Highway road disaster risk prediction method and system based on big data analysis

Abstract

The application discloses a highway road disaster risk prediction method and system based on big data analysis, relates to the technical field of highway road disaster risk prediction, and comprises the following steps: preprocessing original data to obtain original multidimensional time series data flow, shunting the multidimensional time series data to obtain a prediction core feature data flow, performing nonlinear transformation on the prediction core feature data flow after normalization to obtain a three-dimensional floating-point number vector, inputting the three-dimensional floating-point number vector into a radial basis function neural network adopting an expansion factor for prediction to obtain a prediction icing time, dividing a risk grade according to the prediction icing time, generating and publishing structured early warning information, realizing adaptive change of the expansion factor with water depth, significantly improving local fitting precision, accurately predicting the icing time and dividing the risk grade, and providing reliable decision support for highway traffic safety management.

Description

Technical Field

[0001] This invention relates to the field of highway disaster risk prediction technology, and in particular to a method and system for highway disaster risk prediction based on big data analysis. Background Technology

[0002] In recent years, with the rapid development of big data analytics and IoT technologies, road disaster risk prediction has become an important research direction in the field of intelligent transportation. Winter road icing and water accumulation seriously affect driving safety. Traditional early warning methods based on fixed thresholds are difficult to adapt to complex and ever-changing weather and road conditions. Real-time data collection from multi-source sensors combined with machine learning models for disaster risk prediction has become the mainstream technical approach to improve the accuracy and timeliness of early warnings. Among them, radial basis function neural networks are gradually being applied in road icing prediction due to their strong nonlinear mapping capabilities.

[0003] In existing technologies, most methods only perform independent normalization or standardization on the collected environmental parameters, ignoring the physical modulation effect of water depth on the icing process. When water is present, changes in heat conduction efficiency amplify the impact of temperature. At the same time, water weakens the direct cooling effect of wind speed on the road surface. This nonlinear coupling interaction cannot be reflected by traditional independent scaling, making it difficult to accurately capture the changing patterns of icing rate in predictions. In addition, existing radial basis function neural networks usually use a fixed expansion factor, with all hidden layer neurons using the same width parameter. This makes it impossible to adaptively adjust the response range of neurons according to different water depth regions, limiting the ability to learn local features. Furthermore, most systems lack a complete closed-loop process from data acquisition, distribution, feature transformation to prediction, resulting in insufficient real-time performance and deployability. Summary of the Invention

[0004] The technical problem solved by this invention is that the nonlinear coupling effect of multiple environmental factors is difficult to reflect, especially the physical modulation effect of water depth on ambient temperature and wind speed, which cannot be reflected by independent normalization; at the same time, the dynamic modulation effect of water depth on the icing process cannot be adaptively reflected; in addition, the traditional radial basis function neural network has a fixed expansion factor, which cannot be adaptively adjusted with the change of water depth region, resulting in insufficient local fitting ability.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for predicting highway disaster risks based on big data analysis, comprising the following steps: Step S1: Preprocess the raw data to obtain the raw multidimensional time series data stream; Step S2: The multidimensional time series data is split to obtain the predicted core feature data stream. After normalizing the predicted core feature data stream, a nonlinear transformation is performed to obtain a three-dimensional floating-point vector. Step S3: Input the three-dimensional floating-point vector into a radial basis function neural network with an expansion factor for prediction to obtain the predicted icing time; Step S4: Classify risk levels based on predicted icing time, and generate and release structured early warning information.

[0006] As a preferred embodiment of the highway disaster risk prediction method based on big data analysis described in this invention, step S1 specifically includes: Step S101: The raw data includes water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. Step S102, preprocessing the raw data specifically includes: The current time is obtained from the microcontroller as a timestamp, forming a set of physical quantities; The microcontroller encapsulates the timestamps and values ​​of each physical quantity in the set of physical quantities into a multidimensional time series record in a fixed field order. The fields are ordered as follows: timestamp, water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. The multidimensional time series records are converted into a byte stream in field order via Bluetooth. A communication frame header including a synchronization header and length information is added before the byte stream, and an accumulation and check bytes are added at the end to obtain the original multidimensional time series data stream.

[0007] As a preferred embodiment of the highway disaster risk prediction method based on big data analysis described in this invention, step S2 specifically includes: Step S201: The multidimensional time series data is split into two streams to obtain the prediction core feature data stream and the monitoring auxiliary data stream; The traffic diversion specifically includes: Seven fields were extracted from the multidimensional time series data: timestamp, water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. The water depth, ambient temperature, and wind speed values ​​from the seven fields are used as the core predictive feature data to form a core predictive feature data stream. The ambient relative humidity, road surface temperature, and rainwater temperature from the seven fields are used as monitoring auxiliary data to form a monitoring auxiliary data stream. Step S202: The three fields of water depth, ambient temperature and wind speed in the predicted core feature data stream are normalized by the min-max normalization method to obtain temporary normalized values. The temporary normalized values ​​include the temporary normalized water depth, the temporary normalized ambient temperature, and the temporary normalized wind speed. Step S203: Perform nonlinear interactive transformations on the temporary normalized water depth, temporary normalized ambient temperature, and temporary normalized wind speed values ​​respectively to obtain the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values. Nonlinear interactive transformation specifically includes: The temporary normalized water depth is taken as the final normalized water depth; Calculate the product of the temporary normalized water depth and the first interaction coefficient, and record it as the first product. Calculate the sum of 1 and the first product to obtain the enhancement factor. Calculate the product of the temporary normalized ambient temperature and the enhancement factor to obtain the intermediate value of the ambient temperature. Cut the intermediate value of the ambient temperature to the [0,1] interval and use it as the final normalized ambient temperature. Calculate the product of the temporary normalized water depth and the second interaction coefficient, denoted as the second product. Calculate the difference between 1 and the second product to obtain the weakening factor. Calculate the product of the temporary normalized wind speed value and the weakening factor to obtain the intermediate wind speed value. This intermediate wind speed value is within the closed interval [0,1] and is used as the final normalized wind speed value.

[0008] As a preferred embodiment of the highway disaster risk prediction method based on big data analysis described in this invention, step S2 further includes: Step S204: Combine the final normalized water depth, the final normalized ambient temperature, and the final normalized wind speed to obtain a three-dimensional floating-point vector. The ambient relative humidity, road surface temperature, and rainwater temperature in the monitoring auxiliary data stream will be used as monitoring and display data.

[0009] As a preferred embodiment of the highway disaster risk prediction method based on big data analysis described in this invention, step S3 specifically includes: Step S301: Construct a radial basis function neural network. The structure of the radial basis function neural network includes an input layer, a hidden layer, and an output layer. The input layer consists of 3 neurons; The output layer comprises one neuron; The number of neurons in the hidden layer is H, where H is an integer greater than 1; Each hidden layer neuron corresponds to a pre-defined center vector; Each center vector is a three-dimensional vector. The first dimension of the three-dimensional vector is the final normalized water depth component, the second dimension is the final normalized ambient temperature component, and the third dimension is the final normalized wind speed component. The connection weights from the hidden layer to the output layer are preset. .

