CN113947259B - GRU neural network-based method for predicting speeds of drivers in different styles - Google Patents
GRU neural network-based method for predicting speeds of drivers in different styles Download PDFInfo
- Publication number
- CN113947259B CN113947259B CN202111321247.XA CN202111321247A CN113947259B CN 113947259 B CN113947259 B CN 113947259B CN 202111321247 A CN202111321247 A CN 202111321247A CN 113947259 B CN113947259 B CN 113947259B
- Authority
- CN
- China
- Prior art keywords
- driving
- vehicle speed
- gru
- data
- driving style
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 78
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 46
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 230000008569 process Effects 0.000 claims description 33
- 230000001133 acceleration Effects 0.000 claims description 19
- 230000006870 function Effects 0.000 claims description 16
- 238000013145 classification model Methods 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 12
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 238000003064 k means clustering Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 238000012804 iterative process Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 description 12
- 230000006399 behavior Effects 0.000 description 6
- 125000004122 cyclic group Chemical group 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 238000010276 construction Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000006698 induction Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007621 cluster analysis Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000009022 nonlinear effect Effects 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B60—VEHICLES IN GENERAL
- B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
- B60W40/10—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models related to vehicle motion
- B60W40/105—Speed
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Business, Economics & Management (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Quality & Reliability (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Operations Research (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Automation & Control Theory (AREA)
- Transportation (AREA)
- Mechanical Engineering (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a vehicle speed prediction method for drivers of different styles based on GRU neural networks, which comprises the following steps: dividing a driving database to obtain a plurality of driving sub-databases corresponding to different driving styles; respectively constructing GRU vehicle speed prediction models corresponding to different driving styles; respectively extracting training data from each driving sub-database, preprocessing, and generating a training sample sequence database by sliding time by time according to a time sequence; training GRU vehicle speed prediction models corresponding to driving styles by using a training sample sequence database; and judging the current driving style at certain intervals, and introducing a GRU vehicle speed prediction model corresponding to the current driving style to predict the vehicle speed when the current driving style is continuously unchanged in a preset time period. According to the invention, different GRU vehicle speed prediction models are constructed for the prediction tasks under different driving style states, and the prediction results of the driving style prediction models are not distinguished by comparison, so that the stability and the prediction precision of prediction can be effectively improved.
Description
Technical Field
The invention relates to the technical field of automobile speed prediction, in particular to a driver speed prediction method of different styles based on GRU neural networks.
Background
Model Predictive Control (MPC) in the field of automobile control is thoroughly and widely studied at present, and the MPC not only considers current state information in the use process, but also makes a plan in advance by further utilizing the prediction information obtained by the prediction model. All MPC algorithms have a common definition requirement that a model is required to predict future development of variables over a limited range. Current research for automotive control systems has focused mainly on predicting vehicle state variables, which in most cases can be attributed to speed predictions.
Speed prediction generally refers to estimating a speed value over a time domain by reasonable analysis of some historical data. So far, speed prediction models have been developed into two main categories: parametric methods and non-parametric methods. Since the speed of the main vehicle is influenced by the driving mode of the driver and the surrounding environment, the evolution of the speed of the vehicle has strong random and nonlinear properties, and the speed prediction by a non-parameter method has become the current main trend. In recent years, with the optimization of algorithms and the rapid increase of computing power, deep learning methods regain extensive attention and acquire great achievements in various engineering fields. The vehicle speed prediction is taken as a typical prediction task, the evolution of the vehicle speed along with time is dynamic, and the original cyclic neural network (RNN) algorithm and the variation thereof are introduced into the construction of a network neural network prediction model, so that the internal law of the vehicle speed change can be effectively analyzed and memorized.
The most widely applied cyclic neural network in the current vehicle speed prediction field is a long-short-term memory network (LSTM), and the cyclic neural network has the defects of too much parameter and incapability of parallel training while having good performance, so that a large amount of training data cannot be processed well in a short time, and therefore, a more efficient training model is required to be introduced to complete the vehicle speed prediction task. On the other hand, the conventional vehicle speed prediction process only considers the states of the vehicle and the surrounding vehicles, and ignores the influence of the driving characteristics of the driver on the vehicle speed. Therefore, it is highly desirable to provide a vehicle speed prediction method with higher efficiency while ensuring prediction accuracy.
