CN111768635A - A Coupling Robust Tensor Decomposition-Based Approach for Occasional Traffic Congestion Detection - Google Patents

Abstract

The invention discloses a coupling robustness tensor decomposition-based sporadic traffic jam detection method, which comprises the steps of firstly, collecting various traffic variable data; then according to the characteristics of various traffic data, constructing the traffic data into tensor models with the same size and the same order number; then, carrying out filling pretreatment on the lost data; then constructing a Bayesian robustness tensor decomposition model, providing a Bayesian robustness tensor decomposition model with a self-adaptive rank, and describing shared sparse distribution of different types of traffic data from a probability angle, thereby constructing a coupled robustness decomposition model of multiple types of traffic data; and finally, designing a solving method to realize high-precision rapid accidental traffic jam detection.

Description

A Coupling Robust Tensor Decomposition-Based Approach for Occasional Traffic Congestion Detection

Technical field:

The invention relates to the field of intelligent transportation, in particular to a method for detecting occasional traffic congestion based on coupled robustness tensor decomposition.

Background technique:

With the increase of traffic demand and the intensification of traffic congestion, traditional methods have been unable to solve the problem of congestion. Therefore, in recent years, intelligent transportation has been vigorously developed, and traffic congestion detection is part of the Advanced Traveler Information System (ATIS). Therefore, for traffic Research on congestion detection has also attracted extensive interest from scholars. Traffic congestion is divided into periodic traffic congestion and occasional traffic congestion. In comparison, occasional traffic congestion has a greater impact on traffic travelers, because it causes unexpected delays and seriously disturbs people. path selection decision. Since traffic accidents usually lead to occasional traffic congestion, people take it for granted that traffic accident detection is a necessary prerequisite for traffic congestion detection. traffic congestion. Since the 1870s, many surveillance-based traffic accident detection algorithms (AIDs) have been proposed. Subsequently, it was proposed to detect traffic accidents from a statistical point of view using Bayesian and standard normal deviation algorithms. There are also some researchers using pattern recognition algorithms for detection, such as the California algorithm. The California algorithm is also one of the earliest detection algorithms. Although there is a big lack of accuracy, it has also been improved in subsequent research. Besides, some data-driven neural network algorithms (ANNs) are also applied to event detection.

However, traffic accidents are not a complete substitute for occasional traffic jams, as traffic accidents are only one of many causes of occasional traffic jams. Although the causes of occasional traffic congestion vary, the effects of these causes on the distribution of various traffic states are similar. Therefore, the real-time traffic status can be used as a guide for occasional traffic congestion, and whether there is occasional traffic congestion can be judged through changes in the traffic status. Therefore, knowing the distribution and pattern of traffic state data such as traffic flow and speed under normal conditions is the key to detecting occasional traffic congestion. traffic congestion. In order to extract the normal traffic state data distribution and pattern, many existing studies have focused on setting the threshold, that is, given an initial preset value, when the actual value of the observed traffic data is greater than the preset value , it can be considered that there is occasional traffic congestion. Therefore, traditional accidental traffic detection generally includes four parts: data collection, data preprocessing, threshold design of normal traffic state data in a single time mode/dimension, and accidental congestion detection. The main process is shown in Figure 1. First, the road traffic status data (usually road density data) is collected; then the original data is preprocessed, including filtering out outliers, filling missing data, etc. If it is detector data, It is also necessary to eliminate the low detection rate data; then, according to historical experience or statistics and other methods, with time as the independent variable and traffic state data as the dependent variable, set the upper limit of the threshold value of the traffic state data under normal conditions; Traffic status data is compared to normal thresholds to detect when and where occasional traffic jams occur. However, the cyclical travel habits of commuters not only bring about the distinction between morning and evening peaks and non-peaks, but also make the traffic state have a weekly pattern, and the traffic on weekdays is significantly higher than on weekends. A single threshold determined by traffic state data such as traffic flow, speed, etc. may not accurately represent their normal distribution and patterns. also. The traffic data of adjacent paths and location points are correlated, which means that there is a strong spatial correlation in the traffic data. Therefore, a model that can fully utilize the distribution and spatial correlation of traffic data between days and weeks is needed to detect occasional traffic congestion.