[0010] As a preferred embodiment of the highway disaster risk prediction method and system based on big data analysis described in this invention, step S3 further includes: Step S302 involves inputting the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values ​​from the three-dimensional floating-point vector into a radial basis function neural network for processing, and outputting a predicted icing time. Specifically, this includes: The final normalized water depth, final normalized ambient temperature, and final normalized wind speed values ​​of the three-dimensional floating-point vector are input through the input layer to obtain the input three-dimensional floating-point vector. The input three-dimensional floating-point vector is transformed through the hidden layer using the Gaussian radial basis function to obtain the output value of each hidden layer neuron, expressed as: ; in, This represents the input vector, i.e., a three-dimensional floating-point vector. This represents the index number of the hidden layer neurons. =1,2,…, ,in, This represents the number of neurons in the hidden layer. Indicates the first The output values ​​of each hidden layer neuron, i.e., the activation values ​​of the Gaussian radial basis functions. Indicates the first The center vector of each hidden layer neuron, and the input vector The same dimensions Represents the input vector With the center vector The square of the Euclidean distance between them Indicates the first The expansion factor of each hidden layer neuron; where the expansion factor of each hidden layer neuron... Based on the dynamic calculation of the normalized water depth component, the expression is: ； ; in, Indicates the expansion factor. Represents the center vector The first dimension component is the water depth component.

[0011] As a preferred embodiment of the highway disaster risk prediction method and system based on big data analysis described in this invention, wherein: the activation values ​​of the Gaussian radial basis functions of each hidden layer neuron are... The predicted freezing time is obtained by performing linear weighted summation through the output layer; The expression for linear weighted summation is: ; in, Indicates the connection weights. This indicates the predicted freezing time, and the final predicted freezing time takes the value max(0, ).

[0012] As a preferred embodiment of the highway disaster risk prediction method and system based on big data analysis described in this invention, step S4 specifically includes: Step S401, classifying risk levels based on predicted icing time, specifically includes: When the predicted icing time is greater than 60 minutes, it is classified as a Level IV warning. When the predicted freezing time is greater than 30 minutes and less than or equal to 60 minutes, it is judged as a Level III warning; When the predicted icing time is greater than 10 and less than or equal to 30, it is judged as a level 2 warning; When the predicted freezing time is less than or equal to 10 hours, it is classified as a Level 1 warning. The risk levels are divided into four warning levels: Level 4, Level 3, Level 2, and Level 1.

[0013] As a preferred embodiment of the highway disaster risk prediction method and system based on big data analysis described in this invention, step S4 further includes: Step S402: Obtain the predicted icing time, risk level, and environmental relative humidity, road surface temperature, and rainwater temperature from the monitoring and display data; The predicted icing time, risk level, ambient relative humidity, road surface temperature, and rainwater temperature are combined into structured early warning information. By releasing structured early warning information through the traffic management platform, we can obtain road disaster risk warning information after the release.

[0014] A method and system for predicting highway disaster risks based on big data analysis, including a processing module, a transformation module, a prediction module, and a classification module; The processing module preprocesses the raw data to obtain the raw multidimensional time series data stream; The transformation module splits the multidimensional time series data to obtain the predicted core feature data stream, and performs a nonlinear transformation on the normalized predicted core feature data stream to obtain a three-dimensional floating-point vector. The prediction module inputs a three-dimensional floating-point vector into a radial basis function neural network with an expansion factor to predict the icing time. The system is divided into modules, risk levels are classified according to the predicted freezing time, and structured early warning information is generated and released.

[0015] The beneficial effects of this invention are as follows: By introducing a nonlinear interactive transformation based on water depth for the first time, the invention utilizes water depth to enhance ambient temperature and reduce wind speed, respectively, and encodes physical knowledge into feature processing, thus solving the defect that traditional independent normalization cannot reflect coupling relationships; the expansion factor of the radial basis function neural network is designed to dynamically change with the water depth component of the center vector, realizing adaptive change of the expansion factor with water depth, significantly improving local fitting accuracy; through the dual physical knowledge embedding of features and expansion factors, the invention accurately predicts icing time and classifies risk levels, providing reliable decision support for highway traffic safety management. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a highway disaster risk prediction method based on big data analysis, provided as an embodiment of the present invention.

[0017] Figure 2 This is a basic flowchart of a highway disaster risk prediction system based on big data analysis, provided as an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for predicting highway disaster risks based on big data analysis is provided, comprising the following steps: Step S1: Preprocess the raw data to obtain the raw multidimensional time series data stream; Step S2: The multidimensional time series data is split to obtain the predicted core feature data stream. After normalizing the predicted core feature data stream, a nonlinear transformation is performed to obtain a three-dimensional floating-point vector. Step S3: Input the three-dimensional floating-point vector into a radial basis function neural network with an expansion factor for prediction to obtain the predicted icing time; Step S4: Classify risk levels based on predicted icing time, and generate and release structured early warning information.

[0020] In one embodiment, step S1 converts the raw data into a raw multidimensional time series data stream to ensure the standardization and traceability of subsequent analysis. Specifically, this includes: preprocessing the raw data to obtain a raw multidimensional time series data stream, resulting in standardized multidimensional time series data with unified timestamps and verification protection, laying the foundation for feature extraction; step S2 separates the predicted core features from the raw multidimensional time series data stream and enhances the physical correlation between features through nonlinear interactive transformation, thereby improving the input representation capability of the radial basis function neural network. Specifically, this includes: splitting the multidimensional time series data to obtain a predicted core feature data stream; normalizing the predicted core feature data stream and performing a nonlinear transformation to obtain a three-dimensional floating-point vector; compressing the three key quantities—water depth, ambient temperature, and wind speed—to the [0,1] interval and introducing a nonlinear transformation of water depth on temperature and wind speed, making the input vector more consistent with the actual icing physical process; step S3 uses a radial basis function neural network with a dynamic expansion factor to fit complex nonlinearities. The mapping relationship enables accurate prediction of icing time. Specifically, it includes: inputting a three-dimensional floating-point vector into a radial basis function neural network with an extended factor for prediction, obtaining the predicted icing time; by adaptively varying the extended factor with water depth, the radial basis function neural network has differentiated response sensitivity to the icing process under different water conditions, improving prediction accuracy; step S4, converting the continuous prediction results into discrete risk levels and forming publishable early warning information, specifically including: classifying risk levels according to the predicted icing time, generating and publishing structured early warning information, and combining the icing time threshold with auxiliary monitoring data into a structured message, facilitating rapid response and dissemination by the traffic management platform; the entire design framework starts from multi-source sensor data acquisition, and through diversion, nonlinear transformation, adaptive radial basis function neural network prediction to graded early warning, constructs a complete highway icing disaster risk prediction link, realizing end-to-end automated processing from raw data to execution of early warning information, significantly improving the prediction accuracy and early warning timeliness of road icing risk.

