CN110544373B - A method of truck warning information extraction and risk identification based on Beidou Internet of Vehicles - Google Patents

Abstract

The invention relates to a method for extracting and risk identification of truck warning information based on the Beidou Internet of Vehicles, comprising: step 1, obtaining original data related to vehicle early warning through a vehicle network vehicle terminal provided with a Beidou positioning system; The data is preprocessed, step 3, extraction of key variables of truck warning information, step 4, clustering of vehicle safety risks, step 5, discriminant analysis, and step 6, risk identification. The present invention can give the warning frequency per unit driving mileage and the warning frequency per unit driving time for a certain vehicle/driver, and then can provide the judgment function of the discriminant analysis based on the historical data of the Internet of Vehicles, so as to realize the identification of its risk or predict.

Description

A method of truck warning information extraction and risk identification based on Beidou Internet of Vehicles

technical field

The invention belongs to the application field of traffic safety and is applied to the transportation industry, in particular to a method for extracting early warning information and risk identification of trucks based on the Beidou Internet of Vehicles.

Background technique

On April 12, 2019, the Ministry of Transport announced the 2018 Statistical Bulletin on the Development of the Transportation Industry. The data shows that in 2018, my country's annual road freight volume was 39.569 billion tons, an increase of 7.3%, and the road freight turnover was 7,124.921 billion ton-kilometers, an increase of 6.7%. . In recent years, my country's road freight volume and cargo turnover are still gradually increasing, but road traffic safety problems have become increasingly prominent, and road traffic safety accidents are still a major hidden danger in the development of my country's transportation industry. According to the statistics of the Traffic Management Bureau of the Ministry of Public Security, at the end of 2016, there were 194 million vehicles in the country, including 13.5177 million trucks (referred to as trucks), accounting for 7.0% of the total number of vehicles. In 2016, there were 50,400 road traffic accidents involving trucks in the country. , accounting for 30.5% of the total number of automobile liability accidents in the country, which is much higher than the proportion of the number of trucks in the total number of vehicles. It can be seen that the traffic safety of trucks is particularly important, and relevant research on how to reduce truck traffic accidents is imminent.

Traditional road traffic safety research is mostly based on traffic accident data to build accident frequency or accident severity models to reveal the characteristics of traffic accidents and their key influencing factors from four aspects: people, vehicles, roads, and the environment. With the rapid development of the Internet and computer technology, traffic data collection methods are more diversified and intelligent, and the application of Internet of Vehicles technology in traffic safety has been developed accordingly. Most of the existing traffic safety research based on the Internet of Vehicles involves model methods, devices and systems for vehicle early warning. For example, the patent CN109615879A is an early warning model for abnormal vehicle speed based on the Internet of Vehicles, the patent CN109584630A is a vehicle lane change early warning device and an early warning method based on the Internet of Vehicles, CN109147279A is based on the Internet of Vehicles. , CN109584631A is an anti-collision warning method and system based on the Internet of Vehicles. However, there are few related technologies on how to improve road traffic safety by using the early warning information of truck on-board devices based on the Internet of Vehicles.

At present, the application of Internet of Vehicles technology in the field of transportation has broad prospects. The Ministry of Transport stipulates that intercity freight vehicles must be connected to the Internet, and real-time supervision and monitoring of the operation of freight vehicles through the Internet of Vehicles can improve road traffic safety. At present, freight vehicles in many provinces and cities use the Beidou Internet of Vehicles system, and require trucks to install corresponding safety warning devices. The Beidou Internet of Vehicles system uses advanced Beidou positioning, navigation, sensing, control and other technologies to build a vehicle-to-vehicle system with vehicles as nodes to monitor and store the vehicle's driving track, speed, time, mileage and alarms in the driving state in real time. Early warning information, etc. Among them, vehicle early warning information mainly includes speeding warning, fatigue driving warning, collision warning, rollover warning, etc. However, there is still a lack of relevant technologies on how to extract and utilize these early warning information to improve road traffic safety. The early warning mostly occurs before the vehicle violates regulations, indicating that there is a certain potential risk. The early warning is to remind drivers to regulate their driving behavior and reduce the occurrence of traffic accidents and violations. Driver traffic accidents or violations are deterministic dangerous behaviors, with a relatively low probability of occurrence, and the randomness of whether the event occurs or not; while early warning is a potentially dangerous behavior, with a higher probability of occurrence, a larger amount of data, and more information about the warning. A comprehensive and in-depth reflection of the driver's driving risk. Accordingly, the present invention proposes a truck warning information extraction and risk identification method based on the Beidou Internet of Vehicles. Based on the Internet of Vehicles technology, the original data related to vehicle early warning is obtained, the data is processed, and the key variables of the early warning information are extracted. Clustering and discriminant analysis of the risk level of vehicles/drivers can accurately identify high-risk vehicles/drivers, which can urge drivers to develop good driving habits and improve road traffic safety.

SUMMARY OF THE INVENTION

In view of the defects existing in the prior art, the present invention proposes a method for extracting early warning information and risk identification for trucks based on the Beidou Internet of Vehicles.