Most of the early threshold-based detection methods take a single traffic data or data vector as the research object, and a vector can only represent one aspect of traffic data at most, such as sky pattern characteristics, weekly pattern characteristics or spatial characteristics, but not the characteristics of traffic data. Consider these properties together. In recent years, many researchers have begun to use matrices to express the spatiotemporal correlation of traffic data, that is, to construct the traffic data into a matrix model instead of a simple vector. Among them, the BRPCA-based detection method assumes that the distribution of sporadic traffic events is sparse in time and space, and the normal traffic data distribution is low-rank, because they prove that traffic observations have strongly correlated temporal and spatial patterns, It shows periodicity, and the traffic data observations of adjacent upstream and downstream sections show strong correlation. The experimental results based on the BRPCA method demonstrate that outliers such as delays can be well detected and tracked by the low rank and sparsity of traffic data. In addition, Berk Anbaroglu et al. proposed in 2014 the detection of occasional traffic congestion in urban road networks, they also mentioned that it is necessary to make full use of the spatiotemporal correlation of traffic state variables. The states are aggregated and then determine when, where and how much occasional traffic jams occur. They demonstrate that taking full advantage of the spatiotemporal correlation of traffic states can improve the detection accuracy. To sum up, as far as the general trend is concerned, combining multiple traffic state data and making full use of their spatiotemporal correlations can greatly improve the detection accuracy. However, the second-order matrix can only utilize two modal characteristics at the same time, and the traffic data is a natural multi-modal data. To make full use of its multi-modal characteristics, the data model construction still needs to be improved. In terms of incomplete traffic data processing, the inventors demonstrated in 2013 that a tensor model can take full advantage of the potential multimodal nature of traffic data and thus outperform many matrix-based models. Therefore, the inventors consider combining the low-rank characteristics of traffic data and the advantages of tensor models to construct a robust tensor decomposition model to extract low-rank and sparse parts of traffic data, thereby obtaining spatiotemporal location information of occasional traffic congestion.

After the selection of the model is completed, how to make full use of various traffic data is also very important for congestion detection. Most of the existing studies are based on a single traffic state variable, but the impact of occasional traffic on the traffic state is many , and the changes of different states correspond to different types of congestion. Therefore, to improve the detection accuracy of occasional traffic congestion, multiple traffic state variables should be combined. Therefore, on the basis of constructing a robust decomposition model of traffic tensor data, the inventor considers coupling multiple traffic state variables to realize the coupling decomposition of multiple traffic variables.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide an occasional traffic congestion detection method based on coupled robust tensor decomposition, so as to solve the problems existing in the traffic congestion detection in the prior art.

To achieve the above object, the present invention adopts the following technical solutions:

A method for sporadic traffic congestion detection based on coupled robust tensor decomposition, the method comprising:

Collect data on various traffic variables;

constructing a tensor model according to the plurality of traffic variable data;

Filling and preprocessing the missing data in the multiple traffic variable data;

Build a coupled robust decomposition model of multiple traffic variable data;

Occasional traffic jam detection is performed according to the coupled robust decomposition model.

Further, constructing a tensor model according to the multiple traffic variable data includes:

According to the characteristics of traffic state parameters, a tensor model of the following functions is constructed:

Among them, N represents the number of detectors, M represents the number of weeks, W represents 7 days in a week or 5 working days, and T represents time.

Further, the construction of a coupled robust decomposition model for multiple traffic variable data includes:

The coupled robust sub-tensor solution is performed on various traffic state variable data; the sparse part obtained by the coupled robust tensor decomposition is identified as the spatiotemporal location and size information of sporadic traffic occurrences.

Further, the function of coupling robust sub-tensor solutions to various traffic state variable data is:

表示量或者观测数据构建成的张量模型，

表示低秩部分；

表示稀疏部分，ε_i表示数据整体白噪声，其中

为三种数据共有的稀疏分布；in,

Represents a tensor model constructed from quantity or observation data,

represents the low-rank part;

represents the sparse part, ε _i represents the overall white noise of the data, where

is a sparse distribution common to the three types of data;

1＜r＜R；1＜n＜N，其中R表示CP秩，N表示张量的阶数；Further, the matrix factor obtained by CP decomposition of the low-rank part obeys a normal distribution

1<r<R;1<n<N, where R represents the CP rank and N represents the order of the tensor;

的每个元素

取值为0-1，服从伯努利分布

其共轭先验

服从Beta分布

稀疏部分

的每个元素服从正态分布

白噪声服从正态分布

Further, the sparse part

each element of

The value is 0-1, obeying Bernoulli distribution

its conjugate prior

obey Beta distribution

sparse part

Each element of is normally distributed

White noise follows a normal distribution

标识为偶发性交通拥堵发生的时空位置，稀疏部分

表示与

对应时空位置的偶发性交通拥堵大小。Further, the sparse part obtained by decomposing the coupling robustness tensor is identified as spatiotemporal location and size information of occasional traffic occurrences, wherein the sparse part is

Identifies spatiotemporal locations where occasional traffic jams occur, sparse parts

means with

The size of the occasional traffic jam corresponding to the spatiotemporal location.