[0021] Step S1 specifically includes: Step S101: The raw data includes water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. Step S102, preprocessing the raw data specifically includes: The current time is obtained from the microcontroller as a timestamp, forming a set of physical quantities; The microcontroller encapsulates the timestamps and values ​​of each physical quantity in the set of physical quantities into a multidimensional time series record in a fixed field order. The fields are ordered as follows: timestamp, water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. The multidimensional time series records are converted into a byte stream in field order via Bluetooth. A communication frame header including a synchronization header and length information is added before the byte stream, and an accumulation and check bytes are added at the end to obtain the original multidimensional time series data stream.

[0022] In one embodiment, step S1 transforms the collected raw data into a structured, verifiable time-series data stream, providing a standardized input basis for subsequent data diversion and prediction. This step is the data entry point for the entire disaster risk prediction chain, ensuring the integrity, time alignment, and transmission reliability of the raw data. Specifically, the raw data in step S101, i.e., the raw physical quantities, includes water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. The acquisition method specifically includes: measuring the time difference between transmission and reception using an ultrasonic sensor to obtain the round-trip time; calculating the product of the round-trip time and the speed of sound to obtain the total propagation distance; calculating the ratio of the total propagation distance to 2 to obtain the distance from the sensor to the water surface; and subtracting this distance from the sensor's preset installation height to obtain the water depth. The speed of sound is set to 340 meters per second, based on the following experiment: using the square root relationship formula between the speed of sound and temperature in air. Using temperature (T) in Celsius as the independent variable and sound speed (c) as the dependent variable, the method is effective within the range of -10°C to 40°C. At four test points with ambient temperatures of -10°C, 0°C, 20°C, and 40°C, a fixed distance of 2.000 meters was calibrated using a standard metal ruler. The round-trip time was obtained by measuring the time difference between transmission and reception using an ultrasonic sensor. Distance calculations were performed using the actual sound speed calculated by this formula and the fixed sound speed of 340 meters per second. Ten measurements were taken at each temperature point, and the average value was taken. The results show that at -10°C, the actual sound speed is approximately 325.4 meters per second, and the fixed sound speed measurement is 1.957 meters, with an error of -2.15%. At 40°C, the actual sound speed is approximately 354.9 meters per second, and the fixed sound speed measurement is 2.043 meters, with an error of +2.15%. The maximum absolute error at all test points is less than 5%, meeting the requirements for road water accumulation monitoring; therefore, this setting is feasible. Ambient temperature and relative humidity are obtained by reading temperature and humidity sensors via I²C bus; an analog voltage is output from a wind speed transmitter, for example, [0,10] volts corresponds to [0,50] meters per second, and this analog voltage is converted to wind speed using a linear formula after analog-to-digital conversion; a thermocouple senses the road surface temperature to generate a thermoelectric potential, which is converted to a digital value by the microcontroller through the MAX6675 chip, and then read and converted to road surface temperature via the SPI bus; rainwater temperature is collected through a contact temperature probe; and water depth, ambient temperature, and ambient temperature are obtained. After obtaining the relative humidity, wind speed, road surface temperature, and rainwater temperature, the current time is read from the microcontroller as a timestamp. The timestamp adopts the Unix timestamp format, that is, the number of seconds that have elapsed since January 1, 1970, 00:00:00, represented by a four-byte unsigned integer in little-endian byte order, forming a set of physical quantities. This step unifies the data of the six physical quantities—water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature—to the same time base, eliminating time alignment errors between the original data. Step S102 will... The physical quantity set is organized into a structured record with a fixed format, which facilitates stable parsing and distribution in subsequent steps. This step establishes a unified data contract, reducing the risk of data misalignment and parsing ambiguity. Specifically, in step S102, the microcontroller encapsulates the timestamps and values ​​of each physical quantity in the physical quantity set into a multi-dimensional time series record according to a fixed field order. The field order is as follows: timestamp, water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. Each physical quantity is represented by a two-byte signed integer. The values ​​are as follows: water depth is in millimeters, ranging from [0, 50] millimeters. Each millimeter is magnified tenfold to represent a value of 1. For example, a water depth of 12.3 millimeters is stored as 123. Ambient temperature is in 0.1 degrees Celsius, for example, 23.4 degrees Celsius is stored as 234. Ambient relative humidity is in 0.1 percent, for example, 78.5 percent is stored as 785. Wind speed is in 0.01 meters per second, for example, 12.34 meters per second is stored as 1234. Road surface temperature is in 0.1 degrees Celsius. Rainwater temperature is in 0.The unit is 1 degree Celsius; all multi-byte values ​​are stored in little-endian byte order, forming a structured record with a fixed format, ensuring that the position, length, and meaning of each field are completely determined, avoiding data errors caused by disordered field order or inconsistent byte parsing; frame boundary and integrity checks are added to wireless transmission to ensure that data can still be accurately received and identified after transmission in the Bluetooth channel. This step solves the problems of packet loss and bit error in wireless communication and ensures the reliability of the original data. Specifically, it includes: converting the multidimensional time series record into a byte stream according to the field order via Bluetooth, firstly, the timestamp (4 bytes), water depth (2 bytes), ambient temperature (2 bytes), ambient relative humidity (2 bytes), and wind speed value (2 bytes). The data payload consists of 16 bytes: one byte for road surface temperature, two bytes for rainwater temperature, and two bytes for rainwater temperature, concatenated sequentially. A communication frame header is then added before the byte stream. The frame header includes a synchronization header and length information. The synchronization header is set to two hexadecimal values, 0xAA and 0x55, based on the serial communication synchronization mode. The anti-interference capability of this combination is determined experimentally, specifically by randomly inserting single-bit flips and random byte interference into the original multidimensional time-series data stream. Missynchronization is defined as the receiver misidentifying a non-frame header byte as a frame header or missing the true frame header. Single-byte synchronization headers of 0xAA and 0x55, and a two-byte combination of 0xAA and 0x55, are used as frame headers respectively. Each synchronization header... A total of 1000 frames were tested. During the transmission of each frame, the receiver was recorded as to whether it correctly identified the frame start position. The number of missynchronized frames under each synchronization header was counted, and the missynchronization rate was obtained by dividing the number of missynchronized frames by the total number of test frames (1000). The experiment showed that the number of missynchronized frames with the single-byte synchronization header 0xAA was 32, and the number of missynchronized frames with 0x55 was 31, with a missynchronization rate of approximately 3.2%. The number of missynchronized frames with the double-byte combination 0xAA and 0x55 was 0, with a missynchronization rate of less than 0.1%. Therefore, this double-byte combination can effectively resist random interference in the data stream and achieve accurate frame start positioning. The length information is one byte, indicating the number of bytes of subsequent data payload. The data payload consists of a four-byte timestamp and a watermark. The depth (two bytes), ambient temperature (two bytes), ambient relative humidity (two bytes), wind speed (two bytes), road surface temperature (two bytes), and rainwater temperature (two bytes) are concatenated sequentially, totaling sixteen bytes. Therefore, this byte is fixed at sixteen. A checksum byte is added to the end of the byte stream. This checksum byte is calculated by summing all bytes from the synchronization header to the end of the data payload, modulo 256. Specifically, it's the sum of the length information byte and the sixteen data payload bytes. The lower eight bits of the sum are used as the checksum byte, resulting in twenty bytes. These twenty bytes are then sequentially combined into a complete byte sequence to obtain the original multidimensional time series data stream. Step S1 enables the wirelessly transmitted data to have clear frame boundary identification and integrity verification capabilities, reducing the negative impact of packet loss and bit errors on the accuracy of subsequent predictions. From the raw data acquisition of sensors such as ultrasonic, temperature and humidity, wind speed, thermocouples and contact probes, to timestamp alignment, fixed field encapsulation, and then to the generation of Bluetooth frames with synchronization headers, length information and cumulative checksums, the high-fidelity acquisition and standardized transmission of raw data are fully realized, providing a reliable, traceable and time-aligned raw multidimensional time series data stream for subsequent feature splitting and radial basis function neural network prediction.

