Risk superposition-based dangerous driving risk quantitative evaluation and early warning grading method

Abstract

The application discloses a dangerous driving risk quantitative evaluation and early warning grading method considering risk superposition, relates to the field of aircraft driving safety, and comprises the following steps: selecting pilot dangerous driving single-source risk representation indexes and quantifying, constructing a pilot dangerous driving single-source risk quantitative evaluation model and a pilot dangerous driving multi-source risk superposition model; predicting pilot dangerous driving comprehensive risk; grading pilot dangerous driving comprehensive risk; issuing a grading early warning signal to the pilot according to a prediction structure; and comprehensively analyzing pilot dangerous driving behavior characteristics by constructing a pilot dangerous driving risk quantitative evaluation model and grading early warning of the result, so that efficient perception, prediction and early warning of pilot dangerous driving risk can be realized, and therefore, the aircraft accident rate is reduced, and the aviation transportation safety is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aircraft driving safety, and particularly relates to a dangerous driving risk quantitative evaluation and early warning grading method considering risk superposition. BACKGROUND

[0002] With the development of China's economy and aviation industry, the aviation transportation industry has become an important part of China's transportation industry, and the aircraft has become an important means of transportation for human beings. Therefore, how to identify, detect, quantitatively evaluate the dangerous driving risk of pilots, and reasonably control the early warning, has important practical significance for regulating the driving behavior of pilots and ensuring the safe operation of aviation.

[0003] At present, the researches on pilot fatigue state and various emotional states at home and abroad mostly adopt medical means to evaluate the fatigue degree and emotional level of individuals by detecting the physiological and biochemical indexes of pilots. Although this method has high accuracy and clear monitoring indicators, it is not suitable for adoption due to harsh test conditions and complex process. At the same time, the current research also lacks analysis on the transformation and superposition mechanism of the dangerous driving risk of pilots, the accuracy of the measured risk is low, and the comprehensive risk of pilots at a certain time in the future cannot be effectively predicted, resulting in that the early warning control measures are not timely, and the flight safety cannot be guaranteed.

[0004] In order to solve the problems of harsh test conditions, complex process and untimely early warning control measures, the present application aims to provide a dangerous driving risk quantitative evaluation and early warning grading method considering risk superposition. SUMMARY

[0005] In order to make up for the shortcomings of the prior art, the present application provides a dangerous driving risk quantitative evaluation and early warning grading method considering risk superposition, which effectively evaluates the fatigue driving risk, emotional driving risk and comprehensive risk of pilots according to the video image information of the pilot state based on the strong correlation between the facial features of pilots and their fatigue driving state and emotional driving state. The test condition is simple and easy to realize, and the aviation transportation safety level is improved by real-time evaluation and early warning.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] The dangerous driving risk quantitative evaluation and early warning grading method considering risk superposition comprises the following steps:

[0008] (1) The dangerous driving includes fatigue driving and emotional driving. Considering the strong correlation between the facial features of pilots and their fatigue driving state and emotional driving state, important measurement indicators representing fatigue driving risk and emotional driving risk are determined from the facial features;

[0009] (2) Based on the face geometry features and 68 key points of the human face, important measurement indexes representing the fatigue driving risk and the emotional driving risk in step 1 are quantitatively analyzed;

[0010] (3) A pilot dangerous driving single-source risk quantitative evaluation model is constructed, and the pilot dangerous driving single-source risk quantitative evaluation model obtains a pilot fatigue driving risk value and an emotional driving risk value according to the quantified indexes;

[0011] (4) A pilot dangerous driving multi-source risk superposition model is constructed based on the inclusion-exclusion principle and the time series idea, and the pilot dangerous driving multi-source risk superposition model realizes the quantitative evaluation of the superposition effect of the pilot dangerous driving multi-source risk according to the pilot fatigue driving risk value and the emotional driving risk value;

[0012] (5) A pilot dangerous driving comprehensive risk prediction model is constructed, real-time pilot facial images are obtained, and the pilot fatigue driving risk value FR, the emotional driving risk value ER and the superposition risk value are obtained according to steps (1) to (4). After the fatigue driving risk value FR, the emotional driving risk value ER and the superposition risk value are standardized, they are input into the prediction model as characteristic variables to obtain the pilot dangerous driving comprehensive risk prediction value;