Beneficial effects:

The invention first collects various traffic variable data; then according to the characteristics of various traffic data, it is constructed into a tensor model of the same size and order; then the missing data is filled and preprocessed; Rod tensor decomposition model, a Bayesian robust tensor decomposition model of adaptive rank is proposed, and the common sparse distribution of different types of traffic data is described from the perspective of probability, so as to construct a coupled robust decomposition of various traffic data Finally, a solution method is designed to achieve high-precision and rapid detection of occasional traffic congestion.

Description of drawings

1 is a general flow chart of a method for detecting occasional traffic jams in the prior art center;

Fig. 2 is the sporadic congestion detection process based on coupled Bayesian robust tensor decomposition proposed by the present invention;

3 is a reference diagram of a tensor model of the same size of various traffic variable data in an embodiment of the present invention;

4 is a general form reference diagram of a robust tensor decomposition model in an embodiment of the present invention;

FIG. 5 is a reference diagram of a model framework for accidental traffic congestion detection based on coupled robust tensor decomposition in an embodiment of the present invention.

Detailed ways

The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

In view of the shortcomings of the traditional method, the present invention proposes an occasional congestion detection method based on coupled Bayesian robust tensor decomposition. The technical route is shown in Figure 2. First, various traffic variable data (traffic flow, road density and speed, etc.); then according to the characteristics of various traffic data, it is constructed into a tensor model of the same size and order; then the missing data is filled and preprocessed; then a Bayesian robust tensor is constructed Decomposition model, a Bayesian robust tensor decomposition model of adaptive rank is proposed, and the common sparse distribution of different types of traffic data is described from the perspective of probability, so as to construct a coupled robust decomposition model of various traffic data; the final design A solution method to achieve high-precision and rapid detection of occasional traffic congestion.

(1) Traffic data collection

The present invention obtains traffic flow data, road density data and speed data from April 1, 2014 to March 2, 2015 from the California highway database PeMS ( http://pems.dot.ca.gov/) , The time interval is 5min, and the road density is the time density.

(2) Construction of tensor models of various traffic state variable data

According to the characteristics of traffic state parameters, its tensor model can generally be built as

Among them, N represents the number of detectors, M represents the number of weeks, W represents 7 days in a week or 5 working days, and T represents time. The fourth-order tensor models of the three types of traffic data here are shown in Figure 3. When only the data of one detector is considered, the detector dimension is 1, which means it becomes a third-order tensor model. Such tensor models are able to fully characterize and exploit the spatiotemporal multimodal correlation of traffic data.

(3) Preprocessing of missing data

Due to force majeure factors such as weather and facilities, the collected traffic data is usually incomplete, which affects subsequent data analysis. In response to this problem, the inventor uses the traffic data tensor filling method proposed in 2013 to preprocess the missing data.

(4) Coupling robust sub-tensor solutions for various traffic state variable data

After constructing various traffic data into tensor models, it is necessary to construct suitable robust tensor decomposition models and methods to obtain their low-rank (normal traffic data distribution) and sparse parts (sporadic traffic degree distribution), and at the same time The coupling distribution among various data is also extracted.

First, a suitable robust tensor decomposition model that generalizes to coupling needs to be studied. The general form of the robust tensor decomposition model is shown in Figure 4. Observed tensor represents the tensor model constructed from measurement or observation data, lowrank represents the low-rank part obtained by decomposition, Sparse represents the sparse part obtained by decomposition, and Noise represents normal white noise distribution. In order to make the model more interpretable, the inventor designs and adopts a Bayesian robust tensor decomposition model to describe the meaning of traffic data and its decomposed parts from a probabilistic point of view. And an adaptive rank increase and decrease method is proposed to quickly obtain the rank of the low rank part.

