CN110168520A

CN110168520A - A kind of intelligence road traffic method for detecting abnormality

Info

Publication number: CN110168520A
Application number: CN201780050719.8A
Authority: CN
Inventors: 杜豫川; 邓富文; 常光照; 王勤; 葛以衡; 严军
Original assignee: Shanghai Intelligent Transportation Co ltd; Tongji University
Current assignee: Shanghai Intelligent Transportation Co ltd; Tongji University
Priority date: 2016-12-30
Filing date: 2017-12-30
Publication date: 2019-08-23
Also published as: GB2587588A; GB2582532A; CN109716414A; GB201909409D0; WO2018122585A1; GB2572717B; GB2572717A; GB2569924A; CN109844832B; WO2018122806A1; GB201905907D0; GB2582532B; WO2018122805A1; CN109923595B; WO2018122802A1; CN109923595A; CN109791729B; CN109844832A; GB202100341D0; GB2588556B

Abstract

A kind of intelligence road traffic method for detecting abnormality, includes the following steps: to establish space-time sub-district；The pretreatment of historical trajectory data；Historical trajectory data analysis and RNN model training；Abnormality detection；Abnormal severe quantization characterization；System performance evaluation.Using the vehicle-mounted GNSS positioning device of the Floating Car travelled on road, the trace information of vehicle is recorded, it, being capable of intellectualized detection urban highway traffic anomalous event by the analysis mining to these trace informations.

Description

A kind of intelligence road traffic method for detecting abnormality

Technical field

The invention belongs to Vehicle Detection technical fields.Particularly, the present invention relates to a kind of urban highway traffic exception real-time detection methods.By the vehicle-mounted GNSS positioning device of Floating Car, the spatial positional information of its different moments can be obtained, by data prediction, map match and data fusion, the travel speed historical data training R N model based on specific space-time unique；According to the difference between RNN model predication value and real-time travel speed true value, urban highway traffic anomalous event can be effectively identified.Background technique

Traffic abnormal incident detection is one of important component and core function of intelligent transportation system of urban traffic control.Traffic abnormal incident mainly includes traffic accident, vehicle throws paving, the damage of lorry junk, road traffic facility or failure and other cause the special event of traffic flow disorder.Such event be easy to cause traffic congestion, road section capacity to reduce, and the normal operation of entire road traffic system is influenced when serious.It is detected by traffic abnormal incident, traffic administration person can be made to understand traffic abnormity information in time, and take induction appropriate and control measure, reduce the adverse effect of traffic abnormal incident.

Traffic abnormal incident detection can be divided into manual type and automated manner.Manual type includes that cruiser, emergency call are reported with video monitoring etc., and due to consumption manpower and material resources and real-time is poor, is unable to satisfy the needs of traffic administration.Automated manner realizes that basic principle is to identify traffic abnormal incident by detecting the variation of different location road traffic flow by automatic incident detection (AID, Automated Incidence Detection) algorithm.Currently used AID algorithm includes pattern-recognition class algorithm (such as Califorma algorithm, Mo Nika algorithm), statistical forecast class algorithm (such as exponential smoothing, Kalman filtering algorithm), traffic flow model algorithm (such as McMaster algorithm) and intelligent recognition algorithm (such as artificial neural network, fuzzy logic algorithm).

But there is the disadvantages of requirement to facility is high, computation complexity is high, can not do further judgement to the situation of unusual condition in current detection method.The present invention establishes historical traffic slip condition database and real-time traffic states database, by the traffic flow character difference of both analyses reflection, identifies traffic abnormal incident using taxi, the track data of the vehicle-mounted GNSS positioning device passback of bus.This method have the characteristics that real-time it is good, can parallel processing, discrimination it is high and checkout facility is required it is low, suitable for there is the detection of urban highway traffic anomalous event under the data environment of real-time Floating Car location data.

Currently, monitoring for traffic abnormal incident, there is following representative art:

One U.S. Patent application, US 20160148512 disclose the theory of constitution and implementation method of a kind of traffic abnormal incident detection and reporting system.The system is made of sensor, communication module, mobile processing module and user interactive module.Sensor is used to acquire the related data of vehicle-surroundings；Communication module is used to send this vehicle data and receive the data of nearby vehicle；It moves the data that processing module is used to handle and analyze associated vehicle in a certain region and generates traffic incident report；User interactive module can provide traffic incident report as user.The program is a kind of traffic abnormal incident detection technique based on Che Che and bus or train route communication network, can differentiate anomalous event using the various information of sensor acquisition.However, enforcement difficulty is larger since sensor, communication unit need separately installed debugging；Mobile processing unit processing capacity is limited；Mobile and fixed message receiving end is needed simultaneously, and for system itself there are probability of malfunction, reliability is bad.

One Chinese patent application, 104809878 A of CN disclose a kind of method using bus GPS data detection urban highway traffic abnormality.The program obtains section delay time at stop index according to GPS historical data, obtains instantaneous velocity, period average speed, weighting sliding average speed and Duo Che average speed according to GPS current data, is detected using canonical variate analysis algorithm abnormal.This scheme does not need newly Increase checkout facility, implements convenience.But the characteristics of excessively simplifying for the characterization of traffic situation, traffic abnormity situation can not be analyzed and the origin cause of formation;Foundation is lacked to the division of traffic scene, fails to consider the influence that the factors such as weather change traffic situation.Summary of the invention

In order to more clearly illustrate the contents of the present invention, the technical term being involved in first is explained as follows:

Floating Car: also referred to as probe vehicles.Refer to and is mounted with vehicle carried pick device and travels bus and taxi on urban road.

GNSS: Global Navigation Satellite System (Global Navigation Satellite System).Including GPS, GLONASS, GALILEO and Beidou satellite navigation system etc..

Space-time sub-district: the section divided according to two dimensions of time and space is reflected whithin a period of time, the situation in certain spatial dimension.Some time segment, such as 0:00-0:10,0:10-0:20 ..., referred to as time sub-district were divided by one day；It is several space segments, such as 121.58 ° of E-121.59 of longitude by the implementation region division of urban highway traffic abnormality detection⁰Region between E, 31.16 ° of N-31.17 ° of N of latitude, i.e. space sub-district；The space-time segment that the intersection of any one time sub-district and any one space sub-district is formed, referred to as space-time sub-district, such as 121.58 ° of E-121.59 of longitude⁰E, 31.16 ° of N-31.17 of latitude⁰The space-time segment of region between N in 0:00-0:10.

Historical trajectory data: historical trajectory data is long time integration and the track data of storage in the database.Historical trajectory data is the data of dynamic change, needs to be updated in time, and periodically does and handle and analyze again, to guarantee the accuracy of historical traffic feature extraction.The data of each space-time sub-district can be with parallel processing to improve efficiency.It may be simply referred to as historical data in the present invention.

Real-time track data: real-time track data are the track data set in a current time nearest time section.It may be simply referred to as real time data in the present invention.

Traffic situation: the general name of the comprehensive condition of traffic circulation in certain time, certain space.

Traffic abnormity: the case where traffic flow disorder that the events such as paving, the damage of lorry junk, road traffic facility or failure cause, is thrown in traffic accident, vehicle.

