An intelligent monitoring and safe passing system for expressway

Abstract

The application relates to the technical field of expressway safety management, and discloses an expressway intelligent monitoring safe passing system. A data acquisition module of the system is used for acquiring historical accident records of multiple monitoring points of an expressway; a reference point determination module selects a reference monitoring point of a similar accident mode for each monitoring point based on the historical accident records; a risk assessment module assesses a risk assessment value of each monitoring point according to a traffic flow change trend and an accident frequency of the reference monitoring point; an association analysis module analyzes an initial association degree between traffic flow and accidents of all monitoring points to obtain a comprehensive association degree; a feature extraction module extracts non-risk related features and calculates feature saliency; a time point selection module selects key time points based on the feature saliency and the comprehensive association degree; and a safety monitoring module trains a safe passing prediction model by using data of the key time points, monitors the expressway in real time, and outputs a risk early warning.

Description

Technical Field

[0001] This invention relates to the field of highway safety management technology, specifically to a highway intelligent monitoring and safe passage system. Background Technology

[0002] Current highway safety monitoring primarily employs a combination of fixed threshold alarms and post-event statistical analysis. Existing technologies typically assess risk based on independent data from individual monitoring points, lacking consideration of the overall road network's interconnectedness. This isolated analysis model fails to identify potential similarities in accident patterns across different road sections, resulting in inefficient data utilization. Traditional methods focus on obvious risk factors such as speeding and congestion, neglecting predictive information potentially contained in non-traditional features like weather conditions, time periods, and vehicle type distribution. Model training commonly utilizes data from all time periods, failing to differentiate the varying contributions of data from different time periods to prediction accuracy. Historical data contains a large amount of redundant information with low correlation to accidents; directly using this data to train models reduces prediction accuracy.

[0003] Existing accident prediction models often use raw data directly, lacking mechanisms for filtering based on feature saliency and time point importance. Monitoring systems require intelligent analysis methods capable of identifying correlations between accident patterns across road segments, mining deep predictive features, and optimizing training data selection. A major challenge in highway safety management is extracting effective information from massive amounts of monitoring data to build accurate accident prediction models. Existing systems have significant shortcomings in data processing, failing to fully utilize the spatial correlation characteristics of road network data and effectively identifying truly valuable training samples. This results in prediction models often exhibiting problems such as response lag or high false alarm rates, making it difficult to meet the needs of real-time safety monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent monitoring and safe passage system for highways to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an intelligent monitoring and safe passage system for highways, the system comprising: The data acquisition module is used to collect historical accident records from multiple monitoring points on the highway. The reference point determination module is used to select reference monitoring points with similar accident patterns for each monitoring point based on historical accident records; The risk assessment module is used to evaluate the risk assessment value of each monitoring point based on the traffic flow change trend and accident frequency of the reference monitoring points. The correlation analysis module is used to analyze the initial correlation between traffic flow and accidents at all monitoring points, and correct the initial correlation to obtain the comprehensive correlation. The feature extraction module is used to extract non-risk-related features from accident data and calculate the significance of the features; The time point selection module is used to filter key time points based on feature significance and overall correlation. The safety monitoring module uses data from key time points to train a safe passage prediction model, and applies the prediction model to monitor highways in real time and output risk warnings.

[0006] Preferably, the risk assessment module is specifically used for: for a target monitoring point, collecting traffic flow data from all its reference monitoring points during the target time period, calculating the difference between the maximum and minimum traffic flow values ​​for each reference monitoring point as the fluctuation amplitude, and performing inverse normalization on the fluctuation amplitude to obtain a traffic flow stability index; simultaneously, calculating the dispersion of the number of accidents at all reference monitoring points as the accident volatility; aligning the traffic flow stability index sequence and the accident volatility sequence, calculating the Pearson correlation coefficient between the two sequences and converting it into a weight value to obtain the risk assessment weight; and multiplying the risk assessment weight by the traffic flow stability index to obtain the risk assessment value of the target monitoring point.

[0007] Preferably, the correlation analysis module analyzes the correlation strength between traffic flow and accidents at all monitoring points, including: The system acquires the average traffic flow data of all monitoring points during the target time period to form a traffic flow distribution sequence; it also acquires the number of accidents at all monitoring points to form an accident distribution sequence; the system sorts the traffic flow distribution sequence and the accident distribution sequence to generate an ordered traffic flow sequence and an ordered accident sequence; it uses a dynamic regularization algorithm to calculate the alignment distance between the two ordered sequences and converts the alignment distance into an initial value for the association strength; and it smooths the initial value for the association strength to obtain the initial association degree.

[0008] Preferably, the process of correcting the initial correlation degree to obtain the comprehensive correlation degree in the correlation analysis module includes: The monitoring point numbers are sorted in descending order based on the number of accidents at each monitoring point to generate an accident number sequence. The monitoring point numbers are then sorted in ascending order based on traffic flow data to generate a traffic flow number sequence. The consistency of numbers at the same location is compared between the accident number sequence and the traffic flow number sequence, and the proportion of consistent numbers is calculated as the number matching degree. Consistent numbers are removed from both the accident number sequence and the traffic flow number sequence to obtain the remaining accident number sequence and the remaining traffic flow number sequence. The risk assessment values ​​of the monitoring points corresponding to the remaining numbers are combined into a new sequence, and the similarity score of the new sequence is calculated as the assessment matching degree. The number matching degree and the assessment matching degree are weighted and fused to generate a correlation correction factor. The correlation correction factor is multiplied by the initial correlation degree to obtain the comprehensive correlation degree.

[0009] Preferably, the feature extraction module is specifically used for: For the target monitoring point, obtain the accident distribution sequence of all its reference monitoring points to form a reference accident sequence; compare the reference accident sequence with the accident distribution sequence of all monitoring points, and use the dynamic regularization algorithm to calculate the deviation between the two as the risk distribution deviation; calculate the average risk distribution deviation of all monitoring points, and perform reverse mapping on the average value to obtain the significance of non-risk features.

[0010] Preferably, the time point selection module is specifically used for: For each time point, the overall correlation is divided by the significance of non-risk features to obtain a ratio value; the ratio value is normalized to generate a risk assessment importance score; an importance threshold is set, and time points with a risk assessment importance score greater than the threshold are marked as key time points.

