Driving risk identification method and system based on driving factor analysis

Abstract

The invention discloses a driving risk identification method and system based on driving factor analysis, and the method comprises the steps: information obtaining: obtaining a data source which comprises a user database of driver and vehicle attribute information, a driving resume record database and a driving accident record database; database query: extracting a database field according to the unique driver identification number; data integration extraction: carrying out statistics on extracted database fields to form modeling features, and combining the modeling features into a user portrait wide table with a driver identification number as a unique record line index; performing low-risk driving modeling: extracting factors and factor loads of the user portrait wide table through factor analysis, and establishing a logistic regression model of driving risks according to the factors; and low-risk driver identification pushing: calculating a risk factor value and a low-risk tendency value of the driver and judging whether the risk factor value and the low-risk tendency value belong to a low-risk category. The driving risk identification accuracy can be improved, and reliable interpretation can be provided for the method and the system.

Description

Technical Field

[0001] The present invention relates to the technical field of driving risk identification, and in particular to a driving risk identification method and system based on driving factor analysis. Background Art

[0002] As cars become standard equipment for every person, driving risk has become one of the most concerning risks for drivers. When buying a car, they must purchase compulsory traffic insurance and commercial insurance as needed. Auto insurance has become one of the most important businesses in the insurance industry, and assessing the driver's driving risk has become a top priority in terms of reasonable pricing of insurance products.

[0003] The current driving risk identification is mainly based on the perspective of road traffic safety, and is implemented in terms of timely reminders to drivers, behavior improvement, and traffic accident prevention. The potential accident risks of driving are identified by modeling the attribute characteristics of people, vehicles, and roads. However, the group of drivers who meet the model and have high risks is large, but very few of them actually have accidents. Moreover, most of the high-risk attributes that the model focuses on are not something that drivers can change, which is contrary to the original intention of traffic safety prevention. A more reasonable implementation method is to use the driving risk identification model to screen out driver groups, and the possibility of potential accidents in a specific group is as low as possible. Such groups are also what the insurance industry is concerned about.

[0004] Given that current risky driving identification focuses on estimating the risk probability of individual drivers, it is necessary to conduct in-depth and comprehensive exploration of high-risk factors. The differences between individuals are very large, and the accident samples are sparse, making it difficult to effectively make a unified model based on a simple linear or nonlinear model. As a result, the accuracy of risky driving identification in the industry is very low and the application effect is generally poor.

[0005] Positive assessment of driving risks not only has a low accuracy rate, but also makes it difficult to achieve the goal of improving driver safety behavior based on risk notification. Many times, these risks seem insignificant to drivers. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a driving risk identification method and system based on driving factor analysis. Through the method of collaborative feature extraction factors, not only the accuracy of driving risk identification can be improved, but also reliable explanation can be provided for the implementation of the method and system.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A driving risk identification method based on driving factor analysis, comprising:

[0009] Information acquisition: acquiring data sources, including a user database of driver and vehicle attribute information, a driving history record database, and a driving accident record database;

[0010] Database query: Extract database fields based on the unique driver identification number, including matching the associated user database to extract driver and vehicle attribute information, matching the associated driving history database to extract the driver's driving record history for a specified time period, and matching the associated driving accident record database to extract accident records that occurred within a specified time period;

[0011] Data integration extraction: Count the extracted database fields to form modeling features, and merge them into a user portrait wide table with the driver identification number as the unique record row index;

[0012] Driving low risk modeling: extracting factors and factor loadings of the user portrait wide table through factor analysis, and establishing a logistic regression model of driving risk based on the factors;

[0013] Low-risk driver identification push: The driver's risk factor value is calculated based on the modeling characteristics and factor loadings, and then the risk factor value is brought into the logistic regression model to calculate the driver's low-risk tendency value. If it exceeds the threshold, it is judged to belong to the low-risk category and classified into the low-risk driving population.

[0014] Furthermore, in the low-risk driving modeling, establishing a logistic regression model of driving risk according to factors includes:

[0015] Based on the updated factor values, a logistic regression model is established with the low-risk label as the modeling prediction target;

[0016] Use cross entropy loss to calculate the logistic regression model prediction loss;

[0017] Update regression weights and factor loading parameters based on loss backpropagation;

[0018] Update the factor value and iterate continuously until the prediction loss of the logistic regression model is lower than the threshold.

[0019] Furthermore, in the described low-risk driving modeling, the factor risk is evaluated based on the response weight of the factor in the logistic regression model, and the effectiveness of the regression model is verified by comparing with the actual driving risk.