Traffic abnormity seriousness: i.e. the seriousness of traffic flow disorder is the difference of traffic flow and traffic flow character after traffic abnormity generation under normal condition.

Traffic abnormity index: the measurement of traffic abnormity seriousness.Range is 0 ~ 10, and numerical value is bigger, and traffic abnormity is more serious.

Traffic environment: all extraneous summations influenced with strength of road traffic participant are acted on.Traffic activity including condition of road surface, means of transportation, geomorphological features, meteorological condition and other traffic participants.

Map match: by geographical coordinate and the associated process of city road network.

Peak hour flow: the maximum value of the hour magnitude of traffic flow in certain urban road section one day.

Response variable: the variable to be changed according to independent variable, also referred to as dependent variable.

RNN: i.e. Recognition with Recurrent Neural Network (Recurrent Neural Network), it is a kind of artificial neural network of node orientation connection cyclization, the internal state of this network can show dynamic time sequence behavior, can use its internal memory to handle the list entries of arbitrary sequence.

Elman-RNN: a kind of RNN network structure, referring to " A RNN that learns to count))

Training process: the process reduced by iterating to calculate the model error for making neural network on training dataset come optimization neural network parameter.It is a set of based on Floating Car track record system the purpose of the present invention is establishing, using history GNSS location data and real-time GNSS location data, in conjunction with the scheme of traffic environment information identification road traffic anomalous event.In order to achieve the above object, the present invention provides the following technical scheme that

Implementation of the invention on condition that: carry the Floating Car (taxi, bus etc.) of GNSS track recorder；With Mass storage, It calculates, the data center of real-time task processing capacity.

The scope of application of the invention is: the urban road (including surface road and overpass) for having above-mentioned Floating Car to pass through.

Implementation steps of the invention include:

1) it determines the space-time unique of detection and establishes space-time sub-district.

Based on actual application demand, the time range and spatial dimension for needing to carry out traffic abnormal incident detection are determined.Time range can be set as whole day, i.e. 0:00-24:00;Also it can be set as a certain specific period, such as to detect the traffic abnormity time of this period of 17:00-20:00, then it will test time range and be set as 17:00-20:00, only enumerate a particular example here, there are also a lot of other situations, no longer illustrate one by one herein.Spatial dimension can be set as some administrative region of a city, such as Beijing, Shanghai City, Huangpu District etc. according to administrative division；Some urban function region, such as certain city central business district, industrial area etc. can also be set as according to city's spatial structure.

The foundation of space-time sub-district refers to that the time range that will test is divided into several smaller time slices, and the spatial dimension that will test, i.e. the implementation region of urban highway traffic abnormality detection are divided into several smaller space segments.The foundation of space-time sub-district can use a variety of experience division methods, including equidistant space-time partitioning and non-equidistant space-time partitioning.

2) data prediction.

GNSS location data is subjected to data cleansing, data integration, data conversion, data regularization, improves the structuring degree of data.GNSS, i.e. Global Navigation Satellite System positioning system, being can at the earth's surface or any place of terrestrial space is to provide the space base radio-navigation positioning system of round-the-clock three-dimensional coordinate and speed and temporal information.It mainly includes the GPS (Global Positioning System) in the U.S., the big global navigation positioning system of the Beidou satellite navigation system four of the GLONASS (Global Navigation Satellite System) of Russia, the GALILEO of European Union and China, at the same further include the QZSS of Japan, the WASS of the area navigations positioning system such as IRNSS of India and the U.S., Japan the satellite positionings enhancing system such as MSAS.In order to establish unified data distribution standard in different navigation positioning system equipment, National Marine Electronics association has formulated unified NEMA (National Marine Electronics Association) communications protocol, with the data broadcasting of specification GNSS.Therefore, each Member Systems in GNSS, such as GPS, GLONASS etc. possess consistent data distribution format, therefore do not need to convert data format although being established and being safeguarded by country variant and mechanism respectively.

In selected spatial dimension, there are many vehicles of installation GNSS positioning device, common are taxi, bus, goods vehicle, private car etc..Based on current city traffic data application status, in practical application, usually selecting urban taxi is data source of the Floating Car as accident detection system.

It include some unreasonable information in the GNSS location information of acquisition, in order to guarantee that traffic abnormity state-detection differentiates the accuracy of result, it is necessary first to be screened to propose abnormal data, guarantee the reliability of data.These abnormal datas include: the data fallen in except detection space-time unique, the obvious spatial position jump beyond zone of reasonableness.So-called " jumping the obvious spatial position beyond zone of reasonableness ", is exemplified below it.If the anchor point that 10:30:00 one day moment Floating Car positioning device uploads is denoted as A, the anchor point that moment on the same day 10:30:30 Floating Car positioning device uploads is denoted as B, position A is 1500 meters at a distance from the B of position, the travel speed that the Floating Car is so calculated accordingly is at least 180km/h, have exceeded common sense, it therefore is a kind of abnormal spatial position jump, Data processing should be rejected.

3) quick map match.

By pretreated GNSS location data, need to combine city road network data that GNSS anchor point is projected to city map, establishes the matching relationship in anchor point and section by map-matching algorithm, and correct positioning drift bring error.

The electronic map of current each geographic area is all more full and accurate, and this electronic map can derive from the GIS-Geographic Information System in city, naturally it is also possible to come be originated from other modes and by way of.These electronic maps portray in detail to urban road information, if by dividing available trunk section.By the way that by information such as distance, angles, anchor point is matched on section, thereby realizes and location information is matched in actual geographical environment. 4) matching of the expression in Floating Car path and different vehicle path.

Under the premise of giving one group of terminus, the path of vehicle may not be unique.If complicated urban road network contains trunk section, these different sections are numbered, for example, section is expressed as Ll, L2 etc..Road may there are two different driving directions, in this case, it should two different driving directions are expressed as two different sections, give different section numbers.

The intersection point in section in city road network usually can be used in given beginning and end.The path of certain known floating vehicle travelling is now needed from the routing information that other Floating Cars have been sent, and path identical with the Floating Car path is selected, thus the same group of paths between obtaining beginning and end.

5) sampling of data.

In the location data of Floating Car, include the information such as position coordinates, automobile's instant velocity, record time.In the urban highway traffic method for detecting abnormality based on floating car data that this patent proposes, sampling of data refers to that partial data is filtered out from whole floating car datas carries out subsequent analysis processing, and this screening is computing capability based on data center and the required precision proposed in advance and carries out.Based on different computing capability and required precision, different sampling of data methods can be used.For example, the computing capability when data center is stronger, and it is higher to the required precision of detection when, can carry out comprehensive processing using whole Floating Car location datas as process object and analyze；And when the computing capability of data center is limited, it is assumed that current data center can be in 1 minute, 500 datas are handled to each space sub-district, and actual conditions are that each space sub-district can produce 2000 Floating Car location datas at 1 minute, 500 datas can be so randomly selected from 2000 datas to be analyzed, to within the scope of the computing capability of data center, obtain the limited processing result of precision.

It according to the difference to floating car data Land use systems, can be sampled for the different attribute of floating car data, such as travel speed and journey time etc..The urban highway traffic method for detecting abnormality based on floating car data proposed in this patent, using progress urban highway traffic abnormality detection based on travel speed.Therefore, sampling of data, which refers to, is sampled travel speed.