In order to achieve the above purpose, the technical scheme adopted in the present invention is:

A method for extracting and risk identification of truck warning information based on Beidou Internet of Vehicles, comprising the following steps:

Step 1, obtain the original data related to vehicle early warning through the vehicle network vehicle terminal equipped with the Beidou positioning system, the original data includes: mileage information, safety warning information, status data, and the status data includes vehicle ID, ACC status, upload time, etc. ;

Step 2, preprocessing the original data

Use Python programming technology to preprocess and filter the original data; before analyzing the original data, it is necessary to clean and organize the data to improve the data quality; data cleaning includes: filling the missing values in the data, identifying the outliers in the data and redundant data; missing values are mainly manifested in the lack of attribute values, and outliers are mainly manifested in that the value of a single attribute is too large or too small. Combined with the characteristics of the data stored in the vehicle terminal, the original data is preprocessed. The specific methods are as follows:

Step 2.1, operation of missing data values: use Python to perform programming operations, introduce the os and numpy modules in Python, define the required functions, execute the main function, operate the text files stored in the vehicle terminal, and delete the text with missing attribute values. file to ensure the integrity of the attributes;

Step 2.2, operation of abnormal data value: abnormal data value includes: abnormal mileage, abnormal warning status, abnormal upload time;

1) Handling of abnormal mileage: First, traverse all data files to calculate the mileage of the vehicle on the day, secondly, make a cumulative distribution map of the mileage of the vehicle on the day, and determine the excessive and small value points of the mileage of the vehicle on the day; finally, remove the The travel record of the vehicle's mileage that is too large or too small on the day;

2) Abnormal handling of early warning status: Count the warning duration of each warning bit, find out the duration of a single warning of the warning bit, and delete the warning status with obvious errors;

3) Exception handling of upload time: Calculate the time difference between adjacent upload points, and eliminate the points where the difference between adjacent "upload time" remains unchanged or is less than zero;

Step 2.3, operation of redundant data: traverse all the records of the vehicle's trip on the day, delete the repeated upload and the records with smaller data scale of the day, specifically: compare the upload records, delete the repeated upload records, and repeat the execution until Traverse all data files; for records with small data scale, count the travel time of the vehicle on the day, and delete the travel records less than 15min;

Step 3, extraction of key variables of vehicle warning information

According to the historical travel data of the vehicle, two key variables in the vehicle driving warning information are extracted: the warning frequency per unit mileage of the vehicle and the warning frequency per unit driving time of the vehicle; first, the total warning of the specific warning position of each vehicle within the statistical period T days frequency, T is a positive integer; secondly, the total mileage of each vehicle in the statistical period T days; thirdly, the total driving time of each vehicle in the statistical period T days; then, calculate the warning frequency and unit of each vehicle in the unit mileage The frequency of early warning of travel time; the vehicle ID is used as the unique identification code, and the travel record information of the same ID vehicle in different time periods is statistically summarized; the specific steps are:

Step 3.1, the frequency of early warning of each vehicle at a specific early warning position in the statistical period T days

Taking the travel warning record of the vehicle as the object, firstly, the vehicle ID is used as the unique identification code to count the warning frequency of each warning bit of each vehicle in one day, and then the selected specific warning bits are accumulated to obtain each warning bit. The total warning frequency of a vehicle at a specific early warning position within a day, and finally the daily warning frequency of the same ID vehicle at a specific early warning position in the period T day is accumulated to obtain the total early warning frequency of each vehicle in the specific early warning position within the period T day;

Step 3.2, the total mileage of each vehicle in the statistical period T days

Driving mileage is to record the mileage change of the vehicle dashboard, reflecting the driving distance of the vehicle; using the vehicle ID number as the unique identification code, the mileage of vehicles with the same ID is accumulated, and finally the total mileage of each vehicle within the period T is obtained;

Step 3.3, the total travel time of each vehicle in the statistical period T days

The vehicle travel time does not include the time lost by the vehicle due to waiting or delay. The principle of vehicle travel time extraction is: first calculate the total travel time of the vehicle, and then calculate the parking time. The time difference between the two is the total travel time of the vehicle; then, Taking the vehicle ID as the unique identification code, the running time of vehicles with the same ID within the period T is accumulated to obtain the total running time of each vehicle within the period T;

Step 3.4, based on the warning frequency of each vehicle within the specific warning position within the period T obtained in step 3.1 and the total mileage of each vehicle within the period T obtained in step 3.2, divide the two to obtain the warning frequency per vehicle mileage ; Based on the total travel time of each vehicle in the period T in the time period T obtained by step 3.1 and the total travel time of each vehicle in the period T of the specific early warning position and step 3.3, the two are divided to obtain the early warning frequency of the vehicle unit travel time;

Step 4. Clustering of vehicle safety risks

The warning frequency of the vehicle's unit driving mileage and the warning frequency of the vehicle's unit driving time are used as clustering objects to classify the risk level; the two-dimensional data is clustered based on the AGNES hierarchical clustering algorithm;

Specifically:

1) Determine the input sample set O={(WFM ₁ , WFT ₁ ), (WFM ₂ , WFT ₂ ),..., (WFM _n , WFT _n )} and the number of clusters Z value, where WFM _i and WFT _i respectively represent the warning frequency per unit driving mileage and the warning frequency per unit driving time of vehicle i, where i=1,2,...,n, n is the number of samples, that is, the total number of vehicles or drivers;