[0013] (6) The dangerous driving comprehensive risk values of a plurality of pilots are obtained, the pilot dangerous driving comprehensive risk values are subjected to cluster analysis, the pilot dangerous driving comprehensive risk is classified, and the threshold value of each risk level is obtained;

[0014] (7) According to the pilot dangerous driving comprehensive risk value obtained in step (5) and the threshold value of each risk level, the risk level of the pilot is judged, when the risk level of the pilot reaches the required early warning, the pilot dangerous driving comprehensive risk is classified according to the early warning signal, and control measures are taken.

[0015] In the present application, the important measurement indexes representing the fatigue driving risk include the degree of closing eyes, the frequency of yawning and the frequency of nodding;

[0016] The important measurement indexes representing the emotional driving risk include the eye corner up angle, the mouth corner up angle, the eyebrow cohesion degree, the eye opening degree and the mouth opening degree.

[0017] In the present application, the construction of the pilot dangerous driving single-source risk quantitative evaluation model comprises:

[0018] Step 3.1, the judgment criteria of each index are set, and the threshold value is divided into four levels, different levels correspond to different scores, and four risk levels correspond to different scores, and the default first risk dangerous coefficient is the largest;

[0019] Step 3.2, according to the index score, a single source risk model is established, and the formula is:

[0020]

[0021] In the formula: λ i represent the index score, the range is [0, 1], i=1, 2, 3…n, and Risk represents the fatigue driving risk or emotional driving risk.

[0022] In the application, a pilot dangerous driving multi-source risk superposition model is constructed based on the principle of inclusion and the idea of time series, and the steps are as follows:

[0023] Step 4.1, in a continuous period of time, the facial data of Q pilots during flight driving is collected;

[0024] Step 4.2, the fatigue driving risk value FR and the emotional driving risk value ER of the pilots in the Q group data are calculated respectively by using formula (13), and the data amount M of the first level fatigue driving risk, the data amount N of the first level emotional driving risk, and the data amount K of the first level fatigue driving risk and the first level emotional driving risk are counted.

[0025] Step 4.3, the pilot dangerous driving multi-source risk superposition probability is calculated based on the principle of inclusion, and the formula is:

[0026] P(FR∪ER)=P(FR)+P(ER)-P(FR∩ER)=M / Q+N / Q-K / Q (14)

[0027] Thus, the quantitative evaluation of the pilot dangerous driving multi-source risk superposition effect is realized.

[0028] In the application, the pilot dangerous driving comprehensive risk prediction model adopts the XGBoost machine learning model, and the pilot dangerous driving comprehensive risk prediction function expression is:

[0029]

[0030] In the formula: is the pilot dangerous driving comprehensive risk prediction value of the tthiteration, is a characteristic variable of the pilot dangerous driving comprehensive risk, which is composed of the fatigue driving risk value FR, the emotional driving risk value ER and the superposition risk value after standardization, is a prediction function about .

[0031] In the application, the pilot dangerous driving comprehensive risk prediction function expression is iterated for many times, so that the target function Lt reaches a set value, and the final prediction function is obtained;

[0032] The objective function is composed of a loss function and a regularization function that suppresses the model complexity, expressed as:

[0033]

[0034] Wherein: Lt is the objective function value of the tth iteration, is the loss function of a single sample, representing the error between the predicted value of the pilot's dangerous driving comprehensive risk and the true value, is the overall loss function of all samples,

[0035] is the sum of the regularization terms of the first t iterations, representing the complexity of the model decision tree;

[0036] The objective function is expanded by the second-order Taylor expansion, and the expanded objective function is expressed as:

[0037]

[0038] In the formula: g i , h i are the first derivative and the second derivative, is the loss function of a single sample in the t-1th iteration.