6) historical trajectory data analysis and the training of R N model.

So-called historical trajectory data refers to the accumulative Floating Car track data in long-term urban highway traffic operation.Using history Floating Car track data, urban highway traffic characteristic model can establish, for reflecting the general characteristic of urban transportation operation.Urban highway traffic characteristic model mentioned here can refer to certain specific indexs, such as average speed, weighted mean velocity etc.；It can also refer to certain various statistical model, such as the probability distribution of travel speed.Previous many models, the traffic characteristic (such as history average speed) in some section or region is indicated using single index, although the application of this mode is easy, precision is not high, sensibility is poor, tends not to play good effect in traffic abnormity state-detection.Therefore, this patent is proposed for each space-time sub-district, and the training of RNN model is carried out with the situation of change of traffic characteristic variable.The R N model can provide the predicted value of subsequent time traffic characteristic variable in period traffic characteristic variable true values several before providing.

Collectable traffic characteristic variable, including travel speed and journey time etc., this patent reflect traffic circulation feature using travel speed, and old friend leads to characteristic variable and refers to travel speed.

7) analysis of real-time track data and feature extraction.

So-called real-time track data refer to the track data of Floating Car in the traffic circulation within not far a period of time at current time.Using real-time Floating Car track data, the variation dynamic of traffic characteristic can be grasped, for reflecting the instant characteristic of Current traffic operation.This patent describes Current traffic feature using the travel speed of current space-time sub-district.

8) abnormality detection.

The thought of system mode abnormality detection is proposed that the abnormal conditions used by system on monitoring system record of the audit can detecte out and violate event that is safe, may causing system exception by Dennrng earliest.This Model Independent that Dennrng is established is the abnormality detection model in a kind of universal significance in any specific system, application environment, system vulnerability, fault type.The model includes 5 main body, object, record of the audit, profile, exception record and active rule parts.Profile is the main body that indicates with measurement and statistical model relative to object Normal behaviour.The model of Dennrng defines 3 kinds of measurements, that is event counter, intervalometer, resource measurement device, and 5 kinds of statistical models are proposed, model, mean value and standard deviation model, multivariate model, Markov process model and time series models can be operated.The model that Denning is proposed passes through the analysis to system audit data, set up the normal behaviour feature contour based on statistics of system body, when detection, Audit data in system is compared with the normal behaviour feature contour of established main body, if dissimilar parts are more than some threshold value, it is taken as an anomalous event.The model has established the basis of abnormality detection, and many method for detecting abnormality and system developed later are grown up based on it.

In recent years in the development process of abnormality detection technology, the method for introducing more artificial intelligence, to improve the performance of abnormality detection.The method of these artificial intelligence mainly includes data mining, artificial neural network, Fuzzy Evidence Theory etc..The method of data mining is used to determine that in a large amount of data acquisition system, what feature is most important.The technology simply enumerates all normal modes for mainly seeking a kind of definition that normal mode is more succinct in abnormality detection as traditional method for detecting abnormality.Being introduced into for data digging method enables detection system only by the main feature in identification normal mode, it will be able to synoptically include the normal mode for not including in training data.Artificial neural network abnormality detection problem is regarded as in the statistics abnormality detection that a general data classification problem is spoken of in front, and user behavior data is divided into two classes according to certain statistical criteria: i.e. abnormal behaviour and normal behaviour.Since there are certain difficulties when extracting, being abstracted audit example for Statistics-Based Method, it may cause large error, some probability distribution are necessarily dependent upon it is assumed that generally requiring the measurement for portraying user behavior with feeling by rule of thumb, so introducing the clustering method of artificial neural network.Artificial neural network has self study adaptive ability, neural network is trained with the sample point for representing normal users behavior, pass through repeated multiple times study, neural network can extract the mode of normal user or system activity from data, and it is encoded in network structure, when detection, Audit data is passed through into the neural network that succeeds in school, can decision-making system it is whether normal.Since there is abnormal judgment criteria certain ambiguity such as to establish a kind of intrusion detection frame model based on fuzzy expert system so Fuzzy Evidence Theory is introduced in exception, can preferably reduce false dismissed rate and false alarm rate.

This patent proposes a kind of abnormality detection scheme based on Recognition with Recurrent Neural Network, particularly, it is a kind of abnormality detection scheme using RNN (Recognition with Recurrent Neural Network), R N model based on historical data is inputted to real-time track data, obtain the gap between predicted value, then comparison prediction value and true value.The program uses deep learning method, can automatically update model over time, has stronger adaptive ability.

9) abnormal severity quantification characterization and exception information publication.

The seriousness of traffic abnormity situation should be issued by concise mode to the public, to avoid possible congestion regions, improve the operational efficiency of urban transportation.The severity of unusual condition is characterized with traffic abnormity index, range 0-10, wherein 0 indicates without exception, 10 indicate Height Anomalies.

Abnormal generation position projects on electronic map, and is published by forms such as Intelligent mobile equipment APP.

10) system performance evaluation.

The evaluation of system performance refers to the accuracy of evaluation traffic abnormity state-detection, and evaluation index includes rate of false alarm and rate of failing to report.Rate of false alarm and the lower performance for showing system of rate of failing to report are better.In the step 1), the division of space-time sub-district can specifically use following methods:

11) equidistant space-time partitioning.Determine the segment scale of time dimension, time slice span is fixed value, usually takes 30mm as a time slice；Determine the segment scale of Spatial Dimension, space segment span is fixed value, usually takes the space lattice of 200mX 200m as a space segment.

12) non-equidistant space-time partitioning.2km/km is greater than for road mileage²Or peak hour flow is greater than the downtown area of 1000/hour, takes the time slice of 30min and the space segment of 200m X 200m, is less than 2km/km for road mileage²Or peak hour flow takes the time slice of 30min and the space segment of 400mX 400m less than the city suburbs of 1000/hour. The step 3) comprising the following steps:

It 31) is a certain size grid by the Spacial domain decomposition of required processing, the range of each mesh region is represented by={ (x_s,y_s)\x_s ^[x_r,x_r+1),y_s ,^,₊₁) ' each mesh region includes several sections, it is R the set expression in these sections_S, every section in the set in the section is expressed as ij, and assigns number for each section；

32) determine the mesh region where anchor point, and utilize distance and bearing angle, the section ^ matching scheme where certain anchor point A is searched in the set in section includes:

321) single-point matching scheme:

Detection range point A nearest section meets when the difference for meeting the deflection at driving direction angle and section ij of point A is less than threshold value |-| <, matching is completed, the threshold value is 2.5 °, 5 °, 10 ° etc. desirable；If being unsatisfactory for | <, section is deleted in search space and continues searching other sections, until meeting t condition.Matching process is as shown in Figure 3.

