the standard deviation of the temperature deviation at the point to be measured. Let i be the neighborhood temperature deviation at the point to be measured. This represents the average temperature deviation within the neighborhood. Step 7: Calculate the Z-score of the temperature difference at the monitoring point using the Z-score method: .

[0011] In the above formula, The Z-score is the temperature deviation at monitoring point i. This represents the average temperature deviation within the neighborhood. Let be the standard deviation of the temperature deviation at point i to be measured; Data points with |Zi| greater than 2 are selected as special points of road surface temperature prone to icing. Among these special points, those with Zi less than -2 are identified as special points prone to icing, and those with Zi greater than 2 are identified as special points of heat. Finally, adjacent special points of road surface temperature prone to icing are integrated into a special point band with a length greater than or equal to 1 kilometer. This completes the calculation process of determining special points of road surface temperature prone to icing based on the XGBoost temperature prediction model, using the improved Z-score algorithm combined with the spatial features of the thermal spectrum map.

[0012] As a preferred option, the calculation process for screening special temperature points on the road surface prone to icing is as follows: the standardization coefficient is calculated using the following formula:

[0013] In the above formula: The original temperature of point i is collected by a mobile road condition sensor. Let i be the set of all data collection points within a radius of 1.5 km around point i. Let N(i) be the number of points. This represents the average temperature difference within the neighborhood. The standard deviation of the neighborhood temperature difference This represents the temperature value of the k-th local region. This represents the temperature value of the l-th local region adjacent to the k-th local region.

[0014] As a preferred approach: Based on the XGBoost temperature prediction model, after determining the specific temperature points on the road surface prone to icing using the improved Z-score algorithm combined with the spatial features of the thermal spectrum map, the XGBoost temperature prediction model is deployed and dynamically updated. The deployment and dynamic update process is as follows: The XGBoost temperature prediction model was deployed to an ARM Cortex-A53 architecture edge node and quantized and compressed using the TensorRT algorithm. The XGBoost temperature prediction model receives real-time data every 5 minutes, automatically triggers the identification process, and outputs the latitude and longitude of special points and their risk levels. The recognition accuracy is calculated every 24 hours. If the accuracy is less than 80% for three consecutive days, an incremental update is initiated. The model weights are updated after 5,000 new samples are added. The update time is less than or equal to 30 minutes. After deployment and dynamic updating through the XGBoost temperature prediction model, the continuous process of accurate location, monitoring and real-time early warning of special temperature points on the road surface prone to icing is completed.

[0015] A system for monitoring and acquiring temperature-specific points on roads prone to icing, based on thermal spectral mapping and XGBoost, is disclosed. This system enables the monitoring and acquisition of temperature-specific points on roads prone to icing, utilizing thermal spectral mapping and XGBoost. The system comprises multiple fixed weather stations, multiple road surface sensors, multiple mobile composite sensors, and an online platform for acquiring road icing and snow detection data. The fixed weather stations are distributed along the sides of the road surface, the road surface sensors are arranged on the road surface, and the mobile composite sensors are mounted on vehicles. Each vehicle carries at least one mobile composite sensor, which includes a connector, a temperature sensor, a humidity sensor, and a road surface condition sensor. The connector is detachably connected to the vehicle. The temperature sensor, humidity sensor, and road surface condition sensor are respectively mounted on the connector, with the probes of these sensors all facing the road surface. The fixed weather stations, road surface sensors, and mobile composite sensors are electrically connected to the online platform for acquiring road icing and snow detection data.

[0016] Compared with existing technologies, this invention provides a method for monitoring and collecting data on specific temperature points on roads prone to icing based on thermal spectral maps and XGBoost, which has the following beneficial effects: This invention proposes a method for monitoring and collecting temperature anomalies on roads prone to icing based on thermal spectral maps and XGBoost. This method standardizes the quantitative and accurate process for determining these anomalies. By proposing a meteorological interference removal mechanism that combines thermal spectral maps and XGBoost, it balances anomaly identification accuracy with environmental adaptability. A dual-track model—a macro-meteorological modeling and micro-thermal spectral analysis model—is constructed, namely the XGBoost temperature prediction model. The XGBoost algorithm quantifies the impact of ambient temperature and relative humidity on road surface temperature, calculating the temperature difference between the predicted temperature and the measured temperature to remove atmospheric interference. Combined with the spatial continuity characteristics of thermal spectral maps, it accurately captures local temperature anomalies at bridges, shaded curves, and other special points. Experimental verification demonstrates that this method achieves a 92% accuracy rate in identifying temperature anomalies on roads prone to icing, a 37% improvement over traditional temperature threshold methods, while reducing the false alarm rate to below 8%.