[0011] Preferably, the safety monitoring module is specifically used for: Traffic flow and accident data from all monitoring points at key time points are collected to form a training dataset. The training dataset is used to build a support vector machine model, and the model parameters are optimized through grid search. After training, a safe passage prediction model is obtained.

[0012] Preferably, the safety monitoring module is further used for: Real-time traffic flow data from highway monitoring points is collected, and traffic flow characteristics corresponding to key time points are extracted. The traffic flow characteristics are input into a safe passage prediction model, and the model outputs the probability of accident risk. When the probability of risk exceeds a preset threshold, a risk warning signal is generated.

[0013] Preferably, the reference point determination module is specifically used for: For the target monitoring point, calculate the absolute difference between the number of accidents occurring at that point and the number of accidents occurring at all other monitoring points; set a threshold for the difference, and select monitoring points whose absolute difference is less than the threshold as reference monitoring points.

[0014] Preferably, the alignment distance calculation using the dynamic regularization algorithm in the association analysis module includes: interpolating the ordered traffic flow sequence and the ordered accident sequence to make the two sequences have the same length; calculating the Euclidean distance matrix of the interpolated sequences; finding the minimum cumulative distance through the dynamic regularization path and using the minimum cumulative distance as the alignment distance; mapping the alignment distance to the range of zero to one to obtain the initial value of the association strength.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Based on historical accident records, reference monitoring points with similar accident patterns were selected for each monitoring point, establishing a risk correlation network between road segments. The risk of the target monitoring point was assessed by analyzing traffic flow trends and accident frequencies at the reference monitoring points, enabling cross-regional transmission and complementarity of risk information. This method fully utilizes richer data resources within the road network, especially for road segments with fewer accident samples, leveraging empirical data from similar road segments to improve the reliability of risk assessment. Analyzing the initial correlation between traffic flow and accidents at all monitoring points, and then correcting for this correlation to obtain a comprehensive correlation, allows for a more accurate characterization of common accident patterns and their intrinsic relationship with traffic flow parameters. This spatial correlation analysis overcomes the limitations of independent operation at each monitoring point in traditional systems, resulting in better global consistency in risk assessment results.

[0016] Extracting features traditionally considered irrelevant to risk from accident data, such as traffic flow characteristics at different times and specific weather patterns, and calculating the significance of these features, helps in discovering new predictive indicators. Selecting key time points based on feature significance and overall correlation allows model training to focus on data segments with greater information content. This data selection mechanism improves the quality of training samples and avoids interference from a large amount of low-value data. Training the safe passage prediction model using the selected key time point data allows the model to focus more on learning the key patterns that truly affect driving safety, improving prediction accuracy. Applying the optimized prediction model during real-time monitoring enables earlier identification of potential risk situations, allowing more time for preventative measures. This model training method based on data quality optimization significantly improves the predictive model's adaptability to complex traffic scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent monitoring and safe passage system for highways as described in this invention. Figure 2 A flowchart for the risk assessment module; Figure 3 This is a flowchart of the association strength analysis in the association analysis module; Figure 4 A comparison chart of the initial and comprehensive correlation between traffic flow and accidents at highway monitoring points; Figure 5 Confusion matrix diagram for predictive model of safe passage for intelligent monitoring of highways. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1 This invention provides an intelligent monitoring and safe passage system for highways. The system includes: a data acquisition module, a reference point determination module, a risk assessment module, a correlation analysis module, a feature extraction module, a time point selection module, and a safety monitoring module. The data acquisition module collects historical accident records from multiple monitoring points on the highway. These records include basic information such as the time, location, and type of the accident. The reference point determination module selects reference monitoring points with similar accident patterns for each monitoring point based on historical accident records, thereby establishing a similarity network of accident patterns. The risk assessment module evaluates the risk assessment value of each monitoring point based on the traffic flow change trend and accident frequency of the reference monitoring points. This value reflects the risk level of the monitoring point within a specific time period. The correlation analysis module analyzes the initial correlation between traffic flow and accidents at all monitoring points and obtains a comprehensive correlation through correction processing to more accurately characterize the correlation strength between traffic flow and accidents. The feature extraction module extracts non-risk-related features from the accident data and calculates feature significance. These features help identify non-risk factors in accident patterns. The time point selection module filters key time points based on feature significance and comprehensive correlation, representing moments of high risk or drastic change. The safety monitoring module uses data from key time points to train a safe passage prediction model, and applies the model to monitor highways in real time, outputting a risk warning signal when a risk is detected.

[0020] Example 1: See Figure 2 In practical implementation, for the target monitoring point, it is first necessary to collect traffic flow data from all its reference monitoring points within the target time period. Traffic flow data includes parameters such as flow rate, speed, and density, and is stored in time series format. Each reference monitoring point corresponds to an independent sequence. The target time period can be adjusted according to actual monitoring needs, such as using daily, weekly, or monthly units to ensure data coverage of a representative period. After collection, the difference between the maximum and minimum traffic flow values ​​for each reference monitoring point is calculated as the fluctuation amplitude. The fluctuation amplitude reflects the range of traffic flow changes within the target time period. The calculation requires traversing the traffic flow sequence of each reference monitoring point, identifying peaks and troughs in the sequence, and calculating the absolute difference. A larger fluctuation amplitude indicates more severe traffic flow fluctuations. In practical implementation, the calculation of fluctuation amplitude needs to handle potential outliers or missing data, using interpolation or data removal methods to ensure data integrity and the accuracy of subsequent indicators.

[0021] The fluctuation amplitude is inversely normalized to obtain the traffic flow stability index. The inverse normalization process uses a min-max scaling method, mapping the original fluctuation amplitude value to the 0-1 range and then inverting it. This results in larger fluctuation amplitudes corresponding to smaller index values, thus a smaller traffic flow stability index value indicates more unstable traffic flow. In practice, the normalization process needs to determine the minimum and maximum values ​​based on the fluctuation amplitude set of all reference monitoring points and perform a linear transformation. The transformation formula is: Stability Index = 1 - (Fluctuation Amplitude - Minimum Value) / (Maximum Value - Minimum Value), ensuring that the index value range is between 0 and 1 for easy comparison across monitoring points. Simultaneously, the dispersion of the number of accidents at all reference monitoring points is calculated as the accident volatility. Dispersion is measured by standard deviation or variance; higher accident volatility indicates stronger randomness in accident occurrences. The calculation requires summarizing the number of accidents at each reference monitoring point within the target time period to form an accident frequency sequence, and then applying statistical formulas to obtain the discrete values. In practice, standard deviation is preferred because it directly reflects the degree of data dispersion. During calculation, it must be ensured that the accident frequency data has been cleaned and free of duplicates or errors.