[0039] In the application, the RM-K-means++ algorithm is used in step six to perform cluster analysis on the pilot's dangerous driving comprehensive risk value, and by introducing risk factors, random initialization and multiple iteration optimization, the pilot's dangerous driving comprehensive risk is divided into first-level risk, second-level risk, third-level risk and fourth-level risk:

[0040] Step 6.1, initialization stage:

[0041] Obtain the dangerous driving comprehensive risk values of multiple pilots in different states as a point set, randomly select a data point as the first cluster center, and for each remaining data point R i , calculate the distance D(R i ) with the nearest selected cluster center, and take D(R i ) as the weight, and select the next cluster center according to the probability, and the probability formula is as follows:

[0042]

[0043] In the formula: j represents the number of clusters; D(R i ) is greater, the probability of being selected as a cluster center is greater, that is, P(R i ) is greater, repeat the above steps until k initial cluster centers are selected;

[0044] Step 6.2, risk factor adjustment:​

[0045] In selecting the initial clustering center, the risk consequence severity σ is introduced as a risk factor to modify the distance, and the modified distance D σ (R i ) is calculated by the formula:

[0046] D σ (R i )=D(R i )×(1+ασ(R i )) (19)

[0047] Wherein, α is an adjustment parameter, σ(R i ) is the risk value of the data point R i ;

[0048] Step 6.3, iterative optimization:

[0049] The distribution stage: each data point is assigned to the nearest cluster center to form k clusters;

[0050] The update stage: calculate the centroid of each cluster and update the clustering center, and the centroid calculation formula is:

[0051]

[0052] Wherein, μ j is the centroid of the jth cluster, C j is all data points in the jth cluster;

[0053] Repeat the distribution and update stage, and use the Euclidean distance to calculate and minimize the sum of the Euclidean distance from each risk sample point R i to the centroid μj, until the clustering center no longer changes or reaches the preset iteration number, and the Euclidean distance calculation formula is:

[0054]

[0055] Wherein, d(R,μ) is the Euclidean distance, R ji is the ith data point in the jth cluster;

[0056] Step 6.4, finally output the cluster division C={C1,C2,C3,C4}.

[0057] Compared with the prior art, the beneficial effects of the present application are:

[0058] 1. Compared with the prior art, the present application proposes a single-source risk quantification model based on the facial features of pilots and a superimposed risk quantification model based on the principle of inclusion-exclusion and the idea of time series, which can effectively evaluate the fatigue driving risk, emotional driving risk and comprehensive risk of pilots according to the video image information of the pilot state, thereby improving the safety level of air transportation.

[0059] 2. Compared with the prior art, the present application proposes a risk grading method and grading early warning control strategy based on the RM-K-means++ algorithm, which classifies the prediction results of the comprehensive risk into first-level risk, second-level risk, third-level risk and fourth-level risk, and can realize comprehensive and differentiated control of the dangerous driving behavior of pilots.

[0060] 3. Compared with the prior art, the present application proposes a pilot dangerous driving risk quantification evaluation and early warning grading method covering "risk quantification, risk superposition, risk prediction and risk grading", which realizes the perception, prediction and early warning of the two single-source risks of fatigue driving risk and emotional driving risk generated by the dangerous driving of pilots in a short time and the comprehensive risk superimposed by the two risks, and an integrated risk evaluation system improves the safety of aircraft flight in all directions. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The method flowchart of the present application.

[0062] Figure 2 The human face 68 key point diagram in the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments.

[0064] In combination with Figure 1 , a dangerous driving risk quantification evaluation and early warning grading method considering risk superposition of the present application is described in detail. The dangerous driving behavior of pilots refers to fatigue driving and emotional driving in the aircraft take-off and landing stage. The risk superposition refers to the risk superposition effect caused by the simultaneous occurrence of fatigue driving and emotional driving of pilots with the change of space-time characteristics. The dangerous driving risk quantification evaluation and early warning grading method of pilots includes the following steps:

[0065] Step 1, selecting pilot dangerous driving single-source risk representation index: considering the strong correlation between the facial features of pilots and their fatigue driving state and emotional driving state, the present application selects the degree of closing eyes P 80 , the frequency of yawning F Yawn , the frequency of nodding Fnod As an important measure of fatigue driving risk, the eye corner up angle θ eye , the mouth corner up angle θ mouth , the eyebrow cohesion degree d brow , the eye opening degree h eye , and the mouth opening degree h mouth are selected as important measure indicators of emotional driving risk. The above facial feature information can be obtained by collecting video images of pilots during flight through a camera.