2) Using the bottom-up clustering strategy, take each object O _i in the sample set O as a sample cluster Φ _i , calculate the distance between any two sample clusters Φ _c and Φ _h and compare the distances, where c ≠h, find the two nearest sample clusters Φ _h and Φ _c as a new set of sample clusters, Φ _v =Φ _h ∪Φ _c , where c, v, h are positive integers, and the values are all less than or equal to n;

3) Cluster distance metric function

The size of the proximity between the two clusters is jointly determined by the two clusters, and the average distance is used to calculate the aggregation degree between any two sample clusters, and the aggregation degree is used to represent the similarity of the two sample clusters;

G=(WFM _g , WFT _g ), Q=(WFM _q , WFT _q ) (6)

In the formula: Φ _h , Φ _c respectively represent a sample cluster, |Φ _h | and |Φ _c | represent the number of elements in sample clusters Φ _h , Φ _c respectively, G and Q represent sample clusters Φ _h , Φ respectively For a sample in _c , WFM _g represents the warning frequency per unit mileage of vehicle g, WFT _g represents the warning frequency per unit driving time of vehicle g, WFM _q represents the warning frequency per unit driving mileage of vehicle q, and WFT _q represents the driving frequency of vehicle q per unit The warning frequency of time, dist(G, Q) represents the Euclidean distance between the two samples of G and Q;

4) Compare the average distance between each sample cluster calculated in 3), merge two clusters with the closest distance based on the cluster merging principle, continuously update and merge to form a new cluster, and perform cluster division again;

5) Termination condition judgment

According to the set Z value of the number of clusters, if the number of clusters is equal to the Z value, no more clustering is required, the clustering is terminated, and the Z risk level is obtained;

Step 5, Discriminant Analysis

Step 5.1, Identify categorical and discriminant variables

Divide the risk level according to the risk level of the vehicle, and divide it according to the clustering result of step 4, which is divided into 1 level, 2 level, ... As the discriminant variable of the discriminant analysis, the risk level of the vehicle is taken as the category variable of the discriminant analysis;

Step 5.2, the establishment of the discriminant function, determine the quantitative relationship between the categorical variable and the discriminant variable according to the sample data, and use the Fisher discriminant criterion to establish the Fisher discriminant function;

Step 6, Risk Identification

According to the established Fisher discriminant function, classify and discriminate new samples

1) Calculate the center of the category to which the sample point belongs in the Y space; (2) For the new sample, calculate its Fisher discriminant function value Y ₀ , construct the distance function W(Y ₀ ) between Y ₀ and the center of each category, and calculate the difference between Y ₀ and the center of each category. The distance between the centers of each category; (3) Use the distance discrimination method to determine the category to which it belongs.

On the basis of the above scheme, the too small value of mileage in step 2.2 is determined by the cumulative distribution of daily mileage, which is less than 2% of the overall sample; the method of judging the too large value of mileage is determined according to the 3σ principle of basic statistical knowledge .

On the basis of the above scheme, in step 3.2, the formula for calculating the total mileage of each vehicle within the period T is as follows:

Wherein, M _i is the total mileage of vehicle i in the period T; m _ij is the mileage of vehicle i on the jth day, i=1, 2...n; j=1,2...T.

On the basis of the above scheme, in step 3.3, the total travel time of the vehicle: is the duration of the on-board instrument from the start to the end of the recording, and the time difference between the start time of the vehicle's recording start point and the end time of the stop recording is the total time of the vehicle travel; parking Time: The position where the mileage is continuous is the position where the vehicle is stationary. Find the continuous point of the mileage record, count the duration of the continuous point, and get the stop time of the vehicle.

On the basis of the above scheme, in step 5.2, the Fisher discriminant function is:

y=b ₁ x ₁ +b ₂ x ₂ +…+b _p x _p (8)

In the formula, b _α is the discriminant coefficient, α = 1, 2, ..., p, y is a certain dimension of the sample in the low-dimensional Y space.

On the basis of the above scheme, in step 6, the distance function W(Y ₀ ) is as follows:

分别表示Y空间中第e、f类样本的中心点，

表示Y空间中第e类和f类所有样本的中心点，

∑^-1表示第e、f类协方差矩阵的逆矩阵；其中e＝1,2,…,Z，f＝1,2,…,Z，e≠f；In the formula,

represent the center points of the e-th and f-class samples in the Y space, respectively,

represents the center point of all samples of class e and class f in Y space,

∑ ^-1 represents the inverse matrix of the covariance matrix of types e and f; where e=1,2,…,Z, f=1,2,…,Z, e≠f;

When W(Y ₀ )>0, the new sample point belongs to the e-th class.

Description of drawings

The present invention has the following accompanying drawings:

Fig. 1 is the flow chart of the present invention;

Figure 2 is a flowchart of data outlier screening;

Fig. 3 is the clustering result diagram of vehicle risk;

Figure 4 shows the spatial distribution of sample points in the discriminant function.

Detailed ways

The present invention will be further described in detail below in conjunction with accompanying drawings 1-4.

Step 1, data acquisition.

Obtain vehicle ID, update upload time, vehicle safety warning status, and mileage record through the vehicle-connected vehicle terminal of Beidou positioning system.