[0023] Step S2 specifically includes: Step S201: The multidimensional time series data is split into two streams to obtain the prediction core feature data stream and the monitoring auxiliary data stream; The traffic diversion specifically includes: Seven fields were extracted from the multidimensional time series data: timestamp, water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. The water depth, ambient temperature, and wind speed values ​​from the seven fields are used as the core predictive feature data to form a core predictive feature data stream. The ambient relative humidity, road surface temperature, and rainwater temperature from the seven fields are used as monitoring auxiliary data to form a monitoring auxiliary data stream. Step S202: The three fields of water depth, ambient temperature and wind speed in the predicted core feature data stream are normalized by the min-max normalization method to obtain temporary normalized values. The temporary normalized values ​​include the temporary normalized water depth, the temporary normalized ambient temperature, and the temporary normalized wind speed. Step S203: Perform nonlinear interactive transformations on the temporary normalized water depth, temporary normalized ambient temperature, and temporary normalized wind speed values ​​respectively to obtain the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values. Nonlinear interactive transformation specifically includes: The temporary normalized water depth is taken as the final normalized water depth; Calculate the product of the temporary normalized water depth and the first interaction coefficient, and record it as the first product. Calculate the sum of 1 and the first product to obtain the enhancement factor. Calculate the product of the temporary normalized ambient temperature and the enhancement factor to obtain the intermediate value of the ambient temperature. Cut the intermediate value of the ambient temperature to the [0,1] interval and use it as the final normalized ambient temperature. Calculate the product of the temporary normalized water depth and the second interaction coefficient, denoted as the second product. Calculate the difference between 1 and the second product to obtain the weakening factor. Calculate the product of the temporary normalized wind speed value and the weakening factor to obtain the intermediate wind speed value. This intermediate wind speed value is within the closed interval [0,1] and is used as the final normalized wind speed value.

[0024] Step S2 also includes: Step S204: Combine the final normalized water depth, the final normalized ambient temperature, and the final normalized wind speed to obtain a three-dimensional floating-point vector. The ambient relative humidity, road surface temperature, and rainwater temperature in the monitoring auxiliary data stream will be used as monitoring and display data.

[0025] In one embodiment, step S2 involves separating the most critical features for icing prediction from the original multidimensional time series data stream and enhancing the physical correlation between features through normalization and nonlinear interactive transformation, thereby providing a high-quality input vector for the radial basis function neural network. This step eliminates redundant monitoring auxiliary information, allowing the prediction to focus on the three core variables of water depth, ambient temperature, and wind speed, while retaining ambient relative humidity, road surface temperature, and rainwater temperature as auxiliary display data, thus balancing prediction accuracy and information completeness. Specifically, step S201 involves splitting the multidimensional time series data stream to obtain a prediction core feature data stream and a monitoring auxiliary data stream. The splitting specifically includes: extracting time from the multidimensional time series data. The system uses seven fields: water depth, ambient temperature, relative humidity, wind speed, road surface temperature, and rainwater temperature. Water depth, ambient temperature, and wind speed are used as core predictive features to form a core predictive feature data stream. Relative humidity, road surface temperature, and rainwater temperature are used as auxiliary monitoring data to form an auxiliary monitoring data stream. This step clearly distinguishes between the core features used for prediction and the auxiliary monitoring data used only for monitoring and display, reducing the input dimension of the subsequent radial basis function neural network and lowering computational complexity. In step S202, the water depth, ambient temperature, and wind speed fields in the core predictive feature data stream are normalized using the min-max normalization method to obtain temporary normalized values. The normalized values ​​include temporary normalized water depth, temporary normalized ambient temperature, and temporary normalized wind speed. The normalization process specifically includes: for each field, calculating the minimum and maximum values ​​of the field in all historical data; if the maximum value of the field is equal to the minimum value, then all normalized values ​​of the field are set to 0; otherwise, calculating the difference between each measured value of the field and the minimum value of the field to obtain the first difference; calculating the difference between the maximum value and the minimum value of the field to obtain the second difference; calculating the ratio of the first difference to the second difference to obtain the temporary normalized water depth, temporary normalized ambient temperature, and temporary normalized wind speed values. These temporary normalized values ​​are all within [0,1]. This step unifies physical quantities of different dimensions to the same numerical range. In step S203, nonlinear interactive transformations are performed on the temporary normalized water depth, temporary normalized ambient temperature, and temporary normalized wind speed values ​​to obtain the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values. The nonlinear interactive transformation specifically includes: directly using the temporary normalized water depth as the final normalized water depth; calculating the sum of 1 and the first product to obtain the enhancement factor; calculating the product of the temporary normalized ambient temperature and the enhancement factor to obtain the intermediate ambient temperature value; truncating the intermediate ambient temperature value to the [0,1] interval to obtain the final normalized ambient temperature; wherein, if the intermediate value is greater than or equal to 1, it is taken as 1, and if it is less than or equal to 0, it is taken as 0, so that the final result is within [0,1] and is used as the final normalized ambient temperature.The product of the temporary normalized water depth and the second interaction coefficient is calculated and denoted as the second product. The difference between the first and second products is calculated to obtain the weakening factor. The product of the temporary normalized wind speed value and the weakening factor is calculated to obtain the median wind speed value, which naturally lies within [0,1] and is used as the final normalized wind speed value. The first interaction coefficient is a preset constant with a value range of (0,0.5], and the second interaction coefficient is also a preset constant with a value range of (0,0.5]. The basis for setting this value range is: through a grid search experiment, the historical dataset was randomly divided into 7:3 ratios. The training and validation sets were used to test all 100 combinations of the first and second interaction coefficients from 0.05 to 0.5, with a step size of 0.05. For each coefficient combination, a radial basis function neural network was trained on the training set, and the root mean square error (RMSE) between the predicted and measured icing times was calculated on the validation set. Experimental results show that when both coefficients are less than 0.5, the RMSE on the validation set is between 1.2 and 1.5 minutes, with a fluctuation range of no more than 0.3 minutes. When either coefficient exceeds 0.5, for example, when the first interaction coefficient is 0.55 and the second interaction coefficient is less than 0.5, the RMSE is significantly reduced. At a value of 0.5, the root mean square error (RMSE) reaches 2.8 minutes, an increase of 133% compared to the lower limit of 1.2 minutes. Furthermore, the RMSE corresponding to all coefficient combinations exceeding 0.5 is greater than 2.5 minutes, an increase of over 100% compared to the lower limit of 1.2 minutes. This indicates that when the first or second interaction coefficient exceeds 0.5, the adjustment magnitude of the enhancement or weakening factor will more than double the RMSE, leading to a significant decrease in prediction accuracy. Therefore, the value range is determined to be (0, 0.5]. This step uses water depth to perform nonlinear correction on temperature and wind speed; the deeper the water, the greater the impact on temperature. The contribution of ice is enhanced while the inhibitory effect of wind speed is weakened, more realistically reflecting the actual icing physical process; in step S204, the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values ​​are combined to obtain a three-dimensional floating-point vector; at the same time, the ambient relative humidity, road surface temperature, and rainwater temperature in the monitoring auxiliary data stream are used as monitoring display data for subsequent warning information release. This step generates a three-dimensional input vector that can be directly processed by the radial basis function neural network, while retaining the auxiliary monitoring data for result display, thus improving the information content of the warning information; Step S2 accurately extracts the core prediction features from the original multidimensional time series, eliminates the influence of dimensions through min-max normalization, and then adaptively and nonlinearly adjusts the temperature and wind speed through water depth. Finally, it outputs a three-dimensional floating-point vector with clear physical meaning and a standardized value range, laying a reliable data foundation for the high-precision prediction of the subsequent radial basis function neural network.