[0017] The method for monitoring and collecting temperature-specific points prone to icing on road surfaces based on thermal spectral maps and XGBoost in this invention employs an improved Z-score algorithm based on neighborhood moving average. This algorithm automatically identifies temperature-specific points prone to icing on road surfaces within the thermal spectral map. The calculation and filtering process of this invention is complete and accurate. First, a neighborhood with a radius of 0.1° is defined centered on each road surface temperature monitoring point, and the average temperature of all points within the neighborhood is calculated. Then, the temperature difference between the monitoring point and the average temperature of the neighborhood is calculated, and a standardized Z-value is calculated based on the neighborhood temperature fluctuation. Next, local low-temperature specific points are filtered out using a Z-value less than -2 as a threshold. Finally, adjacent specific points are integrated into a specific point band with a length greater than or equal to 1 kilometer. This completes the process of identifying regional temperature anomalies such as entire bridge sections and continuous shaded areas, making the judgment process more complete and accurate, and avoiding missed detections. Experiments show that this invention can reduce the anomaly omission rate in bridge sections and other related areas from 35% to 8%, significantly improving the complete identification capability of anomalies in complex road sections compared to the isolated forest single-point detection algorithm.

[0018] The road surface icing temperature special point monitoring and acquisition system based on thermal spectral map and XGBoost in this invention is a hardware system that combines road surface and control terminal. The structure is reasonable and simple. The arrangement of multiple fixed meteorological stations, multiple road surface sensors, multiple mobile composite sensors and road surface ice and snow detection data online acquisition platform in the road surface icing temperature special point monitoring and acquisition system based on thermal spectral map and XGBoost can be determined according to specific requirements, and the arrangement is diverse and flexible. Attached Figure Description

[0019] Figure 1 This is an on-site installation view of the fixed weather station and road surface sensor in the road surface icing temperature special point monitoring and acquisition system based on thermal spectral map and XGBoost. Figure 2 A schematic diagram of a sample of a mobile composite sensor; Figure 3 This is a schematic diagram of an online data collection platform for monitoring road ice and snow. Figure 4 This diagram shows the installation location of the mobile composite sensor on the vehicle. Figure 5 A distribution view for identifying special points of road surface icing temperature that integrate Xgboost and improved Z-Score methods; Figure 6 This is a histogram showing the distribution of temperature deviations at specific points. Figure 7 This is a scatter plot of the temperature deviation distribution at specific points. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Specific implementation method one: Combining Figures 1 to 7 This implementation method describes a method for monitoring and collecting data on temperature-sensitive points prone to icing on road surfaces based on thermal spectral maps and XGBoost. It involves acquiring a composite three-dimensional dataset of meteorological, road surface, and location data as multi-source data. An XGBoost temperature prediction model is then established, combining multi-source data with quantified meteorological factors. Based on this model, an improved Z-score algorithm combined with spatial features of the thermal spectral map is used to accurately locate, monitor, and provide real-time early warnings for temperature-sensitive points prone to icing on road surfaces. The XGBoost temperature prediction model is established through a multi-source data acquisition system that integrates multi-source data fusion and intelligent algorithms. Specifically, the XGBoost temperature prediction model can continuously process the accurate location, monitoring, and real-time early warning of temperature-sensitive points prone to icing on road surfaces.

[0022] This method utilizes a road surface icing-prone temperature-specific point monitoring and acquisition system based on thermal spectral maps and XGBoost to acquire relevant data. This system includes multiple fixed weather stations, multiple road surface sensors, multiple mobile composite sensors, and an online platform for collecting road surface ice and snow detection data. The fixed weather stations are distributed along the sides of the road surface, the road surface sensors are positioned on the road surface, and the mobile composite sensors are mounted on vehicles. Each vehicle is equipped with at least one mobile composite sensor, which includes a connector and a temperature sensor. Sensor 1, humidity sensor 2, and road condition sensor 3 are detachably connected to the vehicle via a connector. The temperature sensor 1, humidity sensor 2, and road condition sensor 3 are respectively mounted on the connector. The probes of the temperature sensor 1, humidity sensor 2, and road condition sensor 3 are all oriented towards the road surface. Multiple fixed weather stations, multiple road surface sensors, and multiple mobile composite sensors are electrically connected to the online data acquisition platform for road ice and snow detection. The temperature sensor 1, humidity sensor 2, and road condition sensor 3 are all existing sensors, and their working principles are consistent with those of existing temperature sensors, humidity sensors, and road condition sensors.

