CN115796017A

CN115796017A - Interpretable traffic cognition method based on fuzzy theory

Info

Publication number: CN115796017A
Application number: CN202211453961.9A
Authority: CN
Inventors: 安吉尧; 钱欣姣; 赵谨; 刘清钦; 陈佳丽
Original assignee: Hunan University
Current assignee: Hunan University
Priority date: 2022-11-21
Filing date: 2022-11-21
Publication date: 2023-03-14

Abstract

The invention relates to an interpretable traffic cognition method based on a fuzzy theory, and belongs to the technical field of artificial intelligence. The traffic scene real-time traffic data is organized into an external attribute feature matrix, a traffic cognition feature matrix and an adjacency matrix, the preprocessed external attribute feature matrix is sent into a fuzzy reasoning mechanism, the attribute influence feature matrix is output after feature calculation, the matrix, the traffic cognition feature matrix and the adjacency matrix are input into a graph convolution neural network together, then the matrix, the traffic cognition feature matrix and the adjacency matrix are input into a time characteristic capture network based on a gating cycle unit, and finally a prediction result is output and used for attribute influence fuzzy classification and traffic cognition work. The method can efficiently extract the time and space dependence characteristics among a plurality of roads in the traffic scene, has higher traffic data cognition performance and interpretability, and solves the problems of low transparency and poor interpretability of a deep neural network and insufficient consideration of external attribute characteristics of the traffic scene in the traffic cognition process.

Description

Interpretable traffic cognition method based on fuzzy theory

Technical Field

The invention relates to an interpretable traffic cognition method based on a fuzzy theory, and belongs to the technical field of intelligent traffic, fuzzy logic and artificial intelligence.

Background

In recent years, intelligent transportation systems are gradually changing from perception intelligence to cognition intelligence, and intelligent transportation cognition is a type of research for providing decision-making help for the intelligent transportation systems and aims to make better trip decisions, reduce traffic jam and improve traffic operation efficiency. The method for recognizing the intelligent traffic comprises the steps of traffic perception result decision, traffic flow prediction, traffic speed prediction, travel route planning, driving behavior analysis, traffic decision support and the like. The intelligent traffic system realizes intelligent sensing by using roadside sensors (such as induction loops, radars, cameras and the like), acquires traffic data, further analyzes and infers flow data, traffic data and traffic conditions, and researchers can improve the cognitive ability of the intelligent traffic system and finally relieve the traffic pressure of the intelligent traffic system. Although a large number of researchers have focused on the research work of the intelligent transportation system, the interpretability problem of the intelligent transportation cognition and the cognitive method thereof is still pending. Traditional traffic cognition methods are realized based on statistical models, the models utilize traffic data laws to perform traffic cognition work, but the statistical models excessively simplify complex and random traffic scenes, so that the model cognition performance is poor. With the development of deep learning, researchers began to use deep neural networks such as LSTM, convolutional neural networks to learn traffic, but such models only consider the temporal or spatial characteristics of traffic data, and not the spatiotemporal characteristics of traffic data at the same time. Researchers then try to combine the time capture model and the space capture model to capture the spatiotemporal characteristics of the traffic data, but due to the black box characteristics of the model, the model has low interpretability, and the application of the model in the field with high requirements on safety is limited. The introduction of the visualization method improves the interpretability of the deep learning model, but the method provides lower interpretability and does not improve the interpretability of the model. Therefore, the interpretable traffic cognition method based on the fuzzy theory can effectively improve the integral interpretability of the model and solve the problem that the processing process of the deep learning model is invisible to developers.

In addition, after a fuzzy inference mechanism is introduced, some external characteristic attributes in a traffic scene, such as weather, holidays, time periods, traffic congestion indexes and the like, are considered in the traffic cognition process, so that how to express the influence of the external characteristic attributes on traffic data and how to capture the interaction relation among different characteristic attributes are also innovation points of the patent. According to the method, some fuzzy rules are established according to expert experience, external attribute features act on a traffic cognition process based on guidance of the expert experience, and the problem of low transparency of a deep learning model is solved. The traditional traffic awareness method pays attention to the influence of historical traffic data on current traffic data, but the influence can be greatly different in different traffic scenes, so that the traffic awareness process considering the influence of external attributes is more accurate than that paying attention to the historical traffic data singly. The method is based on the fusion of different external attribute characteristics, further analyzes the influence of the attribute characteristics on the traffic data, fully excavates the action relation of the attribute characteristics on the traffic data, and utilizes the decision coefficient to realize more effective traffic cognition. And finally, according to the fuzziness and randomness of the influence of the attribute characteristics and the subjectivity of classification standards, introducing a fuzzy set theory, further constructing a fuzzy membership function, carrying out fuzzy classification on the extracted influence characteristics, and constructing three types of fuzzy sets with small, medium and large influence.

Disclosure of Invention

In order to solve the problems, the invention provides an interpretable traffic cognition method based on a fuzzy theory. The processing procedure of the interpretable method can be presented to a developer in a simple and easy form, and is proposed by researchers in the deep learning field aiming at the black box characteristic of the deep learning model, and the interpretability of the deep learning model is expected to be improved by some methods. The method is realized by embedding a fuzzy inference mechanism in a deep learning model, and the fuzzy rule in the fuzzy inference mechanism can be directly expressed by a meaning, so that the method is called to be interpretable.

The fuzzy inference mechanism designed by the invention can well capture the influence of the attribute characteristics on traffic data and the action relationship among different attribute characteristics, wherein the constructed fuzzy rule improves the interpretability of the whole model. The problem of capturing the space-time characteristics of the traffic data is solved by using a graph convolution neural model and a gating circulation unit in the traffic data processing part. And constructing attribute influence characteristics by using the characteristics that different attribute characteristics have different influences on traffic data, and introducing a fuzzy theory to perform fuzzy classification on the influence characteristics.

The invention is realized by the following technical scheme, which comprises the following steps:

step1, an input data processing algorithm combining a fuzzy inference mechanism and a graph convolution neural network model:

the method comprises the steps of describing historical data of all roads of a traffic scene, describing an adjacency matrix of interactive features among the roads, describing a traffic jam index of external features, a time period, a holiday, weather, traffic flow change and an interpretable traffic data cognitive algorithm;

step2, the structure design of the graph convolution neural network model combined with the fuzzy inference mechanism comprises three parts:

the first part is a fuzzy reasoning mechanism based on a fuzzy theory, which consists of a fuzzy membership function, a fuzzy rule and a defuzzification function and can effectively extract the influence of external attributes on traffic data; a traffic cognition feature tensor composed of the traffic data tensor and the traffic congestion index tensor is combined with the output of the first part to serve as input data of a second part of the model;

the second part is a graph convolution neural network which can directly carry out graph convolution operation on the traffic cognition related tensor; the graph convolution neural network can more efficiently extract the spatial correlation between the road traffic data; the output of the second part is used as a spatial feature and is input into the network of the third part;

the third part is a deep neural network based on a gating cycle unit, the network further processes the spatial features to extract the temporal features, and the temporal features are finally processed into predicted traffic data and used for subsequent traffic cognitive work;

step3, training a graph convolution neural network model;

setting corresponding model parameters and a training environment, and training the model;

step 4, performing various traffic cognition experiments based on the trained model;

step 5, constructing attribute influence description characteristics based on the traffic cognition experiment result;

and 6, carrying out attribute influence fuzzy classification based on the influence description characteristics.