Step 2, data preprocessing

Step 2.1 traverses all data files, deletes travel records that do not meet the requirements, deletes text files lacking attribute values, ensures the integrity of attributes, and deletes records uploaded repeatedly by the vehicle terminal;

Step 2.2 Traverse all data files, calculate the mileage of the vehicle on the day, the cumulative warning frequency of each warning position on the day, and the time difference between the start and the end of the day. According to the cumulative distribution map of the mileage of all vehicles, determine the excessive and small value points of the vehicle mileage; remove the records that the mileage of the day is too large or too small (the points with too small values are less than 2 of the overall sample based on the frequency of the mileage on the day). % to determine, and the point of excessively large value is defined according to the basic knowledge of statistics and using the 3σ principle), in addition, the travel records whose time difference between the start and the end of the day is less than 0 are excluded;

Step 3. Extraction of key variables of truck travel

Step 3.1, count the warning frequency of each vehicle in the period T. Taking the daily travel log of the vehicle as the unit, the statistics of the warning frequency are calculated separately according to the warning type (speed warning, fatigue driving warning, collision warning, etc.) Accumulate the total warning frequency at the selected warning position on the current day. Assuming that the Internet of Vehicles data records S pieces of early warning information of length L for each vehicle in one day, where L is the total number of early warnings, that is, the total number of types of early warnings of on-board equipment, the early warning information can be expressed as:

s＝1,2,…,S,l＝1,2,…,L，in,

s=1,2,…,S,l=1,2,…,L,

即为车辆当天出行记录下第l位的预警频次。再对选定的几个特定预警位的预警频次进行累加，得到每辆车当天在选定预警位的预警总频次。假如选定的是前三位预警位，则车辆一天内在选定预警位的预警总频次为

最后，再对相同ID车辆时段T天内每天(在选定预警位的)预警频次进行累加，得到T天内该车辆(在选定预警位的)预警总频次。If the warning frequency a _l of the lth warning position of a vehicle in one day is counted, there are:

That is, the warning frequency of the 1st place in the travel record of the vehicle on the day. Then, the warning frequencies of several selected specific warning positions are accumulated to obtain the total warning frequency of each vehicle at the selected early warning positions on that day. If the first three warning positions are selected, the total warning frequency of the selected warning positions in one day is:

Finally, the warning frequency of each day (at the selected warning position) for the same ID vehicle within T days is accumulated to obtain the total warning frequency of the vehicle (at the selected warning position) within T days.

Step 3.2, the total mileage of each vehicle in the statistical period T is calculated. The change of driving mileage is to record the mileage change of the vehicle dashboard, reflecting the driving distance of the vehicle. Specifically, the vehicle ID number is used as the unique identification code to accumulate the mileage of the same vehicle ID on the day, and finally the total mileage of each vehicle in the period T is obtained; the mileage of the day can be calculated from the "mileage" attribute field of the vehicle. Obtained, that is, based on the difference between the mileage at the starting point of the mileage record and the end point of the mileage record on the current day.

Step 3.3, the total travel time of each vehicle in the T day period is counted. Vehicle travel time does not include time lost by the vehicle due to waiting or delays. The principle of vehicle travel time extraction is to first calculate the total travel time of the vehicle, and then calculate the parking time, and the time difference between the two is the travel time of the vehicle. 1) The total travel time of the vehicle: the total travel time of the vehicle is the duration from the start of the recording to the end of the on-board instrument, and the time difference between the start of the vehicle’s recording start and the end of the stop recording is the total travel time of the vehicle; 2) Parking time: The position where the mileage is continuous is the position where the vehicle is stationary. Find the continuous point of the mileage record, count the duration of the continuous point, and get the stop time of the vehicle; 3) The vehicle’s travel time on the day: subtract the parking time from the total travel time obtained 4) The total driving time of the vehicle within T days: similar to the total vehicle mileage statistics, with the vehicle ID as the main key, the total driving time of the vehicle is obtained by accumulating the driving time of the same ID vehicle.

Step 3.4, the warning frequency per unit driving mileage and the warning frequency per unit driving time, calculate the warning frequency per unit driving mileage (WFM) and Warning frequency per unit driving time (WFT);

WFM=TW/TM (3)

WFT=TW/TT (4)

Among them, WFM (Warning Frequency per Mileage)—warning frequency per unit of driving mileage, WFT (Warning Frequency per Travel Time)—warning frequency per unit driving time, TW (Total Warning Frequency)—total number of warnings, TM (Total Mileage)—total Travel mileage, TT (Total Travel Time)—total travel time.

Step 4: Clustering of vehicle safety risks, based on the AGNES (AGglomerative NESting) hierarchical clustering algorithm, the sample vehicles are classified into early warning levels. Specifically: 1) Determine the input sample set O={(WFM ₁ , WFT ₁ ), (WFM ₂ , WFT ₂ ),...,(WFM _n , WFT _n )} and the number of clusters Z, where (WFM _i , WFT _i ) (i=1, 2, . . . , n) respectively represent the warning frequency under the unit driving mileage of vehicle i and the warning frequency under the unit driving time. 2) Using the bottom-up clustering strategy, take each object O _i in the sample set O as a sample cluster Φ _i , calculate the distance (c≠h) between any two sample clusters Φ _c and Φ _h , and compare each Distance, find the two sample clusters Φ _h and Φ _c with the closest distance as a new set of sample clusters, Φ _v =Φ _h ∪Φ _c . 3) Clustering cluster distance metric function. The size of the proximity between the two clusters is jointly determined by the two clusters. This time, the average distance (also known as the average-linkage method) is used to calculate the aggregation degree between any two sample clusters, which is used to represent the two sample clusters. similarity measure.