[0023] The structure and working principle of the fixed weather station in this embodiment are consistent with the structure and working principle of the existing weather stations used on the side of the road.

[0024] Specific Implementation Method Two: This implementation method is a further limitation of Specific Implementation Method One. The specific steps of the method for monitoring and collecting special temperature points on roads prone to icing in this implementation method are as follows: before multi-source data fusion, a multi-source data collection process is carried out. Fixed weather stations are set up in multiple areas prone to temperature anomalies. Air temperature data with a range of -40~+60℃ and an accuracy of ±0.1℃, relative humidity data with a range of 0~100%RH and an accuracy of ±2%, and road surface temperature data with a range of -40~+60℃ and an accuracy of ±0.1℃ are collected from the fixed weather stations as the three-source raw data. The three-source raw data are preprocessed to construct a three-dimensional dataset that combines meteorological, road surface, and location data.

[0025] In this embodiment, the fixed meteorological station collection area is mainly concentrated at key locations on highways, namely, fixed meteorological stations are set up in areas prone to temperature anomalies, such as bridges, roadbed transition sections, and the periphery of towns. The continuous monitoring period for collecting parameters including air temperature, relative humidity, and road surface temperature data is no less than one winter, covering typical working conditions such as low temperatures and snow in winter. This is used to provide meteorological and road surface temperature benchmark data to support the subsequent reliability verification of mobile data.

[0026] The road surface icing temperature special point monitoring and acquisition system based on thermal spectral maps and XGBoost also includes a mobile thermal spectral device. This device uses a mobile multi-parameter road meteorological sensor as the core acquisition device. The device is mounted on winter maintenance vehicles such as snowplows and patrol vehicles, and the installation position is selected below the front of the vehicle, 15-20cm from the road surface. The acquired parameters include longitude and latitude, spatial parameters with an accuracy of ±0.5m, road condition parameters, and auxiliary meteorological parameters. The acquisition time is concentrated between 1:00 and 3:00 am, covering different weather conditions such as sunny, light snow, and cloudy. The maintenance vehicle speed is controlled at 80km / h. The acquisition frequency is 1 time / second, and the data is synchronously stored on a local SD card and in the cloud.

[0027] In this embodiment, the road surface condition parameters are: actual road surface temperature -40~+60℃, accuracy ±0.1℃; water layer thickness 0~5mm, ice layer thickness 0~2mm, snow layer water equivalent 0~1mm, all with accuracy ±0.1mm; and wet skid coefficient 0.09~0.82, accuracy ±0.01.

[0028] Specific Implementation Method 3: This implementation method is a further limitation of Specific Implementation Method 1. In this implementation method, the preprocessing process of the three source data includes outlier removal, missing value imputation, reliability verification, and continuous processing of data alignment and structuring. Outlier removal is adopted The criterion processes raw data from three sources. When the deviation of a parameter value from the mean of the data for the time period containing that parameter value exceeds [a certain threshold], [the criterion applies]. At that time, the median of five consecutive adjacent samples is used to replace the error caused by sensor drift, thereby completing the outlier removal process. When the air temperature data σ=3.2℃, the air temperature data values ​​less than -20℃ or greater than 4℃ are judged as outlier data and removed. Missing value imputation uses signal interruption data with a missing rate of less than or equal to 5% as missing data. The missing data is classified and processed as follows: data with a duration of less than 30 seconds is classified as short-term missing data and is imputed using linear interpolation; data with a duration of more than or equal to 30 seconds is classified as long-term missing data and is imputed using the KNN algorithm to ensure the integrity of the three-dimensional dataset. Reliability verification involves comparing air temperature, relative humidity, and road surface temperature data collected by mobile composite sensors with data from fixed weather stations at the same time and location. Data with average errors of less than or equal to 0.7℃ for air temperature, less than or equal to 6% for relative humidity, and less than or equal to 0.5℃ for road surface temperature are used as the initial data. The errors must comply with GB / T33695-2017 "Technical Requirements for Highway Traffic Meteorological Monitoring Facilities" to ensure the reliability and usability of the collected data. Data Alignment and Structuring: The data to be used is categorized into mobile thermal spectrum data and fixed meteorological data based on timestamps and latitude / longitude coordinates. The mobile thermal spectrum data and fixed meteorological data are then correlated and aligned to form a structured dataset containing seven input features: air temperature, relative humidity, road surface temperature, water and snow thickness, latitude / longitude, sampling time, and road segment type. The data is in CSV format for ease of subsequent model training and algorithm processing.