Disclosure of Invention
In view of the above, the invention provides a vehicle speed prediction method for drivers of different styles based on GRU neural networks, which constructs different door control cyclic neural network (GRU) vehicle speed prediction models for prediction tasks in different driving style states, and can effectively improve the stability and the prediction precision of prediction by comparing prediction results of the driving style prediction models without distinguishing the prediction results.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
a driver speed prediction method of different styles based on GRU neural network comprises the following steps:
S1, dividing a driving database by adopting a pre-constructed driving style classification model to obtain a plurality of driving sub-databases corresponding to different driving styles;
S2, respectively constructing GRU vehicle speed prediction models corresponding to different driving styles;
S3, respectively extracting training data from each driving sub-database, preprocessing, and generating a training sample sequence database by sliding time by time according to a time sequence;
S4, training the GRU vehicle speed prediction model corresponding to the driving style by using the training sample sequence database;
and S5, judging the current driving style by adopting the driving style identification model every certain time, and introducing a GRU vehicle speed prediction model corresponding to the current driving style to predict the vehicle speed when the current driving style is continuously unchanged in a preset time period.
Compared with the prior art, the invention discloses a vehicle speed prediction method for drivers of different styles based on GRU neural networks, which has the following beneficial effects:
1. by introducing the more optimized GRU with the network structure, the training time of the model is reduced by more than 10% on the premise of not changing the original prediction precision, the model prediction time is reduced, and the model prediction efficiency is improved.
2. On the premise of considering the states of the vehicle and the peripheral vehicles, the influence of the driving characteristics of the driver on the change of the vehicle speed is introduced, different vehicle speed prediction models are built for the prediction tasks under different driving style states, different GRU vehicle speed prediction models are introduced along with the change of the driving style of the driver, and the prediction stability and the prediction precision of the prediction are effectively improved by comparing the prediction results of the non-passing driving style prediction models.
Further, the method further comprises the following steps:
And S6, updating the actual vehicle speed information with larger error and the history information thereof in the prediction process into the training sample sequence database, and retraining and parameter adjustment are carried out on each GRU vehicle speed prediction model by using the updated training sample sequence database after a preset time interval.
Further, the method further comprises the following steps: s1 comprises the following steps:
s12, performing preliminary clustering on the characteristic data in the driving database by using a self-organizing map neural network;
S13, clustering the primary clustering results again by using a K-means clustering method to obtain a plurality of driving sub-databases corresponding to different driving styles; the driving sub-databases include aggressive driving databases, intermediate driving databases and conservative driving databases.
Further, the method further comprises the following steps: s12 comprises the following steps:
s1201, carrying out normalization processing on the characteristic data in the driving database to obtain dimensionless scalar;
s1202, introducing an SOM model, and setting the geometric shape in the SOM model to be rectangular;
S1203, randomly initializing the weight of the SOM model, and initializing the node weight into a smaller random number;
S1204, setting an initial neighborhood radius sigma, and determining nodes to be contained in the winning neighborhood according to the neighborhood radius sigma;
S1205, setting a proximity function type, and calculating the respective updated amplitude of each node through the proximity function;
S1206, setting an initial learning rate alpha, and determining the amplitude of each iteration weight update;
S1207, randomly taking an input sample X i, and performing a first iteration; traversing each node in the competition layer, calculating the similarity between X i and the node, and selecting the node with the smallest distance as a winning node;
S1208, updating Weight of nodes in a radius range with a winning node as a center and sigma in a winning neighbor;
s1209, carrying out attenuation treatment on sigma and alpha after each iteration is completed;
S1210: after completing one iteration, returning to S1207 until the set iteration number is met, and obtaining a preliminary clustering result.
Further, the method further comprises the following steps: s13 comprises the following steps:
S1301, randomly selecting 3 samples { mu 1,μ2,μ3 } from the preliminary clustering result as initial centroids;
S1302, calculating the distance between each sample Y i and the centroid mu j, namely dis= |Y j-μj||2, and classifying the sample Y j and the nearest centroid mu j into a class;
s1303 updating centroid Wherein ci is cluster clustering in the iterative process, i is {1,2,3};
and S1304, returning to S1302 until the maximum iteration number or the mass center updating amplitude is smaller than the threshold value, and finally obtaining the driving sub-databases of the three driving styles.
Further, the method further comprises the following steps: the training data extracted in S3 includes input data and output data, wherein the input data includes: the system comprises data information continuously changing along with time and state type information discretely distributed along with time, wherein output data are vehicle speed signals; the data information continuously varying with time includes: historical vehicle speed, historical acceleration, engine speed, engine torque and relative distance between the front vehicles; the state category information distributed discretely over time includes gear information.
Further, the method further comprises the following steps: the preprocessing in S3 includes: and carrying out differential and normalization processing on the data information which continuously changes along with time.
Further, the method further comprises the following steps: s3, preprocessing data, taking 15S as a period, taking a 10S historical data sequence X＝{X1,X2,X3,X4,X5,X6,X7,X8,X9,X10} as an input of a GRU vehicle speed prediction model training process, and taking a vehicle speed signal value Y= { Y 1,y2,y3,y4,y5 } of 5S as an output of the GRU vehicle speed prediction model; ,Xj＝{x1,x2,x3,x4,x5,x6},x1,x2,x3,x4,x5,x6 respectively represent six characteristics of historical vehicle speed, historical acceleration, engine speed, engine torque, relative distance between front vehicles and gear information; y 1,y2,y3,y4,y5 is the vehicle speed value for the next 5 seconds.