[0022] The traffic flow stability index series and the accident volatility series are aligned. Alignment ensures a one-to-one correspondence between the time points and consistent series length. In practice, alignment methods include timestamp matching or interpolation compensation. For example, if the time granularity of the series differs, downsampling or upsampling is required to ensure they have the same time dimension. After alignment, the Pearson correlation coefficient between the two series is calculated to assess the linear correlation strength between traffic flow stability and accident volatility. The Pearson correlation coefficient ranges from -1 to 1, with positive values ​​indicating a positive correlation and negative values ​​indicating a negative correlation. The calculation requires first obtaining the mean and squared difference of the series before applying the correlation coefficient formula. In practice, the correlation coefficient calculation requires that the number of series elements be equal and that there are no missing values. Data imputation is performed if necessary. After calculation, the Pearson correlation coefficient is converted into weight values. Conversion methods include taking the absolute value or normalization to ensure the weight values ​​are between 0 and 1. For example, adding 1 to the correlation coefficient and dividing by 2 makes it fall into the 0-1 range. A larger weight value indicates a more significant correlation.

[0023] The risk assessment weights are multiplied by the traffic flow stability index to obtain the risk assessment value of the target monitoring point. Before the multiplication operation, it is necessary to ensure that the weights and index dimensions are consistent, and standardization should be performed if necessary. The risk assessment value integrates the influence of traffic flow dynamics and accident patterns; a higher value indicates a higher risk level. In specific implementation, the multiplication process is performed element-wise, meaning each time point corresponds to a product value. The average or weighted sum of the product series is then used as the final assessment value, which is stored in the database for subsequent modules to use. In some embodiments, the traffic flow data collection frequency can be set to per minute or per hour to meet real-time requirements. The calculation of fluctuation amplitude can be optimized using a sliding window method to dynamically update the amplitude value. Optionally, the mean absolute deviation can also be used as an alternative indicator to calculate the dispersion of accident fluctuations, enhancing robustness. It is understood that the calculation of the Pearson correlation coefficient assumes a linear relationship in the sequence; if the data is highly nonlinear, the Spearman correlation coefficient can be considered.

[0024] In some embodiments, the target time period can be dynamically adjusted based on historical accident peak periods or seasonal traffic flow changes, making the assessment more closely aligned with actual risk patterns. In the reverse normalization process, the mapping function can also employ logarithmic or exponential transformations to adapt to different data distribution characteristics. Optionally, the sequence alignment method can be combined with a dynamic time warping algorithm to handle the alignment problem of non-equal-length sequences, improving flexibility. It is understood that the transformation of risk assessment weights needs to avoid division-by-zero errors, and numerical stability is ensured by adding a small constant. In specific implementations, the module output also includes the confidence interval of the assessment value, calculated through resampling or Monte Carlo methods, providing a measure of uncertainty in risk estimation. In specific implementations, the traffic flow stability index sequence and accident volatility sequence are stored in a time-series database, supporting fast querying and updating; the alignment operation uses database connection functionality to achieve timestamp matching. In specific implementations, the calculation process of the Pearson correlation coefficient includes data standardization, covariance calculation, and ratio calculation; each step requires verification of the data normality assumption and transformations when necessary.

[0025] Example 2: See Figure 3The implementation of the correlation analysis module includes analyzing the initial correlation between traffic flow and accidents at all monitoring points, and correcting the initial correlation to obtain a comprehensive correlation. The calculation of the initial correlation first requires obtaining the average traffic flow data of all monitoring points over the target time period. The average traffic flow data is processed by time aggregation of the raw traffic flow data to form a traffic flow distribution sequence. Time aggregation can be performed on an hourly, daily, or weekly basis to ensure the sequence reflects the overall distribution characteristics of traffic flow. In practice, the calculation of the average traffic flow data is based on cleaned raw data, removing outliers and missing points, and using an arithmetic mean or weighted average method to generate a traffic flow value sequence for each monitoring point. The sequence length is determined by the target time period; for example, data from the past 30 days can be selected as the basis for analysis. Simultaneously, the number of accidents at all monitoring points is obtained. The number of accidents is aggregated according to the same time period to form an accident distribution sequence. Accident data includes accident type and severity, but the number of accidents is used as a quantitative indicator during aggregation to simplify analysis. In practice, the construction of the accident distribution sequence must ensure time alignment, meaning the number of accidents at each monitoring point corresponds to the same time point in the traffic flow data, avoiding time misalignment that could affect the correlation analysis.

[0026] The traffic flow distribution sequence and accident distribution sequence are sorted based on numerical values ​​to generate ordered traffic flow and accident sequences. The sorting operation transforms the original sequences into monotonically increasing or decreasing sequences for easier comparison later. In practice, quicksort or mergesort algorithms are used to ensure efficiency while maintaining the original relationships between sequence elements. The generation of ordered sequences requires retaining monitoring point index information for subsequent correction steps. A dynamic warping algorithm is used to calculate the alignment distance between two ordered sequences. Dynamic warping can handle cases with inconsistent sequence lengths by finding the minimum cumulative path to achieve sequence alignment. In practice, the application of dynamic warping includes sequence interpolation, distance matrix calculation, and path finding. Sequence interpolation ensures the two sequences have the same length, often using linear interpolation methods. The distance matrix is ​​calculated based on Euclidean distance to determine the difference between each pair of points. Path finding uses dynamic programming to solve for the minimum cumulative distance. Alignment distance reflects the morphological differences between two sequences; a smaller distance indicates a stronger association. The alignment distance is converted into an initial value for association strength. Conversion methods include mapping to the 0-1 range, for example, using the formula: Initial value = 1 / (1 + Distance), making the initial value larger as the distance smaller. The initial association strength value is then smoothed using moving averages or Gaussian filtering to eliminate random fluctuations, yielding an initial association degree. This initial association degree serves as a preliminary estimate of the association strength.