Then, the coupling relationship between multiple traffic state data is considered, that is, when occasional traffic congestion occurs, different traffic state data share the same anomalous sparse distribution. On the basis of robust tensor decomposition, the inventor expresses the sparse part as the element product of a 0-1 identification tensor and a normal distribution tensor, and further builds a coupled robust tensor decomposition model such as the formula ( 1) shown:

表示量或者观测数据构建成的张量模型，

表示低秩部分(正常分布)，通过低秩分布采用CP(CANDECOMP/PARAFAC)分解得到；

表示稀疏部分(偶发性交通拥堵)，ε_i表示数据整体白噪声，其中

为三种数据共有的稀疏分布。然后对各部分的概率分布进行先验赋值和后验求解。对于每一种数据，根据自然规律，假设:

Represents a tensor model constructed from quantity or observation data,

Represents the low-rank part (normal distribution), which is obtained by decomposing the low-rank distribution using CP (CANDECOMP/PARAFAC);

represents the sparse part (occasional traffic congestion), ε _i represents the overall white noise of the data, where

is a sparse distribution common to all three types of data. Then, the probability distribution of each part is assigned a priori and solved a posteriori. For each kind of data, according to the laws of nature, suppose:

1＜r＜R；1＜n＜N，其中R表示CP秩，N表示张量的阶数；

的每个元素

取值为0-1，服从伯努利分布

其共轭先验

服从Beta分布

的每个元素服从正态分布

白噪声服从正态分布

其他超参数都服从正态分布的共轭Gamma分布：v～Γ(c₀,d₀)，γ～Γ(e₀,f₀)。值得注意的是，为了提取数据的低秩特性，本发明采用多伽马过程来描述CP分解模型中的超对角参数λ_r，即

δ_l～Γ(a₀,1),a₀＞1,r＝1,…,R；当秩r增大时，τ_r的秩迅速增长，

逐渐趋近于0，从而使得λ_r趋近于0，进而保证了数据的低秩特点。The matrix factors obtained by the CP decomposition of the low-rank part obey the normal distribution

1<r<R;1<n<N, where R represents the CP rank and N represents the order of the tensor;

each element of

The value is 0-1, obeying Bernoulli distribution

its conjugate prior

obey Beta distribution

Each element of is normally distributed

White noise follows a normal distribution

All other hyperparameters obey the conjugate Gamma distribution of the normal distribution: v～Γ(c ₀ ,d ₀ ), γ～Γ(e ₀ ,f ₀ ). It is worth noting that in order to extract the low-rank characteristics of the data, the present invention adopts a multi-gamma process to describe the superdiagonal parameter λ _r in the CP decomposition model, namely

δ _l ~Γ(a ₀ ,1), a ₀ >1, r=1,...,R; when the rank r increases, the rank of τ _r increases rapidly,

Gradually approaching 0, so that λ _r approaching 0, thus ensuring the low-rank characteristics of the data.

After constructing a robust tensor decomposition model, all that needs to be done is to solve the posterior distribution of each part; first, the overall joint probability is expressed as shown in equation (2), and then the linear operation properties of the Gaussian distribution and The posterior distribution is obtained by the Gibbs sampling method, and its update process is shown in equations (4)-(10).

The posterior posterior distribution of each variable is:

(i)

(ii)

(iii)

(iv)

(v)

(vi)

(vii)

(viii)

Among them, λ _r can be set to increase or decrease the threshold. When all λ _r are greater than a certain threshold, the rank will be automatically increased; on the contrary, when one of λ _r is smaller than the threshold, the rank will be reduced by 1, and the rank corresponding to λ _r will be discarded. quantity. Therefore, the adaptive rank can be obtained on the basis of ensuring the low rank characteristic.

(5) Identify the sparse part obtained by decomposing the coupled robustness tensor as the spatiotemporal location and size information of sporadic traffic occurrences

标识为偶发性交通拥堵发生的时空位置，利用稀疏部分

表示与

对应时空位置的偶发性交通拥堵大小。As shown in Figure 5, taking the coupling of two kinds of traffic data as an example, the method of the present invention can extract the low-rank part and the sparse part of the traffic data.

Identify spatiotemporal locations where occasional traffic jams occur, using sparse parts

means with

The size of the occasional traffic jam corresponding to the spatiotemporal location.

(6) Experimental effect

On the basis of model establishment, the inventor conducts experiments on real data, takes the historical records of aperiodic traffic congestion events given by the traffic management department as the true value, and compares the number of accidental congestion events correctly detected by the detection method with the true value The ratio of numbers is the standard of detection accuracy. The traditional congestion detection method Standard Normal Deviates (SND), the new method Coupled BPRCA method in 2014, the new method RSTD method in 2015 and the method proposed by the present invention are compared experimentally. The parameter adjustment standard of the contrast experiment is to make each The false detection rate (false positive: the detection method detects occasional congestion, but no occasional congestion event occurs in the true value) is basically equal. When the data of all days in a total of 48 weeks is used by one detector, the detection results are shown in Table 1. The method of the present invention has achieved good results, followed by BPRCA. When considering two adjacent detectors, the detection results are shown in Table 2, and the overall detection rate has decreased, but the tensor-based methods (RSTD and coupled BRPTF) have achieved better results; the effect of the BPRCA method is significantly better than drops a lot with one detector. Analysis of the reasons shows that there is an on-ramp between the two detectors (some drivers may choose to exit the expressway when an abnormal event occurs), which reduces the influence of the time and location of traffic congestion between the two adjacent detectors.