322) point sequence matching scheme: X

This programme is suitable for high frequency floating car data.Floating Car GNSS data frequency acquisition is expressed as f₀=l, by the time upper point POHO adjacent with A), Pfc+i.；) it is defined as the 1- neighbor point of Α, P04-2i_Q;), P 4+2i_Q) it is defined as the 2- neighbor point of A, and so on, then P (t_A- kk), Pfc+ be defined as A /t- neighbor point./_QWhen < lHz ,/t=l or 2 is taken.If take distance A and A /t- neighbor point apart from the smallest section ^ and calculate A and A neighbor point driving direction angle mean value ^^ meet | .- | <, complete matching；Otherwise, other sections are searched for, until meeting | ^.- | <.

33) linear equation (then approximation is split as straight line if curve section) for utilizing section calculates projection coordinate of the GNSS anchor point on section, reduces because of GNSS positioning drift bring error.Specific method uses GNSS anchor point linear projection method are as follows:

Determine the linear equation (if section is curve, being divided into several linear sections) of section ^: y-y ,=k (x-x) wherein slope are as follows: Projection straight line equation are as follows: y-y_ASolving projection coordinate P is

k² +\

k²y_A +y_i +kx_A -kc_i

k² +\

After map matching process, in conjunction with the time stamp data of anchor point coordinate, anchor point is matched to step 5) described in space-time sub-district specifically can be using one of following methods:

51) bulk sample this programme of velocity information.By whole travel speed data of each in a space-time sub-district Floating Car, constitute overall.Implementation method is the travel speed for calculating each car in space-time sub-district ξ :=+-' ' ", wherein 4,₂... 4-be space-time sub-district ξ Distance between interior the 1st and the 2nd GNSS anchor point ... ..., " -1 with " at a distance between a GNSS anchor point ... t_nFor the 1st in space-time sub-district ... ..., " the timestamp of a GNSS anchor point；Data in each space-time sub-district are not screened, a set ^ is constituted, are used for subsequent processing.

52) the time smoothing sampling plan of velocity information.Specified time fragment length, the same time slice number of data upper limit；The speed data in a space-time sub-district in time each time slice is searched for, if speed data item number is more than the upper limit in time slice, pending data sample is added in the data of random capping item number.Implementation method is the travel speed for calculating each car in space-time sub-district: v_f = ² "²'³+ ...+- ' ' ", wherein₂... it is the 1st distance between the 2nd GNSS anchor point in space-time sub-district ... ..., the " -1 and n-th

Distance between GNSS anchor point is the 1st in space-time sub-district ... ..., " the timestamp of a GNSS anchor point；Specified time fragment length t_p, the same time slice number of data upper limit;^_αχ;The speed data in a space-time sub-district in time each time slice is searched for, if speed data item number is more than the upper limit in time slice;^ takes at random；^_∞Data is added ^ and is used for subsequent processing.The step 6) comprising the following steps:

61) network structure for determining R N, used here as the basic structure of Elman-RNN neural network, network structure is as shown in Figure 4, Figure 5.Network specifically includes consisting of part:

611) input layer: according to the feature of neural network, input layer is each example of historical data to be trained, since input data here is one-dimensional data flow, the time series data that travel speed data i.e. in the sub-district of space are constituted on time dimension, therefore, input layer number and output layer neuron number are both configured to 1.

612) hidden layer: in the design of neural network, the number setting of hidden layer not yet is finalized, and is typically necessary through many experiments come the number of the final hidden neuron for determining network model, it is proposed that value is 5 ~ 8.

613) output layer: the purpose for establishing neural network is output predicted value, i.e., the value at next moment is predicted by the value at the first two moment.Output layer output valve is predicted value, therefore output layer number is set as 1.

614) context level (Context layers): in Elman-Network, context level is used to save the output of previous moment hidden layer, therefore context level neuron number and hidden layer number take identical value.

615) bias node: input layer and output layer respectively contain a bias node, and initial value is disposed as 0.

62) basic parameter of Elman-RNN neural network model is set:

It include two parameters, the respectively number of plies of hidden layer neuron number and context level in Elman network structure.Hidden layer neuron number selects preferable result by test of many times, it is proposed that be set as 5 ~ 8；Since a context level saves the hidden layer output valve at previous moment, it is proposed that be set as 1.

Need to be arranged parameter learning rate when carrying out the training of model using back-propagation algorithm.The size of learning rate determines the variation degree of weight in neural network iterative process, big learning rate can be such that algorithm restrains rapidly, but local solution may be fallen into, small learning rate makes convergence speed of the algorithm slower, but can guarantee and converge to global minima, therefore, under normal conditions, the value of learning rate should choose smaller value, to ensure the performance and stability of model, it is proposed that be set as 0.01 to 0.8.

In the simulated annealing module of algorithm, need to be arranged the number of iterations of initial temperature, final temperature and each temperature value.A possibility that value of initial temperature has an important influence algorithm performance, and initial temperature is bigger, and the number of iterations of algorithm is more, converges to minimum value is also bigger, but time cost is also bigger.Similarly, when initial temperature is lower, the performance of algorithm will be affected, but time cost is also less.In practical applications, the setting of initial temperature needs to be selected by multiple repetition test.Final temperature refers to during temperature decline, lower limit temperature set by termination algorithm.The number of iterations of each temperature value is more, and the number that generation may solve is more, and converging to global minima is worth possibility also bigger.It is recommended that setting 10 for initial temperature⁵, 10-are set by final temperature², the number of iterations of each temperature is arranged It is 100.

63) training of R N model is carried out.The input of R N model is time series data={ ν that travel speed is constituted₂, ... , v_n, it exports as the Elman-R N neural network model with optimized parameter.The step of model training, is as follows:

631) read data from data file or database, and data be combined into (input, output) right, i.e. ^ O^H^) form, to carry out model training.

632) Elman-RNN neural network is created, wherein, input layer number is 1, hidden layer neuron number is 5 ~ 8, output layer neuron number is 1, and context level saves the output of a moment hidden layer, and neuron number is identical as hidden layer, since the output in training data is all positive number, Sigmoid activation primitive is used.

633) R N model is initialized, setting input layer arrives hidden layer, hidden layer to output layer, and the weight connected between neuron is random value, the bias unit of input layer and hidden layer is respectively set, and be initialized as 0.

634) it is back-propagation algorithm that main training algorithm, which is arranged, back-up algorithm is simulated annealing, carry out training pattern using the mixed strategy of the two algorithms, in the training process, if the error rate of model does not decline after certain training, or the amplitude of decline is less than the minimum value of setting, then using simulated annealing training；Using Greedy strategy, if the error rate of model does not improve after certain training, abandons updating weight, be last training result by the weights resetting of model.

635) before starting training, the network weight and error rate of the "current" model that Greedy strategy needs are saved.Save the error rate for the "current" model that mixed strategy needs.Start to train, for context level and input layer data, respectively multiplied by corresponding weight, in addition biasing, the output of hidden layer is calculated by activation primitive.Layer data is hidden multiplied by corresponding weight, in addition biasing, by activation primitive, obtains the output of model.The difference for calculating reality output and model output calculates gradient using back-propagation algorithm by Feedback error to hidden layer and input layer.Weight increment is calculated according to the gradient of calculating, weight is updated, saves updated weight.Duplication hidden layer is output to context level.

64) after single training, model is handled using following strategy:

641) Greedy strategy: if this training error rate does not improve, restore weight and error rate is last value.