The input data in step1 is specifically defined as follows:

step 1.1, historical traffic data and prediction data of all roads of the predicted traffic scene:

road historical data matrix X with traffic scene _traffic Consists of traffic data for τ historical time steps:

wherein the traffic data for each time step is

Which consists of traffic data of N roads in a traffic scene at the time t, wherein N is the number of the roads,

denotes S ^t The vector is an Nx 1 vector, and elements in the vector are real numbers;

suppose predicted data matrix X 'of traffic scene' _traffic The traffic data for T predicted time steps consists of:

the traffic data format of each time step is the same as the historical traffic data, wherein tau is the historical time step, T is the time of predictionThe length of the interval step, N is the number of the roads,

the size of the representation matrix is T multiplied by N, wherein elements are real numbers;

the problem formatting for traffic data prediction is represented as: x' _traffic ＝f(X _traffic )；

Traffic data generally refers to data describing a traffic scene and having spatiotemporal characteristics, including traffic speed, traffic flow, and traffic congestion conditions;

step 1.2, describing an adjacency matrix of the interactive features between roads:

an undirected graph G = { V, E } is used for describing the connection condition of roads in a traffic scene, and a node set V is represented as V = { V } ₁ ,v ₂ ,…,v _N N is the number of roads in the traffic scene;

when the roads are communicated with each other, the edges representing the road interaction relationship are connected; thus the set of edges E expressed as E = { v _i v _j I is more than or equal to 1 and less than or equal to N, j is more than or equal to 1 and less than or equal to N, wherein v is _i v _j Representing the communication between the node i and the node j;

the element level of adjacency matrix a is defined as:

the formula shows that when the node i is communicated with the node j, the weight of the corresponding edge is 1, otherwise, the weight is 0; a. The _ij The size of (1) is NxN, wherein N is the number of roads;

step 1.3, describing a feature matrix of external attributes:

setting a feature matrix X describing external attributes _feature Including a traffic congestion index matrix X _TTI Time period matrix X _time Vacation matrix X _holiday Weather matrix X _weather Traffic flow change matrix X _flowChange ：

X _feature ∈{X _TTI ,X _time ,X _holiday ,X _weather ,X _flowChange },

Wherein the traffic congestion index matrix

Calculated from historical traffic speeds, represents the degree of congestion of the road, tau represents the length of the time step, N represents the number of roads in the traffic scene,

the size of the expression matrix is tau multiplied by N, all elements are real numbers, and the calculation formula of the traffic congestion index is as follows:

wherein V _free For free passage speed, V _current Is the current speed;

time period setting matrix

The element level of the matrix is defined as:

wherein [ t _n ,t _n+1 ]Is a continuous time section, the traffic data in the time section has similar change state, in the actual calculation process, different time sections are represented by different positive integers, tau represents the time step, N represents the number of roads,

the matrix size is represented as tau multiplied by N, wherein elements are real numbers, ij represents that the element is positioned at the ith row and the jth column of the matrix, and N represents that 24 hours of a day is divided into N time periods;

vacation setting matrix

The element level of the matrix is defined as:

this formula indicates that when the date is a weekend or a statutory holiday,

is 1, otherwise is 0, tau represents a time step, N represents the number of links,

the size of the matrix is represented as tau multiplied by N, wherein elements are real numbers, and ij represents that the element is positioned in the ith row and the jth column of the matrix;

weather matrix

The element level of the matrix is defined as:

the weather conditions contained in the weather matrix are represented as:

X _{weatherCondition} ∈{sunny,cloudy,light rain,medium rain,heavy rain}

different weather conditions are expressed by different positive integers, the influence degree of the same weather condition on traffic data is similar, tau represents a time step, N represents a road length,

setting traffic flow change matrix

The element level of the matrix is defined as:

wherein [ flow _k ,flow _k+1 ]Is a continuous flow change interval, the influence degree of the flow change value in the interval on the traffic data is similar, tau represents a time step, N represents the number of roads,

the size of the matrix is represented as tau multiplied by N, wherein elements are real numbers, ij represents that the elements are positioned in the ith row and the jth column of the matrix, and k represents that the traffic flow change data are divided into k intervals;

step 1.4, an interpretable traffic data cognitive algorithm:

the processing process of the interpretable finger model to the data is presented to developers in a simple and understandable form, and is specifically represented by that fuzzy rules in a fuzzy inference mechanism are represented by intuitive IF-THEN statements, wherein the definition of four important fuzzy rules is as follows:

fuzzy rule 1: IF flowChange is zero AND time is seven the effect is small

Fuzzy rule 2: IF flowChange is third AND time is seven the effect is large

Fuzzy rule 3: IF holitray is a holitray AND time is the same that the effect is super large ge

Fuzzy rule 4: IF weather is fog same kind of effect is middle

The fuzzy rule 1 indicates that the influence on the traffic data is small when the traffic flow change is around 0 and the time is between seven and nine points, the fuzzy rule 2 indicates that the influence on the traffic data is large when the traffic flow change is around 30 and the time is between seven and nine points, the fuzzy rule 3 indicates that the influence on the traffic data is very large when the date is holiday and the time is between nine and eleven points, and the fuzzy rule 4 indicates that the influence on the traffic data is medium when the weather is foggy. The numerical words, adjectives and nouns in the fuzzy rule, such as zero, seven, small, holiay, foggy and the like, represent a fuzzy interval after fuzzy division, and numerical values in the interval have similar characteristics; the high interpretability of the fuzzy rule in the fuzzy inference system solves the problem of black boxes in a deep learning model, namely the problems of invisible processing, low transparency and poor safety.

In the step2, a fuzzy inference mechanism based on a fuzzy theory can process multidimensional external attribute tensor, and can explainably extract the influence of the external attribute on traffic data; the graph convolution neural network can process three-dimensional tensors, can more efficiently extract spatial correlation among different roads in a traffic scene, and extracts time correlation among traffic data by combining gate control cycle units;

input matrix with fuzzy inference mechanism

From the external attribute matrix X _time ，X _holiday ，X _weather ，

X _flowChange Composition, which is defined as follows:

X _fuzzy ＝ω ₁ ·X _time +ω ₂ ·X _holiday +ω ₃ ·X _weather +ω ₄ ·X _flowChange

wherein omega ₁ ,ω ₂ ,ω ₃ ,ω ₄ The weights of the time period matrix, the holiday matrix, the weather matrix and the traffic change matrix respectively represent the influence degree of different external attributes on the traffic cognitive data, the larger the weight is, the larger the influence on the traffic cognitive data is, the tau represents the time step length, the N represents the number of roads,

the size of the expression matrix is tau multiplied by N, wherein elements are real numbers;

output matrix of fuzzy inference system

It is defined as follows:

where m is the number of fuzzy rules, k is the number of Gaussian membership functions, X _fuzzy Is a three-dimensional feature matrix, mu, composed of an external attribute matrix _i Is the width, σ, of the ith Gaussian membership function _i Is the center of the ith Gaussian membership function; x _Effect The element in (2) is a signed number, the larger the absolute value of the numerical value is, the larger the influence of the external attribute on the traffic cognitive data is, the positive number represents the influence on the traffic cognitive data when the traffic flow increases, and the negative number represents the influence on the traffic cognitive data when the traffic flow decreases;

indicating the influence of external attributes on traffic awareness data at time t,

setting attribute matrix related to traffic cognition

From X _traffic And X _TTI Composition, which is defined as follows: x _cognition ＝ω ₁ ·X _traffic +ω ₂ ·X _TTI