[0029] Specific Implementation Method Four: This implementation method is a further limitation of Specific Implementation Methods One, Two, or Three. In this implementation method, the structured dataset formed after continuous processing of outlier removal, missing value imputation, reliability verification, and data alignment and structuring is sequentially divided, optimized, and verified to form an XGBoost model. The process of establishing the XGBoost temperature prediction model is as follows: The process of dividing the structured dataset involves using air temperature and relative humidity collected by a fixed weather station as input features and road surface temperature as output features to construct a dataset with a total sample size of 100,000 or more. The dataset is then divided into a training set and a test set in an 8:2 ratio. The split training and test sets are subjected to hyperparameter optimization to configure a parameter grid with 120-180 boosting trees, a maximum tree depth of 2-4, and a learning rate of 0.05-0.15. The XGBoost temperature prediction model is formed by grid search with the goal of minimizing the mean absolute error (MAE). The optimized XGBoost temperature prediction model was validated using five-fold cross-validation to evaluate its performance. The training set determination coefficient R² was greater than or equal to 0.97, and the test set MAE was less than or equal to 0.8℃, ensuring that the XGBoost temperature prediction model has the ability to accurately quantify the impact of meteorological factors on road surface temperature.

[0030] In this implementation, the structured dataset is sequentially divided, optimized, and validated. Reliability is ensured by inputting features and performing spatiotemporal consistency verification. The verification method is that the deviation of similar parameters such as air temperature and relative humidity collected by mobile equipment and fixed weather stations at the same time and in the same area must not exceed a set threshold. If the threshold is exceeded, the equipment calibration status and collection environment are re-verified, and abnormal data is removed or corrected.

[0031] The process of dividing the structured dataset is as follows: using air temperature and relative humidity collected by fixed weather stations as input features and road surface temperature as output features, a dataset with a total sample size of more than or equal to 100,000 is constructed, and the dataset is divided into training set and test set in an 8:2 ratio. The training and test sets were split and hyperparameters were optimized. A parameter grid was configured with 120-180 boosting trees, a maximum tree depth of 2-4, and a learning rate of 0.05-0.15. The optimal parameter combination was determined by grid search with the goal of minimizing the mean absolute error (MAE) to be 150 boosting trees, a maximum tree depth of 3, and a learning rate of 0.1. The optimized dataset model was validated using five-fold cross-validation to evaluate its performance. The training set determination coefficient R² was greater than or equal to 0.97, and the test set MAE was less than or equal to 0.8℃, ensuring that the model could accurately quantify the impact of meteorological factors on road surface temperature.

[0032] Specific Implementation Method Five: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, or Four. In this implementation method, the calculation steps for determining the special points of road surface icing temperature based on the XGBoost temperature prediction model using the improved Z-score algorithm combined with the spatial features of the thermal spectrum map are as follows: Step 1: For each data point in the heat map P i A neighborhood is constructed with a radius of 0.1°. N i As the target neighborhood, count the number of valid data points within the target neighborhood. M Ensure that the value of M is within the range of 5 to 100; Step 2: Calculate the average temperature within the target neighborhood. The calculation formula is as follows: ; In the above formula, The average temperature within the target neighborhood. The number of data collection points in the neighborhood. Let i be the set of all data collection points within a radius of 1.5 km around point i. The road surface temperature at each sampling point within the neighborhood; Step 3: Input the air temperature and relative humidity data collected by the mobile thermal spectrometer into the XGBoost temperature prediction model, and output the predicted road surface temperature. T i ; Step 4: Use the difference between the actual temperature and the predicted temperature at each data collection point as a characteristic to determine the special points of road surface prone to icing. The formula for calculating the characteristics of special points of road surface prone to icing is as follows: ; In the above formula, Let i be the neighborhood temperature difference at the point to be measured. The predicted pavement temperature at point i using the XGBoost temperature prediction model. The average temperature within the target neighborhood; Step 5: Calculate the average temperature difference at specific points on the road surface prone to icing. The calculation formula is as follows: ; In the above formula, This represents the average temperature deviation within the neighborhood. The neighborhood temperature deviation at the point i to be measured; Step Six: Calculate the standard deviation of the point temperature difference at specific points on the road surface prone to icing. The calculation formula is: ; In the above formula, Let i be the standard deviation of the temperature deviation at the point to be measured. Let i be the neighborhood temperature deviation at the point to be measured. This represents the average temperature deviation within the neighborhood. Step 7: Calculate the Z-score of the temperature difference at the monitoring point using the Z-score method: .