642) mixed strategy: if the error rate of model does not decline after this training, or the amplitude of decline is less than the minimum value set, then using simulated annealing training.The primary training of simulated annealing, substantially steps are as follows: calculate the error score of "current" model first, later to all weights and bias of "current" model, add a random number ", obtain new weight and bias, in which:

, , 0.5— Random .

add = * temp

startTemp

In formula, Random is random number, and startTemp is initial temperature, and temp is Current Temperatures；Updated model error score is calculated, if new error score is less than error current score, illustrates that new weight has improvement to the performance of model, then saves new weight, otherwise abandon；By Current Temperatures multiplied by a fixed ratio r atio to reduce temperature:

\o stopTemp I startTemp)

ratio = exp

cycles - 1

In formula, stopTemp is final temperature, and cycles is once trained the number of iterations.It repeats above procedure cycles times.

65) stop condition judges: if being greater than the threshold value of setting to the number that the improvement degree of model is less than minimum value, algorithm is terminated.It repeats 63), 64), 65；) step until algorithm terminate, obtain trained R N model.Training process is referring to Fig. 6.The step 7) specifically can be using one of following methods:

71) time window averaging method.The travel speed data obtained using sampling of data calculate current spatial sub-district travel speed in current time The mean value of sub-district/_rt ,_rt)=" ^, the characterization as real-time traffic situation.

N 7

72) rolling time horizon averaging method.The travel speed data obtained using sampling of data, calculate current spatial sub-district travel speed nearest M time sub-district mean value/_rt(_irt)=~ ^ ￡ $ ^,, 3 ~ 5 wherein are taken as M-, the characterization as real-time traffic situation.

The step 8) comprising the following steps:

81) sequence for the point that outlier detection time series data subscript to be dealt with is sorted in ascending order_Vl,v₂,...,v„.In order to be fitted to data, data are expressed as (1) ..., (", 1) in plane coordinates in the form of two-dimensional points.

82) in order to detect abnormal point numerical, the data of RNN model prediction and the difference of truthful data need to be calculated, the specific implementation process is as follows:

821) the known model ^ being fitted according to historical data, the predicted value of digital simulation model

822) difference of computation model predicted value and true value: [^, ^]=| ^- ^.The step 9) comprising the following steps:

91) by the VELOCITY DISTRIBUTION difference criteria of each space-time sub-district turn to 0 ~ 1 standardization numerical value "_ί:

diff^ -vam diff)

One max diff^-min diff,

92) the traffic abnormity index of each space-time sub-district is calculated 10;

93) regional location by abnormal index higher than 5 projects on electronic map, and the forms such as Intelligent mobile equipment APP are published to society, so that driver avoids potential congestion points, improve the traffic efficiency of urban highway traffic.The step 10) comprising the following steps:

101) rate of failing to report of traffic abnormity state is calculated:

A = ><100%

102) rate of false alarm of traffic abnormity state is calculated:

=^χ100%

n

It is that false positive event is total in the unit time to fail to report total number of events in the unit time in above two formula, ".For the abnormal time sum actually occurred in the unit time.The present invention has the advantage that compared to the similar technique in same field

(1) existing Floating Car operation data (GNSS track data) is made full use of, pass through historical traffic feature extraction and real-time traffic Study on Trend, the variation that traffic behavior occurs is detected, urban highway traffic anomalous event real-time, low cost, intellectualized detection may be implemented；

(2) situation of change of traffic characteristic is established into R N model, more careful can comprehensively reflects traffic circulation rule, avoids one-sidedness, the unstability for characterizing traffic characteristic using single index, the reliability of detection is higher； (3) it can be adaptively adjusted model parameter, reflect slight change of the traffic circulation rule under big time scale；

(4) through the inspection of real data, urban highway traffic abnormality detection technology proposed by the present invention based on floating car data, it can be realized the higher accident detection of accuracy, verification and measurement ratio is more than 90%, rate of failing to report is lower than 15%, rate of false alarm is lower than 20%, achieves good detection effect, can be applied to urban transportation intelligent management, service.Detailed description of the invention

Particular content and advantage of the invention will be apparent and should be readily appreciated that in conjunction with the following drawings, in which:

Fig. 1 shows element and basic principle schematic of the invention；

Fig. 2 shows overall procedure schematic diagram of the present invention in implementation process；

Fig. 3 shows quickly map-matching algorithm embodiment schematic diagram of the invention；

Fig. 4 shows the structure diagram of RNN in the present invention；

Fig. 5 shows the detail of construction of RNN in the present invention；

Fig. 6 shows the flow diagram for carrying out R N training；

Fig. 7 shows the flow diagram that the present invention is carried out abnormality detection by the comparison of RNN model and truthful data.Specific embodiment

The purpose of the present invention, technical solution and advantage are clearly stated in order to be more clear, specific embodiments of the present invention are described in detail below.

As shown in Fig. 1, total system framework of the invention includes: vehicle-mounted GNSS track recorder, data center, GNSS satellite and the communication system that Floating Car is carried.GNSS herein includes any similar navigation satellite positioning system such as GPS, GLONASS, GALILEO, Beidou, IRNSS, QZSS.The GNSS track recorder that the Floating Cars such as taxi, bus are carried, with the location information of certain sample frequency/(generally requiring family 0.1Hz) record vehicle each time point under steam, and location information is sent to by data center by GPRS mobile communications network (wireless network communication techniques such as WCDMA, TD-LTE also can be used, but cost will correspondingly increase) in real time.Data center establishes history road traffic features database by special algorithm by data prediction, data fusion；For nearest received real time data, real-time traffic property data base is established；By the mapping relations of historical data base and real-time data base, whether abnormal Current traffic feature is differentiated, and visualized by processing terminal and generate traffic abnormal incident report.

Referring to fig. 2, including acquisition and storage GNSS track data establishes space-time sub-district to the overall procedure of scheme, historical traffic feature extraction, real-time traffic feature extraction, anomalous identification and etc..Acquisition and storage GNSS track data is the data basis of entire scheme, and since data magnitude is huge, Ying Caiyong distributed storage scheme has mature technology for distributed storage at present, is not the contents of the present invention.Space-time sub-district is established, basic assumption is that have identical traffic characteristic in a certain specific region, specific time period, it is assumed that this is by long-term observation, is blanket.Historical traffic feature extraction, its principle is that travel speed is calculated, utilizes a large amount of history travel speed data of same space-time sub-district using GNSS track data, training Elman-RNN model utilizes the changing rule of neural network model characterization traffic characteristic.Real-time traffic feature extraction, principle are that the travel speed data in current slot are carried out to processing analysis, obtain and implement traffic characteristic statistic.Anomalous identification is that the prediction codomain true value for obtaining the Elman-RNN model of history travel speed training is compared, it is determined whether traffic abnormal incident occurs.

The combination of the implementation method according to summary of the invention, it is as follows to provide embodiment.Embodiment one

Step 11, using equidistant space-time partitioning, determine the segment scale of time dimension, time slice span is fixed value, usually takes 30mm As a time slice；Determine the segment scale of Spatial Dimension, space segment span is fixed value, usually takes the space lattice of 200mX200m as a space segment.