Wherein ω is ₁ ,ω ₂ Respectively traffic data matrix X _traffic And traffic congestion index matrix X _TTI The weight of (2) represents the influence degree of the historical traffic data and the traffic jam index on the traffic cognition data, the larger the weight is, the larger the influence on the traffic cognition data is,

output of layer I in graph convolution neural network

Is defined as:

C _l+1 ＝σ(L _sym C _l W _l )

C ₀ ＝X，

wherein L is _sym Is normalized Laplacian, D is degree matrix calculated from adjacency matrix, C _l+1 Is the output of the l-th layer of convolution, A is an adjacency matrix describing the road node relationship, X is a feature matrix describing the traffic features and external attributes, W _l Is a weight matrix of the l-th layer, the used activation function is a nonlinear activation function Relu, tau represents a time step, N represents the number of road nodes,

after the processing of the graph convolution neural network, a spatial feature matrix is generated, the spatial feature capture of the traffic cognitive data is completed at the moment, and then the spatial feature capture is input into a gate control circulation unit, and the gate control circulation unit resets a gate at a time t

Updating door

And candidate hidden states

Is defined as follows:

R ^t ＝σ(X ^t W _xr +H ^t-1 W _hr +b _r )

Z ^t ＝σ(X ^t W _xz +H ^t-1 W _hz +b _z )

where h is the number of hidden units,

is the input feature at time t, P is the number of extrinsic feature attributes,

is the hidden state at the last time step,

are two weight matrices for the reset gate,

is a bias matrix of the reset gates and,

is to update the two weight matrices of the gate,

is the bias matrix of the update gate, the activation function used by the reset gate and the update gate is the sigma d function,

are two weight matrices for the candidate hidden states,

is a bias matrix for the candidate hidden state, the activation function used by the candidate hidden state is a tanh function,

the matrix is used for representing the size of each matrix, wherein elements are real numbers;

the adjacency matrix and the feature matrix generate predicted traffic data for subsequent traffic cognitive work after passing through the graph convolution layer and the gating circulation unit.

The training data in the step3 are set as follows:

training optimizer Adam, learning rate 0.001, number of gated round robin units 64, batch size bat hsize =32, time step 12 for historical traffic awareness data, and loss function defined as follows:

wherein t is _f Is to predict time step, Y _pred Is predicted traffic data, Y _true Is real traffic data;

using a public data set Los-loop and a real taxi data set SZBZ near the Shenzhen north station as a data set for model training, and dividing the data set into a training set and a testing set according to a certain proportion; and packaging data by using a user-defined Dataloader object, intelligently and iteratively processing the input characteristic matrix, and outputting a data tensor with a corresponding format to the model for training.

The multi-class traffic cognition experiment in the step 4 is set as follows:

the first type of traffic cognition experiment is to verify the influence of different external characteristic attributes on traffic data, and comprises four different experimental settings: the first is that the input data only comprises historical traffic data; the second adds vacation data and weather data on the basis of the first; the third is to add time period data and traffic flow change data on the basis of the second; fourthly, adding the traffic jam index data on the basis of the third method;

the second type of traffic cognition experiment is to verify the cognitive performance of the method in different time periods, and comprises three different experiment settings: the first is to predict traffic data for the next 10 minutes; the second is to predict traffic data for 30 minutes into the future; the third is to predict traffic data 60 minutes in the future;

the prediction results under the experimental settings are used for subsequent traffic cognition work.

In the step 5, attribute influence description characteristics are constructed for the traffic prediction data, and attribute intersection is carried outThe influence of the data is directly reflected on the influence of the method on the prediction accuracy of the traffic data, and the prediction data under the experimental settings are obtained through different experimental settings in the step 4; determining the coefficient R ² Used to measure the predicted performance under different experimental settings, it is defined as follows:

wherein Y is _t In order to be the real traffic data,

in order to be able to predict the traffic data,

the average value of the real traffic data is obtained, and n is the total number of the traffic data; r is ² The value of (2) is between 0 and 1, the closer to 1, the better the regression fitting effect is, the prediction performance of the method under different experimental settings is shown, and the subsequent fuzzy classification work is also facilitated.

In the step6, the description characteristics are subjected to attribute influence fuzzy classification, and a decision coefficient R is used ² Describing the influence degree of the attribute characteristics on traffic data prediction, fully considering the attribute influence characteristics under the background with a traffic scene, introducing a fuzzy set theory due to the fuzziness and randomness of influence and the subjectivity of classification standards, further carrying out fuzzy classification on the obtained influence characteristics, and constructing three fuzzy sets with small influence, medium influence and large influence, R ² Describing the degree of membership to the fuzzy sets by a membership function mu; selecting a Gaussian membership function, and determining a threshold parameter lambda based on a heuristic method ₁ And λ ₂ 。

The invention has the beneficial effects that:

(1) The interpretable traffic cognition method based on the fuzzy theory can capture the influence of the attribute characteristics on traffic cognition and the interaction relation among different attribute characteristics by using a novel data processing fuzzy method, and solves the problem that the interaction relation among different characteristic attributes cannot be extracted by a traditional traffic cognition model.

(2) According to the invention, the fuzzy inference mechanism is embedded in the deep learning model, so that the defects of low interpretability and invisible processing process of the deep learning model are overcome.

(3) The attribute features are divided into external attribute features and traffic cognition related attribute features for processing, the influence of different types of attribute features on traffic cognition is fully considered, different processing methods are adopted for different types of attribute features, and the performance of extracting the influence of the attribute features by a model is improved.

(4) The graph convolution neural network and the gated cyclic network constructed by the method simultaneously consider the time characteristic and the space characteristic of the traffic cognitive data, and improve the cognitive performance of the model on the traffic data.

(5) The influence characteristics based on the traffic data constructed by the invention fully consider the relation between the traffic data and the environment, and more comprehensively depict the traffic cognition process. And introduces fuzzy sets to describe the fuzziness, randomness and subjectivity of attribute influence.

(6) The invention provides a training method of a specific deep neural network model, a construction process of a fuzzy rule in a fuzzy inference mechanism, relevant hyper-parameters and implementation details of an internal structure of the model. The accuracy and the efficiency of model training are ensured, and the phenomena of under-fitting and over-fitting in the model training process can be effectively avoided.

Drawings

FIG. 1 is a block diagram of the present invention.

FIG. 2 is a fuzzy inference mechanism for capturing attribute impact features.

FIG. 3 is a graph convolution neural network for capturing spatial features of traffic data.