[0033] In the above formula, The Z-score is the temperature deviation at monitoring point i. This represents the average temperature deviation within the neighborhood. Let be the standard deviation of the temperature deviation at point i to be measured; Data points with |Zi| greater than 2 are identified as special points indicating road surface icing temperatures. These special points are then further categorized. Points with values ​​greater than 0 are identified as special points prone to freezing. Points with values ​​less than 0 are identified as thermal special points. Finally, adjacent road surface temperature special points that are prone to icing are integrated into a special point band with a length greater than or equal to 1 kilometer. This completes the calculation process of determining road surface temperature special points that are prone to icing based on the XGBoost temperature prediction model, using the improved Z-score algorithm combined with the spatial features of the thermal spectrum map.

[0034] Specific Implementation Method Six: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, or Five. In this implementation method, the calculation process for screening special temperature points on the road surface prone to icing is as follows: The standardization coefficient is calculated using the following formula:

[0035] In the above formula: The original temperature of point i is collected by a mobile road condition sensor. Let i be the set of all data collection points within a radius of 1.5 km around point i. The number of midpoints in Ni This represents the average temperature difference within the neighborhood. The standard deviation of the neighborhood temperature difference This represents the temperature value of the k-th local region. This represents the temperature value of the l-th local region adjacent to the k-th local region.

[0036] Specific Implementation Method Seven: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, or Six. In this implementation method, after determining the special points of road surface icing temperature based on the XGBoost temperature prediction model using the improved Z-score algorithm combined with the spatial features of the thermal spectrum map, the XGBoost temperature prediction model is deployed and dynamically updated. The deployment and dynamic update process is as follows: The XGBoost temperature prediction model was deployed to an ARM Cortex-A53 architecture edge node and quantized and compressed using the TensorRT algorithm. The XGBoost temperature prediction model receives real-time data every 5 minutes, automatically triggers the identification process, and outputs the latitude and longitude of special points and their risk levels. The recognition accuracy is calculated every 24 hours. If the accuracy is less than 80% for three consecutive days, an incremental update is initiated. The model weights are updated after 5,000 new samples are added. The update time is less than or equal to 30 minutes. After deployment and dynamic updating through the XGBoost temperature prediction model, the continuous process of accurate location, monitoring and real-time early warning of special temperature points on the road surface prone to icing is completed.

[0037] In the specific implementation of this invention, a multi-source data acquisition system is first constructed, which coordinates fixed weather stations and mobile road condition monitoring equipment. The fixed weather station is used to acquire macro-meteorological parameters such as air temperature and relative humidity, while the mobile road condition monitoring equipment captures the spatial distribution characteristics of road surface temperature through thermal spectral mapping technology, thus compensating for the technical deficiency of insufficient spatial resolution of weather station data. Secondly, the XGBoost model is introduced to quantify the impact of meteorological factors on road surface temperature. By calculating the deviation between the predicted temperature output by the model and the measured temperature of the thermal spectrum, atmospheric environmental interference is eliminated, highlighting the local temperature anomalies caused by road surface structure and terrain undulations, providing an accurate data foundation for identifying special temperature points on the road surface that are prone to icing.

[0038] Specific Implementation Method Eight: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, Six, or Seven. In its implementation, this method involves equipment selection and deployment. Environmental sensors and road surface temperature sensors are selected as fixed devices and deployed along the target highway at intervals of 5-8 km on bridges and sharp bends. Mobile road condition sensors with integrated GPS are selected and mounted 18 cm below the front of the maintenance vehicle. Edge devices are procured, deployed at maintenance stations, and connected to 4G modules. All devices are calibrated using standard equipment before activation. The selected sensors and their sensing parameters are shown in Table 1.