Step 12 carries out data prediction, and GNSS location data is carried out data cleansing, data integration, data conversion, data regularization, improves the structuring degree of data.

The Spacial domain decomposition of required processing is a certain size grid by step 13, and the range of each mesh region is represented by

A_s = {(x_s,y_s) I x_s e [x_r,x_r+1),y_s e [y_r,y_r+1)}；Determine the mesh region where anchor point, and utilize distance and bearing angle, searches for the section where anchor point；Detection range point A nearest section, takes threshold value ^=2.5 °, when the difference for meeting the deflection at driving direction angle and section ^ of point A is less than threshold value, that is, meets |-| <, complete matching；If being unsatisfactory for |-| <, section is deleted in search space and continues searching other sections, until meeting condition；Utilize the linear equation (then approximation is split as straight line if curve section) in section, calculate projection coordinate of the GNSS anchor point on section, reduce because of GNSS positioning drift bring error, method particularly includes: determine the linear equation of section ^ (if section is curve, then it is divided into several linear sections): y-y,=k (x-x) wherein slope are as follows: projection straight line equation are as follows: solve projection coordinate P are as follows:

k²y_A +y, +kx_A -h,

k² +\

After map matching process, in conjunction with the time stamp data of anchor point coordinate, anchor point is matched to space-time sub-district.

Step 14, vehicle speed datas of being driven a vehicle by the whole of each in a space-time sub-district Floating Car, are constituted overall.It calculates every in space-time sub-district

I d.

The travel speed of vehicle: wherein₂... for the 1st distance between the 2nd GNSS anchor point in space-time sub-district ..., the " -1 with n-th of GNSS anchor point between at a distance from, be the 1st in space-time sub-district ..., " the timestamp of a GNSS anchor point；Data in each space-time sub-district are not screened, a set V is constituted_{, it is used for subsequent processing.

The Elman-RNN neural network that step 15, foundation are made of input layer, hidden layer, output layer, context level, bias node, input layer number is set as 1, hidden layer neuron number is set as 5, output layer neuron number is set as 1, context level neuron number and hidden layer number take identical value, 5 are also configured as, the initial value of bias node is set as 0;The initial parameter of each component part is set, and wherein learning rate is set as 0.1, and initial temperature is set as 10 in simulated annealing module⁵, final temperature is set as 10-², the number of iterations of each temperature is set as 100;Using ^ as input data, the training of neural network is carried out, finally obtains trained R N model.Step 16, real time traffic data is taken on actual time window mean value/_rt ,_rt)=" ^ J, as the characterization for implementing traffic situation.The sequence of step 17, the point that outlier detection time series data subscript to be dealt with is sorted in ascending order_Vl,v₂..., v, it is known that according to the model ^ that historical data is fitted, the predicted value (V) of digital simulation model;The difference of computation model predicted value and true value Step 18, the standardization numerical value α that the VELOCITY DISTRIBUTION difference criteria of each space-time sub-district is turned to 0 ~ 1_ξι:

diff^ -mm(diff)

max [diff、 - min、diff、

Calculate the traffic abnormity index of each space-time sub-district 10.Embodiment two

Step 21, using equidistant space-time partitioning, determine the segment scale of time dimension, time slice span is fixed value, usually takes 30mm as a time slice；Determine the segment scale of Spatial Dimension, space segment span is fixed value, usually takes the space lattice of 200mX200m as a space segment.

Step 22 carries out data prediction, and GNSS location data is carried out data cleansing, data integration, data conversion, data regularization, improves the structuring degree of data.

The Spacial domain decomposition of required processing is a certain size grid by step 23, and the range of each mesh region is represented by A_s = {(x_s,y_s) I x_s e [x_r,x_r+1),y_s e [y_r,y_r+1)}；Determine the mesh region where anchor point, and utilize distance and bearing angle, searches for the section where anchor point；Detection range point A nearest section, takes threshold value ^=2.5 °, when the difference for meeting the deflection at driving direction angle and section ^ of point A is less than threshold value, that is, meets |-| <, complete matching；If being unsatisfactory for |-| <, section is deleted in search space and continues searching other sections, until meeting condition；Utilize the linear equation (then approximation is split as straight line if curve section) in section, calculate projection coordinate of the GNSS anchor point on section, reduce because of GNSS positioning drift bring error, method particularly includes: determine the linear equation (if section is curve, being divided into several linear sections) of section ^:

Y-y ,=k (x-x) wherein slope are as follows: k: projection straight line equation are as follows: y-

Solve projection coordinate P are as follows:

k²+l

k²y_A +y, +kx_A -h,

k²+\

I d.

Step 24, the travel speed for calculating each car in space-time sub-district: wherein₂... 4-1 is the 1st distance between the 2nd GNSS anchor point in space-time sub-district ..., " -1 between n-th of GNSS anchor point at a distance from, ^... is the 1st in space-time sub-district ..., " the timestamp of a GNSS anchor point；The same time slice number of data upper limit of specified time fragment length;^ «;The speed data in a space-time sub-district in time each time slice is searched for, if speed data item number is more than that upper limit p takes p data that ^ is added at random in time slice.

The Elman-RNN neural network that step 25, foundation are made of input layer, hidden layer, output layer, context level, bias node, input layer number is set as 1, hidden layer neuron number is set as 5, output layer neuron number is set as 1, context level neuron Number and hidden layer number take identical value, are also configured as 5, the initial value of bias node is set as 0;The initial parameter of each component part is set, and wherein learning rate is set as 0.1, and initial temperature is set as 10 in simulated annealing module⁵, final temperature is set as 10-², the number of iterations of each temperature is set as 100;Using ^ as input data, the training of neural network is carried out, finally obtains trained R N model.Step 26, real time traffic data is taken on actual time window mean value/_rt ,_rt)=" ^ J, as the characterization for implementing traffic situation.The sequence of step 27, the point that outlier detection time series data subscript to be dealt with is sorted in ascending order_Vl,v₂..., v, it is known that according to the model ^ that historical data is fitted, the predicted value of digital simulation model (V);The difference of computation model predicted value and true value " I

Step 28, the standardization numerical value that the VELOCITY DISTRIBUTION difference criteria of each space-time sub-district is turned to 0 ~ 1:

diff^ -vam diff)

One max diff^-min diff,

Calculate the traffic abnormity index of each space-time sub-district 10.Embodiment three

Step 31, using non-equidistant space-time partitioning, 2km/km is greater than for road mileage²Or peak hour flow is greater than the downtown area of 1000/hour, takes the time slice of 30min and the space segment of 200mX200m, is less than 2km/km for road mileage²Or peak hour flow takes the time slice of 30min and the space segment of 400mX400m less than the city suburbs of 1000/hour.

Step 32 carries out data prediction, and GNSS location data is carried out data cleansing, data integration, data conversion, data regularization, improves the structuring degree of data.