[0020] Furthermore, the factor risk is evaluated based on the response weight of the factor in the logistic regression model, and the effectiveness of the regression model is verified by comparing the actual driving risk, including:

[0021] Traverse all factors in the user portrait wide table and sort them in descending order by the absolute value of factor loading;

[0022] For each factor, the top three feature annotation factors were ranked based on factor loadings;

[0023] The absolute value of the regression weights of the logistic regression model is sorted in descending order;

[0024] Take out the top three factors in the factor loading ranking, and combine them with the risk situation in actual driving scenarios to verify whether the factor really explains the corresponding risk. If it can explain it, it means that the model is effective.

[0025] Furthermore, the information acquisition includes:

[0026] Use the gravity acceleration sensor module to upload the vehicle's vector acceleration in the three vector directions of x, y, and z;

[0027] Use the gyroscope sensor module to upload the vehicle's angular velocity in three vector directions: x, y, and z.

[0028] Upload the vehicle's location information in real time through the GPS module;

[0029] Record driver and vehicle attribute information through user database;

[0030] The vehicle accident data is recorded through the vehicle accident database.

[0031] A driving risk identification system based on driving factor analysis, comprising:

[0032] An information acquisition module is configured to acquire a data source, wherein the data source includes a user database of driver and vehicle attribute information, a driving history record database, and a driving accident record database;

[0033] A database query module is configured to extract database fields according to the unique driver identification number, including matching the associated user database to extract driver and vehicle attribute information, matching the associated driving history database to extract the driver's driving record history for a specified time period, and matching the associated driving accident record database to extract accident records that occurred within the specified time period;

[0034] The data integration and extraction module is configured to count the extracted database fields to form modeling features and merge them into a user portrait wide table with the driver identification number as the unique record row index;

[0035] A driving low risk modeling module is configured to extract factors and factor loadings of the user portrait wide table through factor analysis, and establish a logistic regression model of driving risk based on the factors;

[0036] The low-risk driver identification push module is configured to calculate the driver's risk factor value based on the modeling characteristics and factor loading, and then bring the risk factor value into the logistic regression model to calculate the driver's low-risk tendency value. If it exceeds the threshold, it is judged to belong to the low-risk category and classified into the low-risk driving population.

[0037] Furthermore, in the driving low risk modeling module, establishing a logistic regression model of driving risk according to factors includes:

[0038] Based on the updated factor values, a logistic regression model is established with the low-risk label as the modeling prediction target;

[0039] Use cross entropy loss to calculate the logistic regression model prediction loss;

[0040] Update regression weights and factor loading parameters based on loss backpropagation;

[0041] Update the factor value and iterate continuously until the prediction loss of the logistic regression model is lower than the threshold.

[0042] Furthermore, in the driving low-risk modeling module, the factor risk is evaluated based on the response weight of the factor in the logistic regression model, and the effectiveness of the regression model is verified by comparing with the actual driving risk.

[0043] Furthermore, the factor risk is evaluated based on the response weight of the factor in the logistic regression model, and the effectiveness of the regression model is verified by comparing the actual driving risk, including:

[0044] Traverse all factors in the user portrait wide table and sort them in descending order by the absolute value of factor loading;

[0045] For each factor, the top three feature annotation factors were ranked based on factor loadings;

[0046] The absolute value of the regression weights of the logistic regression model is sorted in descending order;

[0047] Take out the top three factors in the factor loading ranking, and combine them with the risk situation in actual driving scenarios to verify whether the factor really explains the corresponding risk. If it can explain it, it means that the model is effective.

[0048] Furthermore, the information acquisition module includes:

[0049] The gravity acceleration sensor module is configured to upload the vehicle's vector acceleration in three vector directions: x, y, and z;

[0050] The gyroscope sensor module is configured to upload the vehicle's angular velocity in three vector directions: x, y, and z;

[0051] The GPS module is configured to upload the vehicle's location information in real time;

[0052] A user database configured to record driver and vehicle attribute information;

[0053] The vehicle accident database is configured to record the driving accident data of the vehicle.