Further, the method further comprises the following steps: the GRU vehicle speed prediction model in the S2 comprises a GRU layer and a full connection layer;
the GRU layer comprises an update gate, a reset gate, a candidate state and a hidden state; the GRU layer internal expression is as follows:
Zt＝σ(Wz[ht-1,xt]+bz)
rt＝σ(Wr[ht-1,xt]+br)
Wherein x t is an input vector at the current moment, h t-1 is a hidden state vector at the previous moment, W z and b z are respectively a weight matrix and a bias vector of an update gate, sigma is a sigmoid function, Z t is an activation vector of the update gate, r t is an activation vector of a reset gate, wr and b r are respectively a weight matrix and a bias vector of the reset gate, Z t is a candidate state vector, W h and b h are respectively a weight matrix and a bias vector of control candidate state information, and h t is a hidden state vector at the current moment;
and the full connection layer converts the calculation result output by the GRU layer into a final calculation result of the required dimension.
Further, the method further comprises the following steps: in S5, the determining process for the current driving style is: continuously collecting driving speed information, acceleration information and following distance information in the driving process, and extracting vehicle speed distribution characteristics, acceleration derivative values along with time and average following relative distance values in the driving process; and judging the driving style once every 5s, if the driving style is not changed after the driving style lasts for 1min, considering the driving style to be stable, and introducing a GRU vehicle speed prediction model corresponding to the driving style to predict the vehicle speed.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a method for predicting the speed of drivers in different styles based on GRU neural networks.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
As shown in fig. 1, the embodiment of the invention discloses a method for predicting the speed of drivers in different styles based on a GRU neural network, which comprises the following steps:
S1, dividing a driving database based on a pre-constructed driving style classification model to obtain a plurality of driving sub-databases corresponding to different driving styles;
S2, respectively constructing GRU vehicle speed prediction models corresponding to different driving styles;
S3, respectively extracting training data from each driving sub-database, preprocessing, and sliding time by time according to a time sequence to generate a training sample sequence database;
s4, training the GRU vehicle speed prediction model corresponding to the driving style by using a training sample sequence database;
S5, judging the current driving style by adopting a driving style classification model every certain time, and introducing a GRU vehicle speed prediction model corresponding to the current driving style to predict the vehicle speed when the current driving style is continuously unchanged in a preset time period.
The following describes in detail the construction process of the driving style classification model adopted in the embodiment of the present invention.
According to the operation characteristics and behavior trends of different drivers, the driving styles of different drivers can be comprehensively summarized into aggressive drivers, mediocre drivers and conservative drivers. Different types of drivers may have distinct and different driving characteristics during driving, including intensity of acceleration and deceleration, distribution of speed intervals, control of following distance, and so on. General definition: the acceleration and deceleration processes are severe, and the vehicle tends to run at high speed under the same road conditions, and the closer the following distance is, the more aggressive the driver; the acceleration and deceleration processes are mild, the vehicle tends to run in a medium-low speed mode under the same road condition, and the long following distance is defined as a conservative driver; when the two driving styles are defined to be completed, the excessive driving style between the two driver states is defined as a moderate driver. In accordance with the above definition, other reasonable driver behavior characteristic variables with reference value can be introduced for the identification of the driver style.
By collecting a large amount of historical data of the driving process of the driver or directly calling the data information in the existing database, the driving behavior data of different drivers can be more accurately summarized and extracted by combining the related data analysis method, and finally, the driving behavior characteristics which can be used for identifying the style of the driver are obtained. The process completes the data preparation process of driver data identification, and then proper identification models are selected for training, so that good classification effect on historical data can be achieved. Finally, the deployment of the classification model is completed, and the aim of good identification of the newly acquired data can be achieved.
In summary, the whole process of driver driving style recognition is divided into four processes of analysis processing of historical big data, induction and extraction of style recognition feature quantity, selection and training of driving style classification model and re-recognition of driving style after new data is input.
1. Analysis processing of historical big data and induction and extraction of style identification feature quantity
The driving style of the driver can have a great influence on the driving state during driving, which can be directly reflected on the behavior data of the vehicle. A large number of researches show that a driver with aggressive driving style occupies a larger high-speed driving interval in the whole driving process, and meanwhile, the realization of the required vehicle speed can be completed in a short time, which is reflected in that the acceleration and deceleration processes are violent when the speed is changed, and the acceleration and deceleration processes can be defined as that the change rate of the acceleration along with time is larger, and meanwhile, the aggressive driver also tends to drive at a small vehicle distance in a following scene. Based on the above intuitive idea, the present embodiment selects the statistic as shown in table 1 as the feature quantity for identifying the driving style of the driver.