[0027] The process of correcting the initial correlation to obtain the comprehensive correlation involves multiple data processing steps. First, the monitoring point numbers are sorted in descending order based on the number of accidents at each monitoring point, generating an accident number sequence. The order of the numbers in the sequence corresponds to the accident frequency from high to low. In practice, the construction of the accident number sequence must be based on a complete accident dataset. The numbers use unique identifiers for each monitoring point, and the length of the sorted sequence is consistent with the number of monitoring points. Second, the monitoring point numbers are sorted in ascending order based on traffic flow data, generating a traffic flow number sequence. The order of the numbers in the sequence corresponds to the traffic flow from low to high, and the traffic flow data uses the average flow rate as the sorting criterion. Third, the consistency of numbers at the same location is compared between the accident number sequence and the traffic flow number sequence. The proportion of consistent numbers is calculated as the number matching degree. For example, if two sequences have the same number at index i, the count is incremented by one. The final proportion is the number of consistent numbers divided by the total length. In practice, the number consistency comparison needs to traverse each location, implemented using loops or vectorized operations. The calculated proportion is stored as a scalar value. Remove identical numbers from the accident number sequence and traffic flow number sequence to obtain the remaining accident number sequence and remaining traffic flow number sequence. The removal operation is completed by set difference or list filtering to ensure that the remaining sequence does not contain identical numbers.

[0028] The risk assessment values ​​of the remaining monitoring points are grouped into a new sequence. This new sequence contains the risk assessment values ​​of these monitoring points, which are output from the risk assessment module. A similarity score is calculated for the new sequence as the assessment matching degree. The similarity score can be measured by Pearson correlation coefficient or cosine similarity, reflecting the correlation strength of the risk assessment value sequence. In specific implementations, the similarity score calculation requires sequences of equal length. Pearson correlation coefficient is preferentially used for linear relationship assessment, and the sequence values ​​need to be standardized during calculation. The number matching degree and assessment matching degree are weighted and fused to generate a correlation correction factor. The weighted fusion uses a linear combination, for example, correction factor = w1 * number matching degree + w2 * assessment matching degree. The weights w1 and w2 are set according to requirements, such as 0.5 each. The correlation correction factor is multiplied by the initial correlation degree to obtain the comprehensive correlation degree. The multiplication operation is performed element-wise or as a whole. The comprehensive correlation degree is ultimately used in the time point selection module. In some embodiments, the sequence interpolation of the dynamic warping algorithm can use cubic spline interpolation to improve accuracy and adapt to non-linear sequence shapes. Optionally, the alignment distance transformation function can also use an exponential decay form, such as initial value = exp(-distance), to enhance sensitivity to small distances. It's understandable that the calculation of the number matching degree assumes a fixed sequence order; if the number of monitoring points is large, it can be optimized to hash comparison to improve speed.

[0029] In some embodiments, when smoothing the initial correlation, the window size of the moving average can be dynamically adjusted, selecting a 3-point or 5-point average based on the data noise level. The aggregation of accident distribution sequences can incorporate accident severity weights, making the frequency calculation more accurately reflect the risk impact. Optionally, the similarity score calculation for assessing matching degree can be replaced by the reciprocal of the dynamic time-warped distance, directly utilizing sequence morphological similarity. It is understood that the weight settings in weighted fusion can be optimized through grid search, verifying the optimal ratio based on historical data. In specific implementations, the calculation of the comprehensive correlation degree requires iterating over all monitoring points; matrix operations are used during batch processing to reduce loops and improve performance. In specific implementations, the smoothing of the initial correlation degree can be combined with a low-pass filter, such as a Butterworth filter, to preserve trend components. In specific implementations, the calculation of the number matching degree ratio needs to handle division-by-zero errors, returning a default value when the sequence is empty. In specific implementations, the path finding of the dynamic warping algorithm can constrain the curvature coefficient, controlling the flexibility of the alignment path and avoiding overfitting. In specific implementations, the output format of the comprehensive correlation degree is either a time series or a scalar value, with the storage method selected according to the application scenario.

[0030] Optionally, a sliding window mechanism can be introduced to generate traffic flow distribution sequences, updating the sequence content in real time to adapt to flow data. It can be understood that the descending order of accident number sequences can be expanded to multi-indicator sorting, such as combining accident type and frequency. In specific implementations, the calculated sequence for assessing matching degree can cache risk assessment values ​​to reduce duplicate data access. In specific implementations, the weighted fusion of correlation correction factors can normalize matching degree values, ensuring the factor range is between 0 and 1. In specific implementations, the calculation of the Euclidean distance matrix of the dynamic warping algorithm can be parallelized to accelerate large-scale sequence alignment. In specific implementations, the smoothing of initial correlation degree can verify the filtering effect through residual analysis, ensuring signal fidelity. In specific implementations, the multiplication operation of comprehensive correlation degree can introduce nonlinear functions, such as the Sigmoid transform, to enhance the adjustment effect of the correction factors. In specific implementations, the entire correction process can be encapsulated as an independent function, supporting modular calls and parameter configuration.

[0031] See Figure 4This figure visually presents a comparison of the initial and comprehensive correlation scores between traffic flow and accidents at 10 highway monitoring points. The initial correlation score is obtained by calculating the alignment distance between ordered traffic flow sequences and ordered accident sequences using a dynamic regularization algorithm, converting it into an initial correlation strength value, and then smoothing it. The comprehensive correlation score is the final correlation strength obtained by weighted correction based on the initial correlation score, combined with the monitoring point number matching degree and the evaluation matching degree. The figure shows that the comprehensive correlation score for most monitoring points is higher than the initial correlation score, indicating that the weighted fusion correction of number matching degree and evaluation matching degree more accurately depicts the intrinsic connection between traffic flow and accidents. This figure clearly demonstrates the value of the initial correlation score calculation—comprehensive correlation score correction process in the correlation analysis module, providing a reliable basis for the subsequent time point selection module to screen key time points and the safety monitoring module to train prediction models. This helps the highway intelligent monitoring system more accurately identify the correlation patterns between accidents and traffic flow, improving the accuracy of risk assessment and early warning.