G=(WFM _g , WFT _g ), Q=(WFM _q , WFT _q ) (6)

In the formula: Φ _h , Φ _c respectively represent a sample cluster, |Φ _h | and |Φ _c | represent the number of elements in sample clusters Φ _h , Φ _c respectively, G and Q represent sample clusters Φ _h , Φ respectively For a sample in _c , dist(G, Q) represents the Euclidean distance between the two samples of G and Q;

4) Compare the average distance between each sample cluster calculated in 3), merge two clusters with the closest distance based on the cluster merging principle, continuously update and merge to form a new cluster, and divide the cluster again. For example, the distance between cluster Φ ₁ and cluster Φ ₂ is the smallest distance between different clusters, then Φ ₁ and Φ ₂ will be merged to form a new cluster. 5) Termination condition judgment. According to the set Z value of the number of clusters, if the number of clusters is equal to the Z value, no more clustering is required, the clustering is terminated, and the Z risk level is obtained.

Step 5, Discriminant Analysis

Step 5.1, mean test and covariance matrix homogeneity test, in order to ensure that the effect of discriminant analysis is ideal, the means of each discriminant variable under the population of multiple categories should be significantly different, otherwise the probability of giving wrong discriminant results will be high. , usually, the overall mean test should be carried out first, that is, to determine whether the difference between groups of discriminant variables under each category population is significant.

The basic idea of Fisher discriminant analysis is to project first and then discriminate. Projection in discriminant analysis is the key to discriminant analysis. According to the principles of maximizing the dispersion between classes and minimizing the dispersion within a class, high-dimensional data points are projected to low-dimensional data points. , to achieve the maximum separation between classes of samples. Project the p-dimensional X space sample points into r (r<=p)-dimensional Y space. The functional form of Fisher's discriminant function is as follows:

y=b ₁ x ₁ +b ₂ x ₂ +…+b _p x _p (8)

Among them, the coefficient b _α is called the discriminant coefficient, which is the influence of each input variable on the discriminant function. y is some dimension of the sample in the low-dimensional Y space.

By transforming the original data coordinate system, the sample points in the high-dimensional space are transformed into the low-dimensional space. The overall sample points are separated as much as possible through coordinate transformation. When discriminating, first calculate the center of the category to which the sample point belongs in the Y space. For a new sample, calculate its Fisher discriminant function value Y ₀ , and Y ₀ in the Y space and the center of each category The distance of , using the distance discriminant method (Mahilan distance), to determine the category to which it belongs, and to construct the distance function W(Y ₀ ) between Y ₀ and the center of each category,

分别表示Y空间中第e、f类样本的中心点，

表示Y空间中第e类和f类所有样本的中心点，

∑^-1表示第e、f类协方差矩阵的逆矩阵。in,

represent the center points of the e-th and f-class samples in the Y space, respectively,

represents the center point of all samples of class e and class f in Y space,

∑ ^-1 represents the inverse matrix of the covariance matrix of the e and f types.

When W(Y ₀ )>0, the new sample point belongs to the e-th class.

Step 5.2, determine the discriminant factor. According to the clustering results in step 4, the risk levels of vehicles are divided into 1, 2, ... Z, and two key variables of early warning information (warning frequency per vehicle mileage and warning frequency per driving time) are selected as The discriminant factor or discriminant variable takes the risk level of the vehicle as the categorical variable of the discriminant analysis.

Step 5.3, the establishment of the discriminant function. The risk level determined in step 4 is used as the categorical variable of the discriminant analysis, the warning frequency per unit mileage of the vehicle and the warning frequency per unit driving time are used as the discriminant variables, and the relationship between the categorical variable and the discriminant variable is determined according to the existing sample data. Quantitative relationship, using Fisher discriminant function to establish discriminant criteria;

Step 5.3 The above Fisher discriminant analysis can be operated through the discriminant analysis in SPSS software to obtain the discriminant function.

Step 5.4, model result test, judge the interpretation degree of the model by the frequency percentage of confusion matrix or the distribution and position of each sample point in Fisher discriminant function space.

Step 6, risk identification. The judgment and prediction of unknown categories of new data are realized by the discriminant function. For the sample points of the new data, based on the discriminant function, the risk identification of the new sample (vehicle/driver) is realized.

(1) Calculate the center of the category to which the sample point belongs in the Y space; (2) For the new sample, calculate its Fisher discriminant function value and the distance from the center of each category in the Y space; (3) Use distance discrimination to determine the category to which it belongs .

The present invention can provide the warning frequency per unit driving mileage and the warning frequency per unit driving time for a certain vehicle/driver, and then can provide the judgment function of discriminant analysis based on historical data, so as to realize the identification or prediction of its risk.

1 Case data introduction

The data used in the present invention comes from the Internet of Vehicles data provided by a company from September 2017 to October 2017, including dynamic data items: early warning signs, mileage, update upload time, and static data items: vehicle ID (terminal number )Wait. Firstly, data preprocessing and filtering are performed on the original data. Finally, 11,139 original records were filtered and processed to get 10,039 valid records. The data of vehicles with the same ID of different days are merged, and one ID is used as a sample to obtain 862 vehicle samples, which can evaluate the risk of vehicle/driver realization. The final data style is shown in Table 1 below:

The interpretation of each variable is defined as follows:

The first column ID represents the code of the vehicle, which is the unique code to identify the vehicle. Different vehicles have different vehicle IDs, and one vehicle ID corresponds to the same vehicle.