TABLE 1
2. Model selection and training
(1) Selection of identification models
The driving style recognition process of the driver is a process of extracting features aiming at driving data, carrying out cluster analysis aiming at selected features and matching the driving process of the driver with a cluster center, and comprises the following specific steps:
s11, constructing a driving style classification model based on a self-organizing map (SOM) neural network and a K-means clustering method;
s12, performing preliminary clustering on the characteristic data of the driver through a self-organizing map (SOM);
S13, a K-means clustering method is selected to cluster the primary clustering result again, so that final identification of the style of the driver is achieved, and a plurality of driving sub-databases corresponding to different driving styles are obtained.
The SOM neural network has a two-layer structure of an output layer and an input layer, wherein the input layer is determined by input data, so that the network structure is mainly determined by the output layer structure. The output layer of the SOM neural network has a free geometric structure, and can select a one-dimensional structure or a high-dimensional structure. The total node number of the output layer refers to an empirical formula:
where N is the total number of nodes and N is the number of sample points. The output layer of this embodiment is a two-dimensional square structure, and the side length is If the situation that the division cannot be completed occurs, the whole is upwards rounded.
After the SOM neural network completes the preliminary clustering, the K-means method continues to complete the clustering work. In the present embodiment, the style of the driver is classified into three types of aggressive type, intermediate type and conservative type, and thus the cluster center of K-means is set to 3 types.
(2) Training of models
The training process of the SOM neural network comprises the following steps:
And S1201, carrying out normalization processing on the characteristic data in the driving database to obtain dimensionless scalar.
S1202, introducing an SOM model, calling minisom libraries in Python, introducing MiniSom classes, setting the geometric shape in the SOM model as a rectangle, and respectively setting the two side length dimensions as follows
And S1203, randomly initializing the weight of the SOM model, and initializing the node weight into a smaller random number.
S1204, setting an initial neighborhood radius sigma, and determining nodes to be contained in the winning neighborhood according to the neighborhood radius sigma. Nodes within a radius of σ centered around the winning node will be updated. Here σ must be greater than 0, otherwise no neurons will be updated; and sigma cannot be greater than the side length of the 2-dimensional output plane.
S1205, setting a proximity function type, and calculating the updated amplitude of each node through the proximity function. The proximity function is used for determining the influence intensity of the winning node on the neighboring nodes, and the basic idea is that the updating amplitude is larger when the winning node is closer; the farther away the winning node, the smaller the update amplitude. The method is more common to a Gaussian function method and a buffer function method, wherein the update amplitude of the former is continuously reduced along with the increase of the distance from a winning node, and the update amplitude of the latter nodes in the whole winning neighborhood is the same.
S1206, setting an initial learning rate alpha, and determining the amplitude of each iteration weight update. Alpha can be selected from 0 and 1, and the influence of each iteration on the weight update is determined; the smaller the α, the smaller the magnitude of each iteration weight update, the larger the α, and the larger the magnitude of each iteration weight update.
S1207, randomly taking an input sample X i, and performing a first iteration; and traversing each node in the competition layer, calculating the similarity between X i and the node, selecting Euclidean distance representation, and selecting the node with the minimum distance as a winning node.
S1208, updating Weight of nodes in the radius range with the winning node as the center and sigma in the winning neighbor.
Wv(t+1)＝Wv(t)+θ(v,t)·α(t)·(D(i)-Wv(t))。
Wherein s refers to a state corresponding to the current iteration number; v denotes the current node sequence number; w v(s) is the current weight of the v-th node; w v (s+1) refers to the new weight value updated by the v-th node; θ (v, s) is a constraint on the update based on distance from the optimal node, i.e., the return value of the proximity function; alpha(s) refers to the learning rate corresponding to the current state; d (t) refers to the vector to which the current input corresponds.
S1209, performing attenuation processing on σ and α after each iteration is completed. The most common formula for the decay function is:
wherein T represents the current iteration number, t=total iteration number/2, and finally, the following is obtained:
S1210: after completing one iteration (iteration number +1), returning to S1207 until the set iteration number is met, and obtaining a preliminary clustering result.
The process of clustering the preliminary clustering result again by using the K-means clustering method comprises the following steps:
s1301, randomly selecting 3 samples { mu 1,μ2,μ3 } from the primary clustering result as initial centroids;
S1302, calculating the distance between each sample Y i and the centroid mu j, namely dis= |Y j-μj||2, and classifying the sample Y j and the nearest centroid mu j into a class;
s1303 updating centroid Wherein ci is cluster clustering in the iterative process, i is {1,2,3};
and S1304, returning to S1302 until the maximum iteration number or the mass center updating amplitude is smaller than the threshold value, and finally obtaining the driving sub-databases of the three driving styles.