[0066] Step 2, based on facial geometric features and 68 key points of the face 68, as shown in Figure 2 , the pilot dangerous driving single-source risk representation indicators selected in step 1 are quantitatively analyzed:

[0067] Step 2.1, quantitatively analyze the facial indicators representing fatigue state:

[0068] Step 2.1.1, mark 6 key points on the eyes, calculate the eye aspect ratio E AR (the ratio of the Euclidean distance between the longitudinal and transverse landmarks of the eyes), and the calculation formula is:

[0069]

[0070] The calculation formula of the threshold value of E AR is:

[0071] E AR =(E AR,max -E AR,min )(1-X1)+E AR,min (2)

[0072] In the formula: X1 is the proportion of the eyelid covering the pupil area, E AR,max is the maximum eye aspect ratio, and E AR,min is the minimum eye aspect ratio;

[0073] The number of eye closing frames is obtained from E AR , and the eye closing degree P 80 is calculated therefrom, and the calculation formula is:

[0074]

[0075] In the formula: M is the total number of frames in a unit time sequence video, and N is the number of frames with closed eyes in a unit time;

[0076] Step 2.1.2, mark 8 key points on the mouth, calculate the mouth aspect ratio M AR , and the calculation formula is:

[0077]

[0078] M AR The calculation formula of the threshold value is:

[0079] M AR = (M AR,max -M AR,min )(1-X2)+M AR,min (5)

[0080] In the formula, X2 is the ratio of the mouth opening degree to the maximum opening degree, M AR,max is the maximum mouth aspect ratio, and M AR,min is the minimum mouth aspect ratio.

[0081] The yawning frequency can be obtained according to the mouth opening time percentage, and the calculation formula is:

[0082]

[0083] In the formula, T m is the total detection time, and tm is the time of continuous mouth opening.

[0084] The calculation formula of the head posture fatigue feature, i.e., the nodding frequency, is:

[0085]

[0086] In the formula, M is the total number of frames of the sequence video per unit time, and t nod is the number of frames of the head lowering per unit time.

[0087] Step 2.2, quantitatively analyze the facial indicators representing the emotional state:

[0088] Step 2.2.1, calculate the eye corner upward angle θ eye :

[0089]

[0090] In the formula, inner corner is the key point 40 (left eye) or 43 (right eye), and outer corner is the key point 37 (left eye) or 46 (right eye).

[0091] Step 2.2.2, calculate the mouth corner upward angle θ mouth :

[0092]

[0093] In the formula, mouth_center is the key point 63, and mouth_corner_right is the key point 55.

[0094] Step 2.2.3, calculate the eyebrow cohesion degree d brow:

[0095]

[0096] wherein: brow_left_inner is the key point 21, and brow_right_inner is the key point 24;

[0097] Step 2.2.4, calculating the eye opening degree h eye :

[0098] h eye =y upper_syelid -y lower_eyelid (11)

[0099] wherein: upper_eyelid is the key point 38 (left eye) or 45 (right eye), and lower_eyelid is the key point 42 (left eye) or 47 (right eye);

[0100] Step 2.2.5, calculating the mouth opening degree h mouth :

[0101] h mouth =y upper_lip -y lower_lip (12)

[0102] wherein: upper_lip is the key point 52, and lower_lip is the key point 58.

[0103] Step 3, constructing a single-source risk quantification evaluation model for dangerous driving of pilots:

[0104] Step 3.1, setting judgment criteria for each index, and dividing the threshold into four levels, different levels correspond to different scores, and correspond to four levels of risk, and the default first level risk dangerous coefficient is the largest;

[0105] Step 3.2, according to the scores of each index, a single-source risk model is established, and the formula is:

[0106]

[0107] wherein: λ i (i = 1, 2, 3…n) represents the score of each index, and the range is [0, 1], and Risk represents FR (fatigue driving risk) and ER (emotional driving risk);

[0108] Step 4, constructing a multi-source risk superposition model for dangerous driving of pilots based on the inclusion-exclusion principle and time series idea:

[0109] Step 4.1, in a continuous period of time, facial data of Q groups of pilots during flight driving are collected;

[0110] Step 4.2, calculate the fatigue driving risk value FR and the emotional driving risk value ER of the pilot in the Q group data respectively by using formula (13), and count the data amount M of the first level fatigue driving risk, the data amount N of the first level emotional driving risk, and the data amount K of the first level fatigue driving risk and the first level emotional driving risk in the Q group data;

[0111] Step 4.3, calculate the superimposed probability of the pilot dangerous driving multi-source risk based on the inclusion-exclusion principle, and the formula is:

[0112] P(FR∪ER)=P(FR)+P(ER)-P(FR∩ER)=M / Q+N / Q-K / Q (14)

[0113] Thus, the quantitative evaluation of the superimposed effect of the pilot dangerous driving multi-source risk is realized;

[0114] Step 5, adopt the XGBoost (extreme gradient boosting) machine learning model to predict the comprehensive risk of the pilot dangerous driving:

[0115] Step 5.1, the expression of the pilot dangerous driving comprehensive risk prediction function is:

[0116]

[0117] In the formula: is the pilot dangerous driving comprehensive risk prediction value of the tth iteration, is the characteristic variable of the pilot dangerous driving comprehensive risk, which is composed of the fatigue driving risk value FR, the emotional driving risk value ER and the standardized processing of the superimposed risk value, is the prediction function about ;

[0118] Step 5.2, combine the loss function and the regularization of the suppression function model complexity into the objective function, and the expression of the objective function is:

[0119]

[0120] Wherein: Lt is the objective function value of the tth iteration, is the loss function of a single sample, which represents the error between the pilot dangerous driving comprehensive risk prediction value and the true value, is the overall loss function of all samples,

[0121] is the sum of the regularization terms of the first t iterations, which represents the complexity of the model decision tree;

[0122] Step 5.3, perform second-order Taylor expansion on the objective function, and the expression of the expanded objective function is:

[0123]

[0124] wherein g i , h i are the first derivative and the second derivative, respectively, is the loss function of a single sample in the t-1th iteration,

[0125] Step 5.4, after multiple iterations, the objective function Lt reaches a set value, indicating that the function accuracy meets the requirements, and the pilot dangerous driving comprehensive risk prediction function at this time is the final prediction function;

[0126] Step 6, the RM-K-means++ algorithm is used to cluster analyze the pilot dangerous driving comprehensive risk index R, and by introducing the risk factor, random initialization and multiple iteration optimization, the pilot dangerous driving comprehensive risk is divided into first-level risk, second-level risk, third-level risk and fourth-level risk:

[0127] Step 6.1, initialization stage:

[0128] Obtain the dangerous driving comprehensive risk values of multiple pilots in different states as a point set, randomly select a data point as the first cluster center, and for each remaining data point R i , calculate the distance D(R i ) with the nearest selected cluster center, and take D(R i ) as the weight, and select the next cluster center according to the probability, and the probability formula is as follows:

[0129]

[0130] wherein j represents the number of clusters; the point with larger D(R i ) has a larger probability of being selected as a cluster center, i.e. P(R i ) is larger, and the above steps are repeated until k initial cluster centers are selected;

[0131] Step 6.2, risk factor adjustment:

[0132] When selecting the initial cluster center, a risk factor (risk consequence severity) σ is introduced to modify the distance, and the calculation formula of the modified distance D σ (R i ) is as follows:

[0133] D σ (R i )=D(R i )×(1+ασ(R i )) (19) ​

[0134] wherein, a is an adjustment parameter, σ(R i is the risk value of data point R i ;

[0135] Step 6.3, iterative optimization:

[0136] Distribution stage: each data point is assigned to the nearest cluster center, forming k clusters;

[0137] Update stage: calculate the centroid of each cluster and update the cluster center, the centroid calculation formula is:

[0138]

[0139] wherein, μ j is the centroid of the jth cluster, C j is all data points in the jth cluster;

[0140] Repeat the distribution and update stage, use Euclidean distance calculation and minimize the sum of the Euclidean distance of each risk sample point R i to the centroid μ j , until the cluster center no longer changes or reaches the preset number of iterations, the Euclidean distance calculation formula is:

[0141]

[0142] wherein, d(R, μ) is the Euclidean distance, R ji is the ith data point in the jth cluster;

[0143] Step 6.4, finally output cluster division C={C1, C2, C3, C4};

[0144] Step 7, get the pilot dangerous driving comprehensive risk threshold of four risk levels through step 6, use the real-time acquired facial image, use steps 1 to 5 to predict the pilot dangerous driving comprehensive risk, judge which risk level the pilot dangerous driving comprehensive risk belongs to, if the current pilot risk level needs to be warned, then according to the level, a warning is given, control measures are taken, otherwise, real-time monitoring is continued until the end.

[0145] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art, according to the technical solution and the inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, should be covered within the protection scope of the present application.

Claims

1. A method for quantifying and classifying the risk of dangerous driving, taking into account the superposition of risks, characterized in that, Comprise the following steps: (1) dangerous driving including fatigue driving, emotional driving, considering the strong correlation between the facial features of the pilot and his fatigue driving state and emotional driving state, important measurement indicators representing fatigue driving risk and emotional driving risk are determined from the facial features; (2) based on the face geometry features and the 68 key points of the face, the important measurement indicators representing fatigue driving risk and emotional driving risk in step one are quantitatively analyzed; (3) a pilot dangerous driving single-source risk quantitative evaluation model is constructed, and the pilot dangerous driving single-source risk quantitative evaluation model obtains the fatigue driving risk value and the emotional driving risk value of the pilot according to the quantified indicators; (4) a pilot dangerous driving multi-source risk superposition model is constructed based on the inclusion-exclusion principle and the time series idea, and the pilot dangerous driving multi-source risk superposition model realizes the quantitative evaluation of the superposition effect of the pilot dangerous driving multi-source risk according to the fatigue driving risk value and the emotional driving risk value of the pilot; (5) a pilot dangerous driving comprehensive risk prediction model is constructed, the pilot's face image is obtained in real time, and the fatigue driving risk value FR, the emotional driving risk value ER and the superposition risk value of the pilot are obtained according to steps (1) to (4), which are standardized and then input into the prediction model as characteristic variables to obtain the pilot dangerous driving comprehensive risk prediction value; (6) the dangerous driving comprehensive risk values of multiple pilots are obtained, the pilot dangerous driving comprehensive risk values are clustered and analyzed, the pilot dangerous driving comprehensive risk is classified, and the threshold value of each risk level is obtained; (7) according to the pilot dangerous driving comprehensive risk value obtained in step (5) and the threshold value of each risk level, the risk level of the pilot is determined, and when the risk level of the pilot reaches the required warning, a hierarchical warning signal is given according to the risk level of the pilot, and control measures are taken.

2. The method of claim 1, wherein the risk of dangerous driving is quantified and graded by considering the risk superposition, characterized in that, The important measurement indicators representing fatigue driving risk include the degree of closed eyes, the frequency of yawning and the frequency of nodding; The important measurement indicators representing emotional driving risk include the eye corner up angle, the mouth corner up angle, the eyebrow convergence degree, the eye opening degree and the mouth opening degree.

3. The method of claim 1, wherein the risk of dangerous driving is quantified and graded by considering the risk superposition, characterized in that, The construction of the pilot dangerous driving single-source risk quantitative evaluation model comprises: Step 3.1, set the judgment standard for each indicator, divide the threshold value into four levels, different levels correspond to different scores, and correspond to four risk levels, and the default first risk dangerous coefficient is the largest; Step 3.2, according to the scores of each indicator, establish a single-source risk model, the formula is: where: λ i represent the scores of each index, range is [0, 1], i = 1, 2, 3…n, and Risk represents the fatigue driving risk or emotional driving risk.

