Driving risk early warning method based on long-term and short-term driving style characteristics

Abstract

The invention provides a driving risk early warning method based on long-term and short-term driving style features. The method comprises the following steps: constructing long-term and short-term driving feature vectors according to driving behavior data; based on the standardized long-term and short-term driving feature vectors, a clustering method is adopted to determine long-term style feature vectors and long-term and short-term driving risk score labels; constructing a multilayer driving risk assessment model considering long and short term driving styles, and performing fitting training on the assessment model based on the long and short term driving style sample set to obtain an optimal multilayer driving risk assessment model; determining a long-term driving style category according to a long-term driving feature vector collected in real time, inputting the long-term driving style category and a short-term driving feature vector collected in real time into an optimal multilayer driving risk assessment model, and outputting a moving average trend of a scoring time sequence and a fluctuation trend of a driving behavior in a time window according to a predicted driving risk score output in real time. And formulating a two-stage early warning strategy. According to the invention, highly personalized and adaptive risk early warning can be provided.

Description

Technical Field

[0001] The present invention relates to the fields of driving safety assessment and intelligent risk management. In particular, it proposes a multi-layered driving risk assessment model and a two-level warning strategy that combines long-term and short-term driving styles, aiming to improve driving safety and achieve personalized driving risk warning. Background Art

[0002] Driving safety is a core concern in the transportation sector. With the increasing penetration of automobiles, traffic accidents have become one of the leading causes of casualties worldwide. Traditional driving risk assessment methods primarily rely on short-term driving behavior data, such as speed, acceleration, and braking frequency. While this data can reflect a driver's immediate state, it cannot fully reveal their long-term driving habits and behavioral patterns. In recent years, with the development of big data and machine learning technologies, increasing research has focused on how to improve the accuracy of risk assessment by leveraging long-term driving data. Long-term data includes long-term accumulation of driver characteristics, historical driving behavior, and long-term average speed. This information helps identify drivers' stable driving characteristics and potential risk tendencies. However, existing technologies still have shortcomings in integrating short-term and long-term data, processing non-risk assessment relationships, improving model interpretability, and achieving personalized warnings.

[0003] Furthermore, driver behavior is influenced by a variety of factors, including individual psychological state, physiological condition, driving skills, vehicle technical condition, and the external environment. These factors interact to determine the complexity and variability of driving behavior. Therefore, an effective driving risk assessment system must not only be able to process and analyze long-term and short-term data, but also adapt to the dynamic changes in this data, providing real-time, accurate, and interpretable risk assessments and warnings. Existing driving risk assessment systems often use simple threshold judgments or single statistical methods, which are difficult to adapt to the diversity and complexity of driving behavior. For example, a simple speed threshold cannot distinguish safe driving speeds under different road conditions, and a single statistical method cannot capture the temporal changes and pattern characteristics of driving behavior. Furthermore, the lack of personalized warnings makes it impossible for warning systems to provide customized feedback and recommendations based on the specific behavioral characteristics of different drivers. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a driving risk warning method based on long-term and short-term driving style characteristics, which can not only comprehensively analyze the driver's long-term and short-term driving behaviors, but also provide highly personalized and adaptive risk warnings.

[0005] The present invention achieves the above technical objectives through the following technical means.

[0006] A driving risk warning method based on long-term and short-term driving style characteristics:

[0007] Collect historical data related to driving behavior, including vehicle operation and status information, and driver personal and psychological state information;

[0008] Based on the collected data, a long-term driving feature vector and a corresponding short-term driving feature vector are constructed, and outlier cleaning and standardization processing are performed on the driving feature vectors;

[0009] The standardized long-term and short-term driving feature vectors are clustered separately, and long-term driving behaviors are classified into different long-term driving styles to obtain different categories of long-term style feature vectors. Short-term driving behaviors are classified into different clusters, and the long-term and short-term driving risk score labels are comprehensively determined based on the size of the short-term driving behavior cluster and the location of the cluster's centroid.

[0010] The standardized short-term driving feature vector, the clustered long-term style feature vector, and the long-term and short-term driving risk score label are integrated into the long-term and short-term driving style sample set according to the corresponding relationship of the time window;

[0011] A multi-layered driving risk assessment model that considers both long-term and short-term driving styles is constructed. Based on the hierarchical linear regression model framework, the model uses the short-term driving feature vectors and long-term style feature vectors from the short-term and long-term driving style sample sets as observation samples for input. A short-term risk assessment layer that considers short-term driving characteristics and a long-term risk trend layer that considers long-term style characteristics are constructed. The short-term risk assessment layer and the long-term risk trend layer are integrated to obtain a continuous risk latent variable, which is discretized to predict and output a discrete driving risk score.

[0012] A multi-layer driving risk assessment model was trained based on a sample set of long- and short-term driving styles. Model parameters were fitted using the marginal likelihood estimation method, and then the expectation maximization algorithm was used to optimize the parameters to obtain the optimal multi-layer driving risk assessment model.

[0013] The long-term driving style category is determined based on the long-term driving feature vectors collected in real time. The corresponding long-term driving style feature vectors and the short-term driving feature vectors collected in real time are input into the optimal multi-layer driving risk assessment model. Based on the real-time predicted driving risk score, the moving average trend of the output score time series, and the fluctuation trend of driving behavior within the time window, a two-level warning strategy is formulated to issue corresponding warning signals to the driver. In the first-level warning judgment, an output risk score of 1 indicates low risk and no warning is required; an output risk score of 3 indicates high risk and an immediate high-risk warning is generated; an output risk score of 2 indicates low risk and triggers a second-level warning. If the moving average trend of the driving risk score time series continues to rise or the fluctuation trend of driving behavior continues to rise, the warning level is increased from medium risk to high risk. If the moving average trend of the driving risk score time series continues to decline and the fluctuation trend of driving behavior continues to decline, the warning level is reduced from medium risk to low risk.