A detailed analysis of the 7-day laboratory results found that for BRPCA, RSTD, and the method of the present invention, especially the latter two, most of the undetected congestion was distributed on weekends; The traffic data between them are more similar in terms of day patterns, that is, the normal distribution rank is lower, so if working days are considered separately, the detection accuracy may be higher, and for commuters, the most important consideration is the anomaly of working days. congestion. In order to prove whether the conjecture is correct, the present invention conducts experiments on the 48-week data set with only working days, and the results are shown in Tables 3 and 4. Table 3 is the result of considering only one detector, and Table 4 is the result of considering adjacent detectors. From the results of the two detectors, it can be seen that the detection accuracy of the tensor-based method is greatly improved compared with 7 days. The advantages of the method proposed in the present invention are obvious.

Table 1. Detection accuracy of different methods using one detector for all days in 48 weeks

Table 2 The detection accuracy of different methods using the data of all days in 48 weeks of two adjacent detectors

Table 3. Detection accuracy of different methods using one detector for 48 working days

Table 4. The detection accuracy of different methods using the 48-week working day data of two adjacent detectors

Professionals should be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. Interchangeability, the above description has generally described the components and steps of each example in terms of function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of the present invention.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be implemented in hardware, a software module executed by a processor, or a combination of the two. A software module can be placed in random access memory (RAM), internal memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other in the technical field. in any other known form of storage medium.

The above specific embodiments further describe the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (7)

1. A coupling robustness tensor decomposition-based sporadic traffic congestion detection method is characterized by comprising the following steps:

collecting various traffic variable data;

constructing a tensor model according to the multiple traffic variable data;

performing filling preprocessing on lost data in the multiple traffic variable data;

constructing a coupled robustness decomposition model of various traffic variable data;

and carrying out accidental communication congestion detection according to the coupling robustness decomposition model.

2. The method for detecting sporadic traffic congestion based on coupled robust tensor decomposition as recited in claim 1, wherein the constructing a tensor model from the plurality of traffic variable data comprises:

according to the characteristics of the traffic state parameters, a tensor model of the following functions is constructed:

where N denotes the number of detectors, M denotes the number of weeks, W denotes 7 days in a week or 5 days of a weekday, and T denotes the time of day.

3. The method for detecting the sporadic traffic congestion based on the coupled robustness tensor decomposition as recited in claim 1, wherein the constructing the coupled robustness decomposition model of the multiple traffic variable data comprises:

carrying out coupling robustness tensor solution on various traffic state variable data; and identifying the sparse part obtained by decomposing the coupling robustness tensor as space-time position and size information of the sporadic communication.

4. The method for detecting the sporadic traffic congestion based on the coupled robustness tensor decomposition as recited in claim 3, wherein the function of the coupled robustness tensor decomposition on the multiple traffic state variable data is as follows:

wherein,

a tensor model constructed of the quantities of representation or observation data,

represents a low rank portion;

a sparse portion is represented by a graph of,_irepresenting white noise of the data as a whole, wherein

Is a sparse distribution common to all three data.

5. The method of claim 4, wherein the matrix factor obtained by CP decomposition in the low-rank part obeys normal distribution

R is more than 1 and less than R; 1 < N < N, where R denotes CP rank and N denotes the order of the tensor.

6. The method of claim 4, wherein the sparse portion is a portion of the coupling robustness tensor decomposition-based sporadic traffic congestion detection method

Each element of (1)

A value of 0-1, obeying Bernoulli distribution

Its conjugate prior

Obeying Beta distribution

Sparse part

Is subject to a normal distribution

White noise obeys normal distribution

7. The method as claimed in claim 3, wherein the sparse part obtained by decomposing the coupling robustness tensor is identified as the space-time position and size of the occurrence of the accidental trafficInformation, wherein the sparse portion

Marked as the spatio-temporal location of the occurrence of sporadic traffic congestion, sparse part

Is shown and

and the size of the sporadic traffic jam corresponding to the space-time position.

Images