[0026] Step S3 specifically includes: Step S301: Construct a radial basis function neural network. The structure of the radial basis function neural network includes an input layer, a hidden layer, and an output layer. The input layer consists of 3 neurons; The output layer consists of one neuron; The number of neurons in the hidden layer is H, where H is an integer greater than 1; Each hidden layer neuron corresponds to a pre-defined center vector; Each center vector is a three-dimensional vector. The first dimension of the three-dimensional vector is the final normalized water depth component, the second dimension is the final normalized ambient temperature component, and the third dimension is the final normalized wind speed component. The connection weights from the hidden layer to the output layer are preset. .

[0027] Step S3 also includes: Step S302 involves inputting the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values ​​from the three-dimensional floating-point vector into a radial basis function neural network for processing, and outputting a predicted icing time. Specifically, this includes: The final normalized water depth, final normalized ambient temperature, and final normalized wind speed values ​​of the three-dimensional floating-point vector are input through the input layer to obtain the input three-dimensional floating-point vector. The input three-dimensional floating-point vector is transformed through the hidden layer using the Gaussian radial basis function to obtain the output value of each hidden layer neuron, expressed as: ; in, This represents the input vector, i.e., a three-dimensional floating-point vector. This represents the index number of the hidden layer neurons. =1,2,…, ,in, This represents the number of neurons in the hidden layer. Indicates the first The output values ​​of each hidden layer neuron, i.e., the activation values ​​of the Gaussian radial basis functions. Indicates the first The center vector of each hidden layer neuron, and the input vector The same dimensions Represents the input vector With the center vector The square of the Euclidean distance between them Indicates the first The expansion factor of each hidden layer neuron; where the expansion factor of each hidden layer neuron... Based on the dynamic calculation of the normalized water depth component, the expression is: ； ; in, Indicates the expansion factor. Represents the center vector The first component is the water depth component.

[0028] The activation values ​​of the Gaussian radial basis functions of neurons in each hidden layer are... The predicted freezing time is obtained by performing linear weighted summation through the output layer; The expression for linear weighted summation is: ; in, Indicates the connection weights. This indicates the predicted freezing time, and the final predicted freezing time takes the value max(0, ).