[0014] Furthermore, the construction of the long-term driving feature vector includes:

[0015] Select long-term behavioral characteristic variables: driver age X, which represents the driver's personal characteristics long,1 , driving experience X long,2 and family status X long,3 , the driving speed X that characterizes the historical driving behavior characteristics long,4 , driving acceleration X long,5 And the headway X long,6 , the driver's cognitive ability X that represents the driver's psychological state long,7 、Adventurous Psychology X long,8 and emotional state X long,9 ;

[0016] Select long-term external characteristic variables: vehicle safety configuration status X that characterizes the vehicle's own characteristics long,10 、Vehicle Type X long,11 and vehicle status X long,12 , road condition adaptability X that characterizes environmental adaptability long,13 , Traffic Condition AdaptabilityX long,14 and adaptability to the natural environment X long,15 ;

[0017] Calculate the characteristic values ​​of the long-term driving characteristic variables in each time window in chronological order: For the continuous driving characteristic variables (X long,4 ,X long,5 ,X long,6 ,X long,13 ,X long,14 ,X long,15), take its mean value in the time window as the characteristic value. For other long-term discrete driving characteristic variables, take the mode of the observation values ​​of each sampling point in the time window as the characteristic value. The expression of the long-term driving characteristic vector in time window j is:

[0018] Furthermore, the construction of the short-term driving feature vector includes:

[0019] Select short-term behavior characteristic variables: vehicle acceleration X that characterizes the driver's longitudinal operation behavior short,1 , vehicle acceleration X short,2 and headway X short,3 , the vehicle yaw rate X that characterizes the driver's lateral operation behavior short,4 , Lane center offset X short,5 and vehicle center of mass side slip angle X short,6 , driving fatigue state X that characterizes the driver's physiological and psychological state short,7 , concentration X short,8 and emotional state X short,9 ;

[0020] Select short-term external characteristic variables: characterize the traffic flow X in each time window vehicle driving section short,10 、Road type X short,11 , the frequency of emergencies X that characterize emergency driving behavior short,12 , overtaking frequency X that characterizes the behavior characteristics of other vehicles short,13 , Following distance X short,14 and lane-changing frequency X short,15 ;

[0021] Calculate the characteristic value of the short-term driving characteristic variable in each time window in chronological order: For the continuous short-term driving characteristic variable (X short,1 ,...,X short,6 ,X short,12 ,...,X short,15 ), take its mean value in the time window as the characteristic value, for the short-term discrete driving characteristic variable (X short,7 ,X short,8 ,X short,9 ,X short,10 ,X short,11 ), take the mode of the observation values ​​of each sampling point in the time window as the eigenvalue; the short-term driving eigenvector expression in the i-th short-term time window in the j-th long-term time window is:

[0022]

[0023] Furthermore, the normalized long-term driving feature vector W jClustering is done into K clusters, each cluster represents a long-term driving style, and the mean of the long-term driving feature vectors of all sample points in each cluster is counted, and the mean vector of each cluster is As the long-term style feature vector of each sample in the cluster, where j = 1, 2, ..., M; k = 1, 2, ..., K.

[0024] Furthermore, the standardized short-term driving feature vector X ij Clustering into different clusters, where j = 1, 2, ..., M; i = 1, 2, ..., N M The specific rules for comprehensively determining the driving behavior risk level by cluster size and cluster centroid position are:

[0025]

[0026] Assign the risk level of the cluster to which each driving style sample belongs to the sample: risk_score(X ij )=risk_score(C),if X ij ∈C;

[0027] In the formula, cluster_size(C) represents the size of the cluster, C is the number of the cluster, N c is the total number of driving style sample points, Centroid C Indicates the centroid position of the cluster, SafeRegion indicates the safe area, RiskRegion indicates the risk area, and HighRiskRegion indicates the high-risk area;

[0028] SafeRegion:

[0029] Medium-risk area RiskRegion: But the SafeRegion condition is not met;

[0030] High Risk Region:

[0031] Furthermore, if the style samples corresponding to some uncommon or abnormal driving behaviors in the collected historical data cannot be classified into any cluster or the corresponding style sample neighborhood contains fewer than min_samples points, such driving samples will be marked as noise points and assigned the highest risk level:

[0032] risk_score(X ij )=3,if label(X ij )=-1

[0033] Among them, min_samples is the minimum number of neighborhood samples required for a point to become a core point;

[0034] Get each sample group Corresponding long-term and short-term driving risk score label y ij :y ij =risk_score(X ij ).

[0035] Furthermore, the short-term risk assessment layer considering short-term driving characteristics is specifically:

[0036] The short-term driving feature P is used as the input feature vector, and the observation sample of P is X ij , construct continuous risk latent variables

[0037]

[0038] Among them, P ij represents the i-th short-term driving style corresponding to the long-term style feature j, j = 1, 2, ..., M, i = 1, 2, ..., N j β 0j is the intercept term under long-term driving style j, β 1j T is the slope term of the short-term driving feature vector under long-term style j, ε ij represents the random error term.

[0039] Furthermore, the long-term risk trend layer considering long-term characteristics is specifically:

[0040] The intercept term β 0j and the slope term β 1j Constructed as the long-term style feature vector Q j The linear expression of Q j The observation sample is

[0041]

[0042] Among them, γ 00 is the fixed effect intercept, γ 10 is the 15-dimensional fixed effect intercept vector, is the 15-dimensional adjustment coefficient vector of long-term driving characteristics on basic driving risk, Γ 11 is the 15×15 dimensional influence coefficient matrix of long-term driving characteristics on short-term driving characteristics, u 0j Residuals representing heterogeneity of underlying risk between individuals, u 1j Characterize the heterogeneous residuals of short-term behavioral risk sensitivity among individuals.

[0043] Further:

[0044] Integrate the short-term risk assessment layer and the long-term risk trend layer to obtain a continuous risk latent variable The expression is:

[0045]

[0046] Introducing the ordered Probit model, the latent variable Discretized into long-term and short-term driving risk score Y ij :

[0047]

[0048] Among them, Y ij It represents the i-th short-term driving risk score under the j-th long-term driving style, and its observation sample is y ij , μ represents the dividing threshold between medium and high risk scores.