FIG. 4 is a graph of attribute-influenced fuzzy set membership functions.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the accompanying drawings in combination with the embodiments. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention; furthermore, the technical features involved in the different embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

In the invention, the processing procedure of the interpretable method can be presented to developers in a simple and easy manner, and is provided by researchers in the deep learning field aiming at the black box characteristic of the deep learning model, and the interpretability of the deep learning model is expected to be improved by some methods. The method is realized by embedding a fuzzy inference mechanism in a deep learning model, and the fuzzy rule in the fuzzy inference mechanism can be directly expressed by a meaning, so that the method is called to be interpretable.

The invention comprises the following steps:

the method comprises the steps of describing historical data of all roads of a traffic scene, describing a adjacency matrix of interactive characteristics among the roads, describing a traffic congestion index of external characteristics, a time period, a holiday, weather, traffic flow change and an interpretable traffic data cognitive algorithm.

The input data in step1 is specifically defined as follows:

wherein the traffic data for each time step is

Which consists of traffic data for N roads in a traffic scene at time t,n is the number of the roads,

the traffic data format of each time step is the same as the historical traffic data, wherein tau is the historical time step, T is the predicted time step, N is the number of roads,

the problem formatting for traffic data prediction is represented as: x' _traffic ＝f(X _traff )；

an undirected graph G = { V, E } is used for describing the connection condition of roads in a traffic scene, and a node set V is represented as V =

{v ₁ ,v ₂ ,…,v _N N is the number of roads in the traffic scene;

when the roads are communicated with each other, the sides representing the road interaction relationship are connected; thus, the edge set E is denoted as E = { v = { v = _i v _j I is more than or equal to 1 and less than or equal to N, j is more than or equal to 1 and less than or equal to N, wherein v is _i v _j Representing the communication between the node i and the node j;

the element level of adjacency matrix a is defined as:

the formula shows that when the node i is communicated with the node j, the weight of the corresponding edge is 1, otherwise, the weight is 0; a. The _ij The size of (a) is NxN, wherein N is the number of roads;

step 1.3, describing a feature matrix of external attributes:

setting a feature matrix X describing external attributes _feature Including a traffic congestion index matrix X _TTI Time period matrix X _time The vacation matrix X _holiday Weather matrix X _weather Traffic flow change matrix X _flowChange ：

X _feature ∈{X _TTI ,X _time ,X _holiday ,X _weather ,X _flowChange },

Wherein the traffic congestion index matrix

Calculated from the historical traffic speed, represents the congestion degree of the road, tau represents the length of the time step, N represents the number of the roads in the traffic scene,

wherein V _free For free passage speed, V _current Is the current speed;

time period setting matrix

The element level of the matrix is defined as:

wherein [ t _n ,t _n+1 ]Is a continuous timeThe traffic data in the time period has similar change states, in the actual calculation process, different time periods are represented by different positive integers, tau represents a time step, N represents the number of roads,

the size of the matrix is represented as tau multiplied by N, wherein all elements are real numbers, ij represents that the elements are positioned in the ith row and the jth column of the matrix, and N represents that 24 hours of a day is divided into N time periods;

vacation setting matrix

The element level of the matrix is defined as:

this formula indicates that when the date is a weekend or a statutory holiday,

is 1, otherwise is 0, tau represents a time step, N represents the number of lanes,

weather matrix

The element level of the matrix is defined as:

the weather conditions contained in the weather matrix are represented as:

X _{weatherCondition} ∈{sunny,cloudy,light rain,medium rain,heavy rain}

for different weather conditionsThe influence degree of the same weather condition on the traffic data is similar, tau represents a time step, N represents a road length,

setting traffic flow change matrix

The element level of the matrix is defined as:

wherein [ flow _k ,flow _k+1 ]Is a continuous flow change interval, the influence degree of the flow change value in the interval on the traffic data is similar, tau represents the time step, N represents the number of roads,

step 1.4, an interpretable traffic data cognitive algorithm:

fuzzy rule 1: IF flowChange is zero AND time is seven the effect is small

Fuzzy rule 2: IF flowChange is three AND time is seven the effect is large

Fuzzy rule 3: IF holitray is a holitray AND time is the same that the effect is super large

Fuzzy rule 4: IF weather is fog same kind of effect is middle

The fuzzy rule 1 indicates that when the traffic flow change is around 0 and the time is between seven and nine points, the influence on the traffic data is small, the fuzzy rule 2 indicates that when the traffic flow change is around 30 and the time is between seven and nine points, the influence on the traffic data is large, the fuzzy rule 3 indicates that when the date is holiday and the time is between nine and eleven points, the influence on the traffic data is very large, and the fuzzy rule 4 indicates that when the weather is foggy, the influence on the traffic data is medium. The numerical words, adjectives and nouns in the fuzzy rule, such as zero, seven, small, holiay, foggy and the like, represent a fuzzy interval after fuzzy division, and numerical values in the interval have similar characteristics; the high interpretability of the fuzzy rule in the fuzzy inference system solves the problem of black boxes in a deep learning model, namely the problems of invisible processing, low transparency and poor safety.

the third part is a deep neural network based on a gating cycle unit, the network further processes the spatial features to extract the temporal features, and the spatial features are finally processed into predicted traffic data and used for subsequent traffic cognitive work;

in the step2, a fuzzy inference mechanism based on a fuzzy theory can process multidimensional external attribute tensor, and can explainably extract the influence of the external attribute on traffic data; the graph convolution neural network can process three-dimensional tensors, can more efficiently extract spatial correlation among different roads in a traffic scene, and extracts time correlation among traffic data by combining a gating circulation unit;

input matrix with fuzzy inference mechanism

From the external attribute matrix X _time ，X _holiday ，X _weather ，X _flowChange Composition, which is defined as follows:

the representation matrix size is tau x N, wherein elements are real numbers;

output matrix of fuzzy inference system

It is defined as follows:

where m is the number of fuzzy rules, k is the number of Gaussian membership functions, X _fuzzy Is a three-dimensional feature matrix, mu, composed of an external attribute matrix _i Is the width, σ, of the ith Gaussian membership function _i Is the center of the ith Gaussian membership function; x _Effect The element in (1) is a signed number, the larger the absolute value of the numerical value is, the larger the influence of the external attribute on the traffic cognition data is, the positive number represents the influence on the traffic cognition data when the traffic flow increases, and the negative number represents the influence on the traffic cognition data when the traffic flow decreases;

representing the influence of external attributes on traffic awareness data at time t,

setting attribute matrix related to traffic cognition

Wherein omega ₁ ,ω ₂ Respectively traffic data matrix X _traffic And a traffic congestion index matrix X _TTI The weight of (2) represents the influence degree of the historical traffic data and the traffic jam index on the traffic cognition data, the larger the weight is, the larger the influence on the traffic cognition data is,

the representation matrix size is tau x N, wherein elements are real numbers;

output of layer I in graph convolution neural network

Is defined as:

C _l+1 ＝σ(L _sym C _l W _l )

C ₀ ＝X，

Updating door

And candidate hidden states

Is defined as follows:

R ^t ＝σ(X ^t W _xr +H ^t-1 W _hr +b _r )

Z ^t ＝σ(X ^t W _xz +H ^t-1 W _hz +b _z )

where h is the number of hidden units,

is the hidden state at the last time step,

are two weight matrices for the reset gate,

is a bias matrix of the reset gates and,

is to update the two weight matrices of the gate,

is the bias matrix of the update gate, the activation function used by the reset gate and the update gate is the sigmoid function,

are two weight matrices for the candidate hidden states,

and the adjacency matrix and the characteristic matrix generate predicted traffic data for subsequent traffic cognitive work after passing through the graph convolution layer and the gating circulation unit.

