Improvement machine learning method suitable for vehicle-mounted short haul connection net safe early warning
Technical field
The present invention relates to a kind of improvement machine learning methods suitable for vehicle-mounted short haul connection net safe early warning, belong to vehicle
Carry Communication Network Technique field.
Background technology
Vehicle-mounted short haul connection (Vehicle to X:V2X) network be by radio communication, the short distances such as GPS/GIS, sensing
Car (CAN-Controller Area Network), bus or train route (Vehicle-2-RSU) from communication technology realization, workshop
(Vehicle-2-Vehicle), vehicle is outer between (vehicle-2-Infrastructure), people's vehicle (Vehicle-2-Person)
Communication.
In V2X networks, device node includes mainly vehicle-mounted node (OBU:On board Units) and trackside node
(RSU:Road-Side Units).Each vehicle-mounted node periodic broadcast includes the information such as vehicle unique mark, geographical location
Message (hereinafter referred to as heartbeat message), so, the heartbeat message that vehicle-mounted node and trackside node are broadcasted by receiving surroundings nodes can
The information such as geographical location, signal transmission characteristics with vehicle-mounted node around acquisition.
Security application is the main application of Che in V2X-vehicle communication, and four classes are broadly divided into according to the difference of scene:Straight trip turns
To, intersection, lane change.Typically have:Intersection assists (IMA:Intersection Movement Assist), preceding touch
Hit early warning (FCW:Forward Collision Warning), early warning (CLW out of control:Control Losing Warning), it is tight
Sub- brake lamp (the EEBL of urgent telegram:Electronic Emergency Brake Light), no thoroughfare early warning (DNPW:Do Not
Pass Warning), turn left auxiliary (LTA:Left Turn Assist) etc., above-mentioned security application is mainly by giving warning in advance
Mode is realized.Since road structure and environment are complicated, the relative position and speed of each vehicle are changeable, and easily by weather condition
It influences, how according to the heartbeat message of vehicle-mounted node, it is a technological difficulties to carry out reasonable, accurate safe early warning.
With artificial intelligence (AI:Artificial Intelligence) technology development, the security application of V2X is transported
With machine learning algorithm, behavior is judged by the practical driving of continuous learner driver, and according to vehicle-state, condition of road surface,
The factors such as weather conditions realize that rational safe early warning is possibly realized.
Boosting algorithms are a kind of common machine learning algorithms, can significantly improve the classification capacity of grader.
Traditional boosting algorithms are to be weighted trained Weak Classifier to sample, after whole sample trainings, obtain strong classification
Device.It, should be in conjunction with history training result, again according to dynamic training result due to the security application in vehicle-mounted short haul connection net
Constantly learnt, thus, traditional boosting algorithms cannot achieve accurate, rational safe early warning.
Invention content
In view of above-mentioned purpose, the object of the present invention is to provide a kind of changing suitable for vehicle-mounted short haul connection net safe early warning
Into machine learning method, the weight of each Weak Classifier is updated in conjunction with history training result data and dynamic training result data, with
This determines that strong classifier, each vehicle-mounted node realize safe early warning using the newer strong classifier of dynamic, it is practical can to obtain satisfaction
Reasonable, the accurate safe prediction effect of condition of road surface.
To achieve the above object, the present invention uses following technical scheme:
A kind of improvement machine learning method suitable for vehicle-mounted short haul connection net safe early warning includes the following steps:
S1:Trackside node handles the heartbeat message of each vehicle-mounted node of acquisition, generates sample data;
S2:Based on multiple Weak Classifiers of the sample data training with different classifications pattern;
S3:The weight of each Weak Classifier is adjusted according to the dynamic training result data of each Weak Classifier;
S4:According to weight shared by each Weak Classifier, strong classifier is determined;
S5:The strong classifier that trackside node is determined to vehicle-mounted node periodic broadcast.
Further,
The sample data includes state sample data eTWith result sample data eR, the Weak Classifier is according to state sample
Notebook data and result sample data obtain training result data.