[0049] Furthermore, long-term and short-term driving feature vectors are collected in real time, and the long-term style feature vector is obtained based on the mean vector of the long-term driving style cluster. The short-term driving feature vector and the long-term style feature vector are input into the optimal multi-layer driving risk assessment model, and the predicted driving risk score is output in real time to judge the risk level. Finally, a two-level warning mechanism is used to generate a warning strategy. In the initial stage of model operation, if the long-term driving feature data of the current driver for one week cannot be obtained in time, the long-term driving feature vector collected in real time is initialized to the historical driving data of drivers with similar age and driving experience. After one week, it is replaced by the long-term driving feature data actually collected for the current driver within one week.

[0050] The beneficial effects of the present invention are:

[0051] (1) The present invention integrates multidimensional data sources (such as vehicle dynamics, driver physiological state, and environmental conditions) and multidimensional time scales (short-term acceleration data such as 5 to 30 seconds and long-term average speed data such as weekly or monthly average speed data). It captures the complex interactions between data through hierarchical linear regression and ordered probit models, thereby improving the predictive ability for high-risk scenarios.

[0052] (2) The present invention proposes a multi-layered driving risk assessment model that considers both short-term and long-term driving style characteristics. The short-term risk assessment layer mainly measures the risk level of short-term driving characteristics under different driving scenarios, and the long-term risk trend layer mainly captures the differences in risk sensitivity of short-term driving characteristics among individuals under different long-term driving styles, comprehensively integrating the immediate fluctuations and long-term trend characteristics of driving behavior and style.

[0053] (3) In terms of driving warning, the present invention provides a more comprehensive risk warning by combining short-term real-time warnings with long-term trend assessments. The present invention utilizes a two-level warning mechanism. The system can identify potential dangerous trends through long-term data analysis even when short-term dangerous behaviors are not yet apparent. This mechanism ensures the comprehensiveness and accuracy of warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic diagram of characteristic variables of long-term and short-term driving styles of the present invention;

[0055] Figure 2 Schematic diagram of a multi-layer driving risk assessment model based on the fusion of long-term and short-term driving styles of the present invention;

[0056] Figure 3 A schematic diagram of a driving warning mechanism combining long-term and short-term driving styles according to the present invention;

[0057] Figure 4 This is a schematic diagram of the real-time process of driving warning of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0059] like Figure 3 As shown, the present invention provides a driving risk warning method based on long-term and short-term driving style characteristics, which specifically includes the following steps:

[0060] Step 1: Data collection and driving style data preprocessing

[0061] (1) First, historical data related to driving behavior is collected, including vehicle operation information such as speed, acceleration, and headway distance through vehicle sensors and monitoring tools; then, driver personal and psychological state information is obtained through driver profile information, questionnaire assessment, and physiological signal collection equipment, as well as vehicle and road information through vehicle inspection reports and driving monitoring systems.

[0062] (2) In the vehicle controller, based on the collected driving behavior data (including vehicle operation and status information, driver personal and psychological state information), long-term driving feature vectors (from a single complete trip to driving habit characteristics across weeks / months) and corresponding short-term driving feature vectors (instantaneous operation characteristics with a time window of 5 to 30 seconds) are constructed.

[0063] 1) Construction of historical long-term driving feature vector: The construction of long-term driving feature vector involves dividing the driving behavior data of each driver collected in (1) into time window length T q (Recommended time window length T q= 1 week) to obtain a total of M' long-term driving feature time windows. The variable categories and values ​​are shown in Figure 1 and Tables 1 and 2.

[0064] Table 1 Long-term behavioral characteristic variables

[0065]

[0066]

[0067] Table 2 Long-term external characteristic variables

[0068]

[0069] Long-term behavioral characteristic variables, including: driver age X, which represents the driver's personal characteristics long,1 , Driving experience (driving age) X long,2 and family status X long,3 (values ​​range from 0 to 2, where 0 indicates single, 1 indicates married, and 2 indicates having children); driving speed X that characterizes historical driving behavior characteristics long,4 , driving acceleration X long,5 And the headway X long,6 ; Driver cognitive ability X that represents the driver's psychological state long,7 (value ranges from 1 to 3, where 1 represents low level, 2 represents medium level, and 3 represents high level), adventurous psychology X long,8 (values ​​range from 1 to 3, where 1 represents low risk, 2 represents medium risk, and 3 represents high risk) and emotional state X long,9 (The value is 0,1, where 0 represents a negative state and 1 represents a positive state).

[0070] Long-term external characteristic variables: specifically including: vehicle safety configuration status X that characterizes the vehicle's own characteristics long,10 (values ​​1, 2, 3, where 1 represents basic configuration, 2 represents medium configuration, and 3 represents high configuration), vehicle type X long,11 (values ​​1, 2, 3, 4, where 1 represents a car, 2 represents a truck, 3 represents a bus, and 4 represents a special vehicle) and vehicle status X long,12 (values ​​1, 2, 3, where 1 indicates a poor state, 2 indicates a fair state, and 3 indicates a good state); and road condition adaptability X, which represents environmental adaptability based on the driver's reaction time under different conditions. long,13 , Traffic Condition AdaptabilityX long,14 , adaptability to natural environmentX long,15 .

[0071] Long-term time window eigenvalue processing: Calculate the eigenvalue of the long-term driving characteristic variable in each time window in chronological order.long.4 , X 10ng5 , X long，6 , X long.13 , X long,14 , X long.15 , take its mean value in the time window as the characteristic value, and take X long,4 For example, the observation value of the rth sampling point in time window j is (There are m sampling points in the time window), then the characteristic value is For other long-term discrete driving characteristic variables, the mode of the observation values ​​at each sampling point in the time window is taken as the characteristic value; finally, the long-term driving characteristic vectors of M' time windows are obtained, where the expression of the long-term driving characteristic vector in time window j is:

[0072]

[0073] 2) Construction of historical short-term driving feature vector: The construction of short-term driving feature variables involves the collection of data in the long-term time window j according to the time window length t q Segmentation (recommended time window length t q =1min), to obtain a total of N' j Short-term driving characteristic time window, variable categories and values ​​are shown in Figure 1 and Tables 3 and 4.