(3) Checksum improvement of output results
Judging whether the three clustering centers can divide the driving style well according to the three clustering center characteristic values obtained by final clustering, if the clustering result of the driving style type confusion exists, considering a method of adding the driving data of the driver again for training and adding new identification characteristic values to improve the identification result of the driving style, and obtaining a final driving style classification model.
3. Re-recognition of driving style after new data input
After the perfect driving style classification model is obtained, the driving style of the newly input driving data can be identified in real time in the driving process of the driver.
The above-mentioned S1 to S6 will be further described below.
S1, dividing a driving database into 3 sub-databases by using the constructed driving style classification model, and respectively corresponding to the aggressive driver driving database, the moderate driver driving database and the conservative driver driving database.
S2, respectively constructing GRU vehicle speed prediction models corresponding to different driving styles.
The speed prediction of the invention refers to predicting the speed distribution after a period of time domain through reasonable analysis of historical related variable data. The speed of the vehicle in the normal running state has strong randomness and nonlinearity, so that a nonparametric model is adopted to excavate the speed change rule and predict the future speed. In this embodiment, a gated neural network (GRU) is used to predict future vehicle speed.
The GRU neural network screens and inherits the historical data through a gate mechanism to obtain a speed change rule contained in the historical information, so that accurate prediction of the vehicle speed is realized.
The gate control based cyclic neural network model comprises a GRU layer and a full connection layer.
The GRU layer structure comprises an update gate, a reset gate, a candidate state and a hidden state; the GRU hidden layer internal expression is as follows:
Zt＝σ(Wz[ht-1,xt]+bz)
rt＝σ(Wr[ht-1,xt]+br)
wherein x t is an input vector at the current time, h t-1 is a hidden state vector at the previous time, W z and b z are respectively a weight matrix and a bias vector of an update gate, σ is a sigmoid function, Z t is an activation vector of the update gate, r t is an activation vector of a reset gate, wr and b r are respectively a weight matrix and a bias vector of the reset gate, Z t is a candidate state vector, W h and b h are respectively a weight matrix and a bias vector of control candidate state information, and h t is a hidden state vector at the current time.
The full connection layer converts the calculation result output by the GRU layer into a final calculation result of the required dimension, and in the embodiment of the invention, the output dimension of the full connection layer is 5.
Setting an error evaluation function of the GRU vehicle speed prediction model as root mean square error, wherein the expression is as follows:
Wherein m represents the total number of predictions, Representing a speed predictor, y i represents a speed true value.
And S3, respectively extracting training data from each driving sub-database, preprocessing, and sliding time by time according to the time sequence to generate a training sample sequence database.
The training data is mainly divided into two parts, one part is input data used for generating a predicted value through calculation in the feedforward process, the other part is a label part used for calculating a predicted error and further back-propagating and updating network weights, the output part is a vehicle speed signal value in the embodiment, and the input part is a variable signal related to the vehicle speed signal value.
Variables that affect the change in vehicle speed can be divided into two categories: the first type is data information which continuously changes along with time, the most representative is historical vehicle speed information, the information does not have a fixed value range, the change amplitude is large, the fluctuation is strong, and the evolution rule inside the data and the internal correlation with the speed change are required to be mined. The second type of information is state type information which is distributed discretely along with time, such as gear information during driving, the data has a very clear value range, the same value type has continuity in time, the change of different states is step, and the information can be used as a type constraint condition of speed change. In this embodiment, the continuous input data includes historical vehicle speed data, historical acceleration data, preceding vehicle relative distance data, engine torque data, and engine rotational speed data; the discrete input data includes gear information.
For the discrete data and the continuous data described above, the following processing can be made: the continuous input information which continuously changes along with time, including the historical vehicle speed data, the historical acceleration data, the vehicle engine speed data, the engine torque data and the relative vehicle distance data of the front vehicle, needs to be subjected to reasonable differential and normalization processing, so that the fitting and convergence of the following models are facilitated; the discrete information is used as constraint limiting information, special processing is not needed, and the model can be trained in the form of original label data.
The magnitude of the variable value ranges of the input data of different features has great difference, and the distance between sample points of the data space is dominated by the individual feature values, so that all data needs to be mapped into the same scale for the input information of multiple features. All data in the input data are positive values, and the embodiment normalizes the data to the [0,1] interval. For each feature vector, the normalized formula is as follows:
Wherein x scale is the normalized value, x min is the minimum value of the feature data, x max is the maximum value of the feature, and all the data in the feature are mapped to the [0,1] range after processing.