[0054] The beneficial effects of the present invention are:

[0055] Compared with the current driving risk identification technology, the present invention avoids the entanglement of high-risk behaviors with complex and changeable scenario factors, and directly models the low-risk behaviors of drivers. Since drivers have basic safety awareness, high-risk driving behaviors are rare events, with large individual differences and small event samples. Whether linear models or nonlinear models are used, the accuracy and application value of driving risk identification are very low. The present invention takes into account the ubiquity and commonality of low-risk behaviors, and uses linear models to model, which not only ensures high accuracy and stability but is also simple and easy to use. The present invention uses factor analysis to achieve a model with strong explanatory power, which is also a first in driving risk identification technology. This is very friendly for engineers to understand and use the model, and has high application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is one of the flow charts of the driving risk identification method provided by the present invention.

[0057] Figure 2 This is the second flow chart of the driving risk identification method provided by the present invention.

[0058] Figure 3 This is a structural block diagram of the driving risk identification system provided by the present invention.

[0059] Figure 4 This is a schematic diagram of the application of the driving risk identification system provided by the present invention. DETAILED DESCRIPTION

[0060] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, the specific implementation methods of the present invention are now described. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0061] Example 1

[0062] Given that current risky driving identification focuses on estimating the risk probability of individual drivers, it is necessary to conduct in-depth and comprehensive exploration of high-risk factors. The differences between individuals are very large, and the accident samples are sparse, making it difficult to effectively make a unified model based on a simple linear or nonlinear model. As a result, the accuracy of risky driving identification in the industry is very low and the application effect is generally poor.

[0063] Now let's look at the problem of driving risk identification from another angle, that is, screening low-risk drivers. Although the accident situations of each person are different, all low-risk drivers have consistent characteristics, and there are enough target samples, so the modeling effect is naturally better. Based on the integration of continuously recorded three-axis acceleration and three-axis gyroscope data, GPS positioning data, and road condition map data, the attribute characteristics of people, vehicles, roads, etc. are extracted, and then the factor analysis method is used to combine the attribute characteristics and form risk explanation factors, and the risk factors are used to model and evaluate driving risks. It is also worth mentioning that positively evaluating driving risks not only has a low accuracy rate, but also makes it difficult to achieve the purpose of improving driver safety behavior based on risk notification. In many cases, these risks are not worth mentioning in the eyes of drivers. The insurance industry pays more attention to low-risk drivers, and provides a large number of preferential insurance packages for this purpose, while giving non-low-risk drivers much higher insurance premiums, thereby forcing drivers to move closer to low-risk driving characteristics, truly achieving the original purpose of guiding improved driving behavior and traffic accident prevention.

[0064] Based on this, this embodiment provides a driving risk identification method based on driving factor analysis, such as Figure 1 As shown, including:

[0065] Information acquisition: Acquisition of data sources, including a user database of driver and vehicle attribute information, a driving history record database, and a driving accident record database;

[0066] Database query: Extract database fields based on the unique driver identification number, including matching the associated user database to extract driver and vehicle attribute information, matching the associated driving history database to extract the driver's driving record history for a specified time period, and matching the associated driving accident record database to extract accident records that occurred within a specified time period;

[0067] Data integration extraction: Count the extracted database fields to form modeling features, and merge them into a user portrait wide table with the driver identification number as the unique record row index;

[0068] Driving low-risk modeling: extract factors and factor loadings of the user portrait wide table through factor analysis, and establish a logistic regression model of driving risk based on the factors;

[0069] Low-risk driver identification push: The driver's risk factor value is calculated based on modeling characteristics and factor loadings, and then the risk factor value is brought into the logistic regression model to calculate the driver's low-risk tendency value. If it exceeds the threshold, it is judged to belong to the low-risk category and classified into the low-risk driving population.

[0070] like Figure 2 As shown, the driving risk identification method of this embodiment can be implemented by the following steps:

[0071] a. Based on the driver's residence information, randomly sample the driver group in the designated target location without duplication to form a modeling sample group. The sample is accompanied by the driver's identification number, gender, age, driving model, and age of the vehicle;

[0072] b. According to the identification number of the sample driver, the driving record of the driver in the past 90 days is matched from the database, which contains GPS latitude and longitude, timestamp, speed, acceleration sensor data, and gyroscope sensor data;

[0073] c. According to the identification number of the sample driver, the driving accident records of the sample driver in the past 90 days are matched from the accident database, and the risk labels of the samples are summarized and counted;

[0074] d. Sort by record timestamp and divide the 90-day continuous driving record data into six sections of 15 days each, including six data blocks: I1, I2, I3, I4, I5, and I6;

[0075] e. Based on data block I1, traverse by driver number, summarize and extract driving behavior characteristics and road section characteristics, and statistically process the remaining five data blocks I2-I6 in the same process;