[0074] Table 3 Short-term behavioral characteristic variables

[0075]

[0076]

[0077] Table 4 Short-term external characteristic variables

[0078]

[0079] Short-term behavior characteristic variables, specifically including: vehicle acceleration X that characterizes the driver's longitudinal operation behavior in each time window short,1 , vehicle acceleration X short,2 and headway X short,3 ; Vehicle yaw rate X, which represents the driver's lateral operation behavior short,4 , Lane center offset X short,5 and vehicle center of mass side slip angle X short,6 ; and the driving fatigue state X that characterizes the driver's physiological and psychological state based on pupil changes and facial micro-expression dominant state within the driver's time window short,7 (values ​​range from 0 to 1, where 0 represents no fatigue and 1 represents complete fatigue), concentration X short,8(values ​​0 and 1, where 0 represents low concentration and 1 represents high concentration) and emotional state X short,9 (The values ​​are 0 and 1, where 0 represents a negative state and 1 represents a positive state).

[0080] Short-term external characteristic variables, specifically including: characterizing the traffic flow X in each time window vehicle driving section short,10 、Road type X short,11 (values ​​range from 1 to 5, where 1 represents expressways, 2 represents urban main roads, 3 represents urban secondary roads, 4 represents urban branch roads, and 5 represents rural roads); and the frequency X of emergencies representing emergency driving behaviors (such as sudden braking, sharp turns, etc.) within the road section R (m) where the vehicle is located in each time window short,12 ; Overtaking frequency X that characterizes the behavior characteristics of other vehicles short,13 , Following distance X short,14 and lane-changing frequency X short,15 .

[0081] Short-term time window eigenvalue processing: Calculate the eigenvalue of the short-term driving characteristic variable in each time window in chronological order. short,1 ,...,X short,6 ,X short,12 ,...,X short,15 ), take its mean value in the time window as the characteristic value, and use X short,1 For example, the observation value of the sth sampling point in the i-th short-term time window in the long-term time window j is (There are n sampling points in the short-term time window), then the characteristic value is recorded as For short-term discrete driving characteristic variables (X short,7 ,X short,8 ,X short,9 ,X short,10 ,X short,11 ), take the mode of the observation values ​​of each sampling point in the time window as the characteristic value. Make the same division for the M' long-term driving characteristic time windows in 1), and finally get the common The short-term driving feature vector of time windows is expressed as follows:

[0082]

[0083] 3) Cleaning and standardization of driving characteristic outliers: In order to make driving characteristics suitable for data analysis and model training, X' ij and W j'Perform outlier cleaning to detect data points that deviate from the normal range, remove sensor noise and error values, and perform data standardization to convert feature values ​​to a unified scale to ensure consistency of model input.

[0084] 4) Classification of long-term driving style and driving behavior risk level: The standardized long-term and short-term driving feature vectors are clustered using the DBSCAN clustering method. The long-term and short-term driving feature vectors reflect long-term driving behavior and short-term driving behavior, respectively. Long-term driving behavior is classified into different long-term driving styles, and short-term driving behavior is classified into different driving risk levels.

[0085] First, the standardized long-term driving feature vector W j (j=1,2,…,M) as clustering samples, draw a K-distance graph, and determine the appropriate eps value and min_samples value by observing the "inflection point" in the distance graph; the inflection point refers to the point where the curve in the K-distance graph becomes steep, indicating the transition from a dense area to a sparse area; the eps value is defined as the maximum distance between two sample points. When the distance between two points is less than eps, one point is considered a neighbor of the other point; min_samples is the minimum number of neighborhood samples required for a point to become a core point. The minimum number of neighborhood samples is an empirical value and is generally 3. According to the eps value and min_samples value, DBSCAN is used to cluster long-term driving features: DBSCAN(W,eps w , min_samples w ), cluster the long-term driving feature samples into K clusters, each cluster represents a long-term driving style, and count the mean of the long-term driving feature vectors of all sample points in each cluster, and calculate the mean vector W of each cluster. k (k=1,2,…,K) is used as the long-term style feature vector of each sample in the cluster.

[0086] Then use the standardized short-term driving feature vector X ij (j=1,2,…,M; i=1,2,…,N M ) as clustering samples, and use DBSCAN to cluster short-term driving features according to the same steps: DBSCAN(X,eps X , min_samples X ), clustering short-term driving behaviors into different clusters. Further, different clusters are classified as low risk, medium risk, or high risk. The specific steps for cluster risk classification are as follows:

[0087] Assume that L is the cluster label list of driving style samples, C is the cluster number (excluding noise points), and the number of driving feature samples in each cluster is counted to determine the cluster size. The formula for calculating the cluster size is: where Nc is the total number of driving feature sample points, is an indicator function, when L i =C a , the value is 1, otherwise it is 0.

[0088] Cluster size is used as a criterion for risk level. Larger clusters represent more common and safe driving behavior patterns; small clusters or isolated points usually represent high-risk behaviors. The driving risk scoring rules for each cluster are defined as:

[0089]

[0090] Among them, the numbers 1, 2, and 3 represent low, medium, and high risks respectively.