Step3, training a graph convolution neural network model;

setting corresponding model parameters and a training environment, and training the model; the training data in the step3 are set as follows: training optimizer Adam, learning rate 0.001, number of gated loop units 64, batch size batchsize =32, time step 12 of historical traffic awareness data, and loss function defined as follows:

using a public data set Los-loop and a real taxi data set SZBZ near the Shenzhen north station as a data set for model training, and dividing the data set into a training set and a testing set according to a certain proportion; and packaging data by using a self-defined Dataloader object, intelligently and iteratively processing the input characteristic matrix, and outputting a data tensor of a corresponding format to the model for training.

And 4, performing various traffic cognition experiments based on the trained model.

The multi-class traffic cognition experiment in the step 4 is set as follows:

the first type of traffic cognition experiment is to verify the influence of different external characteristic attributes on traffic data, and includes four different experimental settings: the first is that the input data only comprises historical traffic data; the second kind adds vacation data and weather data on the basis of the first kind; the third is adding time section data and traffic flow change data on the basis of the second; the fourth is to add the traffic jam index data on the basis of the third;

the second type of traffic cognition experiment is to verify the cognitive performance of the method in different time periods, and comprises three different experiment settings: the first is to predict traffic data for the next 10 minutes; the second is to predict traffic data for 30 minutes in the future; the third is to predict traffic data 60 minutes in the future;

And 5, constructing attribute influence description characteristics based on the traffic cognition experiment results.

In the step 5, the attribute influence description is constructed on the traffic prediction dataThe influence of the attribute on the traffic data is directly reflected on the influence of the method on the prediction accuracy of the traffic data, and the prediction data under the experimental settings are obtained through different experimental settings in the step 4; determining the coefficient R ² Used to measure the predicted performance at different experimental settings, it is defined as follows:

wherein Y is _t In order to be the real traffic data,

in order to be able to predict the traffic data,

the average value of the real traffic data is obtained, and n is the total number of the traffic data; r ² The value of (2) is between 0 and 1, the closer to 1, the better the regression fitting effect is, the prediction performance of the method under different experimental settings is shown, and the subsequent fuzzy classification work is also facilitated.

In the step6, the description characteristics are subjected to attribute influence fuzzy classification, and a decision coefficient R is used ² Describing the influence degree of the attribute characteristics on traffic data prediction, fully considering the attribute influence characteristics under the background with a traffic scene, introducing a fuzzy set theory due to the fuzziness and randomness of influence and the subjectivity of classification standards, further carrying out fuzzy classification on the obtained influence characteristics, and constructing three fuzzy sets with small, medium and large influence, R ² Describing the degree of membership to the fuzzy sets by a membership function mu; selecting a Gaussian membership function and heuristically determining a threshold parameter lambda ₁ And λ ₂ 。

The invention will be described in more detail below with reference to the accompanying drawings. Like elements in the various figures are denoted by like reference numerals. For purposes of clarity, the various features of the drawings are not necessarily to scale; preferred embodiments of the present invention will be further described with reference to the accompanying drawings, in which figures 1 to 4 are shown:

step1: model input data processing

Before the model training, the raw data needs to be processed into a required form. Carrying out fuzzy classification processing on external attribute features related to traffic cognition to generate an external attribute feature matrix which is used as input of a fuzzy inference mechanism; and the road traffic data of the traffic cognition is observed and calculated to generate an adjacency matrix and a traffic cognition characteristic matrix, and then the adjacency matrix and the traffic cognition characteristic matrix are used as the input of the graph convolution neural network together with the attribute influence characteristic matrix output by the fuzzy inference mechanism. The following are introduced separately:

(1) Road traffic data

The original road traffic data is a two-dimensional table, each row is a data sampling point and comprises information such as a timestamp, a road identification serial number, vehicle average speed and traffic flow change, and slicing is carried out according to time (the length of each segment is a history step plus a prediction step). The historical traffic data of the road of the traffic scene is composed of tau road speed data of historical time steps:

wherein the velocity data at each time step is

It consists of the speeds of N roads in the traffic scene at time t. The predicted speed of the traffic scene consists of speed data for T predicted time steps:

the speed data format of each time step is the same as the historical speed data; the problem formatting for speed prediction is represented as: x' _traffic ＝f(X _traffic )。

The traffic data sampling frequency is 10 minutes, and a comparison experiment shows that the historical speed duration is 120 minutes, namely the prediction effect is better when 12 historical time steps are performed; and predicting the speed duration, and respectively trying 10 minutes, 30 minutes and 60 minutes, namely 1,3 and 6 predicted time steps. So in the experiment, τ =12, t = {1,3,6}, and the total step size of each data segment is 13, 15, and 18, respectively. N is the total number of roads used in the experiment, where N =12;

(2) Fuzzy classification process

Before processing the extrinsic attribute data into a required extrinsic attribute feature matrix, fuzzy classification processing needs to be performed on the data. Firstly, carrying out digital normalization processing on data in a semantic form, and then carrying out fuzzy classification according to a normalized digital matrix. The fuzzy classification processing method solves the problems of non-uniform external attribute data types, large data span, non-digitalization and the like by a skillful fuzzification strategy, and determines the generalization, specialization and aggregation properties of different types of data;

deterministic classification requires that an element be assigned to a particular class, i.e., the element belongs to either a class or not, and the element belongs to and only belongs to a class, with no ambiguity, similar to the concept of a collection in mathematics. Although it is possible to digitize non-digitized data and control the data within a relatively small range, the data in real life scenes is often not accurately described, and thus the characteristics of the data cannot be well described using deterministic classification.

The fuzzy classification process is to re-classify based on the probability that an element belongs to a certain class and control the classification result to be in the range of 0 to 1. A classification result of 0 indicates that the element definitely does not belong to the class, a classification result of 1 indicates that the element definitely belongs to the class, a value between 0 and 1 indicates that the element may belong to the class, and the larger the numerical value, the larger the probability of belonging to the class. In addition, the fuzzy classification processing can be performed by means of classification functions, and the classification functions include a fuzzy gaussian membership function, a fuzzy large value transformation function, a fuzzy small value transformation function, a fuzzy linear transformation function, a fuzzy MS large value transformation function, a fuzzy MS small value transformation function, a fuzzy adjacent value transformation function, and the like. Developers can select a proper classification function according to data characteristics and expert experience, wherein a fuzzy Gaussian membership function is used;

(3) Adjacency matrix, traffic cognition feature matrix and external attribute feature matrix

For an undirected graph G = { V, E } describing connectivity of roads in a traffic scene, a set of nodes V is denoted as V = { V = ₁ ,v ₂ ,…,v _N And N is the number of roads in the traffic scene. When roads are connected to each other, edges representing road interactions should be connected. The edge set E is represented as E = { v = { [ v ] _i v _j I is more than or equal to 1 and less than or equal to N, j is more than or equal to 1 and less than or equal to N, wherein v _i v _j Representing the communication between the node i and the node j;

the element level of adjacency matrix a is defined as:

the formula represents that when the node i is communicated with the node j, the weight of the corresponding edge is 1, otherwise, the weight is 0; a. The _ij The size of (1) is NxN, wherein N is the number of roads;

traffic congestion index matrix

Is calculated from the historical traffic speed and represents the congestion degree of the road.