[0076] f. The six feature statistics output in step d are matched according to the driver number, and feature statistics are refined to form the input features of the model;

[0077] g. Associate the data output from step a with the label data output from step b by driver number, and use the WOE technique to convert gender, age, driving model, and vehicle age into continuous variables;

[0078] h. According to the driver number, the data tables output from steps f and g are associated to form a feature wide table containing a risk label field;

[0079] i. Take out the characteristic fields in the wide table output in step h, use factor analysis techniques to estimate the number of factors and calculate factor loadings as initial loading values;

[0080] j. After extracting the factor expression of the features using factor loading, the driver risk modeling is done using logistic regression technology;

[0081] k. Evaluate the risk of the factors based on the response weights of the factors in the regression model and compare them with the actual driving risk to verify the effectiveness of the regression model;

[0082] l. Estimate driving risk based on the regression model. If the risk value reaches the specified threshold, the driver is classified as a low-risk group.

[0083] Preferably, step a specifically includes:

[0084] a1. Take out the driver sample population that meets the specified place of residence, assuming the sample size is N;

[0085] a2. Assume that the required sample size is M, where M < N. In the first round, randomly select one from N samples with a probability of 1 / N. In the second round, randomly select one from the remaining N - 1 samples with a probability of 1 / (N - 1), and repeat this method until a total of M samples are drawn.

[0086] Preferably, step c specifically includes:

[0087] c1. According to SQL and database query techniques, match the driving accident records in the accident database in the past 90 days by driver identification number. In the accident database, the accident types include minor accidents such as minor scratches, major accidents, single - vehicle or two - vehicle accidents;

[0088] c2. If there are no accidents of any type in the records in the past 90 days that match the driver number, it is a low - risk and marked as 1, otherwise marked as 0.

[0089] Preferably, step e specifically includes:

[0090] e1. For the currently obtained data block I1, group by driver identification number and screen out the groups with a total driving mileage less than 10 kilometers within the group;

[0091] e2. For the groups obtained from data block I1, count the proportion of night driving mileage, the proportion of highway driving mileage, the proportion of days with continuous driving duration greater than 4 hours, the proportion of driving mileage on roads with a width less than two lanes, the proportion of weekend driving mileage, the proportion of rural driving mileage, the proportion of downtown driving mileage, the proportion of the number of emergency deceleration stops relative to the number of stops, the proportion of high - speed turns relative to the number of turns, and the proportion of rapid acceleration starts relative to the number of stop starts within the group;

[0092] e3. Similarly, perform steps e1 - e2 on data block I2;

[0093] e4. Similarly, perform steps e1 - e2 on data block I3;

[0094] e5. Similarly, perform steps e1 - e2 on data block I4;

[0095] e6. Similarly, perform steps e1 - e2 on data block I5;

[0096] e7. Similarly, perform steps e1 - e2 on data block I6.

[0097] Preferably, step f specifically includes:

[0098] f1. Based on the driver identification number as the primary key, query the six data blocks output in step e. If no record is matched in a data block, it is set to a null value of Null. With the primary key as the index, for each feature of step e, the corresponding six values ​​are matched;

[0099] f2. Traverse the identification number primary key and the statistical features of step e, and calculate the variance and mean for each set of matching values, where null values ​​are not included in the statistics;

[0100] f3. Use the driver identification number as the record row index, and the variance and mean of each statistical feature as the field column to form a data wide table.

[0101] Preferably, step g specifically comprises:

[0102] g1. The driver identification number is indexed and the data table associated with the output of steps a and b is displayed;

[0103] g2. Traverse the attribute fields in step a, and group and calculate the low-risk WOE values ​​(weight of evidence) by attribute level values;

[0104] g3. The statistical WOE value corresponds to the level value of the replacement attribute.

[0105] Preferably, step i specifically includes:

[0106] i1. Based on the factor analysis method in statistics, the weighted average expression of the wide table feature fields is extracted from the wide table data output in step h;

[0107] i2. Based on the principal component decomposition technology, the number of factors is screened and determined according to the requirement that the information retention ratio exceeds 90%, that is, the number of weighted average expressions in step i1;

[0108] i3. Use the factor loading calculated in step i2 as the initial value of the factor weight, substitute it into the weighted average expression, and calculate the factor value.