The second column WFM represents the warning frequency of the vehicle in the unit driving mileage (every 10km).

The third column WFT represents the warning frequency of the vehicle per unit travel time (hourly).

Table 1 Basic format (part) of organizing sample data

2 Cluster analysis based on case data

The data is processed into the format shown in Table 1, and clustering is performed based on the warning frequency per unit mileage in the second column of the case data and the warning frequency per unit driving time in the third column. First, import our data as shown in Table 1, enter The preset Z value (Z=3) is combined with formula (5) and formula (6) to calculate the distance between sample points, and the clustering result can be obtained by clustering according to the average distance between each sample point. Part of the results obtained by cluster analysis are shown in Table 2. The meanings of the first three columns of data are consistent with those in Table 1. The last column, RL (Risk Level), is the clustering result of the vehicle, indicating the risk level of the vehicle.

Table 2 Clustering results of sample data

Figure 3 presents the cluster analysis results of the sample data. As shown in the figure, the sample points can be divided into three categories. The larger the value of WFM and WFT, the higher the warning frequency of the vehicle, that is, the greater the risk level of the vehicle. Therefore, when the values of WFM and WFT are clustered into three categories, they The risks can be defined as: 1-safe, 2-general, 3-dangerous. The sample points marked with circles in the figure have small WFM and WFT values, so this type of sample has the lowest risk level and is in a safe state; the diamonds represent sample points with large WFM and WFT values, and their risk The highest level is the dangerous state; the triangles represent the sample points whose risk level is the general state.

2 Establishment of discriminant function based on clustering results

The risk level of the vehicle determined in the previous step is used as the category variable of the discriminant analysis, and the warning frequency per unit mileage and the warning frequency per unit driving time of the vehicle are used as the discriminant variables. The quantitative relationship of , using Fisher's criterion to establish a discriminant function.

(1) Determine the discriminant factor

Two key variables of early warning information are selected as discriminant factors, namely the frequency of warning per unit mileage and the frequency of warning per unit driving time as discriminant variables.

(2) Establish a discriminant function

A discriminant function is established to realize the identification and judgment of risky vehicles, and a relevant discriminant function is established according to the driving risk index and risk warning level of each vehicle. Taking the risk level of the vehicle as the category variable of the discriminant analysis, the warning frequency per unit mileage and the warning frequency per unit driving time of the vehicle as the discriminant variables, the Fisher discriminant function is established to analyze the risk intensity of each vehicle, thereby realizing the risk assessment of the vehicle. . In this paper, a risk discriminant function is established based on Fisher's discriminant method to realize risk level determination.

1. The established discriminant function

Table 3 Discriminant function coefficients

Get the discriminant function:

In the formula: x ₁ —WFM (warning frequency per unit driving mileage); x ₂ —WFT (warning frequency per unit driving time).

2. Calculate the center position of the category

The class center positions of the three classes of sample points are calculated as follows:

Table 4 Functions at the center of the categories

3. Test of discrimination ability

In order to test whether the projection of the Fisher discriminant function can well separate the various samples, and to further judge which discriminant function is more important for the interpretation of the discriminant results, it is necessary to calculate the two eigenvalues, the percentage of variance explained, and the cumulative percentage of variance explained. .

Table 5 Summary of calculation results

It can be seen that the ability of the first discriminant function to explain the variance is 100%, while the ability of the second discriminant function to explain the variance is 0%, so the second discriminant function can be omitted. The final discriminant function is obtained as:

y=-1.019-2.619x ₁ +5.426x ₂ (10)

Bring the Fisher discriminant function into the new sample point, then calculate the distance from the center of each category, and use the distance discrimination to determine the category to which it belongs.

If a new sample is obtained through data acquisition (new vehicle or driver travel behavior data, or new travel behavior data of the original sample in the last month), the data of the new sample is as follows: the warning frequency per unit mileage WFM is 0.735 times/ 10km, the warning frequency WFT per unit driving time is 3.101 times/h.

Taking the new sample (0.735, 3.101) as an example, it is first brought into the discriminant function formula (10), and its value is calculated to be 13.882, that is, the value of the new sample mapped to the one-dimensional space sample set Y is 13.882. Then, since the final discriminant function is the single function form of formula (10), according to function 1 in Table 4, it can be known that the values of the center points of each category mapped into the one-dimensional space Y are -0.700, 10.893, and 27.826, respectively. Calculate the distance between the new sample point and the center point of each category in Y, and use the distance discrimination of formula (8). Since the discriminant function is a single function form, the distance can be directly calculated, which are 14.582, 2.988, and -13.946 respectively. It can be seen that after the mapping, the new sample point is closest to the center point of the second category, and it is finally determined that the risk category of the new sample belongs to the second category, that is, the risk level is general.

(4) Result analysis

The distribution and position of each sample point in the Fisher discriminant function space are shown in Figure 4. From the distribution and position of the sample points in the Fisher discriminant function space, the distribution of each category is relatively concentrated, so the discriminant effect is ideal.