The step S3 includes:
1) according to formula (1) to Weak Classifier hiDynamic training result data evaluated;
SW=L0(eT, eR) (1)
Wherein, the quantity of Weak Classifier is N, i ∈ { 1,2 ..., N }, SWIt is Weak Classifier hiWrongheaded state sample
The quantity of data;
2) Weak Classifier h is calculatediMisjudgment ratio:
εi=SW/ST (2)
Wherein, STIt is the total quantity of state sample data;
3) Weak Classifier h is calculatediWeight:
4) Weak Classifier h is calculatediComprehensive weight:
βi=(λ αi+(1-λ)α′i) (4)
Wherein, λ is the allocation proportion of dynamic training result data and history training result data, 0≤λ≤1.
In the step S4, determine that strong classifier is:
The state sample data eT={ car speed, travel direction, two following distances, current time, weather conditions, vehicle
Type }, the result sample data eR={ changes in vehicle speed, two vehicle distances }.
The vehicle-mounted node receives the heartbeat message of the vehicle-mounted node of surrounding, the safe early warning algorithm of the vehicle-mounted node configuration
Module determines the safe condition of the vehicle-mounted node of surrounding according to the strong classifier.
The safe early warning state that the safe early warning algorithm is provided according to the strong classifier is dangerous, carries out sound report
It is alert, and pass through the position of screen display hazardous vehicles and hazard types.
It is an advantage of the invention that:
The present invention the improvement machine learning method suitable for vehicle-mounted short haul connection net safe early warning, trackside node according to
The heartbeat message of each vehicle-mounted node of dynamic acquisition, training Weak Classifier, according to history training result data and dynamic training knot
Weight shared by each Weak Classifier of fruit data update, determines adaptable strong classifier with this, and by the strong classification of acquisition
For device periodically to each vehicle-mounted node broadcasts, each vehicle-mounted node realizes safe early warning using the newer strong classifier of dynamic, can
Obtain reasonable, the accurate safe prediction effect for meeting real road situation.
Description of the drawings
Fig. 1 is the method flow schematic diagram of the present invention.
Fig. 2 is the application scenarios of a specific embodiment of the invention.
Specific implementation mode
Further detailed description is done to the present invention below in conjunction with drawings and examples.
As shown in Figure 1, the improvement machine learning side disclosed by the invention suitable for vehicle-mounted short haul connection net safe early warning
Method includes the following steps:
S1:Trackside node carries out processing to the heartbeat message of acquisition and generates sample data;
Vehicle-mounted node, trackside node include periodically vehicle unique mark, vehicle location, vehicle speed to periodic knot broadcast
The heartbeat message of degree, transmits information.According to the needs of early warning mechanism and learning algorithm, specific information type can be increased.
Trackside node configures learning algorithm L0, sample signal type e={ e1, e2..., enAnd sample algorithm module.Road
Side gusset is basic data with the heartbeat message of each vehicle-mounted node received, and sample algorithm module is according to sample algorithm c={ c1,
c2..., cn, basic data is handled, generation meets sample signal type e={ e1, e2..., enRequire study number
According to.
Sample signal type can be divided according to practical application scene.It is by sample signal in the embodiment of the present invention
Type is divided into two classes, and one kind is state sample data eT={ car speed, travel direction, two following distances, current time, day
Vaporous condition, type of vehicle }, one kind is result sample data eR={ changes in vehicle speed, two vehicle distances }.Wherein, weather conditions
Data can be determined according to the weather conditions information that Weather prediction system is sent to trackside node, and using weather condition data as
One in basic data.Sample algorithm module is handled every basic data according to sample algorithm, mark off need into
The state sample data and result sample data of row study.
S2:Weak Classifier is trained based on sample data;
By each Weak Classifier according to learning algorithm L0Determine specific mode of learning.Weak Classifier is according to state sample data
Training result data are obtained with result sample data, training result data include that history training result data are instructed with dynamic in real time
Practice result data.