[0029] In one embodiment, step S3 utilizes a radial basis function neural network with a dynamic expansion factor to fit the complex nonlinear mapping relationship between the three-dimensional input vector and the icing time, achieving accurate prediction of the icing time. This step is the core of the disaster risk prediction chain, transforming the three-dimensional floating-point vector into a quantitative predicted icing time, providing a basis for subsequent risk level classification. Specifically, step S301 involves constructing a radial basis function neural network. The structure of the radial basis function neural network includes an input layer, a hidden layer, and an output layer. The input layer includes three neurons, used to receive the three-dimensional floating-point vector X, whose three components are the final normalized water depth, the final normalized ambient temperature, and the final normalized wind speed value, respectively. The output layer includes one neuron, used to output the predicted icing time in minutes. The number of neurons H in the hidden layer is set to 12, a value determined experimentally. Specifically, the process involves: randomly dividing the historical dataset into training and validation sets in a 7:3 ratio; for each integer H from 6 to 20, using a fixed random seed and K-means clustering algorithm to determine the center vector; solving for the output weights using ridge regression; training a radial basis function neural network on the training set; and calculating the root mean square error (RMSE) between the predicted and measured icing times on the validation set. Each H value is repeated 5 times, and the average RMSE is taken. Experimental results show that when H is less than 8, the average RMSE is greater than 2.5 minutes; when H is greater than 15, the validation set error increases by more than 20% compared to the training set error, indicating overfitting; when H equals 12, the average RMSE on the validation set is 1.3 minutes, with fluctuations not exceeding 0.1 minutes across multiple experiments, indicating the smallest and most stable error. Therefore, H is set to 12. Each hidden layer neuron corresponds to a pre-defined center vector. The center vector is a three-dimensional vector, and the first dimension of the three-dimensional vector is the final normalized water depth component. The second dimension component is the final normalized ambient temperature component, and the third dimension component is the final normalized wind speed component; the connection weights from the hidden layer to the output layer are denoted as... Step S302 involves inputting the three-dimensional floating-point vector into a radial basis function neural network for processing and outputting the predicted icing time. Specifically, this includes: passing the three-dimensional floating-point vector through the input layer to obtain the input vector X; transforming the input vector X through the hidden layers using Gaussian radial basis functions to obtain the output values ​​of each hidden layer neuron, expressed as: ; in, This represents the input vector, i.e., a three-dimensional floating-point vector. This represents the index number of the hidden layer neurons. =1,2,…, H=12 Indicates the first The output values ​​of each hidden layer neuron, i.e., the activation values ​​of the Gaussian radial basis functions. Indicates the first The center vector of each hidden layer neuron, and the input vector The same dimensions Represents the input vector With the center vector The square of the Euclidean distance between them reflects and The degree of similarity between them Indicates the first The expansion factor of each hidden layer neuron; The dynamic calculation is performed based on the first dimension of the center vector, i.e., the normalized water depth component, and the expression is: ; ; in, Indicates the expansion factor. Represents the center vector The first component, namely the water depth component, with coefficients 0.2 and 2 determined experimentally, specifically includes: The historical dataset was randomly divided into a training set and a validation set in a 7:3 ratio, with the number of hidden layer neurons fixed at H=12. The search range is defined as the coefficient b within [1,3] and the numerator coefficient a within [0.1,0.3], with a step size of 0.01, generating a total of 200 parameter combinations. For each combination (a,b), a radial basis function neural network is constructed and trained on the training set, and the root mean square error (RMSE) between the predicted and measured icing times is calculated on the validation set. Experimental results show that when a=0.2 and b=2, the validation set RMSE reaches a minimum of 1.3 minutes. When a belongs to the closed interval [0.19,0.21] and b belongs to the closed interval [1.9,2.1], the validation set RMSE is within [1.3,1.4] minutes, with the maximum value of 1.4 minutes increasing by approximately 7.7% compared to the minimum value of 1.3 minutes, and the error change is less than 10%. When a deviates from 0.2 by more than 0.03 or b deviates from 2 by more than 0.3, the RMSE increases to more than 1.8 minutes, an increase of more than 38% compared to the minimum value. Under this parameter combination, the expansion factor is... When =0, take 0.2. When a = 1, we take 0.2 / 3 ≈ 0.0667, which results in a smaller expansion factor for samples with greater water depth, a more concentrated radial basis function response, and improved prediction sensitivity. Therefore, we determine a = 0.2 and b = 2; on the historical dataset, we test the denominator with a step size of 0.01. The objective is to minimize the prediction error on the validation set for all combinations where the coefficient of the numerator is in the range [1,3] and the coefficient of the numerator is in the range [0.1,0.3]. Experimental results show that when the numerator is 0.2 and the coefficient in the denominator is 2, the expansion factor is within the range of [1,3]. When =0, take 0.2. When =1, take 0.2 / 3≈0.0667. This range makes the radial basis function response of samples with large water depth more concentrated, thereby improving the prediction sensitivity. The activation values ​​ωh(X) of neurons in each hidden layer are linearly weighted and summed through the output layer to obtain the predicted freezing time. The linear weighted summation expression is as follows: ; in, Indicates the connection weights. This indicates the predicted freezing time, and the final predicted freezing time takes the value max(0, This step ensures that the freezing time is non-negative; by adaptively varying the expansion factor with the water depth, the radial basis function neural network has differentiated response sensitivity to the freezing process under different water depths, significantly improving prediction accuracy. The training process of the radial basis function neural network specifically includes: Historical datasets were acquired. Each sample in the historical dataset included the final normalized water depth, final normalized ambient temperature, and final normalized wind speed, which served as the input vector for the radial basis function neural network (RBFNN). The corresponding measured icing time served as the target output of the RBFNN. The number of hidden layer neurons was preset to H=12. After fixing the random seed, the K-means clustering algorithm was used to cluster the input vectors from the historical data into 12 classes. The cluster centers were then used as the center vectors of the hidden layer neurons. According to each central vector First dimension component The expansion factor of each hidden layer neuron is calculated by computed to the expansion factor expression. The input vector is passed through the input layer and transformed by the Gaussian radial basis function in the hidden layer to obtain the hidden layer output matrix. Using the hidden layer output matrix as the input matrix and the measured value of the target output as the observation value, the connection weights from the hidden layer to the output layer are solved by ridge regression. The regularization parameter λ was set to 0.001, a value determined experimentally. Specifically, the historical dataset was randomly divided into a training set and a validation set at a ratio of 7:3. Within the closed interval [0.0001, 0.1], λ was successively set to 0.0001, 0.001, 0.01, and 0.1 with a step size of 10. For each λ value, a radial basis function neural network was trained, and the root mean square error between the predicted and measured icing times was calculated on the validation set. Simultaneously, the connection weights obtained from five repeated training iterations were recorded. The standard deviation of the validation set is 1.3 minutes when λ=0.001, and the standard deviation of each connection weight in five repeated trainings is less than 0.05, while the standard deviation of the weights corresponding to other λ values ​​is greater than 0.2. Therefore, the error is minimized and the weight fluctuation is minimized when λ=0.001, so 0.001 is chosen. Ridge regression minimizes the sum of the weighted mean square error between the output layer prediction and the target output and the sum of the squares of the connection weights, thus obtaining the trained radial basis function neural network. Step S3 constructs a radial basis function neural network with a dynamic expansion factor. The center vector is determined by K-means clustering and the connection weights are solved by ridge regression. This achieves a high-precision mapping from three-dimensional input features to predicted freezing time. The expansion factor adapts to the water depth, enabling the radial basis function neural network to have differentiated responses to different water depths. The average absolute error of the predicted freezing time is less than 2 minutes, providing a reliable quantitative basis for subsequent risk level classification.

[0030] Step S4 specifically includes: Step S401, classifying risk levels based on predicted icing time, specifically includes: When the predicted icing time is greater than 60 minutes, it is classified as a Level IV warning. When the predicted freezing time is greater than 30 minutes and less than or equal to 60 minutes, it is judged as a Level III warning; When the predicted icing time is greater than 10 and less than or equal to 30, it is judged as a level 2 warning; When the predicted freezing time is less than or equal to 10 hours, it is classified as a Level 1 warning. The risk levels are divided into four warning levels: Level 4, Level 3, Level 2, and Level 1.

[0031] Step S4 also includes: Step S402: Obtain the predicted icing time, risk level, and environmental relative humidity, road surface temperature, and rainwater temperature from the monitoring and display data; The predicted icing time, risk level, ambient relative humidity, road surface temperature, and rainwater temperature are combined into structured early warning information. By releasing structured early warning information through the traffic management platform, we can obtain road disaster risk warning information after the release.