The calculation formula of the traffic jam index is as follows:

wherein V _free For free passage speed, V _current Is the current speed;

traffic awareness feature matrix

From a traffic data matrix X _traffic And a traffic congestion index matrix X _TTI Composition, which is defined as X _cognition ＝ω ₁ ·X _traffic +ω ₂ ·X _TTI Wherein ω is ₁ ,ω ₂ Are each X _traffic And X _TTI The weight of (a) represents the degree of influence of the historical traffic data and the traffic congestion index on the traffic awareness dataThe larger the weight is, the greater the influence on the traffic awareness data is.

The attribute matrix obtained by fuzzy classification processing of external attributes comprises

Extrinsic attribute feature matrix

It is composed of these attribute matrices, which are defined as follows:

wherein ω is ₁ ,ω ₂ ,ω ₃ ,ω ₄ The traffic cognitive data processing method comprises the following steps that weights of a time period matrix, a vacation matrix, a weather matrix and a flow change matrix are respectively used for representing the influence degree of different external attributes on traffic cognitive data, the influence on the traffic cognitive data is larger when the weight is larger, tau represents a time step length, and N represents the number of roads.

Step2: building models

(1) Model integral structure

After the front data processing part processes the input traffic data and external attribute data into a required feature matrix, the data are input into three sub-modules. The external attribute feature data firstly enters a fuzzy reasoning mechanism to generate an attribute influence feature matrix X _Effect Attribute influence feature matrix X _Effect And traffic cognition feature matrix X _{c ion} Together with the input map convolution neural network, captures the spatial features therein and generates a spatial feature matrix. And inputting the spatial characteristic matrix into a module consisting of gate control circulation units for capturing the final time characteristic, and outputting a traffic data prediction result. The overall model structure thus comprises three parts: the first part is a fuzzy inference mechanism for capturing the influence characteristics of the external attributes, which is composed of fuzzy membership function, fuzzy rule and solution of fuzzyThe method comprises the steps of forming a function, and extracting the influence characteristics of external attributes on traffic data and different external attributes; the second part is a graph convolution neural network used for capturing data space characteristics, the input of the graph convolution neural network is a traffic cognition characteristic matrix, an external attribute characteristic matrix and an adjacent matrix, and the space dependence characteristics among different roads can be captured; the third part is a sub-module based on a gating cycle unit, the input of the sub-module is spatial characteristic data processed by a graph convolution neural network, the time dependence characteristic in the data can be further extracted, and the data are finally decoded into predicted road traffic data;

(2) Fuzzy inference mechanism

And the fuzzy inference mechanism is constructed according to the data characteristics and expert experience and is used for extracting the influence characteristics of the external attributes on the traffic data and generating an attribute influence characteristic matrix. Before inputting the fuzzy inference mechanism, different weights can be added to different attribute feature matrices according to data characteristics, and here, the weights of all attribute feature matrices are directly set to 1 in order to simplify the experimental process. Characteristic matrix of each attribute

The probability that the characteristic elements belong to a certain class is obtained through fuzzification processing of a fuzzy Gaussian membership function, so that the input definite value is converted into a probability value in the range of 0 to 1;

fuzzy rules in the fuzzy inference system are determined according to expert experience and attribute characteristics, and the purpose is to enable a model to better capture the characteristics of external attributes and the interaction relation among different attributes;

and (4) performing defuzzification on inference results obtained by fuzzy rule calculation to obtain control output, wherein influence values of all input attributes on traffic data are obtained. Commonly used defuzzification functions include the maximum membership function method, the center of gravity method, and the weighted average method, where the center of gravity method is used. The gravity center method is to take the gravity center of an area enclosed by a membership function curve and a horizontal coordinate as a final output value of fuzzy inference, and is defined as follows:

compared with other two methods, the gravity center method has smoother output reasoning control, even if input data slightly changes, the output data also changes, and the method is more suitable for application scenes;

(3) Spatial characteristic extraction structure

The three-dimensional tensor can be processed by the spatial characteristic extraction structure, namely the graph convolution neural network, and the dynamic interaction characteristic and the spatial dependence characteristic between roads in the traffic scene can be extracted more efficiently. The three-dimensional tensor X is composed of an adjacency matrix A for describing the road node relation, a traffic cognition feature matrix for describing traffic features and an external attribute feature matrix, and the output of the l-th layer of a graph convolution neural network

Is defined as:

C _l+1 ＝σ(L _sym C _l W _l )

C ₀ ＝X,

wherein L is _sym Is normalized Laplacian, D is degree matrix calculated from adjacency matrix, C _l+1 Is the output of the l-th layer convolution, W _l The method is a weight matrix of the l-th layer, the used activation function is a nonlinear activation function Relu, tau represents a time step, and N represents the number of road nodes;

after the adjacency matrix A and the feature matrix X are processed by the graph convolution neural network, a spatial feature matrix is generated, and the spatial feature capture of traffic cognitive data is completed for subsequent feature extraction work;

(4) Temporal characteristic extraction structure

And further extracting the features of the spatial feature matrix output by the convolutional neural network of the previous graph based on the time feature extraction structure of the gating cycle unit. Reset gate of gate control circulation unit at time t

Updating door

And candidate hidden states

Is defined as follows:

R ^t ＝σ(X ^t W _xr +H ^t-1 W _hr +b _r )

Z ^t ＝σ(X ^t W _xz +H ^t-1 W _hz +b _z )

where h is the number of hidden units, where h =64,

is the input feature at time t, P is the number of extrinsic attribute features,

is the hidden state at the last time step,

are two weight matrices for the reset gate,

is a bias matrix of the reset gates and,

is to update the two weight matrices of the gate,

are two weight matrices for the candidate hidden states,

is a bias matrix of the candidate hidden state, and the activation function used by the candidate hidden state is a tanh function;

after the spatial feature matrix is processed by the gating circulation unit, the time characteristics of the spatial feature matrix are further extracted, and the spatial feature and the time characteristics are fused to generate predicted traffic data for subsequent traffic cognitive work.