[0091] Based on the short-term driving feature vector X ij The risk feature space divides the driving behavior into safe area SafeRegion, medium risk area RiskRegion and high risk area HighRiskRegion. The risk feature space includes acceleration feature value and headway characteristic value These reflect the smoothness of driving behavior and the safety of following vehicles. The regional divisions are based on traffic safety standards (Euro NCAP, ISO2631) and driving behavior research, and are specifically defined as follows:

[0092] SafeRegion:

[0093] Medium-risk areaRiskRegion But the SafeRegion condition is not met;

[0094] High Risk Region:

[0095] By setting the centroid position of each cluster to C By comparing with the defined risk areas, the risk scoring rules were revised. The revised principle is: the closer the centroid is to the predefined safe area, the lower the overall risk of the cluster; the closer the centroid is to the predefined high-risk area, the higher the overall risk of the cluster. The final risk scoring rules are:

[0096]

[0097] Thus, the driving risk level of each driving style sample in the cluster is obtained, that is, the risk level of the cluster to which each driving style sample belongs is assigned to the sample: risk_score(X ij )=risk_score(C),if X ij ∈C.

[0098] Since (1) the collected data may contain some uncommon or abnormal driving behaviors, and DBSCAN is a clustering method based on the density of data points, during the DBSCAN clustering process, the style samples corresponding to these special driving behaviors may not be classified into any cluster or the corresponding style sample neighborhood may contain less than min_samples points. Such samples are marked as noise points (labeled as -1) and assigned the highest risk level.

[0099] risk_score(X ij )=3,if label(X ij )=-1

[0100] Integrate the above risk level rules to obtain the long-term and short-term driving risk score label y corresponding to each short-term driving feature sample ij :y ij =risk_score(X ij ).

[0101] 5) The standardized short-term driving feature vectors, the clustered long-term style feature vectors, and the long-term and short-term driving risk score labels are integrated into a long-term and short-term driving style sample set according to the time window correspondence. A one-to-many correspondence is used in the sample, such as all the standardized short-term driving feature vector samples X in the long-term time window j ij (i=1,2,…,N j ) corresponds to the long-term driving feature vector W after normalization of the time window j , while W j The long-term style feature vector corresponding to the cluster to which it belongs X ij Corresponding long-term and short-term driving risk score label y ij , and finally form a sample set:

[0102] (As shown in Table 5).

[0103] Each long-term style feature vector represents the driver's relatively stable long-term driving style in a cycle, and each corresponding standardized short-term driving feature vector X ij Characterizes the differentiated short-term driving style or state caused by changes in road and traffic dynamics (changes in driving scenarios) during driving.

[0104] Table 5 Model sample data example table

[0105]

[0106]

[0107] Step 2: Construct a multi-layer driving risk assessment model considering long-term and short-term driving styles

[0108] (1) Construction of multi-layer driving risk assessment model: Based on the hierarchical linear regression model framework (such as Figure 2 ), using the long-term and short-term driving style feature sample group obtained in step 1 (j=1,2,…,M,i=1,2,…,N M , k=1,2,…,K) is the observation sample of the model input feature vector (P, Q), and a short-term risk assessment layer considering the short-term driving style characteristics P and a long-term risk trend layer considering the long-term driving style characteristics Q are constructed respectively; at the same time, the discretization mechanism of the ordered Probit model is introduced, and the predicted output variable is the long-term and short-term driving risk score Y, and the observation sample of Y is the y obtained in step 1 ij .

[0109] Construction of short-term risk assessment layer considering short-term characteristics: Taking P (15 dimensions) as input feature vector, linear regression is used to construct continuous risk latent variables To measure the risk level of short-term driving characteristics in different driving scenarios:

[0110]

[0111] Among them, P ij (j=1,2,…M) represents the i-th short-term driving style (i=1,2,…,N j ), β 0j is the intercept term under long-term driving style j, representing the basic driving risk level of this long-term driving style; β 1j T is the slope term (15 dimensions) of the short-term driving feature vector under the long-term driving style j, which represents the impact of short-term driving characteristics on the driving risk level under the category of long-term driving style j; ij represents the random error term, reflecting the inevitable interference factors in the actual driving process, and obeys the time autocorrelation AR(1) structure:

[0112] ε ij =ρε i-1,j +η ij (2)

[0113] Where ρ is the autocorrelation coefficient, which is used to capture the time dependence of the risk scores between adjacent driving style samples (adjacent short-term driving feature time windows of each driver under the same long-term driving style), and ||ρ||<1 ensures stationarity; white noise η ab ~N(0,σ 2 ).

[0114] The long-term risk trend layer is constructed by considering the long-term characteristics: the intercept term β in formula (1) is 0j and the slope term β 1j Constructed as the long-term style feature vector Q j A linear expression (15 dimensions) is used to capture the differential impact of long-term driving style on the risk sensitivity of short-term driving characteristics:

[0115]

[0116] Among them, γ 00 is the fixed effect intercept, γ 10 is the 15-dimensional fixed effect intercept vector; is the 15-dimensional adjustment coefficient vector of long-term driving characteristics on basic driving risk, representing the correction effect on the basic risk level; Γ 11 is the 15×15 dimensional influence coefficient matrix of long-term driving characteristics on short-term driving characteristics, representing the regulatory mechanism of long-term driving style on the sensitivity of short-term behavioral risk; u 0j and u 1j represents random effects, where u 0j Residuals representing heterogeneity of underlying risk between individuals, u 1j Characterizing the heterogeneous residuals of short-term behavioral risk sensitivity among individuals, both obey a normal distribution with a mean of 0, that is:

[0117]

[0118] Among them, 0 p represents the zero matrix, p represents the feature dimension (p=15), τ 00 Characterizes the degree of dispersion of basic risk among individuals (the larger the variance, the greater the individual differences in basic risk level among the driver group); τ 01 is the covariance term, which represents the correlation between individual basic risk and short-term behavior risk sensitivity (if it is less than 0, it means that drivers with high basic risk have weaker risk response to short-term behavior); T 11 Characterize the differences in risk responses of different drivers to the same short-term behavior (the larger the variance, the more significant the individual differences in drivers' responses to short-term behavior risks). The expression is as follows:

[0119]

[0120] Covariance matrix Satisfy strict positive definiteness constraints: τ 00 >0,T 11 Zhengding, This constraint ensures that the model can accurately reflect the individual differences of different drivers in terms of basic risk and short-term behavioral risk sensitivity and their interrelationships.