[0032] In one embodiment, step S4 converts the predicted icing time in continuous numerical form into discrete risk levels and combines auxiliary monitoring data to form structured early warning information for direct use by traffic management departments, realizing a closed-loop business process from prediction to early warning. This step is the output end of the entire prediction chain, combining the calculation results of the radial basis function neural network with the on-site auxiliary monitoring data. The clear levels and formats improve the operability and response efficiency of the early warning. Specifically, step S401 divides the risk level according to the predicted icing time. The division rules specifically include: when the predicted icing time is greater than 60 minutes, it is judged as level four. Warnings are issued in four levels: Level 3 warning, Level 4 warning, Level 3 warning, Level 2 warning, and Level 1 warning. The risk levels, from lowest to highest, are: Level 4 warning, Level 3 warning, Level 2 warning, and Level 1 warning. The thresholds of 60 minutes, 30 minutes, and 10 minutes are set based on historical road icing data collected over three consecutive winters (180 days total), recording the actual time from the appearance of water accumulation to complete road icing, and simultaneously tracking traffic accidents during this period. Accident incidence rates were determined experimentally. When the predicted icing time was greater than 60 minutes, the accident incidence rate was less than 1%; when the icing time was between 30 and 60 minutes, the accident incidence rate was 3% to 5%; when the icing time was between 10 and 30 minutes, the accident incidence rate was 8% to 12%; and when the icing time was less than or equal to 10 minutes, the accident incidence rate exceeded 20%. Therefore, thresholds of 60, 30, and 10 minutes were set to clearly define the risk level for each warning level. Step S402 involves acquiring the predicted icing time, risk level, and environmental relative humidity, road surface temperature, and rainwater temperature from the monitored and displayed data. The five data points are combined into a structured early warning information in JSON object format, including five fields: predict_time (in minutes, integer), risk_level (string value, such as Level 1 warning), humidity (percentage integer), road_temp (in 0.1 degrees Celsius, integer), and rain_temp (in 0.1 degrees Celsius, integer). This structured early warning information is published through the message queue interface of the traffic management platform. After receiving it, the platform pushes it to variable message signs and mobile terminals within its jurisdiction to obtain the published road disaster risk early warning information. Step S4 transforms the abstract predicted values ​​into intuitive risk levels and combines them with real-time auxiliary monitoring data to form a complete information package, which facilitates traffic management personnel to quickly judge and take countermeasures such as road closures and salting. This realizes the automated conversion from predicting icing time to graded early warning, and ensures the timeliness and understandability of the early warning through a standardized information release interface, effectively reducing the risk of traffic accidents caused by road icing.

[0033] Example 2, refer to Figure 2 In another embodiment of the present invention, which differs from the first embodiment, a highway disaster risk prediction system based on big data analysis is provided, including a processing module, a transformation module, a prediction module, and a segmentation module; The processing module preprocesses the raw data to obtain the raw multidimensional time series data stream; The transformation module splits the multidimensional time series data to obtain the predicted core feature data stream, and performs a nonlinear transformation on the normalized predicted core feature data stream to obtain a three-dimensional floating-point vector. The prediction module inputs a three-dimensional floating-point vector into a radial basis function neural network with an expansion factor to predict the icing time. The system is divided into modules, risk levels are classified according to the predicted freezing time, and structured early warning information is generated and released.

[0034] In one embodiment, the processing module converts the raw data into a raw multidimensional time series data stream to ensure the standardization and traceability of subsequent analysis. Specifically, this includes: preprocessing the raw data to obtain a raw multidimensional time series data stream, resulting in standardized multidimensional time series data with unified timestamps and verification protection, laying the foundation for feature extraction; the transformation module separates the core predictive features from the raw multidimensional time series data stream and enhances the physical correlation between features through nonlinear interactive transformation, thereby improving the input representation capability of the radial basis function neural network. Specifically, this includes: splitting the multidimensional time series data to obtain a core predictive feature data stream; normalizing the core predictive feature data stream and performing a nonlinear transformation to obtain a three-dimensional floating-point vector; compressing the three key quantities—water depth, ambient temperature, and wind speed—to the [0,1] interval and introducing a nonlinear transformation of water depth on temperature and wind speed, making the input vector more consistent with the actual icing physical process; and the prediction module uses a radial basis function neural network with a dynamic expansion factor to fit complex nonlinearities. The mapping relationship enables accurate prediction of icing time. Specifically, it involves inputting a three-dimensional floating-point vector into a radial basis function neural network (RBF) with an extended factor for prediction, obtaining the predicted icing time. By adaptively varying the extended factor with water depth, the RBF network exhibits differentiated response sensitivity to icing processes under different water depths, improving prediction accuracy. The module segmentation transforms continuous prediction results into discrete risk levels and generates publishable early warning information. Specifically, it involves classifying risk levels based on predicted icing time, generating and publishing structured early warning information, and combining icing time thresholds with auxiliary monitoring data into structured messages for rapid response and dissemination by the traffic management platform. The entire design framework starts from multi-source sensor data acquisition, and through diversion, nonlinear transformation, adaptive RBF neural network prediction to tiered early warning, constructs a complete highway icing disaster risk prediction chain. This achieves end-to-end automated processing from raw data to execution of early warning information, significantly improving the accuracy and timeliness of road icing risk prediction.

[0035] To verify the effectiveness of this scheme, its performance was compared with that of the traditional independent normalization combined with the fixed expansion factor RBF method. The results are shown in Table 1. Table 1. Performance comparison of different methods for predicting freezing time; ; Table 1 compares the performance of our proposed method with that of traditional methods in predicting road icing time. Traditional methods use radial basis function neural networks with independent normalization and fixed expansion factors, and the root mean square error reaches 2.8 minutes when the interaction coefficient exceeds 0.5. Our proposed method reduces the root mean square error to 1.3 minutes and the mean absolute error to less than 2 minutes by using nonlinear interactive transformation (water depth enhances temperature and weakens wind speed) and adaptive variation of the expansion factor with water depth. Moreover, no combination of coefficients shows an error increase of more than 100%, which significantly improves the prediction accuracy.

[0036] This invention introduces a nonlinear interactive transformation based on water depth for the first time, utilizing water depth to enhance ambient temperature and reduce wind speed respectively. By encoding physical knowledge into feature processing, it overcomes the deficiency of traditional independent normalization in failing to reflect coupling relationships. The expansion factor of the radial basis function neural network is designed as a function that dynamically changes with the water depth component of the center vector, enabling the expansion factor to adapt to water depth and significantly improving local fitting accuracy. Through the dual embedding of physical knowledge in features and expansion factors, the invention accurately predicts icing time and classifies risk levels, providing reliable decision support for highway traffic safety management.

[0037] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0038] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention. All data acquisition actions in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located and with the authorization granted by the owner of the corresponding device.

Claims

1. A method for predicting highway disaster risks based on big data analysis, characterized in that, Includes the following steps: Step S1: Preprocess the raw data to obtain the raw multidimensional time series data stream; Step S2: The multidimensional time series data is split to obtain the predicted core feature data stream. After normalizing the predicted core feature data stream, a nonlinear transformation is performed to obtain a three-dimensional floating-point vector. Step S3: Input the three-dimensional floating-point vector into a radial basis function neural network with an expansion factor for prediction to obtain the predicted icing time; Step S4: Classify risk levels based on predicted icing time, and generate and release structured early warning information.

2. The method for predicting highway disaster risks based on big data analysis as described in claim 1, characterized in that: Step S1 specifically includes: Step S101: The raw data includes water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. Step S102, preprocessing the raw data specifically includes: The current time is obtained from the microcontroller as a timestamp, forming a set of physical quantities; The microcontroller encapsulates the timestamps and values ​​of each physical quantity in the set of physical quantities into a multidimensional time series record in a fixed field order. The fields are ordered as follows: timestamp, water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. The multidimensional time series records are converted into a byte stream in field order via Bluetooth. A communication frame header including a synchronization header and length information is added before the byte stream, and an accumulation and check bytes are added at the end to obtain the original multidimensional time series data stream.