Step3: model training

(1) Training data

Training data are sampled at the frequency of 10 minutes, historical data are sampled at 12 time steps, and the time span is 120 minutes; the prediction data are 1,3 and 6 time steps respectively, and the time span is 10 minutes, 30 minutes and 60 minutes. To prevent problems such as gradient explosion during the experiment, the input data is normalized to be between 0 and 1. The training data set accounts for 80%, and the remaining 20% of the data is used as the test data set. For more convincing, a public data set Los-loop measured on the US Los Angeles expressway and a real taxi data set SZBZ near the Shenzhen north station are used as data sets for model training, a user-defined Dataloader object is used for packaging data, an input feature matrix is processed in an intelligent iteration mode, and a data tensor with a corresponding format is output to the model for training. The SZBZ data set is richer than a Los-loop data set, so that the SZBZ data set is used for hyper-parameter adjustment, ablation experiments, noise tests and comparison experiments with other models, and the Los-loop data set is used for comparison experiments of model cognition performance under different prediction durations;

(2) Training environment

The training environment is a laboratory server, 3.19GHz 12 ^th Gen Intel (R) Core (TM) i9-12900K,64GiB memory, video card NVIDIA GeForce RTX 3090 and video memory 128GB. The development language python3.7.11, the deep learning framework tensoflow2.5.0, the development platform used was pycharm2021.2.3, MATLAB R2021b. Training optimizer Adam, learning rate 0.001, number of gated loop units 64, batch size of batch =32, loss function

Wherein t is _f Is a predicted time step, Y _pred Is predicted traffic data, Y _true Is real traffic data;

(3) Adjusting the training times according to the actual experiment requirements and the training effect, and storing a model parameter file once after each round of training;

(4) Model operation process

The overall structure of the model is shown in the figure I. And the characteristic channel of the traffic cognition related data is 3, wherein an external attribute characteristic matrix generates an attribute influence characteristic matrix after passing through a fuzzy reasoning mechanism, the attribute influence characteristic matrix, the traffic cognition characteristic matrix and an adjacent matrix are input into the graph convolution neural network together, and a spatial characteristic matrix output by the graph convolution neural network is input into a subsequent gating cyclic unit. The internal details of the fuzzy inference mechanism are shown in figure two, firstly, the constructed fuzzy rule is utilized to carry out data preprocessing and fuzzy division, the processed data is combined with the fuzzy rule to carry out decision making, namely, feature calculation is carried out, a fuzzy value is obtained after the feature calculation, and finally, a determined value is obtained through a defuzzification function and an attribute influence feature matrix is generated. The input of the graph convolution neural network consists of an attribute influence characteristic matrix, a traffic cognition characteristic matrix and an adjacent matrix, and spatial dependence characteristics of data are extracted and a spatial characteristic matrix is generated through modules of a normalization layer and a convolution layer. Meanwhile, the entrance and the exit of the module are processed by the activation function, so that the problem of gradient disappearance or gradient explosion is avoided. And the number of the hidden units of the last sub-module based on the gated circulation unit is 32, and the spatial characteristic matrix finishes the extraction of the time characteristic of the data after passing through the gated circulation unit and generates a predicted traffic data matrix.

Step 4: predicting traffic data

Inputting the traffic data to be tested into the trained traffic data prediction model in a form consistent with the input during training, i.e. inputting the traffic data to be tested into the trained traffic data prediction model

The model outputs the prediction data as:

the data format at each time step is the same as the historical data. The predicted traffic data can be used for calculating model performance indexes and can also be further used for subsequent traffic cognitive work and attribute influence fuzzy classification work.

Step 5: constructing attribute impact description features for traffic prediction data

The influence of the attribute on the traffic data is directly reflected on the influence of the method on the prediction accuracy of the traffic data, and the traffic data is further recognized by comparing and analyzing different experimental results by utilizing the characteristic. Different experimental settings are obtained by controlling and inputting different attribute characteristics, and through the four steps described above, prediction data under these experimental settings have been obtained. Before the next traffic cognition, attribute influence description characteristics need to be constructed, and a decision coefficient R is adopted by combining the characteristics of traffic data ² Used to measure the predicted performance under different experimental settings, it is defined as follows:

wherein Y is _t In order to be the real traffic data,

for predicted trafficThe data of the data is transmitted to the data receiver,

the average value of the real traffic data is obtained, and n is the total number of the traffic data; r is ² The value of (2) is between 0 and 1, the closer to 1, the better the regression fitting effect is, the prediction performance of the method under different experimental settings is shown, and the subsequent traffic cognition and fuzzy classification work is also facilitated.

Step6: traffic cognition and attribute influence fuzzy classification is carried out based on attribute influence description characteristics,

defining an interpretable traffic awareness method based on a decision coefficient by determining a coefficient R ² Describing the influence degree of the attribute characteristics on traffic cognition, fully considering the attribute influence characteristics in a specific traffic scene, introducing a fuzzy set theory based on the fuzziness, randomness and subjectivity of the attribute influence and classification standards, further carrying out fuzzy classification on the influence characteristics, constructing three types of fuzzy sets with small influence, medium influence and large influence, and realizing cognition in an easily understood and interpretable mode; determining the coefficient R ² Describing the degree of the fuzzy sets with a membership function mu, selecting a fuzzy Gaussian membership function, and determining a threshold parameter lambda based on a heuristic method ₁ And λ ₂ . In particular lambda on the SZBZ dataset ₁ ＝0.3,λ ₂ =0.7. Note that, because the traffic conditions on the road do not change immediately in a short time and the changes in different time periods are significantly different, the cognitive method spans multiple time steps and respectively learns the experimental results at different time steps;

the unified form of the three fuzzy sets is

Wherein the fuzzy Gaussian membership function is defined as:

the width of the function curve is defined by k (k)>0) The center position of the curve is determined by α. Specifically, the membership functions of the three fuzzy sets are respectively

Wherein alpha is ₁ ＝0，α ₂ ＝0.5，α ₃ ＝1，k＝0.7。

After the fuzzy sets and the fuzzy membership function thereof are constructed, the experimental results under different experimental settings are utilized to carry out fuzzy classification and traffic cognition. Specifically, different characteristic attributes are input into the model, experimental processing is performed respectively, and the decision coefficients under different experimental settings are calculated by using the prediction results of the model. And after the decision coefficient is obtained, carrying out fuzzy classification by using the constructed fuzzy membership function, and carrying out traffic cognition according to a classification result to obtain the influence of different attribute characteristics on traffic data.

The above-listed series of detailed descriptions are merely specific illustrations of possible embodiments of the present invention, and they are not intended to limit the scope of the present invention, and all equivalent means or modifications that do not depart from the technical spirit of the present invention are intended to be included within the scope of the present invention.

Claims

1. An interpretable traffic cognition method based on a fuzzy theory is characterized by comprising the following steps:

the first part is a fuzzy reasoning mechanism based on a fuzzy theory, which consists of a fuzzy membership function, a fuzzy rule and a defuzzification function and can effectively extract the influence of external attributes on traffic data; a traffic cognitive characteristic tensor composed of the traffic data tensor and the traffic congestion index tensor is combined with the output of the first part to serve as input data of a second part of the model;

step3, training a graph convolution neural network model;

step 4, performing multi-class traffic cognition experiments based on the trained model;

step 5, constructing attribute influence description characteristics based on the traffic cognition experiment results;

2. The interpretable traffic cognition method based on the fuzzy theory as claimed in claim 1, wherein the input data in the step1 is specifically defined as follows:

wherein the intersection of each time stepGeneral data are

traffic scene prediction data matrix X' _traffic Consists of traffic data for T predicted time steps:

the size of the expression matrix is T multiplied by N, wherein elements are real numbers;