[0039] Table 1. Comparison of Sensor Types and Sensing Parameters

[0040] Specific Implementation Method Nine: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, Six, Seven, or Eight. In this implementation method, during the multi-source data acquisition process, the detailed processing is as follows: fixed equipment collects ambient temperature and relative humidity once per minute and saves them as a CSV file in the format of "timestamp + parameter + device ID"; mobile composite sensors collect road surface temperature, water ice and snow thickness, and latitude and longitude once every 30 seconds and save them as a CSV file in the format of "timestamp + spatial parameter + road surface parameter". During the winter, from 1:00 AM to 3:00 AM, the maintenance vehicle travels at 40-60 km / h to collect data for the entire road section. The data is backed up locally and uploaded to the cloud via HTTPS.

[0041] Specific Implementation Method Ten: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, Six, Seven, Eight, or Nine. The specific details of the dataset partitioning and sample balancing process in this invention are as follows: the data is divided into training set: validation set: test set = 70%: 15%: 15% according to the time series, ensuring that each set covers different weather and road conditions; in order to address the problem of low proportion of non-dry samples such as wet, snow, and ice, the SMOTE algorithm is used for oversampling to make the ratio of dry to non-dry samples close to 1:1.

[0042] Specific Implementation Method Eleven: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, Six, Seven, Eight, Nine, or Ten. The basic configuration of the XGBoost temperature prediction model is to select the XGBoost regression algorithm and integrate 150 decision trees based on CART trees. The training subset of each tree is generated by using Bootstrap sampling at a ratio of 70%. When splitting a node, a feature subset is randomly selected, i.e., the number is equal to the square root of the total number of features. The loss function is set as mean squared error (MSE).

[0043] Specific Implementation Method Twelve: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, Six, Seven, Eight, Nine, Ten, or Eleven. In this implementation method, when optimizing the model parameters, the parameter range is set as follows: see Table Two, maximum tree depth 2~4, learning rate 0.05~0.15, minimum number of sample splits 1~3; the parameters are optimized by combining five-fold cross-validation with grid search, with the goal of minimizing the validation set RMSE, to determine the optimal parameter combination.

[0044] Table 2 Core Parameter Range

[0045] In the process of core feature selection, this invention calculates the "gain value" of each feature in the decision tree split and sorts them by proportion; features with a cumulative contribution of ≥90% are retained, and the final core features are road surface temperature, ambient temperature, relative humidity, temperature change in the previous hour, and water ice and snow thickness.

[0046] In the process of improving the Z-score algorithm to identify special points, this invention uses the XGBoost model to output the predicted road surface temperature and calculates the temperature difference ∆T between the predicted and measured temperatures; it constructs a neighborhood with a radius of 0.1° and calculates the average temperature difference and single-point deviation Dev of the neighborhood; it calculates the standardization coefficient according to Equation 4, and points with |Z|>2 are identified as special points and are given a graded warning according to ∆T.

[0047] In the lightweight deployment of the model, this invention uses TensorRT to quantize the XGBoost model into a 16-bit integer, compressing its size to 35MB; it is deployed on an NVIDIA Jetson Nano edge device to develop an automated inference process, receiving data and outputting special point information every 5 minutes.

[0048] In the process of dynamic model updates, this invention uses 1,000 new data points for incremental training each month, updating only the model weights; when the RMSE of the test set is greater than 1.5% for three consecutive months or the cumulative number of samples exceeds 50,000, the model is fully retrained based on the latest six months of data.

[0049] The data acquisition module of this system consists of a fixed weather station and a mobile thermal spectroscopy device, which realizes multi-source spatiotemporal data acquisition according to the above acquisition method; the model calculation module: deploys the XGBoost prediction model and the improved Z-score algorithm to realize special point identification; the edge deployment module: real-time data processing and result output; the dynamic update module: monitors the model accuracy and automatically triggers incremental updates or full retraining.

Claims

1. A method for monitoring and collecting data on specific temperature points prone to icing on road surfaces based on thermal spectral maps and XGBoost, characterized in that: The pavement icing temperature special point monitoring collection method is to obtain a three-dimensional data set combined with weather, pavement and position as multi-source data collection, establish an XGBoost temperature prediction model combining multi-source data and quantitative weather factors, and determine the accurate positioning, monitoring and real-time early warning process of the pavement icing temperature special point based on the XGBoost temperature prediction model and the improved Z-score algorithm combined with the spatial features of the thermal spectrum map.