3. The method for predicting highway disaster risks based on big data analysis as described in claim 2, characterized in that: Step S2 specifically includes: Step S201: The multidimensional time series data is split into two streams to obtain the prediction core feature data stream and the monitoring auxiliary data stream; The traffic diversion specifically includes: Seven fields were extracted from the multidimensional time series data: timestamp, water depth, ambient temperature, ambient relative humidity, wind speed, road surface temperature, and rainwater temperature. The water depth, ambient temperature, and wind speed values ​​from the seven fields are used as the core predictive feature data to form a core predictive feature data stream. The ambient relative humidity, road surface temperature, and rainwater temperature from the seven fields are used as monitoring auxiliary data to form a monitoring auxiliary data stream. Step S202: The three fields of water depth, ambient temperature and wind speed in the predicted core feature data stream are normalized by the min-max normalization method to obtain temporary normalized values. The temporary normalized values ​​include the temporary normalized water depth, the temporary normalized ambient temperature, and the temporary normalized wind speed. Step S203: Perform nonlinear interactive transformations on the temporary normalized water depth, temporary normalized ambient temperature, and temporary normalized wind speed values ​​respectively to obtain the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values. Nonlinear interactive transformation specifically includes: The temporary normalized water depth is taken as the final normalized water depth; Calculate the product of the temporary normalized water depth and the first interaction coefficient, and record it as the first product. Calculate the sum of 1 and the first product to obtain the enhancement factor. Calculate the product of the temporary normalized ambient temperature and the enhancement factor to obtain the intermediate value of the ambient temperature. Cut the intermediate value of the ambient temperature to the [0,1] interval and use it as the final normalized ambient temperature. Calculate the product of the temporary normalized water depth and the second interaction coefficient, denoted as the second product. Calculate the difference between 1 and the second product to obtain the weakening factor. Calculate the product of the temporary normalized wind speed value and the weakening factor to obtain the intermediate wind speed value. This intermediate wind speed value is within the closed interval [0,1] and is used as the final normalized wind speed value.

4. The method for predicting highway disaster risks based on big data analysis as described in claim 3, characterized in that: Step S2 also includes: Step S204: Combine the final normalized water depth, the final normalized ambient temperature, and the final normalized wind speed to obtain a three-dimensional floating-point vector. The ambient relative humidity, road surface temperature, and rainwater temperature in the monitoring auxiliary data stream will be used as monitoring and display data.

5. The method for predicting highway disaster risks based on big data analysis as described in claim 4, characterized in that: Step S3 specifically includes: Step S301: Construct a radial basis function neural network. The structure of the radial basis function neural network includes an input layer, a hidden layer, and an output layer. The input layer consists of 3 neurons; The output layer comprises one neuron; The number of neurons in the hidden layer is H, where H is an integer greater than 1; Each hidden layer neuron corresponds to a pre-defined center vector; Each center vector is a three-dimensional vector. The first dimension of the three-dimensional vector is the final normalized water depth component, the second dimension is the final normalized ambient temperature component, and the third dimension is the final normalized wind speed component. The connection weights from the hidden layer to the output layer are preset. .

6. The method for predicting highway disaster risks based on big data analysis as described in claim 5, characterized in that: Step S3 also includes: Step S302 involves inputting the final normalized water depth, final normalized ambient temperature, and final normalized wind speed values ​​from the three-dimensional floating-point vector into a radial basis function neural network for processing, and outputting a predicted icing time. Specifically, this includes: The final normalized water depth, final normalized ambient temperature, and final normalized wind speed values ​​of the three-dimensional floating-point vector are input through the input layer to obtain the input three-dimensional floating-point vector. The input three-dimensional floating-point vector is transformed through the hidden layer using the Gaussian radial basis function to obtain the output value of each hidden layer neuron, expressed as: ； in, This represents the input vector, i.e., a three-dimensional floating-point vector. This represents the index number of the hidden layer neurons. =1,2,…, ,in, This represents the number of neurons in the hidden layer. Indicates the first The output values ​​of each hidden layer neuron, i.e., the activation values ​​of the Gaussian radial basis functions. Indicates the first The center vector of each hidden layer neuron, and the input vector The same dimensions Represents the input vector With the center vector The square of the Euclidean distance between them Indicates the first The expansion factor of each hidden layer neuron; where the expansion factor of each hidden layer neuron... Based on the dynamic calculation of the normalized water depth component, the expression is: ； ; in, Indicates the expansion factor. Represents the center vector The first dimension component is the water depth component.

7. The method for predicting highway disaster risks based on big data analysis as described in claim 6, characterized in that: The activation values ​​of the Gaussian radial basis functions of neurons in each hidden layer are... The predicted freezing time is obtained by performing linear weighted summation through the output layer; The expression for linear weighted summation is: ； in, Indicates the connection weights. This indicates the predicted freezing time, and the final predicted freezing time takes the value max(0, ).

8. The method for predicting highway disaster risks based on big data analysis as described in claim 7, characterized in that: Step S4 specifically includes: Step S401, classifying risk levels based on predicted icing time, specifically includes: When the predicted icing time is greater than 60 minutes, it is classified as a Level IV warning. When the predicted freezing time is greater than 30 minutes and less than or equal to 60 minutes, it is judged as a Level III warning; When the predicted icing time is greater than 10 and less than or equal to 30, it is judged as a level 2 warning; When the predicted freezing time is less than or equal to 10 hours, it is classified as a Level 1 warning. The risk levels are divided into four warning levels: Level 4, Level 3, Level 2, and Level 1.

9. The method for predicting highway disaster risks based on big data analysis as described in claim 8, characterized in that: Step S4 also includes: Step S402: Obtain the predicted icing time, risk level, and environmental relative humidity, road surface temperature, and rainwater temperature from the monitoring and display data; The predicted icing time, risk level, ambient relative humidity, road surface temperature, and rainwater temperature are combined into structured early warning information. By releasing structured early warning information through the traffic management platform, we can obtain road disaster risk warning information after the release.

10. A highway disaster risk prediction system based on big data analysis, which is applied to the highway disaster risk prediction method based on big data analysis as described in any one of claims 1-9, characterized in that, It includes a processing module, a transformation module, a prediction module, and a partitioning module; The processing module preprocesses the raw data to obtain the raw multidimensional time series data stream; The transformation module splits the multidimensional time series data to obtain the predicted core feature data stream, and performs a nonlinear transformation on the normalized predicted core feature data stream to obtain a three-dimensional floating-point vector. The prediction module inputs a three-dimensional floating-point vector into a radial basis function neural network with an expansion factor to predict the icing time. The system is divided into modules, risk levels are classified according to the predicted freezing time, and structured early warning information is generated and released.

Images