The Spacial domain decomposition of required processing is a certain size grid by step 33, and the range of each mesh region is represented by

A ={( ' 1 ^[x_r,x_r+1),y_s

Floating Car GNSS data frequency acquisition is expressed as f₀=l, by the time upper point P (t adjacent with A_A-t₀), Pfc+io) it is defined as the 1- neighbor point of A, P (t_A-2t₀), ^04+2) be defined as the 2- neighbor point of A, and so on, then Ρ θ 4-/；), be defined as Α/- neighbor point./_QWhen < lHz, ^=1 or 2 are taken.The neighbor point of distance A and A are taken apart from the smallest section ^ and calculates A's and A

The mean value at ^ neighbor point driving direction angle takes threshold value ^=5 °, if meeting |-| <, complete matching；Otherwise, other sections are searched for, until meeting condition.

Using the linear equation (then approximation is split as straight line if curve section) in section, projection coordinate of the GNSS anchor point on section is calculated, is reduced because of GNSS positioning drift bring error.Method particularly includes:

Determine the linear equation (if section is curve, being divided into several linear sections) in section: y-y ,=k (x-x) wherein slope are as follows: projection straight line equation is

Solve projection coordinate P are as follows: ky_A-k +k²x_{l +}x_A

^P k²+\

v _^y_A +y,

k²+l

Step 34, the travel speed for calculating each car in space-time sub-district :=W .++- ' ' ", wherein₂Gap when ...-^ is

~

The 1st distance between the 2nd GNSS anchor point in area ..., the " -1 between n-th of GNSS anchor point at a distance from, ^... is the 1st in space-time sub-district ..., " the timestamp of a GNSS anchor point；The same time slice number of data upper limit of specified time fragment length;^ «;The speed data in a space-time sub-district in time each time slice is searched for, if speed data item number is more than that upper limit p takes p data that ^ is added at random in time slice.

The Elman-RNN neural network that step 35, foundation are made of input layer, hidden layer, output layer, context level, bias node, input layer number is set as 1, hidden layer neuron number is set as 5, output layer neuron number is set as 1, context level neuron number and hidden layer number take identical value, 5 are also configured as, the initial value of bias node is set as 0;The initial parameter of each component part is set, and wherein learning rate is set as 0.05, and initial temperature is set as 10 in simulated annealing module⁵, final temperature is set as 10-², the number of iterations of each temperature is set as 200;Using ^ as input data, the training of neural network is carried out, finally obtains trained R N model.

Step 36, using rolling time horizon averaging method, the travel speed data obtained using sampling of data, calculate current spatial sub-district travel speed

1

Nearest M time sub-district mean value/_rt(_irt)=~ ^ ￡ $ ^,, wherein M=3, the characterization as real-time traffic situation.

M -N_rt i i

The sequence of step 37, the point that outlier detection time series data subscript to be dealt with is sorted in ascending order_Vl,v₂..., v, it is known that according to the model ^ that historical data is fitted, the predicted value of digital simulation model (V);The difference of computation model predicted value and true value " I

Step 38, the standardization numerical value that the VELOCITY DISTRIBUTION difference criteria of each space-time sub-district is turned to 0 ~ 1:

diff^ -vam diff)

One max diff^-min diff,

Calculate the traffic abnormity index of each space-time sub-district 10。

Claims

Claims

1. a kind of intelligence road traffic method for detecting abnormality, includes the following steps:

1) it establishes space-time sub-district: being divided into some time segment for one day, each time slice is known as a time sub-district；It is several space segments by the implementation region division of urban highway traffic abnormality detection, each space segment is known as a space sub-district；The intersection of any one time sub-district and any one space sub-district is known as space-time sub-district；Implementation method is: determining the segment scale of time dimension, time slice span is fixed value, takes 30mm as a time slice；Determine the segment scale of Spatial Dimension, space segment span is fixed value, takes the space lattice of 200m X 200m as a space segment；

2) pretreatment of historical trajectory data: being the sampling vehicle speed data of historical track by Floating Car GNSS position history data processing；The pretreatment of real-time track data: being the sampling vehicle speed data of real-time track by Floating Car GNSS positioning real time data processing；

3) historical trajectory data analysis and RNN model training: it is regular to carry out tissue for the sampling vehicle speed data of the historical track, and training RNN model obtains history speed characteristic model M_RNN；

The analysis of real-time track data and feature extraction: using the sampling vehicle speed data of the real-time track, the parameter for being able to reflect real-time traffic feature is calculated；

4) abnormality detection: measuring the difference of the history speed characteristic model Yu the real-time traffic feature by otherness index, obtains history and real-time traffic feature difference value；

5) history and real-time traffic feature difference value calculating traffic condition abnormal index abnormal severity quantification characterization: are utilized；

6) system performance evaluation: the accuracy of evaluation traffic abnormity state-detection measures the degree of stability of system operation.

2. intelligence road traffic method for detecting abnormality as described in claim 1, which is characterized in that the pretreatment of historical trajectory data described in step 2) includes:

2a) data structured: Floating Car GNSS position history data are subjected to data cleansing, data integration, data conversion, data regularization, obtain structuring GNSS position history data；

2b) quick map match: in conjunction with city road network data, pass through map-matching algorithm, by structuring GNSS position history data projection to city road network, establish the anchor point in the structuring GNSS position history data and the matching relationship in section, the anchor point in the structuring GNSS position history data and the matching relationship table in the section are obtained, and corrects positioning drift bring error；2c) traffic circulation calculation of characteristic parameters and sampling: traffic circulation characteristic parameter is calculated according to the structuring GNSS position history data, obtain the traffic characteristic data of historical track, and sampling of data is carried out to the traffic characteristic data of the historical track, obtain the sample traffic characteristic of sampling history track.

3. intelligence road traffic method for detecting abnormality as claimed in claim 2, which is characterized in that step 2b) described in quick map match include:

2b 1) it by the Spacial domain decomposition of required processing is a certain size grid, the range of each mesh region is represented by

A = {(^y_s) \ x_s e [„ ,₊₁),^ e[u₊₁), each mesh region includes several sections, is R the set expression in these sections_s, the set R in the section_sIn every section be expressed as and be that each section assigns number；

2b2) determine the mesh region where anchor point, and utilizes distance and bearing angle, the set R in section_sSection where middle search certain anchor point A

GNSS anchor point linear projection method 2b3) is utilized, projection coordinate of the GNSS anchor point on section is calculated.

4. intelligence road traffic method for detecting abnormality as claimed in claim 3, it is characterized in that, step 2b2) using one of following methods: 2b21) single-point matching process: detection range anchor point A nearest section, implementation method is: for certain a road section ij in the set in section when the difference for meeting the deflection at driving direction angle and section ij of point A is less than threshold value, that is, meeting

|-| < when, complete matching；If being unsatisfactory for |-| <, other sections in the set R of section are continued searching, until meeting |-| <；

2b22) point sequence matching process: this programme is suitable for high frequency floating car data；The Floating Car GNSS data time interval of every two adjacent times is expressed as., Floating Car GNSS data frequency acquisition is expressed as/_Q=l/i_Q, the time sheet of certain anchor point A is shown as the time upper point P (t adjacent with the anchor point A_A-t₀), Pfc+if is defined as the 1- neighbor point of A, P (t_A-2h), P04+2i_Q;) it is defined as the 2- neighbor point of certain anchor point A, and so on, then Pi-kt^h is defined as the ^ neighbor point of certain anchor point A；/_QWhen < lHz, ^=1 or 2 are taken;The mean value ^41 of the anchor point and its ^ neighbor point driving direction angle is taken apart from the ^ neighbor point of certain anchor point A and the anchor point apart from a smallest section ij' and calculates, if meeting | ^ mono- | <, complete matching；Otherwise, other sections are searched for, until meeting | .- | < 4.