2. The method according to claim 1, wherein the method is characterized by: The multi-source data collection process is to arrange fixed weather stations in multiple temperature abnormal areas, collect air temperature data, relative humidity data and pavement temperature data as three original data sources through the fixed weather stations, the air temperature data is temperature data with a range of-40~+60℃ and an accuracy of ±0.1℃, the relative humidity data is humidity data with a range of 0~100%RH and an accuracy of ±2%, and the pavement temperature data is temperature data with a range of-40~+60℃ and an accuracy of ±0.1℃, and the three original data sources collected are preprocessed, and the preprocessed data is constructed to form a three-dimensional data set combined with weather, pavement and position.

3. The method according to claim 2, wherein the method is characterized by: The preprocessing process of the three original data sources includes the processes of outlier rejection, missing value filling, reliability verification and data alignment and structuring; The abnormal value elimination is adopted 3 The three-source original data is processed by the criterion. When a parameter value deviates from the average value of the period data where the parameter value is located by more than 2 times, the median of the adjacent five continuous samples is used to replace, the error caused by the sensor drift is reduced, and the processing process of the abnormal value elimination is completed. When the air temperature data σ is 3.2℃, the temperature value of the air temperature data less than-20℃ or greater than 4℃ is determined as the abnormal data, and the elimination processing is performed. The missing value filling is to control the missing rate of interrupted signals to be less than or equal to 5% as missing data, classify the missing data, use linear interpolation method to fill the short-term missing data with a time length less than 30 seconds, and use KNN algorithm to fill the long-term missing data with a time length greater than or equal to 30 seconds, so as to ensure the integrity of the three-dimensional data set; The reliability verification is to compare the air temperature, relative humidity and pavement temperature data collected by the mobile composite sensor with the data of the fixed weather station at the same time and space position, and ensure that the data with an average error of air temperature less than or equal to 0.7℃, an average error of relative humidity less than or equal to 6% and an average error of pavement temperature less than or equal to 0.5℃ are used as the data to be used; The data alignment and structuring are to classify the data to be used into mobile thermal spectrum data and fixed weather data according to the time stamp and latitude and longitude coordinates, align the mobile thermal spectrum data and fixed weather data, and form a structured data set including seven input features of air temperature, relative humidity, pavement temperature, water and ice thickness, latitude and longitude, sampling time and road type.

4. The method according to claim 3, wherein the method is characterized by: The structured data set formed by the continuous processing of outlier rejection, missing value filling, reliability verification and data alignment and structuring is sequentially divided, optimized and verified to form an XGBoost model, and the establishment process of the XGBoost temperature prediction model is as follows: The structured data set is divided into two parts: the air temperature and relative humidity collected by the fixed weather station are used as input features, and the pavement temperature is used as output feature, a data set with a total sample size greater than or equal to 100,000 is constructed, and the data set is divided into a training set and a test set in a ratio of 8:

2. The divided training set and test set are subjected to hyperparameter optimization processing to form a parameter grid of 120-180 trees, a maximum tree depth of 2-4, and a learning rate of 0.05-0.15, and an XGBoost temperature prediction model is formed through grid search with the objective of minimizing the mean absolute error (MAE); The optimized XGBoost temperature prediction model is subjected to verification processing, and five-fold cross-validation is used to evaluate the model performance. The determination coefficient R² of the training set is greater than or equal to 0.97, and the MAE of the test set is less than or equal to 0.8°C, ensuring that the XGBoost temperature prediction model has precise quantitative performance in using meteorological factors to predict road surface temperature.