4. The method of claim 1, wherein the risk of dangerous driving is quantified and graded by considering the risk superposition, characterized in that, Based on the inclusion-exclusion principle and the time series idea, a pilot dangerous driving multi-source risk superposition model is constructed, and the steps are as follows: Step 4.1, in a continuous period of time, collect the face data of Q pilots during flight driving; Step 4.2, calculate the fatigue driving risk value FR and the emotional driving risk value ER of the pilot in the Q group data respectively by using formula (13), and count the data amount M of the pilot in the first level of fatigue driving risk, the data amount N of the pilot in the first level of emotional driving risk, and the data amount K of the pilot in the first level of fatigue driving risk and the first level of emotional driving risk simultaneously; Step 4.3, calculate the superimposed probability of the pilot's dangerous driving multi-source risk based on the inclusion-exclusion principle, and the formula is: P(FR∪ ER)=P(FR)+P(ER)-P(FR∩ ER)=M / Q+N / Q-K / Q (14) Thus, the quantitative evaluation of the superimposed effect of the pilot's dangerous driving multi-source risk is realized.

5. The method of claim 1, wherein the risk of dangerous driving is quantified and graded by considering the risk of superposition, characterized in that, The prediction model of the pilot's dangerous driving comprehensive risk adopts the XGBoost machine learning model, and the pilot's dangerous driving comprehensive risk prediction function expression is: wherein: is the pilot's predicted value of the overall risk of dangerous driving at the tth iteration, is a characteristic variable of the pilot's overall risk of dangerous driving, composed of the fatigue driving risk value FR, the emotional driving risk value ER and the superimposed risk value after standardization processing, is a prediction function for .

6. The method of claim 5, wherein the risk of dangerous driving is quantified by considering the risk of the superposition of the risk of dangerous driving, characterized in that, After several iterations of the pilot dangerous driving comprehensive risk prediction function expression, the target function L t reaches the set value, and the final prediction function is obtained. The objective function is composed of the loss function and the regularization of the model complexity of the suppression function, and the expression is: wherein: L t is the objective function value of the tth iteration, is the loss function of a single sample, representing the error between the predicted value of the comprehensive risk of dangerous driving of the pilot and the true value, is the overall loss function of all samples, is the sum of the regularization terms of the previous t iterations, representing the complexity of the model decision tree; The objective function is second-order Taylor expanded, and the expanded objective function expression is: where: g i , h i are first and second derivatives, is the loss function for a single sample in iteration t-1.

7. The risk superimposed dangerous driving risk quantitative evaluation and early warning grading method according to claim 5, characterized in that, In step six, the RM-K-means++ algorithm is used for clustering analysis of the pilot's dangerous driving comprehensive risk value. By introducing the risk factor, random initialization and multiple iteration optimization, the pilot's dangerous driving comprehensive risk is divided into first level risk, second level risk, third level risk and fourth level risk: Step 6.1, initialization stage: The dangerous driving comprehensive risk values of multiple pilots in different states are acquired as a point set, a data point is randomly selected as a first cluster center, for each remaining data point R i , the distance D(R i ) with the nearest selected cluster center is calculated, and D(R i ) is taken as a weight to probabilistically select a next cluster center, and the probability formula is as follows: In the formula, j represents the number of clusters; D(R i ) is the distance between the point and the cluster center; P(R i ) is the probability that the point is selected as the cluster center; and k is the number of initial cluster centers. The above steps are repeated until k initial cluster centers are selected. Step 6.2, risk factor adjustment: In selecting the initial clustering center, the risk consequence severity σ is introduced as a risk factor to modify the distance, and the modified distance D σ (R i ) is calculated as follows: D σ (R i )=D(R i )×(1+ασ(R i )) (19) where a is an adjustment parameter, s(R i ) is the risk value of data point R i . Step 6.3, iteration optimization: Allocation stage: allocate each data point to the nearest cluster center to form k clusters; Update stage: calculate the centroid of each cluster and update the cluster center, and the centroid calculation formula is: where μ j is the centroid of the jth cluster, i.e., the threshold for the jth level of risk, C j is all data points of the jth cluster. Repeat the assignment and update phase, using Euclidean distance calculation and minimize each risk sample point R i to the centroid μ j The sum of the Euclidean distances, until the cluster center no longer changes or reaches the preset number of iterations, the Euclidean distance calculation formula is: where d(R, μ) is the Euclidean distance, R ji is the i-th data point in the j-th cluster; Step 6.4, finally output the cluster division C={C1,C2,C3,C4}.

Images