[0032] Example 3: The implementation of the feature extraction module involves extracting non-risk-related features from accident data and calculating feature significance. For a target monitoring point, the accident distribution sequence of all its reference monitoring points is obtained. The accident distribution sequence is formed by arranging the number of accidents at the reference monitoring points in chronological order, and the sequence length is determined by the historical analysis period. In specific implementation, the construction of the accident distribution sequence requires querying the accident records of the reference monitoring points from the database, sorting them by timestamp, and aggregating them into a discrete sequence. Each element in the sequence represents the number of accidents within a time unit, which can be a day, week, or month, depending on the required granularity of analysis. The reference accident sequence is compared with the accident distribution sequences of all monitoring points. The accident distribution sequences of all monitoring points are obtained by aggregating the accident data of the entire highway network to form a global accident pattern benchmark. The comparison process needs to ensure that the time range and granularity of all sequences are completely consistent. The dynamic warping algorithm is used to calculate the deviation value between the reference accident sequence and the accident sequence of each monitoring point. The dynamic warping algorithm measures the difference in sequence shape by finding the optimal alignment path, and the deviation value is calculated based on the cumulative distance of the alignment path. In its implementation, the dynamic regularization algorithm includes an initialization phase, a cumulative distance calculation phase, and a path backtracking phase. The distance matrix uses Euclidean distance as the basis function, and the cumulative distance calculation satisfies the dynamic programming recurrence relation D(i,j)=dist(i,j)+min(D(i-1,j),D(i,j-1),D(i-1,j-1)). The deviation values ​​of all monitoring points are averaged to obtain the average deviation value. The averaging method uses an arithmetic mean, summing all deviation values ​​and dividing by the total number of monitoring points. The average deviation value is then normalized by dividing by a reference value, making the normalized deviation value dimensionless. The reference value can be the maximum or median of the deviation values ​​among all monitoring points. The normalized average deviation value undergoes a reverse mapping process, achieved through a nonlinear transformation. Smaller deviation values ​​indicate higher feature significance, thus yielding the non-risk feature significance. The non-risk feature significance reflects the degree of difference between the accident mode of the target monitoring point and the overall network; higher values ​​indicate more prominent non-risk factors. The formula for calculating the significance of a feature is expressed as follows: in: Represents the significance of non-risk characteristics. The arithmetic mean of the dynamic normalization deviation values ​​of all monitoring points. The reference value for the deviation value is the maximum value of the deviation values ​​between all monitoring points.

[0033] The implementation of the time point selection module is based on feature significance and comprehensive correlation to screen key time points. For each time point, the granularity can be hour, day, or month, etc. The comprehensive correlation is divided by the non-risk feature significance to obtain the ratio value. In specific implementation, the ratio value calculation must ensure that both the comprehensive correlation and feature significance have been normalized and are within the same numerical range. The comprehensive correlation comes from the output of the correlation analysis module, and the feature significance comes from the calculation results of the feature extraction module. Before the division operation, it is necessary to check whether the feature significance is zero. A very small positive number ε is added to prevent division by zero error. The ε value is usually set to 1e-8. The larger the ratio value, the stronger the risk correlation of that time point, because high comprehensive correlation corresponds to strong accident correlation, and low feature significance corresponds to high risk specificity. The ratio value is normalized using the min-max scaling method to map the ratio value to the 0-1 range, generating a risk assessment importance score. The normalization process requires first calculating the minimum and maximum values ​​of the ratio values ​​for all time points, and then applying the linear transformation formula: Score = (Ratio value - Minimum value) / (Maximum value - Minimum value). An importance threshold is set, which can be determined based on the quantiles of historical data distribution, for example, using the 80th percentile of the importance score. Time points where the risk assessment importance score is greater than the threshold are marked as key time points. These key time points are used for subsequent model training and represent periods of high risk. In practice, the threshold setting can employ an adaptive method, calculating local thresholds through a sliding window to adapt to dynamic data changes. The marking operation generates a binary sequence or a list of time point indices.

[0034] In some embodiments, the deviation calculation of the dynamic normalization algorithm can introduce a weighting mechanism, assigning higher weights to recent data to enhance timeliness. The construction of the accident distribution sequence can consider weighting based on accident severity, for example, giving a higher weight to major accidents than minor accidents. Optionally, the calculation of the average deviation value can be replaced by a weighted average, allocating weights according to the importance of the monitoring points. It is understood that the reverse mapping process can also adopt an exponential function form, such as S=exp(-average deviation value), to enhance the sensitivity of small deviation values. In specific implementations, the feature saliency calculation process can be parallelized to process multiple target monitoring points, utilizing multi-threading technology to improve computational efficiency. In specific implementations, the normalization of ratio values ​​needs to update the minimum and maximum values ​​in real time, employing a streaming computation method to adapt to continuous data inflow. In specific implementations, the determination of the importance threshold can be combined with a clustering algorithm to automatically identify the natural boundary points of importance scores.

[0035] In some embodiments, the choice of time granularity can be dynamically adjusted based on business needs; for example, hourly granularity can be used during peak periods, and daily granularity during off-peak periods. A sliding time window can be introduced into the calculation of non-risk feature saliency, periodically updating the saliency value to reflect pattern changes. It can be understood that the division operation in the ratio calculation can be replaced with a logarithmic ratio, i.e., log(overall correlation / feature saliency), to improve the value range distribution. In specific implementations, the marking results of key time points can be post-processed, for example, merging adjacent time points to form consecutive risk periods. In specific implementations, the path curvature coefficient of the dynamic warping algorithm can be configured to control the flexibility of sequence alignment.

[0036] In some embodiments, the temporal alignment of accident distribution sequences can be combined with external event data, such as weather or holiday information, to enhance the contextual relevance of sequence comparisons. Ratio calculation can introduce a confidence interval for comprehensive correlation; when the confidence interval is too wide, the time point is weighted less. It is understood that the selection of key time points can be combined with time series anomaly detection algorithms to identify abrupt changes in importance scores as supplementary key points. In specific implementations, the storage format for non-risk feature saliency uses a time series database, supporting fast querying and backtracking analysis. In specific implementations, the normalization of importance scores can use the Z-score standardization method, making the scores follow a standard normal distribution for easier threshold setting. In specific implementations, the computational optimization of the dynamic warping algorithm includes using an early termination strategy, stopping the calculation when the accumulated distance exceeds a set upper limit to improve efficiency.