Specifically, the training sample sequence database is generated by the following steps:
The input data at each time includes six features of [ historical vehicle speed, historical acceleration, engine speed, engine torque, relative distance between preceding vehicles, gear information ], 15 seconds is taken as a period, a historical 10-second historical data sequence X＝{X1,X2,X3,X4,X5,X6,X7,X8,X9,X10} is taken as input, and X j＝{x1,x2,x3,x4,x5,x6 } is the preprocessed vehicle state information. The vehicle speed predicted value y= { Y 1,y2,y3,y4,y5 } of the following 5 seconds is the predicted output, where Y 1,y2,y3,y4,y5 is the vehicle speed value of the following 5 seconds. Each sequence sample is generated by sliding time step by time step, and a training sample sequence database is built.
And S4, training the GRU vehicle speed prediction model corresponding to the driving style by using a training sample sequence database.
And training the corresponding GRU vehicle speed prediction model by using the training sample sequence database which is established by the S3 and aims at different driving styles. The 15 second input sample sequence is flattened into a one-dimensional vector, the size 15 x N f,Nf representing the number of features of the input data, N f = 6 in this embodiment. And (3) inputting data to obtain predicted data, optimizing the uniform square root error loss function by using a random gradient descent (SGD) optimization algorithm, and carrying out back propagation on the result to update network parameters until the GRU prediction model is converged, wherein model training is considered to be completed.
S5, judging the current driving style by adopting a driving style identification model every certain time, and introducing a GRU vehicle speed prediction model corresponding to the current driving style to predict the vehicle speed when the current driving style is continuously unchanged in a preset time period.
S51, judging the driving style of the driver.
In the driving behavior process of a driver, continuously acquiring driving speed information, acceleration information and following distance information of the driver, extracting vehicle speed distribution characteristics, acceleration derivative values along with time and average following relative distance values in the process, judging the driving style of the driver once every 5s by adopting a driving style classification model, keeping the driving style of the driver unchanged for 1min, considering the driving style of the driver to be stable, and introducing a GRU vehicle speed prediction model corresponding to the driving style.
S52, predicting the future vehicle speed
And extracting the past 10s speed data in the collected historical speed data, inputting 10 historical speed values into the GRU speed prediction model at intervals of every second, and outputting 5 future speed values.
S53, driving style supervision and correction
The style characteristics of the driver can change along with time and the environment of the surrounding vehicle, so that the style identification of the driver is continuously carried out in the process of continuously collecting driving data. When the driving style identification result of the driver is changed and lasts for more than 1min, the GRU vehicle speed prediction model is switched, and the step S52 is repeated.
In a further advantageous embodiment, it further comprises:
In a further advantageous embodiment, it further comprises:
S6, online adjustment of model parameters:
and updating the actual vehicle speed information with larger error in the prediction process and the history information thereof into a training sample sequence database, and retraining and parameter adjustment are carried out on each GRU vehicle speed prediction model by using the updated training sample sequence database after a preset time interval. Through on-line adjustment of GRU speed prediction model parameters, optimization and upgrading of the model are finally realized, so that generalization and prediction accuracy of the model are enhanced.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (8)
1. The method for predicting the vehicle speeds of drivers in different styles based on the GRU neural network is characterized by comprising the following steps of:
s1, dividing a driving database based on a pre-constructed driving style classification model to obtain a plurality of driving sub-databases corresponding to different driving styles; s1 comprises the following steps:
s11, constructing a driving style classification model based on a self-organizing map neural network and a K-means clustering method;
s12, performing preliminary clustering on the characteristic data in the driving database by using a self-organizing map neural network;
S13, clustering the primary clustering results again by using a K-means clustering method to obtain a plurality of driving sub-databases corresponding to different driving styles; the driving sub-database comprises an aggressive driving database, a middle-level driving database and a conservative driving database;
S2, respectively constructing GRU vehicle speed prediction models corresponding to different driving styles; the GRU vehicle speed prediction model comprises a GRU layer and a full connection layer;
the GRU layer comprises an update gate, a reset gate, a candidate state and a hidden state; the GRU layer internal expression is as follows:
Zt＝σ(Wz[ht-1,xt]+bz)
rt＝σ(Wr[ht-1,xt]+br)
Wherein x t is an input vector at the current moment, h t-1 is a hidden state vector at the previous moment, W z and b z are respectively a weight matrix and a bias vector of an update gate, sigma is a sigmoid function, Z t is an activation vector of the update gate, r t is an activation vector of a reset gate, wr and b r are respectively a weight matrix and a bias vector of the reset gate, Z t is a candidate state vector, W h and b h are respectively a weight matrix and a bias vector of control candidate state information, and h t is a hidden state vector at the current moment;
the full connection layer converts the calculation result output by the GRU layer into a final calculation result of a required dimension;
S3, respectively extracting training data from each driving sub-database, preprocessing, and generating a training sample sequence database by sliding time by time according to a time sequence;
S4, training the GRU vehicle speed prediction model corresponding to the driving style by using the training sample sequence database;
And S5, judging the current driving style by adopting the driving style classification model every certain time, and introducing a GRU vehicle speed prediction model corresponding to the current driving style to predict the vehicle speed when the current driving style is continuously unchanged in a preset time period.