In the present embodiment, a kind of mode of learning is that intersection is learnt according to two vehicle distances and relative position
Whether danger is classified as, and learning outcome is yes/no.Specifically, it is with m-th of Weak Classifier hmTo state sample data eTInto
Row judges whether current vehicle is dangerous in intersection, with result sample data eRTo Weak Classifier hmJudging result carry out
Verification, if actual vehicle closes on intersection, two vehicles distance is less than preset safe distance, and driver takes braking
Measure makes car speed change significantly, then Weak Classifier hmJudging result be correct, be otherwise mistake.
The training of Weak Classifier and quantity can be configured according to system requirements, and classification mode can be according to difference
Factor combination be configured, cover it is all it can happen that, can be evaluated and tested according to different affecting factors, can also
It is evaluated and tested according to the combination of the different values of all factors.
Since the factors such as the driving habit of driver, different vehicle type, different weather situation can influence Weak Classifier
Evaluation result, it is therefore desirable to adjust the power of each Weak Classifier immediately according to the dynamic training result data of each Weak Classifier
Weight.
S3:The weight that each Weak Classifier is adjusted according to the dynamic training result data of each Weak Classifier, determines strong classifier;
S31:Weight shared by each Weak Classifier is adjusted according to the dynamic training result data of each Weak Classifier;
It specifically includes:
1) according to algorithm shown in formula (1) to Weak Classifier hiDynamic training result data evaluated;
SW=L0(eT, eR) (1)
Wherein, the quantity of Weak Classifier is N, and i ∈ { 1,2 ..., N }, SWIt is Weak Classifier hiWrongheaded state sample
Notebook data eTQuantity.
2) Weak Classifier h is calculatediMisjudgment ratio:
εi=SW/ST (2)
Wherein, STIt is state sample data eTTotal quantity.
3) it is based on dynamic training result data, calculates Weak Classifier hiWeight:
4) Weak Classifier h is calculatediComprehensive weight:
βi=(λ αi+(1-λ)α′i) (4)
Wherein, 0≤λ≤1, λ are the allocation proportion of dynamic training result data and history training result data, for balancing
The ratio of dynamic training result and history training result.
S32:Determine strong classifier.
According to weight shared by determining each Weak Classifier, determine that strong classifier is:
S4:Trackside node periodic broadcast strong classifier, provides safety precaution function.
Trackside node trains Weak Classifier according to the heartbeat message of dynamic acquisition, according to history training result data and moves
State training result data determine weight shared by each Weak Classifier, determine strong classifier, and periodically by the strong classifier H of acquisition
To each vehicle-mounted node broadcasts.
Vehicle-mounted node is configured with safe early warning algoritic module L1, the heartbeat message of the vehicle-mounted node of vehicle-mounted node reception surrounding,
Safe early warning algoritic module L1The safe condition of the vehicle-mounted node of surrounding is determined according to strong classifier H:If safe early warning algorithm L1Root
It is dangerous according to the strong classifier H safe early warning states provided, then carries out audible alarm, and pass through the position of screen display hazardous vehicles
It sets and hazard types, reminds driver to take appropriate measures in advance and avoids danger.
As shown in Fig. 2, in the embodiment of the present invention, the safe early warning algoritic module L that is configured1And strong classifier H can be used for
Evaluate and test the safe condition of the vehicle at intersection.However, the invention is not limited thereto, it can be according to real road situation, weather feelings
The multinomial factor such as condition, type of vehicle, vehicle condition, principle, is determined suitable for practical application request according to the method for the present invention
Algorithm and strong classifier, to provide reasonable, accurate safety precaution function for vehicle-mounted node.
The above is presently preferred embodiments of the present invention and its technical principle used, for those skilled in the art
For, without departing from the spirit and scope of the present invention, any equivalent change based on the basis of technical solution of the present invention
Change, simple replacement etc. is obvious changes, all fall within the protection scope of the present invention.