[0121] Integrate the short-term risk assessment layer and the long-term risk trend layer, comprehensively consider the impact of fixed effects, random effects and explanatory variables on latent variables, and obtain continuous risk latent variables. The expression is as follows:

[0122]

[0123] Introducing the ordered Probit model, the latent variable Discretize into the final long-term and short-term driving risk score Y ij :

[0124]

[0125] Among them, Y ij represents the i-th short-term driving risk score under the j-th long-term driving style, μ represents the demarcation threshold between medium and high risk scores, and μ>0.

[0126] (2) Parameter fitting of multi-layer driving risk assessment model

[0127] According to the long-term and short-term driving risk score label y obtained in (1) ab , and the marginal likelihood estimation method was used to fit the model parameters.

[0128] ① Initialization parameters: In the process of fitting parameters, taking an empirical initial value can ensure the stability of the model to a certain extent, and then gradually optimize it through the iterative process. Set the initial values ​​of all parameters in formulas (2), (4) and (5), among which the adjustment coefficient γ 00 Initial setting is 0, γ 01 , γ 10 and Γ 11 Initially, it is set to zero vector and zero matrix, which means that in the initial stage, it is assumed that the influence of long-term style on the sensitivity of basic risk and short-term behavioral risk is zero, and then it is gradually optimized through iteration. The covariance matrix is ​​initialized to the diagonal matrix Ψ (0) =0.1I 16 , AR(1) parameter initialization ρ=0.5, σ 2 = 1. The risk level classification threshold μ is initially determined based on the distribution characteristics of the sample data. It is recommended to take the value of y in the sample. ij The 33% quantile of the labels.

[0129] ②Construction of marginal likelihood function:

[0130]

[0131] in, is the expected value of the latent variable calculated based on the sample set described in step 1. Risk score probability (r=1, 2, 3, representing low risk, medium risk and high risk respectively), where μ0=-∞, μ1=0, μ2=μ, μ3=+∞, and the probability intervals corresponding to different risk levels are calculated using the standard normal distribution cumulative function Φ. is the error term ε ij The variance of where Σ j =σ 2 (I-ρA) -1 (I-ρA T ) -1 N is the AR(1) process j ×N j The covariance matrix describes the error autocorrelation strength between driving style samples, where matrix A is the adjacency matrix, representing the single-step correlation of the time series:

[0132] Σ j Capture the temporal correlation between driving style samples, adjust the autocorrelation strength by ρ, and ensure that the marginal likelihood function takes into account the dynamic dependence between driving style samples. 0j ,u 1j ) is the 1+p=16-dimensional normal density function:

[0133]

[0134] ③ Adaptive Gaussian integral approximation:

[0135] For the long-term style feature j, in order to accurately approximate the (1+p) multiple integral in the joint marginal likelihood function, the multidimensional Gauss-Hermite quadrature formula is used. The specific approximate formula is:

[0136]

[0137] Among them, w q w r is the preset weight, u 0j,q ,u 1j,r is the integration node, and the Gauss-Hermite quadrature principle is used to select appropriate integration nodes and weights according to the accuracy requirements; Q and R represent u 0j 、u 1j The number of nodes for the Gauss-Hermite integration.

[0138] ④EM algorithm iterative optimization

[0139] Step E: Calculate random effects based on current parameter estimates The posterior distribution of :

[0140]

[0141] Step M: Update the parameters using Newton iteration method:

[0142]

[0143] Among them, H is the Hessian matrix, which reflects the second-order derivative information of the log-likelihood function; is the gradient vector, which indicates the gradient direction of the log-likelihood function at the current parameter point.

[0144] Through the Newton iteration method, the parameters are continuously updated in the direction of increasing the log-likelihood function to gradually approach the optimal parameter value.

[0145] When the relative change rate of log-likelihood of adjacent iterations ||(l (t+1) -l (t) ) / l (t) Less than 10 -5 Or the Euclidean distance of the parameter vector changes by less than 10 -4 When , the parameter estimation process reaches the convergence state and the calculation is terminated. When one of the above two conditions is met, it indicates that the parameter has converged to the optimal value within a certain accuracy range.

[0146] Step 3: Online real-time warning mechanism

[0147] (1) Determine the real-time driving risk based on the driving risk score output in step 2 and perform a first-level real-time warning.

[0148] Obtain the short-term driving feature vector x and the long-term driving feature vector w in real time, and calculate the mean vector of w and the long-term driving style cluster obtained in step 1 (k=1,2,…,K) Euclidean distance, the mean vector corresponding to the shortest Euclidean distance This is the long-term style feature vector of the current driver. The model uses the optimal multi-layer driving risk assessment model described in step 2 as input and outputs a predicted driving risk score Y in real time. In the initial model run, if the current driver's long-term driving characteristic data for a week is not available in time, w is initialized to the historical driving data of drivers of similar age and driving experience from step 1. After one week, this data is replaced with the current driver's actual long-term driving characteristic data collected during the week.

[0149] The risk level is determined in real time based on the predicted driving risk score Y output by the model. When the output risk score Y is 1, it represents low risk and no warning is required; when the output risk score Y is 3, it represents high risk and a high-risk warning is generated immediately; when the output risk score Y is 2, it represents medium risk and a secondary warning is triggered, allowing for a more in-depth assessment of risk changes in the short term.

[0150] (2) Secondary warning discrimination mechanism: The predicted driving risk score Y output by the multi-layer driving risk assessment model in N time steps at the current time t is sorted by time, and the driving risk score time series {Y t-N ,...,Y t-1 ,Y t}, according to the moving average trend of the sequence and the fluctuation trend of driving behavior, the warning level is adjusted based on (1) warning judgment.

[0151] Calculate the driving risk score time series {Y t-N ,...,Y t-1 ,Y t}The moving average MA(Y t ):

[0152]

[0153] Among them, Y b is the predicted driving risk score at time b, and N is the size of the moving window.