5. The method according to claim 4, wherein the method is characterized by: Based on the XGBoost temperature prediction model, the calculation steps for determining the road surface icing temperature special point using the improved Z-score algorithm combined with the spatial features of the thermal spectrum map are as follows: Step one: For each data point Pi in the thermal spectrum map, construct a neighborhood Ni with a radius of 0.1° as the target neighborhood, and count the number of valid data points M in the target neighborhood to ensure that M is in the range of 5-100; Step two: Calculate the average temperature within the target neighborhood, using the formula ; In the above formula, is the average temperature in the target neighborhood, is the number of data collection points in the neighborhood, is the set of all data collection points within a radius of 1.5 km around point i, is the road surface temperature at each collection point in the neighborhood; Step three: Input the air temperature data and relative humidity data collected by the mobile thermal spectrum device into the XGBoost temperature prediction model, and output the predicted road surface temperature Ti; Step four: using the difference between the actual temperature and the predicted temperature of each data collection point as the feature of the special point of the road surface icing temperature, and the calculation formula of the feature of the special point of the road surface icing temperature is ; In the above formula, is the neighborhood temperature difference at the point i to be measured, is the predicted road surface temperature of the XGBoost temperature prediction model at the point i to be measured, is the average temperature within the target neighborhood; Step five: Calculate the point temperature difference mean of the road surface icing temperature special point, and the calculation formula is ； In the above formula, is the mean of the neighborhood temperature deviations, is the neighborhood temperature deviation at the point i to be measured; Step six: Calculate the point temperature difference standard deviation of the road surface icing temperature special point, and the calculation formula is: ； In the above formula, is the standard deviation of the temperature deviation at the point i to be measured, is the neighborhood temperature deviation at the point i to be measured, is the mean value of the neighborhood temperature deviation; Step seven: Calculate the Z-score of the monitoring point temperature difference by the Z-score method: ； In the above formula, Z-score for monitoring the temperature deviation at point i, Mean of the neighborhood temperature deviation, Standard deviation of the temperature deviation at point i to be measured; The data points with |Zi| greater than 2 are screened as special points of pavement icing temperature, and the special points of pavement icing temperature are divided into hot special points and cold special points according to the value of Zi The points greater than 0 are determined as cold special points, The points less than 0 are determined as hot special points, and finally, adjacent special points of pavement icing temperature are integrated into special point bands with a length greater than or equal to 1 km, thereby completing the calculation process of determining special points of pavement icing temperature based on the XGBoost temperature prediction model, the improved Z-score algorithm and the spatial characteristics of the thermal spectrum map.

6. The method according to claim 5, wherein the method is characterized by: The calculation process of screening the road surface icing temperature special point is as follows: calculate the standardized coefficient by the following formula: ； In the above formulae: is the original temperature of point i collected by the mobile road condition sensor, is the set of all data collection points within a radius of 1.5 km around point i, is the number of points in N(i), is the mean of the neighborhood temperature difference, is the standard deviation of the neighborhood temperature difference, represents the temperature value of the kth local region, represents the temperature value of the lth local region adjacent to the kth local region.

7. The method according to claim 1, 2, 3, 4, 5 or 6, characterized in that: After determining the road surface icing temperature special point based on the XGBoost temperature prediction model using the improved Z-score algorithm combined with the spatial features of the thermal spectrum map, the XGBoost temperature prediction model is deployed and dynamically updated. The deployment and dynamic updating process is as follows: Deploy the XGBoost temperature prediction model to the ARM Cortex-A53 architecture edge node and perform quantization compression processing through the TensorRT algorithm; The XGBoost temperature prediction model receives real-time data every 5 minutes, automatically triggers the identification process, and outputs the special point latitude and longitude and risk level; Calculate the identification accuracy every 24 hours, and start incremental updating when the accuracy is less than 80% for three consecutive days. Update the model weight after adding 5000 samples, which takes less than or equal to 30 minutes. After deployment and dynamic updating of the XGBoost temperature prediction model, the continuous processing process of precise positioning, monitoring, and real-time warning of the road surface icing temperature special point is completed.

8. A road surface easy icing temperature special point monitoring and collecting system based on a thermal spectrum map and XGBoost, for realizing a road surface easy icing temperature special point monitoring and collecting method based on a thermal spectrum map and XGBoost, characterized in that: The pavement icing temperature special point monitoring and collecting system comprises multiple fixed weather stations, multiple pavement sensors, multiple mobile composite sensing components and a pavement ice and snow detection data online collecting platform, the multiple fixed weather stations are distributed on the side of the pavement, the multiple pavement sensors are arranged on the pavement, and the multiple mobile composite sensing components are arranged on vehicles, at least one mobile composite sensing component is arranged on each vehicle, the mobile composite sensing component comprises a connecting seat, a temperature sensor (1), a humidity sensor (2) and a pavement state sensor (3), the connecting seat is detachably connected to the vehicle, the temperature sensor (1), the humidity sensor (2) and the pavement state sensor (3) are respectively arranged on the connecting seat, the probe of the temperature sensor (1), the probe of the humidity sensor (2) and the probe of the pavement state sensor (3) are all arranged towards the pavement, and the multiple fixed weather stations, the multiple pavement sensors and the multiple mobile composite sensing components are electrically connected with the pavement ice and snow detection data online collecting platform.

Images