[0037] Optionally, the average deviation value in the feature significance calculation can be replaced with the median deviation value to enhance robustness against outliers. In practice, the output of the time point selection module can include confidence scores for key time points, calculated based on the distribution characteristics of the ratio values. Throughout the implementation process, data quality must be monitored, and an alarm mechanism should be triggered when the incident data missing rate exceeds a threshold. The update cycle for non-risk feature significance can be configured to synchronize with the data acquisition cycle to ensure feature timeliness. The marking results of key time points can be visualized to help managers understand risk time distribution patterns. The implementation of the dynamic warping algorithm can utilize built-in functions of a dedicated time-series database to improve computational performance and accuracy.

[0038] Example 4: The implementation of the safety monitoring module includes training a safe passage prediction model using data from key time points, and applying the prediction model to perform real-time monitoring and risk warnings for highways. The training process begins with collecting traffic flow data and accident data from all monitoring points at key time points. Key time points are derived from the output of the time point selection module and represent high-risk periods. In practice, traffic flow data includes parameters such as flow rate, speed, and density, while accident data includes the number, type, and severity of accidents. Data collection must ensure time alignment, meaning each key time point corresponds to a complete dataset of monitoring points. When forming the training dataset, data is organized by time point as the sample unit. Each sample contains a feature vector and a label. The feature vector is composed of traffic flow parameters, and the label is a binary accident risk identifier; for example, an accident is recorded as 1, and no accident is recorded as 0. In practice, the construction of the training dataset requires data cleaning to handle missing and outlier values. Missing values ​​are filled using interpolation, and outliers are identified and corrected using statistical methods. The dataset is then divided into a training set and a test set, typically in a ratio of 7:3 or 8:2. A Support Vector Machine (SVM) model is built using a training dataset. SVM is a supervised learning algorithm suitable for classification tasks, which finds the optimal hyperplane to separate data categories. In practice, building the SVM model includes selecting a kernel function and setting a penalty coefficient. The kernel function can be a linear kernel, a polynomial kernel, or a radial basis function kernel, and the initial parameters are set empirically. The model parameters are optimized using a grid search, a parameter tuning method that exhaustively searches for preset parameter combinations and evaluates the model performance of each set of parameters. In practice, the parameter space of the grid search includes the kernel function type, the penalty coefficient C, and the kernel function parameter gamma. The parameter range is determined through pre-experiments; for example, C values ​​range from 0.1 to 10, and gamma values ​​range from 0.001 to 1. Cross-validation is used during training to evaluate model performance and prevent overfitting. Cross-validation divides the training set into k folds, using k-1 folds for training and 1-fold for validation. After training, a safe passage prediction model is obtained. Model performance is evaluated using metrics such as accuracy and recall, and after achieving the required metrics, it is serialized and stored.

[0039] When applying the safe traffic prediction model for real-time monitoring, traffic flow data from highway monitoring points is collected in real time. Data collection is achieved through a sensor network deployed on the road, with a sampling frequency consistent with the training data. In implementation, the real-time data stream undergoes preprocessing, including denoising, alignment, and standardization, to ensure data quality compatibility with the training phase. Traffic flow features corresponding to key time points are extracted using methods consistent with the training phase, such as calculating statistical quantities like mean traffic flow and speed variance. These traffic flow features are input into the safe traffic prediction model, which outputs an accident risk probability. The probability value ranges from 0 to 1, representing the likelihood of an accident occurring at the current time. When the risk probability exceeds a preset threshold, a risk warning signal is generated. The preset threshold is set based on historical data recall, such as 0.7 or 0.8. In implementation, the warning signal is output through the monitoring center interface, triggering audible and visual alarms or message pushes to prompt management personnel to intervene. The entire real-time monitoring process must ensure low latency, with data acquisition and warning output completed within seconds. Refer to Table 1 for the feature structure of the training dataset.

[0040] Table 1: Feature Description of Training Dataset In practice, the training of the Support Vector Machine (SVM) model utilizes popular machine learning libraries to achieve a standardized process. The evaluation metric for grid search can be the F1 score, balancing accuracy and recall. The k-value for cross-validation is typically set to 5 or 10 to balance bias and variance. After model serialization, it is deployed to the production environment and updated periodically to adapt to changes in data distribution. Real-time data acquisition supports streaming frameworks such as Apache Kafka for high throughput. The feature extraction module can be configured as a microservice, independently scaling its processing capabilities. After the early warning signal is generated, it can be integrated into the existing monitoring system for automated response. In practice, model prediction results must be logged for subsequent auditing and optimization. During the data cleaning phase, data consistency must be verified, for example, checking whether traffic flow parameters are within reasonable ranges.

[0041] In some embodiments, the label definition of the training dataset can be expanded to multi-class classification, distinguishing accident types or severity levels. The kernel function selection for the support vector machine model can be automatically determined through data distribution analysis; for example, a linear kernel can be used for linearly separable data. Optionally, grid search can be replaced with random search or Bayesian optimization to improve parameter search efficiency. It is understood that a sliding window mechanism can be introduced into the risk probability calculation in real-time monitoring to smooth the output by integrating probabilities from multiple time points. In specific implementations, the model update cycle can be set to monthly or quarterly, adapting to new data through incremental learning. In specific implementations, the warning threshold can be dynamically adjusted, optimizing sensitivity based on real-time feedback. In specific implementations, a feature selection step can be added to the feature extraction process to remove redundant features and improve the model's generalization ability.

[0042] See Figure 5 This figure shows the confusion matrix of the safe passage prediction model in the safety monitoring module, used to evaluate the accuracy of the model in predicting risk levels. The vertical axis of the matrix represents the actual risk level, divided into four categories: low risk, low-medium risk, medium-high risk, and high risk; the horizontal axis represents the predicted risk level, with the categories consistent with the actual risk levels. The value of each cell is a normalized probability, and the color intensity corresponds to the probability magnitude. This confusion matrix intuitively reflects the model's performance in risk level prediction, providing data support for optimizing model parameters and improving the accuracy of real-time risk warnings for highways. It is a key visualization tool for the safety monitoring module to evaluate the reliability of the prediction model.