Contents not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims (6)

1. The utility model provides a method of extraction and risk identification of freight train early warning information based on big dipper car networking which characterized in that includes several following steps:

step 1, obtaining original data related to vehicle early warning through a vehicle-mounted terminal of a vehicle networking provided with a Beidou positioning system, wherein the original data comprises: mileage information, safety early warning information and state data, wherein the state data comprises a vehicle ID, an ACC state and uploading time;

step 2, preprocessing the original data;

preprocessing and screening original data by using a Python programming technology; before analyzing the original data, the original data needs to be cleaned and sorted, so that the data quality is improved; the data cleaning comprises the following steps: filling missing values in the data, and identifying abnormal values and redundant data in the data; the method comprises the following specific steps of preprocessing original data by combining the characteristic of data storage of the vehicle-mounted terminal:

step 2.1, data missing value operation: the method comprises the steps of performing programming operation by using Python, introducing an os module and a numpy module in the Python, defining a required function, executing a main function, operating a text file stored in a vehicle-mounted terminal, deleting the text file lacking an attribute value, and ensuring the integrity of the attribute;

step 2.2, data abnormal value operation: the data abnormal value comprises abnormal mileage, abnormal early warning state and abnormal uploading time;

1) and (3) mileage exception handling: firstly, traversing all data files, calculating the mileage of the vehicle on the day, secondly, making an accumulated distribution map of the mileage of the vehicle on the day, and determining an overlarge value point and an undersize value point of the mileage of the vehicle on the day; finally, removing travel records with overlarge or undersize numerical values in the driving mileage of the vehicle on the day;

2) and (4) abnormal processing of the early warning state: counting the early warning duration of each early warning bit, solving the duration of single early warning of the early warning bit, and deleting the early warning state with obvious errors;

3) and (3) processing an uploading time exception: calculating the time difference between adjacent uploading points, and eliminating the recording points with the difference value of the adjacent uploading time unchanged or less than zero;

step 2.3, redundant data operation: traversing all records of the vehicle trip on the same day, and deleting the repeatedly uploaded records and the records with smaller data scale on the same day, wherein the records specifically comprise: comparing the uploaded records, deleting the repeatedly uploaded records, and repeatedly executing until all data files are traversed; counting the trip time of the vehicle on the day according to the record with smaller data scale, and deleting the trip record less than 15 min;

step 3, extracting key variables of the vehicle early warning information;

according to the historical travel data of the vehicle, two key variables in the vehicle running early warning information are extracted: the early warning frequency of the unit driving mileage of the vehicle and the early warning frequency of the unit driving time of the vehicle; firstly, counting the total early warning frequency of a specific early warning position of each vehicle in a time period of T days, wherein T is a positive integer; secondly, counting the total driving mileage of each vehicle within T days; thirdly, counting the total driving time of each vehicle within the time period T days; then, calculating the early warning frequency of each vehicle in unit driving mileage and the early warning frequency of each vehicle in unit driving time; taking a vehicle ID as a unique identification code, and counting and summarizing travel record information of the same ID vehicle in different time periods; the method comprises the following specific steps:

step 3.1, counting the early warning frequency of each vehicle at a specific early warning position within a time period T days;

taking travel early warning records of vehicles as objects, firstly taking a vehicle ID as a unique identification code, counting the early warning frequency of each early warning position of each vehicle in one day, then accumulating selected specific early warning positions to obtain the early warning total frequency of each vehicle in the specific early warning position in one day, and finally accumulating the early warning frequency of the same ID vehicle in the specific early warning position every day in a time period T to obtain the early warning total frequency of each vehicle in the specific early warning position in the time period T;

step 3.2, counting the total driving mileage of each vehicle within the time period T days;

the driving mileage is to record the mileage change of the instrument panel of the vehicle and reflect the driving distance of the vehicle; accumulating the driving mileage of the same ID vehicle by taking the vehicle ID number as a unique identification code, and finally obtaining the total driving mileage of each vehicle in a time period T days;

step 3.3, counting the total driving time of each vehicle in the period T days;

the vehicle running time does not include the time lost by the vehicle due to waiting or delay, and the principle of vehicle running time extraction is as follows: firstly, calculating the total travel time of the vehicle, and then calculating the parking time, wherein the time difference between the total travel time and the parking time is the total travel time of the vehicle; then, accumulating the running time of the same ID vehicle in the time period T by taking the vehicle ID as a unique identification code to obtain the total running time of each vehicle in the time period T;

step 3.4, dividing the early warning frequency of each vehicle at the specific early warning position in the time period T day obtained in the step 3.1 and the total driving range of each vehicle in the time period T day obtained in the step 3.2 to obtain the early warning frequency of the unit driving range of the vehicle; dividing the early warning frequency of each vehicle at the specific early warning position in the time period T day obtained in the step 3.1 and the total running time of each vehicle in the time period T day obtained in the step 3.3 to obtain the early warning frequency of the vehicle in unit running time;

step 4, clustering the vehicle safety risks;

taking the early warning frequency of the unit driving mileage of the vehicle and the early warning frequency of the unit driving time of the vehicle as clustering objects, and dividing risk grades; clustering the two-dimensional data based on an AGNES hierarchical clustering algorithm;

the method specifically comprises the following steps:

1) determining an input sample set O { (WFM)₁，WFT₁),(WFM₂，WFT₂),...,(WFM_n，WFT_n) And the number of clusters Z value, where WFM_iAnd WFT_iRespectively representing the early warning frequency of i unit driving mileage and the early warning frequency of unit driving time of the vehicle, wherein i is 1,2, …, n is the number of samples, namely the total number of vehicles or drivers;

2) adopting a bottom-up clustering strategy to sample each object O in the set O_iAs a cluster of samples phi_iCalculating any two sample clusters phi_cAnd phi_hComparing the distances, wherein c is not equal to h, and searching two sample clusters phi with the shortest distance_h、Φ_cAs a new set of sample clusters, phi_v＝Φ_h∪Φ_cWherein c, v and h are positive integers, and the values are all less than or equal to n;

3) clustering cluster distance metric function;

the proximity degree between the two clusters is determined by the two clusters together, the average distance is adopted to calculate the aggregation degree between any two sample clusters, and the aggregation degree is used for representing the similarity degree of the two sample clusters;

G＝(WFM_g,WFT_g),Q＝(WFM_q,WFT_q) (6)

in the formula: phi_h,Φ_cRespectively represent a certain sample cluster, | phi_h|、|Φ_c| represents a cluster of samples Φ respectively_h,Φ_cThe number of middle elements, G and Q respectively represent a sample cluster phi_h,Φ_cA certain sample in (1), WFM_gWFT (weighted average) representing the warning frequency of g unit mileage of a vehicle_gIndicating the warning frequency of the vehicle in g units of travel time, WFM_qWFT (weighted average) representing the warning frequency of the vehicle for q units of mileage_qThe early warning frequency of the vehicle Q unit running time is represented, dist (G, Q) represents the Euclidean distance between two samples G and Q;

4) comparing the average distance between each sample cluster obtained by calculation in the step 3), merging two clusters with the closest distance based on a clustering merging principle, continuously updating and merging to form a new cluster, and performing cluster division again;

5) judging a termination condition;

according to the set value of the clustering number Z, if the clustering number is equal to the value Z, clustering is not needed, and clustering is terminated to obtain a Z-type risk level;

step 5, judging and analyzing;

step 5.1, determining a category variable and a discrimination variable;

carrying out risk grade division according to the risk degree of the vehicle, dividing the risk grade into 1 grade, 2 grade and … Z grade according to the clustering result in the step 4, taking the early warning frequency of the unit driving mileage of the vehicle and the early warning frequency of the unit driving time of the vehicle as discrimination variables of the discrimination analysis, and taking the risk grade of the vehicle as a category variable of the discrimination analysis;

step 5.2, establishing a discrimination function, namely determining the quantitative relation between the category variable and the discrimination variable according to the sample data, and establishing the Fisher discrimination function by using a Fisher discrimination criterion;

step 6, risk identification;

classifying and distinguishing the new samples according to the established Fisher distinguishing function;

(1) calculating the center of the category to which the sample point belongs in the Y space; (2) for the new sample, calculating Fisher discriminant function value Y₀Construction of Y₀Distance function W (Y) from the center of each class₀) And calculate Y₀Distance from the center of each category; (3) the category to which the image belongs is determined by a distance discrimination method.

2. The method for extracting early warning information and identifying risks of trucks based on the Beidou Internet of vehicles as claimed in claim 1, wherein the over-small points of mileage in step 2.2 are determined by cumulative daily mileage distribution, and are determined according to less than 2% of the total sample; the judgment method of the excessive value points of the mileage is determined according to the principle of 3 sigma of basic statistical knowledge.

3. The method for extracting early warning information and identifying risks of trucks based on the Beidou Internet of vehicles as claimed in claim 1, wherein in step 3.2, the calculation formula of the total driving range of each truck in the time period T is as follows:

wherein

Wherein M is_iIs the total driving range of vehicle i over a period of T days; m is_ijIs the mileage of vehicle i on day j, i ═ 1,2 … n; j is 1,2 … T.

4. The method for extracting early warning information and identifying risks of trucks based on the Beidou Internet of vehicles as claimed in claim 1, wherein in step 3.3, the total time of vehicle traveling: the time difference between the starting time of the vehicle on day recording and the ending time of the stop recording is the total time of the vehicle trip; parking time: and finding a continuous invariant point of the mileage record, and counting the duration of the continuous invariant point to obtain the stop time of the vehicle.

5. The method for extracting early warning information and identifying risks of trucks based on the Beidou Internet of vehicles as claimed in claim 1, wherein in step 5.2, the Fisher discriminant function is as follows:

y＝b₁x₁+b₂x₂+…+b_px_p(8)

in the formula, x₁,x₂,...,x_pTo discriminate variables, b_αFor the discrimination coefficient, α is 1,2, …, and p, Y is a dimension of the sample in the low-dimensional Y space.

6. The method for extracting early warning information and identifying risks of trucks based on Beidou Internet of vehicles according to claim 5, wherein in step 6, the distance function W (Y) is₀) As follows:

in the formula,

respectively representing the center points of the e-th and f-th class samples in the Y space,

representing the center points of all samples of class e and class f in Y space,

∑^-1an inverse matrix representing the e and f type covariance matrixes; wherein e is 1,2, …, Z, f is 1,2, …, Z, e ≠ f;

when W (Y)₀)>At 0, the new sample point belongs to class e.

Images