[0109] Preferably, step j specifically includes:

[0110] j1. Use the calculated updated factor values ​​to perform logistic regression with the low-risk label prepared in step c as the modeling prediction target;

[0111] j2. Use cross entropy loss to calculate the regression model prediction loss;

[0112] j3. Update regression weights and factor loading parameters based on loss back propagation;

[0113] j4. Update the factor value and repeat steps j1-j3 until the loss value calculated in step j2 is lower than the specified threshold.

[0114] Preferably, step k specifically includes:

[0115] k1. Traverse all factors of the model output in step j and sort them in descending order according to the absolute value of factor loading;

[0116] k2. For each factor, annotate the feature factors based on the top three loadings;

[0117] k3. Sort by the absolute value of the regression weight of the regression model in descending order;

[0118] k4. Take out the factors of the top three regression loads and combine them with the risk situation in actual driving scenarios to verify whether the factors really explain the corresponding risks. If they can explain, it means that the model is effective.

[0119] Example 2

[0120] This embodiment is based on embodiment 1:

[0121] This embodiment provides a driving risk identification method based on driving factor analysis, which specifically includes the following steps:

[0122] STEP 1: According to the driver's residence information, randomly sample the driver group in the designated target location without duplication. Assume that the sample size M is obtained, where the fields include driver identification number, gender, age, driving model, and vehicle age, recorded as Table 1;

[0123] STEP 2: According to the driver identification number in Table 1, the driving records of the driver in the past 90 days are obtained from the database. In addition to the driver identification number, the extracted fields include GPS latitude and longitude, timestamp, speed, acceleration sensor value, and gyroscope sensor value, which are recorded as Table 2;

[0124] STEP 3: According to the driver identification number in Table 1, match and obtain the driving accident records of the driver in the past 90 days from the accident database, and count whether there is zero accident as the risk label. Its fields include driver identification number and low risk label, which is recorded as Table 3;

[0125] STEP 4: According to the GPS longitude and latitude in Table 2, obtain the corresponding road type, number of lanes, whether it is a town road, including the GPS longitude and latitude fields from the map database, which is recorded as Table 4;

[0126] STEP 5: Use GPS longitude and latitude as the primary key to concatenate Table 2 and Table 4, and record the combined table as Table 5;

[0127] STEP 6: Use Table 5 and sort the timestamp field in ascending order to divide the 90-day continuous driving record data into six sections of 15 days each, which are recorded as Table I1, Table I2, Table I3, Table I4, Table I5, and Table I6 respectively;

[0128] STEP 7: In Table I1, group the drivers by their identification numbers and exclude the groups with a total mileage of less than 10 kilometers.

[0129] STEP 8: According to the grouping obtained in STEP 7, the proportion of nighttime mileage, the proportion of highway mileage, the proportion of days with continuous driving time greater than 4 hours, the proportion of mileage with a road width of less than two lanes, the proportion of weekend mileage, the proportion of rural mileage, the proportion of city center mileage, the proportion of sudden deceleration and stop times relative to the number of stops, the proportion of high-speed turns relative to the number of turns, and the proportion of sudden acceleration and start times relative to the number of stops and starts are counted in the record table consisting of these ten statistical fields and the driver identification number as the index field, which is recorded as T1, with a total of N1 rows;

[0130] STEP 9: Similarly, perform STEP7-STEP8 operations on table I2, and the statistical output table is recorded as T2, with a total of N2 rows;

[0131] STEP 10: Similarly, perform STEP7-STEP8 operations on table I3, and the statistical output table is recorded as T3, with a total of N3 rows;

[0132] STEP 11: Similarly, perform STEP7-STEP8 operations on table I4, and the statistical output table is recorded as T4, with a total of N4 rows;

[0133] STEP 12: Similarly, perform STEP7-STEP8 on table I5, and the statistical output table is recorded as T5, with a total of N5 rows;

[0134] STEP 13: Similarly, perform STEP7-STEP8 operations on Table I6, and the statistical output table is recorded as T6, with a total of N6 rows;

[0135] STEP 14: Tables T1-T6 have the same fields. They are spliced ​​vertically according to the data columns corresponding to the fields. Table T1 is spliced ​​with T2, and then spliced ​​with T3. After splicing T6, a data table with N1+N2+N3+N4+N5+N6 rows is obtained. The field names are the same as T1, recorded as Table 6.

[0136] STEP 15: According to Table 6, group by driver identification number, and count the sample variance and mean in each feature field group. At this time, 20 features are extracted from the ten features in Table 6. The driver identification number is also used as the index field to form a record table, which is recorded as Table 7.