2. The method for predicting vehicle speeds of drivers in different styles based on a GRU neural network of claim 1, further comprising:
And S6, updating the actual vehicle speed information with larger error and the history information thereof in the prediction process into the training sample sequence database, and retraining and parameter adjustment are carried out on each GRU vehicle speed prediction model by using the updated training sample sequence database after a preset time interval.
3. The method for predicting the vehicle speed of drivers in different styles based on the GRU neural network according to claim 1, wherein S12 comprises the steps of:
s1201, carrying out normalization processing on the characteristic data in the driving database to obtain dimensionless scalar;
s1202, introducing an SOM model, and setting the geometric shape in the SOM model to be rectangular;
S1203, randomly initializing the weight of the SOM model, and initializing the node weight into a smaller random number;
S1204, setting an initial neighborhood radius sigma, and determining nodes to be contained in the winning neighborhood according to the neighborhood radius sigma;
S1205, setting a proximity function type, and calculating the respective updated amplitude of each node through the proximity function;
S1206, setting an initial learning rate alpha, and determining the amplitude of each iteration weight update;
S1207, randomly taking an input sample X i, and performing a first iteration; traversing each node in the competition layer, calculating the similarity between X i and the node, and selecting the node with the smallest distance as a winning node;
S1208, updating Weight of nodes in a radius range with a winning node as a center and sigma in a winning neighbor;
s1209, carrying out attenuation treatment on sigma and alpha after each iteration is completed;
S1210: after completing one iteration, returning to S1207 until the set iteration number is met, and obtaining a preliminary clustering result.
4. A method for predicting driver vehicle speed in different styles based on a GRU neural network according to claim 3, wherein S13 comprises:
S1301, randomly selecting 3 samples { mu 1,μ2,μ3 } from the preliminary clustering result as initial centroids;
S1302, calculating the distance between each sample Y i and the centroid mu j, namely dis= ||Yj-mu j|| 2, and classifying the sample Y j and the nearest centroid mu j into a class;
s1303 updating centroid Wherein ci is cluster clustering in the iterative process, i is {1,2,3};
and S1304, returning to S1302 until the maximum iteration number or the mass center updating amplitude is smaller than the threshold value, and finally obtaining the driving sub-databases of the three driving styles.
5. The method for predicting vehicle speeds of drivers in different styles based on a GRU neural network according to claim 1, wherein the training data extracted in S3 includes input data and output data, and wherein the input data includes: the system comprises data information continuously changing along with time and state type information discretely distributed along with time, wherein output data are vehicle speed signals; the data information continuously varying with time includes: historical vehicle speed, historical acceleration, engine speed, engine torque and relative distance between the front vehicles; the state category information distributed discretely over time includes gear information.
6. The method for predicting vehicle speeds of drivers in different styles based on a GRU neural network according to claim 5, wherein the preprocessing in S3 comprises: and carrying out differential and normalization processing on the data information which continuously changes along with time.
7. The method for predicting the vehicle speed of different types of drivers based on the GRU neural network according to claim 5, wherein in S3, after preprocessing the data, 15S is taken as a period, a 10S historical data sequence X＝{X1,X2,X3,X4,X5,X6,X7,X8,X9,X10} is taken as the input of the training process of the GRU vehicle speed prediction model, and a vehicle speed signal value Y= { Y 1,y2,y3,y4,y5 } of 5S is taken as the output of the GRU vehicle speed prediction model; ,Xj＝{x1,x2,x3,x4,x5,x6},x1,x2,x3,x4,x5,x6 respectively represent six characteristics of historical vehicle speed, historical acceleration, engine speed, engine torque, relative distance between front vehicles and gear information; y 1,y2,y3,y4,y5 is the vehicle speed value for the next 5 seconds.