Traffic data generally refers to data with spatio-temporal characteristics describing traffic scenes, including traffic speed, traffic flow, traffic congestion conditions;

describing the connection condition of roads in a traffic scene by using an undirected graph G = { V, E }, wherein a node set V is represented as V = { V ₁ ,v ₂ ,…,v _N N is the number of roads in the traffic scene;

when the roads are communicated with each other, the sides representing the road interaction relationship are connected; thus the set of edges E expressed as E = { v _i v _j I is more than or equal to 1 and less than or equal to N, j is more than or equal to 1 and less than or equal to N, wherein v _i v _j Representing the communication between the node i and the node j;

the element level of adjacency matrix a is defined as:

step 1.3, describing a feature matrix of external attributes:

X _feature ∈{X _TTI ,X _time ,X _holiday ,X _weather ,X _flowChange },

Wherein the traffic congestion index matrix

wherein V _free For free passage speed, V _current Is the current speed;

time period setting matrix

The element level of the matrix is defined as:

wherein [ t ] _n ,t _n+1 ]Is a continuous time section, the traffic data in the time section has similar change state, in the actual calculation process, different time sections are represented by different positive integers, tau represents the time step, N represents the number of roads,

vacation setting matrix

The element level of the matrix is defined as:

this formula indicates that when the date is a weekend or a statutory holiday,

weather matrix

The element level of the matrix is defined as:

the weather conditions contained in the weather matrix are represented as:

X _{weatherCondition} ∈{sunny,cloudy,light rain,medium rain,heavy rain}

setting traffic flow change matrix

The element level of the matrix is defined as:

step 1.4, an interpretable traffic data cognitive algorithm:

the interpretable finger model presents the processing process of the data to developers in a simple and understandable form, and is embodied that the fuzzy rules in the fuzzy inference mechanism are represented by intuitive IF-THEN statements, wherein the four important fuzzy rules are defined as follows:

fuzzy rule 1: IF flowChange is zero AND time is seven the effect is small

Fuzzy rule 2: IF flowChange is third AND time is seven the effect is large

Fuzzy rule 4: IF weather is fog same kind of effect is middle

The fuzzy rule 1 indicates that when the traffic flow change is located around 0 and the time is between seven and nine points, the influence on the traffic data is small; the fuzzy rule 2 indicates that when the traffic flow change is located at about 30 and the time is between seven and nine points, the influence on the traffic data is large; fuzzy rule 3 indicates that when the date is holiday and the time is between nine and eleven, the influence on the traffic data is very large; the fuzzy rule 4 indicates that when the weather is foggy, the influence on the traffic data is moderate.

3. The method for interpretable traffic cognition based on the fuzzy theory as claimed in claim 1, wherein the fuzzy inference mechanism based on the fuzzy theory in the step2 can process multidimensional external attribute tensor, and can explainably extract the influence of the external attribute on traffic data; the graph convolution neural network can process three-dimensional tensors, can more efficiently extract spatial correlation among different roads in a traffic scene, and extracts time correlation among traffic data by combining gate control cycle units; input matrix with fuzzy inference mechanism

wherein ω is ₁ ,ω ₂ ,ω ₃ ,ω ₄ The weights of the time period matrix, the holiday matrix, the weather matrix and the traffic change matrix respectively represent the influence degree of different external attributes on the traffic cognitive data, the larger the weight is, the larger the influence on the traffic cognitive data is, the tau represents the time step length, the N represents the number of roads,

output matrix of fuzzy inference system

It is defined as follows:

the expression matrix has a size of τ × N, where the elementsAre all real numbers;

setting attribute matrix related to traffic cognition

From X _traffic And X _TTI Composition, which is defined as follows:

X _cognition ＝ω ₁ ·X _traffic +ω ₂ ·X _TTI

output of layer I in graph convolution neural network

Is defined as:

C _l+1 ＝σ(L _sym C _l W _l )

wherein L is _sym Normalized Laplacian, D is a degree matrix calculated from an adjacency matrix, C _l+1 Is the output of the l-th layer of convolution, A is an adjacency matrix describing the road node relationship, X is a feature matrix describing the traffic features and external attributes, W _l Is the weight matrix of the l-th layer, the activation function used is the nonlinear activation function Relu, T tableShowing the time step, N the number of road nodes,

Updating door

And candidate hidden states

Is defined as follows:

R ^t ＝σ(X ^t W _xr +H ^t-1 W _hr +b _r )

Z ^t ＝σ(X ^t W _xz +H ^t-1 W _hz +b _z )

where h is the number of hidden units,

is the hidden state at the last time step,

is two rights to reset the gateThe weight matrix is a matrix of the weight,

is a bias matrix of the reset gates and,

is to update the two weight matrices of the gate,

are two weight matrices for the candidate hidden states,

the matrix size is expressed, wherein elements are real numbers;

4. The interpretable traffic cognition method based on the fuzzy theory as claimed in claim 1, wherein the training data in the step3 are set as follows:

training optimizer Adam, learning rate 0.001, number of gated loop units 64, batch size batchsize =32, time step 12 of historical traffic awareness data, and loss function defined as follows:

using a public data set Los-loop and a taxi data set SZBZ near a real Shenzhen north station as a data set for model training, and dividing the data set into a training set and a testing set according to a certain proportion; and packaging data by using a user-defined Dataloader object, intelligently and iteratively processing the input characteristic matrix, and outputting a data tensor with a corresponding format to the model for training.

5. The interpretable traffic cognition method based on the fuzzy theory as claimed in claim 1, wherein the multi-class traffic cognition experiment in the step 4 is set as follows:

the first type of traffic cognition experiment is to verify the influence of different external characteristic attributes on traffic data, and includes four different experimental settings: the first is that the input data only comprises historical traffic data; the second kind adds vacation data and weather data on the basis of the first kind; the third is to add time period data and traffic flow change data on the basis of the second; fourthly, adding the traffic jam index data on the basis of the third method;

the second type of traffic cognition experiment is to verify the cognitive performance of the method in different time periods, and comprises three different experiment settings: the first is to predict traffic data for 10 minutes into the future; the second is to predict traffic data for 30 minutes into the future; the third is to predict traffic data 60 minutes in the future;

6. The method as claimed in claim 1, wherein in the step 5, attribute influence description features are constructed on the traffic prediction data, the influence of the attributes on the traffic data is directly reflected on the influence of the method on the prediction accuracy of the traffic data, and the difference in the step 4 is used for realizing the interpretability of the traffic dataSetting experiments to obtain the prediction data under the experimental settings; determining the coefficient R ² Used to measure the predicted performance at different experimental settings, it is defined as follows:

wherein Y is _t In order to be the real traffic data,

in order to be able to predict the traffic data,

7. The method as claimed in claim 1, wherein the step6 comprises performing fuzzy classification of attribute impact on the descriptive features by using a decision coefficient R ² Describing the influence degree of the attribute characteristics on traffic data prediction, fully considering the attribute influence characteristics under the background with a traffic scene, introducing a fuzzy set theory due to the fuzziness and randomness of influence and the subjectivity of classification standards, further carrying out fuzzy classification on the obtained influence characteristics, and constructing three fuzzy sets with small influence, medium influence and large influence, R ² Describing the degree of the fuzzy sets to which the fuzzy sets belong by using a membership function mu; selecting a Gaussian membership function, and determining a threshold parameter lambda based on a heuristic method ₁ And λ ₂ 。