[0154] Similarly, the moving average MA(Y t-1 ). If MA(Y t-1 ) <MA(Y t ), it indicates that the moving average trend of the risk score increases over time, and the potential risk in driving behavior is increasing; if MA(Y t-1 )>MA(Y t ), it indicates that the moving average trend of the risk score decreases over time, and the potential risk of driving behavior gradually decreases.

[0155] The real-time acceleration a and jerk J are organized into acceleration time series data and jerk time series data, respectively, in chronological order. The volatility of the two is calculated to characterize the fluctuation trend of driving behavior within the time window. Volatility is used to capture drastic changes in real-time driving behavior. As time goes by, the volatility slowly increases. This indicates that the potential risk of driving behavior is gradually increasing. Acceleration volatility σ d and the jerk fluctuation σ J The calculation formula is as follows:

[0156]

[0157] Among them, a b is the driving acceleration at time b, is the average driving acceleration in the time window [t-N+1,t], J b is the jerk at time b, is the average value of the jerk in the time window [t-N+1,t].

[0158] Similarly, the acceleration and the fluctuation rate of jerk in the time window [tN, t-1] can be calculated. d (t)>σ d (t-1) and σ J (t)>σ J (t-1), the driving behavior fluctuation trend increases over time, and the potential risk gradually increases; if σ d (t)<σ d (t-1) and σ J (t)<σ J (t-1), the driving behavior fluctuation trend gradually decreases over time, and the potential risk gradually decreases; otherwise, the driving behavior fluctuation trend is non-significant change, and the potential risk does not change significantly.

[0159] (3) Adjustment of the warning mechanism: Based on the (1) first-level warning judgment, combined with the moving average and volatility trends, a corresponding warning signal is issued to the driver.

[0160] When the model outputs the real-time risk score Y at time t t When the risk score is 1 or 3, the system warning level is determined by the first-level warning judgment of (1) and no adjustment is required; when the model outputs the risk score Y t When the driving risk score is 2, t Moving average trend MA(Y t ) increases over time, or the fluctuation trend of driving behavior increases over time, the system will raise the warning level from medium risk to high risk to capture the potential trend of risk accumulation, thereby making more sensitive and real-time driving warnings for driving behavior; on the contrary, if the driving risk score Y t Moving average trend MA(Y t ) and the fluctuation trend of driving behavior decreases over time, the system will reduce the warning level from medium risk to low risk, making the warning more cautious in the short term; if the driving risk score Y t The moving average trend MA(Y t ) decreases over time, and the fluctuation trend of driving behavior is not a significant change, the system maintains the medium risk warning level. Figure 4 .

[0161] The embodiments described are preferred implementations of the present invention, but the present invention is not limited to the above implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention are within the scope of protection of the present invention.

Claims

1. A driving risk warning method based on long-term and short-term driving style characteristics, characterized by: Collect historical data related to driving behavior, including vehicle operation and status information, and driver personal and psychological state information; Based on the collected data, a long-term driving feature vector and a corresponding short-term driving feature vector are constructed, and outlier cleaning and standardization processing are performed on the driving feature vectors; The standardized long-term and short-term driving feature vectors are clustered separately, and long-term driving behaviors are classified into different long-term driving styles to obtain different categories of long-term style feature vectors. Short-term driving behaviors are classified into different clusters, and the long-term and short-term driving risk score labels are comprehensively determined based on the size of the short-term driving behavior cluster and the location of the cluster's centroid. The standardized short-term driving feature vector, the clustered long-term style feature vector, and the long-term and short-term driving risk score label are integrated into the long-term and short-term driving style sample set according to the corresponding relationship of the time window; A multi-layered driving risk assessment model that considers both long-term and short-term driving styles is constructed. Based on the hierarchical linear regression model framework, the model uses the short-term driving feature vectors and long-term style feature vectors from the short-term and long-term driving style sample sets as observation samples for input. A short-term risk assessment layer that considers short-term driving characteristics and a long-term risk trend layer that considers long-term characteristics are constructed. The short-term risk assessment layer and the long-term risk trend layer are integrated to obtain a continuous risk latent variable, which is discretized to predict and output a discrete driving risk score. A multi-layer driving risk assessment model was trained based on a sample set of long- and short-term driving styles. Model parameters were fitted using the marginal likelihood estimation method, and then the expectation maximization algorithm was used to optimize the parameters to obtain the optimal multi-layer driving risk assessment model. The long-term driving style category is determined based on the long-term driving feature vectors collected in real time. The corresponding long-term driving style feature vectors and the short-term driving feature vectors collected in real time are input into the optimal multi-layer driving risk assessment model. Based on the real-time output predicted driving risk score, the moving average trend of the output score time series, and the fluctuation trend of driving behavior within the time window, a two-level warning strategy is formulated to issue corresponding warning signals to the driver; In the first-level early warning judgment, when the output risk score is 1, it means low risk and no early warning is required; When the output risk score is 3, it indicates high risk and a high-risk warning is generated immediately; If the output risk score is 2, indicating medium risk, a secondary warning is triggered: if the moving average trend of the driving risk score time series continues to rise or the fluctuation trend of driving behavior continues to rise, the warning level is raised from medium risk to high risk; If the moving average trend of the driving risk score time series continues to decline and the fluctuation trend of driving behavior continues to decline, the warning level will be reduced from medium risk to low risk.

2. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 1 is characterized in that: The construction of the long-term driving feature vector includes: Select long-term behavioral characteristic variables: driver age X, which represents the driver's personal characteristics long,1 , driving experience X long,2 and family status X long,3 , the driving speed X that characterizes the historical driving behavior characteristics long,4 , driving acceleration X long,5 And the headway X long,6 , the driver's cognitive ability X that represents the driver's psychological state long,7 、Adventurous Psychology X long,8 and emotional state X long,9 ; Select long-term external characteristic variables: vehicle safety configuration status X that characterizes the vehicle's own characteristics long,10 、Vehicle Type X long,11 and vehicle status X long,12 , road condition adaptability X that characterizes environmental adaptability long,13 , Traffic Condition AdaptabilityX long,14 and adaptability to the natural environment X long,15 ; Calculate the characteristic values ​​of the long-term driving characteristic variables in each time window in chronological order: For the continuous driving characteristic variables (X long,4 ,X long,5 ,X long,6 ,X long,13 ,X long,14 ,X long,15 ), take its mean value in the time window as the characteristic value. For other long-term discrete driving characteristic variables, take the mode of the observation values ​​of each sampling point in the time window as the characteristic value. The expression of the long-term driving characteristic vector in time window j is:

3. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 2 is characterized in that: The construction of the short-term driving feature vector includes: Select short-term behavior characteristic variables: vehicle acceleration X that characterizes the driver's longitudinal operation behavior short,1 , vehicle acceleration X short,2 and headway X short,3 , the vehicle yaw rate X that characterizes the driver's lateral operation behavior short,4 , Lane center offset X short,5 and vehicle center of mass side slip angle X short,6 , driving fatigue state X that characterizes the driver's physiological and psychological state short,7 , concentration X short,8 and emotional state X short,9 ; Select short-term external characteristic variables: characterize the traffic flow X in each time window vehicle driving section short,10 、Road type X short,11 , the frequency of emergencies X that characterize emergency driving behavior short,12 , overtaking frequency X that characterizes the behavior characteristics of other vehicles short,13 , Following distance X short,14 and lane-changing frequency X short,15 ; Calculate the characteristic value of the short-term driving characteristic variable in each time window in chronological order: For the continuous short-term driving characteristic variable (X short,1 ,...,X short,6 ,X short,12 ,...,X short,15 ), take its mean value in the time window as the characteristic value, for the short-term discrete driving characteristic variable (X short,7 ,X short,8 ,X short,9 ,X short,10 ,X short,11 ), take the mode of the observation values ​​of each sampling point in the time window as the eigenvalue; the short-term driving eigenvector expression in the i-th short-term time window in the j-th long-term time window is:

4. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 3 is characterized in that: The standardized long-term driving feature vector W j Clustering is done into K clusters, each cluster represents a long-term driving style, and the mean of the long-term driving feature vectors of all sample points in each cluster is counted, and the mean vector of each cluster is As the long-term style feature vector of each sample in the cluster, where j = 1, 2, ..., M; k = 1, 2, ..., K.

5. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 3 is characterized in that The standardized short-term driving feature vector X ij Clustering into different clusters, where j = 1, 2, ..., M; i = 1, 2, ..., N M The specific rules for comprehensively determining the driving behavior risk level by cluster size and cluster centroid position are: Assign the risk level of the cluster to which each driving style sample belongs to the sample: risk_score(X ij )=risk_score(C a ),if X ij ∈C a , X ij is the standardized short-term driving feature vector; In the formula, cluster_size(C a ) represents the size of the cluster, C a is the cluster number, N c is the total number of driving style sample points, Indicates the centroid position of the cluster, SafeRegion indicates the safe area, and RiskRegion indicates the risk area. HighRiskRegion represents a high-risk area; SafeRegion: Medium-risk areas But the SafeRegion condition is not met; High Risk Region:

6. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 5 is characterized in that: If the style samples corresponding to some uncommon or abnormal driving behaviors in the collected historical data cannot be classified into any cluster or the corresponding style sample neighborhood contains fewer than min_samples points, such driving samples will be marked as noise points and assigned the highest risk level: risk_score(X ij )=3,if label(X ij )=-1 Among them, min_samples is the minimum number of neighborhood samples required for a point to become a core point; Get each sample group (X ij ,W j ) corresponding long-term and short-term driving risk score label y ij :y ij =risk_score(X ij ).

7. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 3 is characterized in that: The short-term risk assessment layer considering short-term driving characteristics is specifically: The short-term driving feature P is used as the input feature vector, and the observation sample of P is X ij , construct continuous risk latent variables Among them, P ij represents the i-th short-term driving style corresponding to the long-term style feature j, j = 1, 2, ... M; β 0j is the intercept term under the j-th long-term driving style, β 1j T is the slope term of the short-term driving feature vector under the j-th long-term driving style, ε ij represents the random error term.

8. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 7 is characterized in that: The long-term risk trend layer considering long-term characteristics is specifically: The intercept term β 0j and the slope term β 1j Constructed as the long-term style feature vector Q j The linear expression of Q j The observation sample is Among them, γ 00 is the fixed effect intercept, γ 10 is the 15-dimensional fixed effect intercept vector, is the 15-dimensional adjustment coefficient vector of long-term driving characteristics on basic driving risk, Γ 11 is the 15×15 dimensional influence coefficient matrix of long-term driving characteristics on short-term driving characteristics, u 0j Residuals representing heterogeneity of underlying risk between individuals, u 1j Characterize the heterogeneous residuals of short-term behavioral risk sensitivity among individuals.

9. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 8, characterized in that: Integrate the short-term risk assessment layer and the long-term risk trend layer to obtain a continuous risk latent variable The expression is: Introducing the ordered Probit model, the latent variable Discretized into long-term and short-term driving risk score Y ij : Among them, Y ij It represents the i-th short-term driving risk score under the j-th long-term driving style, and its observation sample is y ij , μ represents the dividing threshold between medium and high risk scores.

10. The driving risk warning method based on long-term and short-term driving style characteristics according to claim 9, characterized in that: Long-term and short-term driving feature vectors are collected in real time. The long-term style feature vector is obtained based on the mean vector of the long-term driving style cluster. The short-term driving feature vector and the long-term style feature vector are input into the optimal multi-layer driving risk assessment model, which outputs a predicted driving risk score and determines the risk level in real time. Finally, a two-level warning mechanism is used to generate a warning strategy. In the initial stage of model operation, if the long-term driving characteristic data of the current driver for one week cannot be obtained in time, the long-term driving characteristic vector collected in real time is initialized with the historical driving data of drivers with similar age and driving experience. After one week, it is replaced with the long-term driving characteristic data actually collected for the current driver within one week.

Images