[0137] STEP 16: Combine Table 1 output from STEP 1 and Table 3 output from STEP 3 with the driver identification number as the primary key. The combined data table is recorded as Table 8.

[0138] STEP 17: For the field age in Table 8, count the frequency percentage of each age level. If the corresponding percentage of a level value is less than 5%, merge it into the nearest level value until the frequency percentage of the merged level is higher than 5%. Assume that there are k level categories V1, V2, … V k ,In addition, the number of risk labels in Table 8 is counted, and there are R1 low-risk labels 1 and R0 low-risk labels 0;

[0139] STEP 18: V1-V k The horizontal categories are grouped and the number of low-risk marks 1 and 0 are counted respectively, denoted as V i1 and V i0 , where i = 1, 2, ..., k, using the WOE technology, substitute the following formula to get the WOE value corresponding to these k levels, recorded as W1, W2, ..., W k ,in:

[0140]

[0141] STEP 19: Based on W i With V i The corresponding relationship is shown in Figure 2. i Map Replacement V i The level value in the level class completes the update of the age attribute value. Similarly, the fields of gender, driving model, and vehicle age are respectively updated through STEP17-STEP18, and the attribute values ​​of the corresponding fields are updated. The new table is recorded as Table 9.

[0142] STEP 20: Use the driver identification number as the primary key to associate and merge Table 7 and Table 9, and record the merged table as Table 10;

[0143] STEP 21: Take out all the characteristic field columns in Table 10, a total of 24 columns, recorded as X1, X2, …, X 24 , respectively count the field means, denoted as u1,u2,…,u 24 , substitute the following formula to calculate the characteristic covariance matrix Σ, and solve the singular value decomposition of the covariance matrix, which is expressed as follows:

[0144]

[0145] STEP 22: In the decomposed covariance matrix, U is the eigenvector of the covariance matrix, λ i ,i=1,2,..,24 is the variance contribution of the i-th factor, and they are arranged in descending order. The top p is selected so that the total variance contribution accounts for more than 95%, and the remaining variance contributions are suppressed to 0. Simplify L and delete redundant factors, where U p is the matrix composed of eigenvectors corresponding to λ, and the specific formula is as follows:

[0146]

[0147] STEP 23: Calculate the output L from STEP 22 p That is the estimated factor loading matrix, with a total of p factors. With the help of factor rotation technology, L p The two columns in the matrix are paired and rotated by θ to maximize the sample variance of the factor loading. The specific rotation formula is as follows, taking the first two columns as an example:

[0148]

[0149] STEP 24: Rotate other factor column pairs in the same way. After traversing all pairs, repeat multiple rounds until the change in factor loading variance is less than the specified threshold. The updated L p is a factor loading with strong explanatory power. The p factors are recorded as f1, f1, …, f p ;

[0150] STEP 25: Based on factor analysis and least squares technology, use feature X i (i=1,2…,24) estimation factor, its expression is as follows, where the coefficient matrix of the estimation expression is recorded as B:

[0151]

[0152] STEP 26: Use the estimated factors to make logistic regression predictions for risk labels, use standard normal distribution random numbers to initialize regression parameters, and then use the cross entropy function to calculate its prediction loss. The specific formula is as follows. Note that there is no constant term when using factors for regression:

[0153]

[0154] The cross entropy loss function of the predicted value p is as follows:

[0155] Loss(β)=-∑(y*ln(p)+(1-y)*ln(1-p))

[0156] STEP 27: Use gradient descent to find the optimal regression parameters and use the updated β i , calculate the gradient function of the parameters, the specific formula is as follows:

[0157]

[0158] Update the regression parameters as follows, where α is the preset learning rate, which is set to 0.001 here:

[0159] β i =β i -α*dLoss / dβi

[0160] STEP 28: Repeat STEP 27 until the loss function value converges or stops after the specified number of rounds. At this time, reset the learning rate to a much smaller value of α = 0.0001, and combine the factor estimation and regression prediction together. It can be regarded as a three-layer neural network with one hidden layer of p nodes, whose activation function is a linear function, and one output layer node, whose activation function is a sigmoid function.