5. intelligence road traffic method for detecting abnormality as claimed in claim 4, which is characterized in that step 2c) described in historical trajectory data pretreatment using one of following methods:

2cl) bulk sample this method: it by whole travel speed data of each in a space-time sub-district Floating Car, constitutes totally, implementation method is the travel speed for calculating each car in space-time sub-district :=W .++- ' ' ", wherein₂... it is space-time sub-district

~

Distance between interior the 1st and the 2nd GNSS anchor point ..., the " -1 between n-th of GNSS anchor point at a distance from, iL.A is the 1st in space-time sub-district ..., " the timestamp of a GNSS anchor point；Travel speed data in each space-time sub-district are not screened, a set ^ is constituted, are used for subsequent processing；

2c2) the methods of sampling of time smoothing: the same time slice number of data upper limit is arranged in specified time fragment length；The speed data in each time slice of space-time sub-district is searched for, if speed data item number is more than the upper limit in time slice, the data of random capping item number are used for subsequent processing, and implementation method is the travel speed for calculating each car in space-time sub-district ξ: V_f = ^{2 +} '³' ' ", wherein₂... for the 1st distance between the 2nd GNSS anchor point in space-time sub-district ..., the " -1 with the " at a distance between a GNSS anchor point, ^ ... is space-time sub-district ^ Nei Shang a ..., " the timestamp of a GNSS anchor point；Same time slice number of data upper limit ^ is arranged in specified time fragment length_∞;The speed data in a space-time sub-district in each time slice of time i-th is searched for, if speed data item number is more than upper limit ρ in time slice_∞αχ, take ρ Μ according to addition ^ and for subsequent processing at random.

6. the intelligent road traffic method for detecting abnormality as described in one of claim 1 to 5, which is characterized in that following steps are taken in the analysis of historical trajectory data described in step 3) and the training of R N model:

3a) establish the basic structure of Elman-RNN neural network, neuron number including input layer is arranged, is arranged the neuron number of hidden layer, the neuron number of output layer is arranged, the neuron number of context level is set, the biasing of input layer and output layer is set Node simultaneously assigns initial value；

The basic parameter of Elman-R N neural network model 3b) is set, parameter learning rate when model training, the initial temperature and final temperature of simulated annealing and the number of iterations of each temperature are carried out including back-propagation algorithm；

3c) using the sampling vehicle speed data of the historical track, the training of RNN model is carried out；

3d) model is handled after single training, the weight including updating neural network, the more new strategy when error rate assessment of model and error rate do not decline；

The stop condition judgement of model training 3e) is carried out, if being greater than the threshold value of setting to the number that the improvement degree of model is less than minimum value, algorithm is terminated, and obtains RNN model M_RNN；

The analysis of real-time track data described in step 3) and feature extraction are using one of following methods:

3f) time window averaging method: using the sampling vehicle speed data of the real-time track, current spatial sub-district travel speed is calculated in mean value/^ (v of current time sub-district_{i rt})=" ^ ￡ v, the expression parameter as real-time traffic feature；

3g) rolling time horizon averaging method: using the sampling vehicle speed data of the real-time track, current spatial sub-district travel speed is calculated in mean value/^ (v of nearest M time sub-district_{f rt}) = ~^ $ _w, wherein 3 ~ 5 are taken, the expression parameter as real-time traffic feature.

7. intelligence road traffic method for detecting abnormality as claimed in claim 6, which is characterized in that step 3a) it comprises the steps of:

3al) 1 is set by the neuron number of input layer;

3a2) 5 ~ 8 are set by the neuron number of hidden layer；

3a3) 1 is set by the neuron number of output layer;

3a4) set identical as hidden layer neuron number for the neuron number of context level；

One bias node respectively 3a5) is set for input layer and output layer, initial value is disposed as 0;

Step 3b) it comprises the steps of:

Parameter learning rate when back-propagation algorithm 3bl) being carried out model training is set as 0.01 ~ 0.8；

3b2) 10 are set by the initial temperature of simulated annealing⁵, the final temperature of simulated annealing is set as 10-²；

3b3) 100 are set by the number of iterations of each temperature.

8. intelligence road traffic method for detecting abnormality as claimed in claim 6, which is characterized in that step 3c) it comprises the steps of:

3d) according to time sequence by the sampling vehicle speed data of the historical track of each space sub-district, and it is the sampling vehicle speed data combination of two of the historical track after sequence is right at (input, output), that is (), ^, ^), ..., ^^, ^) form；

3c2) create Elman-RNN neural network, wherein, input layer number is 1, hidden layer neuron number is 5,6,7 or 8, output layer neuron number is 1, context level saves the output of a moment hidden layer, and neuron number is identical as hidden layer；Use Sigmoid activation primitive；

3c3) setting input layer to hidden layer, hidden layer to the weight connected between output layer neuron be the random value between 0 ~ 1；The bias unit of input layer and hidden layer is respectively set and is initialized as 0;

Two kinds of algorithms of back-propagation algorithm and simulated annealing 3c4) are used, composition mixed strategy carrys out training pattern.

9. intelligence road traffic method for detecting abnormality as claimed in claim 6, which is characterized in that step 3d) using one of following methods:

3dl) Greedy strategy: if the error rate of model does not decline after certain primary training, restore weight and error rate is the value before training；3d2) mixed strategy: if the error rate of model does not decline after certain primary training, or the amplitude of decline is less than the minimum value set, then using simulated annealing training；Training step of simulated annealing is as follows: calculating the error score of "current" model first, value to the weight and bias unit that are connected between all neurons of "current" model later, add a random number ", obtain the value of new weight and bias unit, middle add=- Random) in I startTemp * temp ' formula, Random is random number, and range is greater than 0 less than 1, startTemp is initial temperature, takes 10⁵, temp is Current Temperatures；Updated model error score is calculated, if new error score is less than error current score, illustrates that new weight has improvement to the performance of model, then saves new weight, otherwise abandon；By Current Temperatures multiplied by a fixed ratio r atio to reduce temperature: ratio=ex (log (stopTemp I startTemp)/cycles-1)), in formula, stopTemp is final temperature, takes 10-², cycles is once trained the number of iterations, takes 100;It repeats above procedure cycles times.

10. the intelligent road traffic method for detecting abnormality as described in one of claim 1 to 5, which is characterized in that step 4) abnormality detection the following steps are included:

4a) the sequence v v for the point that the subscript of the sampling vehicle speed data of abnormality detection historical track to be dealt with is sorted in ascending order₂,... ,v_n;Data are expressed as (1 in plane coordinates in the form of two-dimensional points；),. .

4b) calculate predicted value i >=M of RNN model_RNN(), and the difference d of computation model predicted value and true valueⁱff [^v _rn^v _pr ] = \^v _pr - _rt \ .

Step 5) exception severity quantification, which characterizes, includes:

5a) by each space-time sub-district

5b) calculate the traffic abnormity index of each space-time sub-district 10。