[0043] Example 5: For a target monitoring point, calculate the absolute difference between its accident occurrence count and the accident occurrence counts of all other monitoring points. The absolute difference reflects the similarity of accident frequencies. In a specific implementation, assume the highway network contains five monitoring points, numbered M001, M002, M003, M004, and M005. The target monitoring point is M001. In the past year, M001 had 15 accidents, M002 had 18 accidents, M003 had 5 accidents, M004 had 16 accidents, and M005 had 22 accidents. Calculate the absolute difference between M001 and M002 as |15-18|=3, the absolute difference between M001 and M003 as |15-5|=10, the absolute difference between M001 and M004 as |15-16|=1, and the absolute difference between M001 and M005 as |15-22|=7. A difference threshold is set, which can be determined based on historical data distribution. For example, the median of all absolute differences can be used. In the current example, the set of absolute differences is {3, 10, 1, 7}, and the median is 5. Monitoring points with absolute differences less than the threshold of 5 are selected as reference monitoring points. In the example, the difference 3 for M002 is less than 5, and the difference 1 for M004 is less than 5. Therefore, M002 and M004 are selected as reference monitoring points for M001. Reference monitoring points represent a set of monitoring points with accident patterns similar to the target monitoring point, which will be used for subsequent analysis in the risk assessment module.

[0044] The specific steps for calculating the alignment distance using the dynamic regularization algorithm in the association analysis module include sequence preprocessing and distance calculation. The ordered traffic flow sequence and the ordered accident sequence are interpolated to ensure they have the same length. In this implementation, the ordered traffic flow sequence is assumed to be [120, 135, 118], representing the number of vehicles passing through three time points, and the ordered accident sequence is [2, 3], representing the number of accidents at two time points. Due to the inconsistent sequence lengths, interpolation is required. For example, linear interpolation is used for the ordered accident sequence, inserting calculated values ​​at indices 0.5 and 1.5 to generate a new sequence of length 3: [2, 2.5, 3]. The Euclidean distance matrix of the interpolated sequences is then calculated. The Euclidean distance is calculated based on the geometric distance between sequence elements. For example, the distance at the first time point is |120 ​​- 2| = 118, the distance at the second time point is |135 - 2.5| = 132.5, and the distance at the third time point is |118 - 3| = 115, forming the initial distance matrix. The minimum cumulative distance is found through dynamic regularization (DOL). DOL allows sequences to bend and align along the time axis to find the best matching path. In practice, the DOL search uses a dynamic programming algorithm. The cumulative distance matrix is ​​initialized, and the minimum cumulative distance at each position is recursively calculated using the formula D(i,j)=dist(i,j)+min(D(i-1,j),D(i,j-1),D(i-1,j-1)). Finally, the path that minimizes the cumulative distance is found by backtracking. The minimum cumulative distance is used as the alignment distance; a smaller alignment distance indicates higher sequence similarity. The alignment distance is mapped to a range of zero to one to obtain the initial value of the association strength. The mapping function can be a linear transformation or exponential decay, for example, the initial value = 1 / (1+distance), making the initial value in the interval (0,1].

[0045] In practical implementation, the difference calculation of the reference point determination module requires traversing all monitoring points. For large-scale networks, index optimization can be used to accelerate the query. The difference threshold can be set based on statistical methods, such as using the mean plus or minus the standard deviation as the threshold range. In practical implementation, the selection results of reference monitoring points are stored as an adjacency list or matrix structure for easy and fast lookup. The interpolation processing of the dynamic warping algorithm can use spline interpolation or nearest neighbor interpolation to adapt to different sequence characteristics. The calculation of the Euclidean distance matrix can use vectorized operations to improve efficiency and avoid explicit loops. Solving for the minimum cumulative distance requires handling boundary conditions, such as the initialization of the sequence start point. The mapping processing must ensure monotonicity; the smaller the distance, the larger the initial value. In practical implementation, the entire process can process multiple monitoring point pairs in batches, accelerated by using a parallel computing framework.

[0046] In some embodiments, the absolute difference calculation of the reference point determination module can incorporate time weighting, assigning higher weight to recent incidents. The difference threshold can be dynamically adjusted, with real-time thresholds calculated based on a sliding window. Optionally, the selection of reference monitoring points can be combined with clustering algorithms to automatically group similar monitoring points. In specific implementations, the path finding of the dynamic warping algorithm can constrain the curved window, limiting the maximum curvature to control computational complexity. In specific implementations, the alignment distance mapping function can be replaced with a sigmoid function to enhance sensitivity in small distance regions. In specific implementations, the output of the reference point determination module can include a similarity score, rather than just a binary selection result.

[0047] Optionally, the absolute difference calculation can be extended to multi-indicator differences, such as combining accident type and severity. The difference threshold can be optimized through cross-validation to select the threshold that best performs subsequent models. In specific implementations, the interpolation method of the dynamic warping algorithm is configurable, automatically selecting the optimal interpolation strategy based on sequence smoothness. In specific implementations, the calculation of Euclidean distance can be replaced with Manhattan distance or cosine distance to adapt to different data distributions. In specific implementations, the backtracking of the minimum cumulative path can be optimized to an iterative method to reduce memory usage. In specific implementations, the mapping range of the initial value of the association strength can be adjusted, for example, mapping to the [0.1, 0.9] interval to avoid extreme values.

[0048] It is understandable that graph algorithms can be introduced into the similarity judgment of the reference point determination module to construct a monitoring point similarity network. In practical implementation, the computational accuracy of the dynamic regularization algorithm can be improved through floating-point optimization to enhance numerical stability. Outlier handling before sequence interpolation is crucial; obviously erroneous data points need to be filtered out. In practical implementation, the update mechanism for the reference monitoring point set needs to be executed periodically, such as monthly. In practical implementation, a length normalization step can be added to the alignment distance calculation to eliminate the influence of sequence length. In practical implementation, the entire module's implementation needs to consider real-time requirements and optimize computational bottlenecks.

[0049] Optionally, absolute difference calculation can use relative difference ratios to enhance cross-dimensional comparability. In implementation, difference thresholds can be set in zones, with different thresholds used for monitoring points in different areas. In implementation, a dedicated time-series analysis library can be used to implement the dynamic warping algorithm to improve accuracy and efficiency. In implementation, a moving average can be added to the smoothing of the initial association strength value to eliminate random fluctuations. In implementation, the log records of the reference point determination module should be detailed for easy debugging and auditing. In implementation, parameters of the dynamic warping algorithm, such as the curvature coefficient, can be optimized through grid search. In implementation, standardization before sequence alignment can improve the reliability of distance calculation. When the number of monitoring points in the reference point determination module is large, a sampling method can be used to approximate the calculation of absolute differences. In implementation, the parallelization of the dynamic warping algorithm can be accelerated using GPUs, and the calibration of the initial association strength value for large-scale sequence alignment can be corrected later through historical data regression.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart monitoring and safe passage system for highways, characterized in that, The system includes: The data acquisition module is used to collect historical accident records from multiple monitoring points on the highway. The reference point determination module is used to select reference monitoring points with similar accident patterns for each monitoring point based on historical accident records; The risk assessment module is used to evaluate the risk assessment value of each monitoring point based on the traffic flow change trend and accident frequency of the reference monitoring points. The correlation analysis module is used to analyze the initial correlation between traffic flow and accidents at all monitoring points, and correct the initial correlation to obtain the comprehensive correlation. The feature extraction module is used to extract non-risk-related features from accident data and calculate the significance of the features; The time point selection module is used to filter key time points based on feature significance and overall correlation. The safety monitoring module uses data from key time points to train a safe passage prediction model, and applies the prediction model to monitor highways in real time and output risk warnings.

2. The intelligent monitoring and safe passage system for highways as described in claim 1, characterized in that, The risk assessment module is specifically used for: for the target monitoring point, collecting traffic flow data of all its reference monitoring points in the target time period, calculating the difference between the maximum and minimum traffic flow values ​​of each reference monitoring point as the fluctuation range, and performing reverse normalization on the fluctuation range to obtain the traffic flow stability index; at the same time, calculating the dispersion of the number of accidents at all reference monitoring points as the accident volatility. The traffic flow stability index series and the accident volatility series are aligned. The Pearson correlation coefficient between the two series is calculated and converted into a weight value to obtain the risk assessment weight. The risk assessment weight is then multiplied by the traffic flow stability index to obtain the risk assessment value of the target monitoring point.

3. The intelligent monitoring and safe passage system for highways as described in claim 1, characterized in that, The correlation analysis module analyzes the correlation strength between traffic flow and accidents at all monitoring points, including: The system acquires the average traffic flow data of all monitoring points during the target time period to form a traffic flow distribution sequence; it also acquires the number of accidents at all monitoring points to form an accident distribution sequence; the system sorts the traffic flow distribution sequence and the accident distribution sequence to generate an ordered traffic flow sequence and an ordered accident sequence; it uses a dynamic regularization algorithm to calculate the alignment distance between the two ordered sequences and converts the alignment distance into an initial value for the association strength; and it smooths the initial value for the association strength to obtain the initial association degree.

4. The intelligent monitoring and safe passage system for highways as described in claim 3, characterized in that, The correlation analysis module, in which the initial correlation degree is corrected to obtain the comprehensive correlation degree, includes: The monitoring point numbers are sorted in descending order based on the number of accidents at each monitoring point to generate an accident number sequence. The monitoring point numbers are then sorted in ascending order based on traffic flow data to generate a traffic flow number sequence. The consistency of numbers at the same location is compared between the accident number sequence and the traffic flow number sequence, and the proportion of consistent numbers is calculated as the number matching degree. Consistent numbers are removed from both the accident number sequence and the traffic flow number sequence to obtain the remaining accident number sequence and the remaining traffic flow number sequence. The risk assessment values ​​of the monitoring points corresponding to the remaining numbers are combined into a new sequence, and the similarity score of the new sequence is calculated as the assessment matching degree. The number matching degree and the assessment matching degree are weighted and fused to generate a correlation correction factor. The correlation correction factor is multiplied by the initial correlation degree to obtain the comprehensive correlation degree.

5. The intelligent monitoring and safe passage system for highways as described in claim 3, characterized in that, The feature extraction module is specifically used for: For the target monitoring point, obtain the accident distribution sequence of all its reference monitoring points to form a reference accident sequence; compare the reference accident sequence with the accident distribution sequence of all monitoring points, and use the dynamic regularization algorithm to calculate the deviation between the two as the risk distribution deviation; calculate the average risk distribution deviation of all monitoring points, and perform reverse mapping on the average value to obtain the significance of non-risk features.

6. The intelligent monitoring and safe passage system for highways as described in claim 1, characterized in that, The time point selection module is specifically used for: For each time point, the overall correlation is divided by the significance of non-risk features to obtain the ratio value; the ratio value is then normalized to generate a risk assessment importance score. Set an importance threshold and mark the time points where the risk assessment importance score is greater than the threshold as critical time points.

7. The intelligent monitoring and safe passage system for highways as described in claim 1, characterized in that, The security monitoring module is specifically used for: Traffic flow and accident data from all monitoring points at key time points are collected to form a training dataset. The training dataset is used to build a support vector machine model, and the model parameters are optimized through grid search. After training, a safe passage prediction model is obtained.

8. The intelligent monitoring and safe passage system for highways as described in claim 7, characterized in that, The safety monitoring module is also used for: Real-time traffic flow data from highway monitoring points is collected, and traffic flow characteristics corresponding to key time points are extracted. The traffic flow characteristics are input into a safe passage prediction model, and the model outputs the probability of accident risk. When the probability of risk exceeds a preset threshold, a risk warning signal is generated.

9. The intelligent monitoring and safe passage system for highways as described in claim 1, characterized in that, The reference point determination module is specifically used for: For the target monitoring point, calculate the absolute difference between the number of accidents occurring at that point and the number of accidents occurring at all other monitoring points; set a threshold for the difference, and select monitoring points whose absolute difference is less than the threshold as reference monitoring points.

10. The intelligent monitoring and safe passage system for highways as described in claim 3, characterized in that, The correlation analysis module uses a dynamic regularization algorithm to calculate the alignment distance, which includes: interpolating the ordered traffic flow sequence and the ordered accident sequence to make the two sequences have the same length; calculating the Euclidean distance matrix of the interpolated sequences; finding the minimum cumulative distance through the dynamic regularization path and using the minimum cumulative distance as the alignment distance; and mapping the alignment distance to the range of zero to one to obtain the initial value of the correlation strength.

Images