[0161] STEP 29: Similar to the gradient descent method, perform gradient backpropagation on the neural network parameters, use the reset learning rate to update the hidden layer and predicted parameters, repeat the process for the specified number of rounds and stop, to achieve fine-tuning and optimization of the overall model;

[0162] STEP 30: Use the factor explanation output from STEP 24 and the factor weights in the regression model output from STEP 29 to evaluate the role of risk factors in driving risk assessment and verify them with practical experience to prove the effectiveness of the driving risk identification model;

[0163] STEP 31: Use the driving risk identification model to make predictions and calculate the driver's low-risk tendency value. Based on example experience, if it exceeds the threshold of 0.5, it is judged to be in the low-risk category and classified as a low-risk driver.

[0164] Example 3

[0165] like Figure 3 and Figure 4 As shown, this embodiment provides a driving risk identification system based on driving factor analysis, including:

[0166] An information acquisition module is configured to acquire data sources, the data sources including a user database of driver and vehicle attribute information, a driving history record database, and a driving accident record database;

[0167] A database query module is configured to extract database fields according to the unique driver identification number, including matching the associated user database to extract driver and vehicle attribute information, matching the associated driving history database to extract the driver's driving record history for a specified time period, and matching the associated driving accident record database to extract accident records that occurred within the specified time period;

[0168] The data integration and extraction module is configured to count the extracted database fields to form modeling features and merge them into a user portrait wide table with the driver identification number as the unique record row index;

[0169] The driving low-risk modeling module is configured to extract factors and factor loadings of the user portrait wide table through factor analysis, and establish a logistic regression model of driving risk based on the factors;

[0170] The low-risk driver identification push module is configured to calculate the driver's risk factor value based on modeling characteristics and factor loadings, and then bring the risk factor value into the logistic regression model to calculate the driver's low-risk tendency value. If it exceeds the threshold, it is judged to belong to the low-risk category and classified into the low-risk driving population.

[0171] Preferably, the information acquisition module includes: a gravity acceleration sensor module, used to upload the vector acceleration value of the car through the three vector directions of x\y\z; a gyroscope sensor module, used to upload the angular velocity of the car through the three vector directions of x\y\z; and a GPS module, a user database and a vehicle accident database.

[0172] Preferably, the database query module includes: an access module for storing massive historical data through big data cloud technology, and quickly querying and obtaining data that meets specified conditions.

[0173] Preferably, the low-risk driver identification push module includes:

[0174] A low-risk driver identification unit, which outputs the driver's characteristic data based on the data integration and extraction module, calculates the driver's risk factor and evaluates the driver's low-risk tendency value;

[0175] The low-risk driver screening and pushing unit is used to collect the driver identification numbers that meet the threshold conditions output by the low-risk driver identification unit and push them to relevant personnel for verification.

[0176] Example 4

[0177] This embodiment is based on embodiment 1:

[0178] This embodiment provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the driving risk identification method based on driving factor analysis of embodiment 1 are implemented. The computer program may be in source code form, object code form, executable file, or some intermediate form.

[0179] Example 5

[0180] This embodiment is based on embodiment 1:

[0181] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the driving risk identification method based on driving factor analysis of Example 1. The computer program may be in source code form, object code form, executable file, or some intermediate form. The storage medium includes: any entity or device capable of carrying computer program code, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content contained in the storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electric carrier signals and telecommunication signals.

[0182] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

Claims

1. A driving risk identification method based on driving factor analysis, characterized in that: include: Information acquisition: acquiring data sources, including a user database of driver and vehicle attribute information, a driving history record database, and a driving accident record database; Database query: Extract database fields based on the unique driver identification number, including matching the associated user database to extract driver and vehicle attribute information, matching the associated driving history database to extract the driver's driving record history for a specified time period, and matching the associated driving accident record database to extract accident records that occurred within a specified time period; Data integration extraction: Count the extracted database fields to form modeling features, and merge them into a user portrait wide table with the driver identification number as the unique record row index; Driving low risk modeling: extracting factors and factor loadings of the user portrait wide table through factor analysis, and establishing a logistic regression model of driving risk based on the factors; Low-risk driver identification push: The driver's risk factor value is calculated based on the modeling characteristics and factor loading, and then the risk factor value is brought into the logistic regression model to calculate the driver's low-risk tendency value. If it exceeds the threshold, it is judged to belong to the low-risk category and classified into the low-risk driving population.

2. The driving risk identification method based on driving factor analysis according to claim 1 is characterized in that: In the low-risk driving modeling, the logistic regression model of driving risk is established according to the factors including: Based on the updated factor values, a logistic regression model is established with the low-risk label as the modeling prediction target; Use cross entropy loss to calculate the logistic regression model prediction loss; Update regression weights and factor loading parameters based on loss backpropagation; Update the factor value and iterate continuously until the prediction loss of the logistic regression model is lower than the threshold.

3. The driving risk identification method based on driving factor analysis according to claim 1 is characterized in that: In the described low-risk driving modeling, the factor risk is evaluated based on the response weights of the factors in the logistic regression model, and the effectiveness of the regression model is verified by comparing the actual driving risks.

4. The driving risk identification method based on driving factor analysis according to claim 3 is characterized in that: The factor risk is evaluated based on the response weight of the factor in the logistic regression model, and the effectiveness of the regression model is verified by comparing the actual driving risk, including: Traverse all factors in the user portrait wide table and sort them in descending order by the absolute value of factor loading; For each factor, the top three feature annotation factors were ranked based on factor loadings; The absolute value of the regression weights of the logistic regression model is sorted in descending order; Take out the top three factors in the factor loading ranking, and combine them with the risk situation in actual driving scenarios to verify whether the factor really explains the corresponding risk. If it can explain it, it means that the model is effective.

5. The driving risk identification method based on driving factor analysis according to claim 1 is characterized in that: The information acquisition includes: Use the gravity acceleration sensor module to upload the vehicle's vector acceleration in the three vector directions of x, y, and z; Use the gyroscope sensor module to upload the vehicle's angular velocity in three vector directions: x, y, and z. Upload the vehicle's location information in real time through the GPS module; Record driver and vehicle attribute information through user database; The vehicle accident data is recorded through the vehicle accident database.

6. A driving risk identification system based on driving factor analysis, characterized in that: include: An information acquisition module is configured to acquire a data source, wherein the data source includes a user database of driver and vehicle attribute information, a driving history record database, and a driving accident record database; A database query module is configured to extract database fields according to the unique driver identification number, including matching the associated user database to extract driver and vehicle attribute information, matching the associated driving history database to extract the driver's driving record history for a specified time period, and matching the associated driving accident record database to extract accident records that occurred within the specified time period; The data integration and extraction module is configured to count the extracted database fields to form modeling features and merge them into a user portrait wide table with the driver identification number as the unique record row index; A driving low risk modeling module is configured to extract factors and factor loadings of the user portrait wide table through factor analysis, and establish a logistic regression model of driving risk based on the factors; The low-risk driver identification push module is configured to calculate the driver's risk factor value based on the modeling characteristics and factor loading, and then bring the risk factor value into the logistic regression model to calculate the driver's low-risk tendency value. If it exceeds the threshold, it is judged to belong to the low-risk category and classified into the low-risk driving population.

7. The driving risk identification system based on driving factor analysis according to claim 6 is characterized in that: In the driving low risk modeling module, establishing a logistic regression model of driving risk according to factors includes: Based on the updated factor values, a logistic regression model is established with the low-risk label as the modeling prediction target; Use cross entropy loss to calculate the logistic regression model prediction loss; Update regression weights and factor loading parameters based on loss backpropagation; Update the factor value and iterate continuously until the prediction loss of the logistic regression model is lower than the threshold.

8. The driving risk identification system based on driving factor analysis according to claim 6 is characterized in that: In the driving low-risk modeling module, the factor risk is evaluated based on the response weight of the factor in the logistic regression model, and the effectiveness of the regression model is verified by comparing the actual driving risk.

9. The driving risk identification system based on driving factor analysis according to claim 8 is characterized in that: The factor risk is evaluated based on the response weight of the factor in the logistic regression model, and the effectiveness of the regression model is verified by comparing the actual driving risk, including: Traverse all factors in the user portrait wide table and sort them in descending order by the absolute value of factor loading; For each factor, the top three feature annotation factors were ranked based on factor loadings; The absolute value of the regression weights of the logistic regression model is sorted in descending order; Take out the top three factors in the factor loading ranking, and combine them with the risk situation in actual driving scenarios to verify whether the factor really explains the corresponding risk. If it can explain it, it means that the model is effective.

10. The driving risk identification system based on driving factor analysis according to claim 6, characterized in that: The information acquisition module includes: The gravity acceleration sensor module is configured to upload the vehicle's vector acceleration in three vector directions: x, y, and z; The gyroscope sensor module is configured to upload the vehicle's angular velocity in three vector directions: x, y, and z; The GPS module is configured to upload the vehicle's location information in real time; A user database configured to record driver and vehicle attribute information; The vehicle accident database is configured to record the driving accident data of the vehicle.

Images