8. The method for predicting the vehicle speed of drivers in different styles based on the GRU neural network according to claim 1, wherein in S5, the determining process of the current driving style is as follows: continuously collecting driving speed information, acceleration information and following distance information in the driving process, and extracting vehicle speed distribution characteristics, acceleration derivative values along with time and average following relative distance values in the driving process; and judging the driving style once every 5s, if the driving style is not changed after the driving style lasts for 1min, considering the driving style to be stable, and introducing a GRU vehicle speed prediction model corresponding to the driving style to predict the vehicle speed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111321247.XA CN113947259B (en) | 2021-11-09 | 2021-11-09 | GRU neural network-based method for predicting speeds of drivers in different styles |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111321247.XA CN113947259B (en) | 2021-11-09 | 2021-11-09 | GRU neural network-based method for predicting speeds of drivers in different styles |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113947259A CN113947259A (en) | 2022-01-18 |
| CN113947259B true CN113947259B (en) | 2024-05-31 |
Family
ID=79337112
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111321247.XA Active CN113947259B (en) | 2021-11-09 | 2021-11-09 | GRU neural network-based method for predicting speeds of drivers in different styles |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113947259B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115879646B (en) * | 2023-02-15 | 2023-11-07 | 山东捷讯通信技术有限公司 | Reservoir water level prediction method, device, medium and equipment |
| CN117349386B (en) * | 2023-10-12 | 2024-04-12 | 吉玖(天津)技术有限责任公司 | Digital humane application method based on data strength association model |
| CN117131955B (en) * | 2023-10-27 | 2024-01-16 | 北京航空航天大学 | Short-time vehicle speed prediction method considering multiple constraint conditions |
| CN117325875B (en) * | 2023-12-01 | 2024-02-02 | 北京航空航天大学 | Vehicle long-term speed prediction method based on individual driving characteristics |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111301426A (en) * | 2020-03-13 | 2020-06-19 | 南通大学 | Method for predicting energy consumption in future driving process based on GRU network model |
| CN111547064A (en) * | 2020-05-26 | 2020-08-18 | 吉林大学 | Driving style recognition and classification method for automobile adaptive cruise system |
| CN112193245A (en) * | 2020-09-24 | 2021-01-08 | 同济大学 | Deep learning following prediction method considering driver fuzzy perception |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101901801B1 (en) * | 2016-12-29 | 2018-09-27 | 현대자동차주식회사 | Hybrid vehicle and method of predicting driving pattern |
-
2021
- 2021-11-09 CN CN202111321247.XA patent/CN113947259B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111301426A (en) * | 2020-03-13 | 2020-06-19 | 南通大学 | Method for predicting energy consumption in future driving process based on GRU network model |
| CN111547064A (en) * | 2020-05-26 | 2020-08-18 | 吉林大学 | Driving style recognition and classification method for automobile adaptive cruise system |
| CN112193245A (en) * | 2020-09-24 | 2021-01-08 | 同济大学 | Deep learning following prediction method considering driver fuzzy perception |
Non-Patent Citations (1)
| Title |
|---|
| 喻丹.机动车驾驶人行为建模及可靠性分析.全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113947259A (en) | 2022-01-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113947259B (en) | GRU neural network-based method for predicting speeds of drivers in different styles | |
| US20220188568A1 (en) | Methods and systems for mining minority-class data samples for training a neural network | |
| Ramadhan et al. | Parameter tuning in random forest based on grid search method for gender classification based on voice frequency | |
| US7362892B2 (en) | Self-optimizing classifier | |
| Chen et al. | Adaptive feature selection-based AdaBoost-KNN with direct optimization for dynamic emotion recognition in human–robot interaction | |
| Tran et al. | Improved PSO for feature selection on high-dimensional datasets | |
| Badawi et al. | A hybrid memetic algorithm (genetic algorithm and great deluge local search) with back-propagation classifier for fish recognition | |
| Zhang et al. | Fast covariance matching with fuzzy genetic algorithm | |
| CN111539444A (en) | Gaussian mixture model method for modified mode recognition and statistical modeling | |
| Abro et al. | Machine learning classifiers: a brief primer | |
| Wan et al. | A hidden semi-Markov model for chart pattern matching in financial time series | |
| Vijay et al. | Deep Dive on Oversampling and Under Sampling Techniques in Machine Learning | |
| US20220284261A1 (en) | Training-support-based machine learning classification and regression augmentation | |
| Senthilkumar et al. | Prediction of rice disease using modified feature weighted fuzzy clustering (MFWFC) based segmentation and hybrid classification model | |
| EP3968214A1 (en) | Device and method for classifying images and accessing the robustness of the classification | |
| CN108549936A (en) | The Enhancement Method that self organizing neural network topology based on deep learning is kept | |
| CN111783866A (en) | Production logistics early warning information multi-classification method based on improved FOA-SVM | |
| CN115273645B (en) | Map making method for automatically clustering indoor surface elements | |
| CN111639688A (en) | Local interpretation method of Internet of things intelligent model based on linear kernel SVM | |
| Gorokhovatskiy et al. | Vector Quantization, Learning and Recognition in the Space of Descriptors of Structural Features of Images | |
| CN110570048A (en) | user demand prediction method based on improved online deep learning | |
| CN118484540B (en) | Text classification method and system based on feature perception | |
| Smart | Genetic programming for multiclass object classification | |
| Polyakova et al. | Investigation of reference sample reduction methods for ensemble output with fuzzy logic-based systems | |
| Rahman et al. | Improvement of Starling Image Classification with Gabor and Wavelet Based on Artificial Neural Network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |