CN113177258A - Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishing method - Google Patents

Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishing method Download PDF

Info

Publication number
CN113177258A
CN113177258A CN202110351401.1A CN202110351401A CN113177258A CN 113177258 A CN113177258 A CN 113177258A CN 202110351401 A CN202110351401 A CN 202110351401A CN 113177258 A CN113177258 A CN 113177258A
Authority
CN
China
Prior art keywords
phase space
driving
model
time
neural network
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.)
Granted
Application number
CN202110351401.1A
Other languages
Chinese (zh)
Other versions
CN113177258B (en
Inventor
蔡英凤
滕成龙
陈龙
王海
孙晓东
熊晓夏
孙晓强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jiangsu University
Original Assignee
Jiangsu University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Jiangsu University filed Critical Jiangsu University
Priority to CN202110351401.1A priority Critical patent/CN113177258B/en
Publication of CN113177258A publication Critical patent/CN113177258A/en
Application granted granted Critical
Publication of CN113177258B publication Critical patent/CN113177258B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/15—Vehicle, aircraft or watercraft design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Abstract

The invention discloses an automatic driving automobile deep neural network prediction model system based on phase space reconstruction and an establishing method. The method comprises the steps of establishing a vehicle dynamics phase space reconstruction model VDPSRM with historical sequence input and future sequence output on the basis of a planar single-track model PSTM, then establishing an acquisition and storage module RVDDM for collecting and storing real vehicle driving data, calculating false nearest neighbor of the driving data sequence along with time delay under each typical driving condition by using a false nearest neighbor calculation module FNNCM, calculating an embedded dimension m and delay time tau of each time step in the vehicle dynamics phase space reconstruction model, and calculating the delay time tau under each typical driving condition by using an average mutual information calculation module AMICMThe time-delay mutual information of the driving data sequence is obtained, the prediction dimension n in the vehicle dynamics phase space reconstruction model is obtained, and the reconstruction mapping G of the DNNM learning phase space reconstruction model of the deep neural network module is constructedrecAnd an automatic driving automobile deep neural network prediction model with good prediction capability on the future state is established.

Description

Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishing method
Technical Field
The invention relates to the technical field of automatic driving automobile application, in particular to a phase space reconstruction-based automatic driving automobile deep neural network prediction model system and an establishing method thereof.
Background
The automatic driving automobile is a high and new technology product with the cross combination of computers, automation and artificial intelligence, and is expected to be capable of being skillfully and safely driven like a human driver, so that the requirement on the human driver is reduced, and the moving cost is obviously reduced. The automatic driving automobile needs to adapt to various complex conditions, including running on a wet and slippery low-friction road, running on a road with variable curves, or performing emergency obstacle avoidance operation when necessary, so as to meet various requirements on stability, safety, comfort and the like.
An autonomous vehicle is a combination of a robotic driver and a vehicle that should have similar motion characteristics to a manned vehicle to enhance the confidence that a typical driver will be a road partner. An important reason for safe human driving is that the automated driving vehicle also needs to have similar characteristics based on the current running state of the vehicle and good prediction ability for future state, which requires the establishment of a prediction model of the automated driving vehicle. The establishment of the predictive model will tightly integrate planning and control to further improve the overall performance of the autonomous vehicle. However, the vehicle dynamics have delay effects and time dependencies, wherein the delay effects exist between input variables and output variables and between parameters of the input variables, and the time dependencies are expressed by the relation of time series data of the input variables and the output variables, and under the joint action of the delay effects and the time series data, a phase space distortion effect is generated, and the effect is strengthened along with the increase of the nonlinear degree of the vehicle motion. How to establish an automatic driving automobile prediction model under the phase space distortion effect becomes an important difficult problem which needs to be solved urgently at present.
Disclosure of Invention
In order to solve the technical problems, the invention provides a phase space reconstruction-based depth neural network prediction model system and an establishment method for an automatic driving automobile, and the method comprises the steps of firstly establishing a vehicle dynamics phase space reconstruction model with historical sequence information input and future sequence information output based on a phase space reconstruction theory; actual vehicle driving data including typical driving conditions and typical road conditions are then collected and stored, the typical driving conditions including: the method comprises the following steps of straight driving, curve driving, annular driving, S-shaped driving and emergency obstacle avoidance driving, wherein typical road conditions comprise: high friction pavement and low friction pavement; then calculating the time delay mutual information of the driving data sequence under the typical driving working condition by using an average mutual information method, and selecting a prediction dimension in a vehicle dynamics phase space reconstruction model; then, calculating false nearest neighbors of the driving data sequence changing along with the embedding dimension under each typical driving working condition by using a false nearest neighbor method, and selecting the embedding dimension and the delay time of each time step in the vehicle dynamics phase space reconstruction model; and finally, constructing a mapping relation of a deep neural network learning vehicle dynamics phase space reconstruction model based on the selected prediction dimension, the embedded dimension and 3 parameters of the delay time of each time step, and testing and verifying the learning effect of the deep neural network.
The invention discloses a technical scheme of an automatic driving automobile deep neural network prediction model system based on phase space reconstruction, which comprises the following steps: the vehicle dynamic phase space reconstruction model system comprises a vehicle dynamic phase space reconstruction model VDPSRM, a real vehicle driving data collection and storage module RVDDM, a false adjacent point calculation module FNNCM, an average mutual information calculation module AMICM and a deep neural network module DNNM.
The vehicle dynamics phase space reconstruction model VDPSRM is a theoretical model which is established under a phase space reconstruction theoretical framework on the basis of a planar single-track model PSTM, is used for guiding the establishment of an automatic driving automobile deep neural network prediction model, and comprises a reconstructed input phase space, a reconstructed output phase space and reconstruction mapping, and further comprises the following reconstruction parameters: embedding dimension m, prediction dimension n, delay time τ per time step.
The planar monorail model PSTM:
Figure BDA0003002259380000021
equation (1) is a nonlinear coupling model, including the parameters: distance a from front wheel to center of mass, distance b from rear wheel to center of mass and rotational inertia IzMass m, front wheel steering angle delta, longitudinal speed UxTransverse velocity UyThe yaw angular velocity r; front wheel transverse force FyfFront wheel longitudinal force FxfRear wheel transverse force FyrDerivative of yaw rate
Figure BDA0003002259380000022
Derivative of lateral velocity
Figure BDA0003002259380000023
Equation (1) can be simplified to the way of vehicle dynamics mapping:
Figure BDA0003002259380000024
in the formula (2), fbikeRepresenting a vehicle dynamics map, XtIs the input of the vehicle model at time t, (C)f,Crμ) is a tire parameter, where CfIs front wheel cornering stiffness, CrIs the rear wheel cornering stiffness, and μ is the tire ground adhesion coefficient.
Introducing phase space reconstruction parameters, and establishing a vehicle dynamics phase space reconstruction model VDPSRM on the basis of the formula (2):
Figure BDA0003002259380000025
in the formula (3) { (X)t,Yt-τ),(Xt-τ,Yt-2τ),…,(Xt-(m-1)τ,Yt-mτ) Is the reconstructed input phase space, (Y)t,Yt+τ,Yt+2τ,…,Yt+nτ) Is the reconstructed output phase space, GrecIs the reconstruction map.
The real vehicle driving data and storage module RVDDM comprises a tire six-component force measuring system, an S-Motion sensor, a steering wheel corner sensor and a driving data recorder, and is used for collecting and storing front wheel longitudinal force FxfYaw rate r and lateral rate uyLongitudinal speed uxTime series data of a vehicle steering angle δ for collecting and storing typical driving condition data including typical road conditions including: high friction and low friction road surfaces, said typical driving conditions comprising: straight running, curve running, annular running, S-shaped running and emergency obstacle avoidance running.
The false adjacent point calculating module FNNCM is used for selecting a minimum embedding dimension m and a delay time tau of each time step, and in an m-dimensional phase space, a coordinate vector y (t) of a one-dimensional time sequence and an r-order nearest neighbor y(r)Square of euclidean distance of (t):
Figure BDA0003002259380000031
in the formula (6), x (t + k τ) is k +1 elements of the coordinate vector y (t), and x(r)(t + k τ) is the nearest neighbor y of order r(r)K +1 elements of (t).
When added to the phase space in the m +1 dimension, the squared euclidean distance becomes:
Figure BDA0003002259380000032
when the one-dimensional time sequence is not embedded properly in the m-dimensional phase space, the Euclidean distance between the coordinate vector y (t) and the nearest neighbor of the R-th order is higher than the threshold value RtolBoth are called false nearest neighbors; when the one-dimensional time series is properly embedded in the m-dimensional phase space, the Euclidean distance is lower than the threshold value RtolM is the minimum embedding dimension, and the corresponding delay time of each time step is obtained simultaneouslyτ。
For an N-dimensional time series, the squared euclidean distance is calculated as follows:
Figure BDA0003002259380000033
Figure BDA0003002259380000034
in the formula (8), xi(t + k τ) is k +1 elements of the coordinate vector of the ith-dimensional time series,
Figure BDA0003002259380000035
is the k +1 elements of the r-th nearest neighbor of the i-th dimensional time series.
When the N-dimensional time sequence is not properly embedded in the m-dimensional phase space, the Euclidean distance is higher than the threshold value RtolBoth are called false nearest neighbors; when the N-dimensional time series are properly embedded in the m-dimensional phase space, the Euclidean distance is lower than the threshold value RtolAnd m is the minimum embedding dimension, and the corresponding delay time tau of each time step is obtained at the same time.
The average mutual information calculation module AMICM is configured to calculate mutual information between the original time sequence and the translated time sequence. Original ith time series xi(t) and shifted ith time series xi(t-τi) Mutual information of I (x)i(t),xi(t-τi))：
Figure BDA0003002259380000041
In the formula (4), pjk(τi) Is a joint probability distribution, pjAnd pkIs the edge probability distribution, τiIs the delay time of the ith time series.
Average mutual information of the N-dimensional time series:
Figure BDA0003002259380000042
when the average mutual information IAMICMBelow the threshold 1/e (e is a natural constant), the corresponding average delay time divided by the delay time τ of each time step is the prediction dimension n.
And the deep neural network module DNNM is used for realizing reconstruction mapping from the reconstructed input phase space to the reconstructed output phase space. The deep neural network module DNNM learns the input-output relationship of the driving data through the combined structure of the neurons described by the following formula, so as to realize the reconstruction mapping relationship:
Figure BDA0003002259380000043
in the formula (10), XtIs the input vector of the neural network element, htIs an output vector, ctIs a memory state vector, ft、itAnd otActivation vectors, σ, for the forgetting gate, the input gate and the output gate, respectivelyc、σgAnd σhFor the corresponding activation function, W and U are weight matrices, b is a bias vector, and x is a vector multiplication.
The invention provides a method for establishing an automatic driving automobile deep neural network prediction model based on phase space reconstruction, which adopts the technical scheme that the method comprises the following steps:
step 1) carrying out simplified equivalence on an automatic driving automobile to obtain a planar single-track model PSTM;
step 2) introducing a phase space reconstruction parameter, and establishing a vehicle dynamics phase space reconstruction model VDPSRM on the basis of the planar single-track model PSTM, wherein the vehicle dynamics phase space reconstruction model VDPSRM comprises a reconstructed input phase space, a reconstructed output phase space and a reconstructed mapping;
step 3) establishing a real vehicle driving data collecting and storing module RVDDM, and collecting typical driving condition data under typical road conditions, wherein the typical road conditions comprise: high friction and low friction road surfaces, said typical driving conditions comprising: straight running, curve running, annular running, S-shaped running and emergency obstacle avoidance running;
step 4), a false adjacent point calculation module FNNCM is used for solving the minimum embedding dimension m and the corresponding delay time tau of each time step;
step 5) calculating average mutual information I of typical driving condition time sequence by using average mutual information calculation module AMICMAMICMSolving a prediction dimension n;
step 6) based on the reconstruction parameters: embedding the dimension m, the prediction dimension n and the delay time tau of each time step, and constructing a deep neural network module DNNM to realize the reconstruction mapping G from the reconstructed input phase space to the reconstructed output phase spacerec；
The invention is based on a phase space reconstruction theory, establishes a vehicle dynamics phase space reconstruction model VDPSRM with historical sequence information input and future sequence information output on the basis of a planar monorail model PSTM, establishes a real vehicle driving data acquisition and storage module RVDDM, the method is used for collecting and storing real vehicle driving data comprising typical driving conditions and typical road conditions, a false nearest neighbor calculation module FNNCM is used for calculating false nearest neighbors of a driving data sequence along with time delay under each typical driving condition, an embedding dimension m and delay time tau of each time step in a vehicle dynamics phase space reconstruction model are obtained, time delay mutual information of the driving data sequence under each typical driving condition is calculated by an average mutual information calculation module AMICM, a prediction dimension n in the vehicle dynamics phase space reconstruction model is obtained, and a depth neural network module DNNM is constructed to learn a reconstruction mapping G of the vehicle dynamics phase space reconstruction model.recAnd an automatic driving automobile deep neural network prediction model with good prediction capability on the future state is established.
The invention has the beneficial effects that:
1. the method constructs a vehicle dynamic phase space reconstruction model VDPSRM of the automatic driving vehicle, expands the automatic driving vehicle model to a high-dimensional phase space on the basis of a planar single-track model PSTM, and solves the problem of space-time dynamic depiction of the automatic driving vehicle dynamic model;
2. the method constructs a real vehicle driving data collection and storage module RVDDM, and provides a basic basis for obtaining the reconstruction parameters in the vehicle dynamics phase space reconstruction model;
3. the method for acquiring the reconstruction parameters in the vehicle dynamics phase space reconstruction model is established, the prediction dimension n is calculated by using an average mutual information method, the embedding dimension m and the delay time tau of each time step are calculated by using a false adjacent point method, and the practicability and the scientificity of the vehicle dynamics phase space reconstruction model are improved;
4. the invention constructs a method and a device for establishing an automatic driving automobile deep neural network prediction model based on phase space reconstruction, and constructs a reconstruction mapping G of a deep neural network module DNNM learning vehicle dynamics phase space reconstruction modelrecThe method can establish an automatic driving automobile prediction model under the phase space distortion effect caused by the time delay effect and the time dependency, and lays a good foundation for the safe driving and control strategy design of the automatic driving automobile.
Drawings
FIG. 1 is a simplified equivalent diagram of a planar single-track model PSTM for an autonomous vehicle.
Fig. 2 is a vehicle dynamics phase space reconstruction model VDPSRM diagram of an autonomous vehicle.
Figure 3 is a diagram of the real vehicle driving data collection and storage module RVDDM.
Fig. 4 is a flowchart of reconstruction parameter calculation.
FIG. 5 is a diagram of a method and apparatus for establishing a phase space reconstruction based prediction model for an autopilot deep neural network.
Detailed Description
The invention will be further explained with reference to the drawings.
As shown in fig. 1, in a first step, the autonomous vehicle goes through a simplified equivalent to build a planar single-track model PSTM, which is a nonlinear coupled model, comprising the parameters: distance a from front wheel to center of mass, distance b from rear wheel to center of mass and rotational inertia IzMass m, front wheel steering angle delta, longitudinal speed UxTransverse velocity UyThe yaw angular velocity r;front wheel transverse force FyfFront wheel longitudinal force FxfRear wheel transverse force FyrDerivative of yaw rate
Figure BDA0003002259380000061
Derivative of lateral velocity
Figure BDA0003002259380000062
Second, the planar single-track model PSTM is further reduced to a vehicle dynamics mapping mode, wherein fbikeRepresenting a vehicle dynamics map, XtIs the input of the vehicle model at time t, (C)f,Crμ) is a tire parameter;
as shown in fig. 2, the vehicle dynamics phase space reconstruction model VDPSRM is built by introducing phase space reconstruction parameters based on the vehicle dynamics mapping, wherein { (X)t,Yt-τ),(Xt-τ,Yt-2τ),…,(Xt-(m-1)τ,Yt-mτ) Is the reconstructed input phase space, (Y)t,Yt+τ,Yt+2τ,…,Yt+nτ) Is the reconstructed output phase space, GrecIs a reconstruction map;
as shown in fig. 3, the real vehicle driving data collection and storage module RVDDM collects and stores data including typical driving conditions on a high friction road surface and a low friction road surface, respectively, the typical driving conditions including: straight running, curve running, annular running, S-shaped running and emergency obstacle avoidance running;
as shown in fig. 4, the reconstruction parameter calculation process includes an embedding dimension m, a prediction dimension n, and a delay time τ of each time step, and the calculation process is divided into two aspects: on one hand, the average mutual information calculation module AMICM obtains typical driving condition data from the real vehicle driving data collection and storage module RVDDM, and calculates average mutual information IAMICMSolving a prediction dimension n; on the other hand, the false adjacent point calculating module FNNCM obtains typical driving condition data from the real vehicle driving data collecting and storing module RVDDM, calculates the square of the Euclidean distance of the r-order nearest neighbor and the coordinate vector y (t),and a threshold value RtolComparing, and solving the minimum embedding dimension m and the delay time tau of each time step;
as shown in FIG. 5, the automatic driving automobile deep neural network prediction model system based on phase space reconstruction comprises a vehicle dynamics phase space reconstruction model VDPSRM, a real vehicle driving data collection and storage module RVDDM, an average mutual information calculation module AMICM, a false adjacent point calculation module FNNCM and a deep neural network module DNNM.
A vehicle dynamics phase space reconstruction model VDPSRM is a theoretical model which is built under a phase space reconstruction theoretical framework on the basis of a planar single-track model PSTM and is used for guiding the building of an automatic driving automobile deep neural network prediction model, and comprises a reconstructed input phase space { (X)t,Yt-τ),(Xt-τ,Yt-2τ),…,(Xt-(m-1)τ,Yt-m) Reconstructed output phase space (Y)t,Yt+τ,Yt+2τ,…,Yt+nτ) Reconstructing the mapping GrecThe embedded dimension m, the predicted dimension n and the delay time tau of each time step are reconstruction parameters, and the reconstruction parameters are obtained by an average mutual information calculation module AMICM and a false adjacent point calculation module FNNCM according to the data of a real vehicle driving data collection and storage module RVDDM;
the real vehicle driving data collecting and storing module RVDDM comprises a tire six-component force measuring system, an S-Motion sensor, a steering wheel corner sensor and a driving data recorder, and collects and stores front wheel longitudinal force FxfYaw rate r and lateral rate uyLongitudinal speed uxTime series data of the vehicle steering angle delta are sent to an average mutual information calculation module AMICM, a false adjacent point calculation module FNNCM and a deep neural network module DNNM;
the false adjacent point calculating module FNNCM receives typical driving condition data under the typical road condition from the real vehicle driving data collecting and storing module RVDDM, calculates the square of the Euclidean distance and compares the square with a threshold value RtolComparing, and obtaining the minimum embedding dimension m and the corresponding delay time tau of each time step for reconstructing an input phase space;
an average mutual information calculation module AMICM for receiving the typical driving condition data under the typical road condition from the real vehicle driving data collection and storage module RVDDM and calculating the average mutual information IAMICMSolving a prediction dimension n for reconstructing an output phase space;
the deep neural network module DNNM receives the driving data from the real vehicle driving data collection and storage module RVDDM, and learns the input-output relationship of the driving data through the combined structure of neurons, so that the mapping relationship is reconstructed:
the finally formed automatic driving automobile deep neural network prediction model system based on phase space reconstruction comprises a vehicle dynamics phase space reconstruction model VDPSRM, a real vehicle driving data collection and storage module RVDDM, a false adjacent point calculation module FNNCM, an average mutual information calculation module AMICM and a deep neural network module DNNM.
A method for establishing an automatic driving automobile deep neural network prediction model based on phase space reconstruction comprises the following steps:
step 1) simplifying equivalence is carried out on an automatic driving automobile to obtain a planar single-track model PSTM, and then the simplified equivalent is further simplified into a vehicle dynamics mapping mode;
step 2) establishing a vehicle dynamics phase space reconstruction model VDPSRM based on a phase space reconstruction theory, wherein the model comprises a reconstructed input phase space { (X)t,Yt-τ),(Xt-τ,Yt-2τ),…,(Xt-(m-1)τ,Yt-mτ) Reconstructed output phase space (Y)t,Yt+τ,Yt+2τ,…,Yt+nτ) Reconstructing the mapping GrecEmbedding dimension m, predicting dimension n and delay time tau of each time step as reconstruction parameters for determining the structure of a phase space;
step 3) establishing a real vehicle driving data collecting and storing module RVDDM, and collecting typical driving condition data under typical road conditions, wherein the typical road conditions comprise: high friction and low friction road surfaces, said typical driving conditions comprising: the method comprises the following steps of straight line driving, curve driving, annular driving, S-shaped driving and emergency obstacle avoidance driving, and specifically comprises the following steps:
(A) collecting and storing time series data of straight running, curve running, annular running, S-shaped running and emergency obstacle avoidance running on a high-friction road surface;
(B) collecting and storing time series data of straight running, curve running, annular running, S-shaped running and emergency obstacle avoidance running on a low-friction road surface;
step 4) using a false adjacent point calculation module FNNCM to obtain the minimum embedding dimension m and the corresponding delay time tau of each time step, and the specific steps are as follows:
(A) in m-dimensional phase space, calculating the square of Euclidean distances of the time series and the nearest neighbor of r order;
(B) in an m +1 dimensional phase space, calculating the square of the Euclidean distance between the time sequence and the nearest neighbor of the r-th order;
(C) when the Euclidean distance is larger than the threshold value RtolWhen the embedding dimension is increased by 1 and then returned to (A), when the Euclidean distance is less than or equal to the threshold value RtolThen m is the minimum embedding dimension, and the corresponding delay time tau of each time step is obtained at the same time;
step 5) calculating average mutual information I of typical driving condition time sequence by using average mutual information calculation module AMICMAMICMAnd solving a prediction dimension n, and specifically comprising the following steps:
(A) calculating the mutual information of each typical driving condition time sequence and each time delay time sequence in turn;
(B) calculating average mutual information I of mutual information of all time sequencesAMICM；
(C) Calculating to obtain a prediction dimension n according to the average delay time;
step 6) based on the reconstruction parameters: embedding the dimension m, the prediction dimension n and the delay time tau of each time step, and constructing a deep neural network module DNNM to realize the reconstruction mapping G from the reconstructed input phase space to the reconstructed output phase spacerecThe method comprises the following specific steps:
(A) constructing a DNNM structure of a deep neural network module;
(B) learning the input and output mapping relation of data in the real vehicle driving data collection and storage module RVDDM;
(C) verifying and testing the effect of the DNNM;
the specific embodiment of the invention: a vehicle dynamics phase space reconstruction model VDPSRM is formed by using a mathematical model, a real vehicle driving data collection and storage module RVDDM is formed by using a tire six-component force measurement system, an S-Motion sensor, a steering wheel corner sensor and a driving data recorder, a false adjacent point calculation module FNNCM and an average mutual information calculation module AMICM are compiled by using MATLAB/Simulink, a deep neural network module DNNM is compiled by using Python, learning, verification and testing of a mapping relation from an input phase space to an output phase space are realized in a Pythrch frame, and an automatic driving automobile deep neural network prediction model system based on phase space reconstruction and an establishing method are realized.
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 (10)

1. An automatic driving automobile deep neural network prediction model system based on phase space reconstruction is characterized by comprising a vehicle dynamics phase space reconstruction model VDPSRM, a real vehicle driving data collection and storage module RVDDM, a false adjacent point calculation module FNNCM, an average mutual information calculation module AMICM and a deep neural network module DNNM;
the vehicle dynamics phase space reconstruction model VDPSRM is a theoretical model established under a phase space reconstruction theoretical framework on the basis of a planar single-track model PSTM and is used for guiding the establishment of an automatic driving automobile deep neural network prediction model;
the real vehicle driving data and storage module RVDDM is used for collecting and storing typical driving condition data including typical road conditions, comprises a tire six-component force measuring system, an S-Motion sensor, a steering wheel angle sensor and a driving data recorder, and collects and stores front wheel longitudinal force FxfYaw rate r, yawTo velocity uyLongitudinal speed uxTime series data of the vehicle steering angle delta are sent to an average mutual information calculation module AMICM, a false adjacent point calculation module FNNCM and a deep neural network module DNNM;
the false adjacent point calculating module FNNCM is used for selecting the minimum embedding dimension m and the delay time tau of each time step;
the average mutual information calculation module AMICM is configured to calculate mutual information between an original time sequence and a translated time sequence;
and the deep neural network module DNNM is used for realizing reconstruction mapping from the reconstructed input phase space to the reconstructed output phase space.
2. The system of claim 1, wherein the PSTM is a nonlinear coupled model of the following:
Figure FDA0003002259370000011
the parameters are as follows: distance a from front wheel to center of mass, distance b from rear wheel to center of mass and rotational inertia IzMass m, front wheel steering angle delta, longitudinal speed UxTransverse velocity UyThe yaw angular velocity r; front wheel transverse force FyfFront wheel longitudinal force FxfRear wheel transverse force FyrDerivative of yaw rate
Figure FDA0003002259370000012
Derivative of lateral velocity
Figure FDA0003002259370000013
The planar single-track model PSTM is simplified into a vehicle dynamics mapping model:
Figure FDA0003002259370000021
in the formula (2), fbikeRepresenting a vehicle dynamics map, XtIs the input of the vehicle model at time t, (C)f，Crμ) is a tire parameter, where CfIs front wheel cornering stiffness, CrIs the rear wheel cornering stiffness, and μ is the tire ground adhesion coefficient.
3. The system of claim 2, wherein the vehicle dynamics phase space reconstruction model VDPSRM comprises a reconstructed input phase space, a reconstructed output phase space and a reconstruction map, and further comprises the following reconstruction parameters: embedding dimension m, predicting dimension n, and delay time tau of each time step;
the model is a vehicle dynamics phase space reconstruction model VDPSRM established on the basis of an equation (2) by introducing phase space reconstruction parameters:
Figure FDA0003002259370000022
in the formula (3) { (X)t，Yt-τ)，(Xt-τ，Yt-2τ)，…，(Xt-(m-1)τ，Yt-mτ) Is the reconstructed input phase space, (Y)t，Yt+τ，Yt+2τ，…，Yt+nτ) Is the reconstructed output phase space, GrecIs the reconstruction map.
4. The system of claim 1, wherein the typical road conditions comprise: high friction and low friction road surfaces, said typical driving conditions comprising: straight running, curve running, annular running, S-shaped running and emergency obstacle avoidance running.
5. The system of claim 1, wherein the FNNCM is designed as follows:
in m-dimensional phase space, coordinate vector y (t) and r-order nearest neighbor y of one-dimensional time series(r)The squared euclidean distance of (t) is as follows:
Figure FDA0003002259370000023
in the formula (4), x (t + k τ) is k +1 elements of the coordinate vector y (t), and x(r)(t + k τ) is the nearest neighbor y of order r(r)K +1 elements of (t).
When added to the phase space in the m +1 dimension, the squared euclidean distance becomes:
Figure FDA0003002259370000031
when the one-dimensional time series has not been properly embedded in the m-dimensional phase space, the coordinate vector y (t) and the nearest neighbor y of order r(r)(t) Euclidean distance above threshold RtolBoth are called false nearest neighbors; when the one-dimensional time series is properly embedded in the m-dimensional phase space, the Euclidean distance is lower than the threshold value RtolAnd m is the minimum embedding dimension, and the corresponding delay time tau of each time step is obtained at the same time.
For an N-dimensional time series, the squared euclidean distance is calculated as follows:
Figure FDA0003002259370000032
Figure FDA0003002259370000033
in the formula (6), xi(t + k τ) is k +1 coordinate vectors of the ith-dimensional time seriesThe elements are selected from the group consisting of,
Figure FDA0003002259370000034
is the k +1 elements of the r-th nearest neighbor of the i-th dimensional time series.
When the N-dimensional time sequence is not properly embedded in the m-dimensional phase space, the Euclidean distance is higher than the threshold value RtolBoth are called false nearest neighbors; when the N-dimensional time series are properly embedded in the m-dimensional phase space, the Euclidean distance is lower than the threshold value RtolAnd m is the minimum embedding dimension, and the corresponding delay time tau of each time step is obtained at the same time.
6. The system of claim 1, wherein the average mutual information calculation module AMICM is designed as follows:
let the original ith time series xi(t) and shifted ith time series xi(t-τi) Mutual information of is I (x)i(t)，xi(t-τi) Calculated by the formula:
Figure FDA0003002259370000035
in the formula (8), pjk(τi) Is a joint probability distribution, pjAnd pkIs the edge probability distribution, τiIs the delay time of the ith time series;
average mutual information of the N-dimensional time series:
Figure FDA0003002259370000036
when the average mutual information IAMICMBelow the threshold 1/e (e is a natural constant), the corresponding average delay time divided by the delay time τ of each time step is the prediction dimension n.
7. The system of claim 1, wherein the DNNM learns the input-output relationship of the driving data through the combined structure of neurons described by the following formula, so as to realize the reconstructed mapping relationship:
Figure FDA0003002259370000041
in the formula (10), XtIs the input vector of the neural network element, htIs an output vector, ctIs a memory state vector, ft、itAnd otActivation vectors, σ, for the forgetting gate, the input gate and the output gate, respectivelyc、σgAnd σhFor the corresponding activation function, W and U are weight matrices, b is a bias vector, and x is a vector multiplication.
8. The method for establishing the automatic driving automobile deep neural network prediction model based on the phase space reconstruction as claimed in any one of claims 1 to 7, characterized in that a vehicle dynamics phase space reconstruction model of historical sequence information input and future sequence information output is established based on a phase space reconstruction theory; then collecting and storing actual vehicle driving data including typical driving conditions and typical road conditions; then calculating the time delay mutual information of the driving data sequence under the typical driving working condition by using an average mutual information method, and selecting a prediction dimension in a vehicle dynamics phase space reconstruction model; then, calculating false nearest neighbors of the driving data sequence changing along with the embedding dimension under each typical driving working condition by using a false nearest neighbor method, and selecting the embedding dimension and the delay time of each time step in the vehicle dynamics phase space reconstruction model; and finally, constructing a mapping relation of a deep neural network learning vehicle dynamics phase space reconstruction model based on the selected prediction dimension, the embedded dimension and 3 parameters of the delay time of each time step, and testing and verifying the learning effect of the deep neural network.
9. The method for establishing the neural network prediction model of the autopilot based on the phase space reconstruction as claimed in claim 8, is characterized in that the method is implemented by the following steps:
s1, simplifying and equivalence are carried out on the automatic driving automobile to obtain a plane single-track model PSTM, and then the model is further simplified into a vehicle dynamics mapping mode;
s2, establishing a vehicle dynamic phase space reconstruction model VDPSRM based on a phase space reconstruction theory, wherein the model comprises a reconstructed input phase space { (X)t，Yt-τ)，(Xt-τ，Yt-2τ)，…，(Xt-(m-1)τ，Yt-mτ) Reconstructed output phase space (Y)t，Yt+τ，Yt+2τ，…，Yt+nτ) Reconstructing the mapping GrecEmbedding dimension m, predicting dimension n and delay time tau of each time step as reconstruction parameters for determining the structure of a phase space;
s3 builds a real vehicle driving data collection and storage module RVDDM, collecting typical driving condition data under typical road conditions, including: high friction and low friction road surfaces, said typical driving conditions comprising: straight running, curve running, annular running, S-shaped running and emergency obstacle avoidance running;
s4, calculating the minimum embedding dimension m and the corresponding delay time tau of each time step by using a false adjacent point calculation module FNNCM;
s5 average mutual information I of typical driving condition time sequence is calculated by using average mutual information calculation module AMICMAMICMSolving a prediction dimension n;
s6 based on the reconstruction parameters: embedding the dimension m, the prediction dimension n and the delay time tau of each time step, and constructing a deep neural network module DNNM to realize the reconstruction mapping G from the reconstructed input phase space to the reconstructed output phase spacerec。
10. The method for building the neural network prediction model of the autopilot based on the phase space reconstruction as claimed in claim 9,
the specific steps of S3 are as follows:
s3.1, collecting and storing time sequence data of straight line driving, curve driving, annular driving, S-shaped driving and emergency obstacle avoidance driving on a high-friction road surface;
s3.2, collecting and storing time sequence data of straight line driving, curve driving, annular driving, S-shaped driving and emergency obstacle avoidance driving on a low-friction road surface;
the specific steps of S4 are as follows:
s4.1, calculating the square of the Euclidean distance between the time sequence and the nearest neighbor of the r-order in an m-dimensional phase space;
s4.2, calculating the square of the Euclidean distance between the time sequence and the nearest neighbor of the r-order in the m + 1-dimensional phase space;
s4.3 when the Euclidean distance is larger than the threshold value RtolWhen the embedding dimension is increased by 1 and then returned to S4.1, when the Euclidean distance is less than or equal to the threshold value RtolThen m is the minimum embedding dimension, and the corresponding delay time tau of each time step is obtained at the same time;
the specific steps of S5 are as follows:
s5.1, sequentially calculating mutual information of each typical driving condition time sequence and each time delay time sequence;
s5.2 calculating the average mutual information I of all time series mutual informationAMICM；
S5.3, calculating according to the average delay time to obtain a prediction dimension n;
the specific steps of S6 are as follows:
s6.1, constructing a DNNM structure of the deep neural network module;
s6.2, learning the input and output mapping relation of data in the real vehicle driving data collection and storage module RVDDM;
s6.3, verifying and testing the effect of the DNNM module.
CN202110351401.1A 2021-03-31 2021-03-31 Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishment method Active CN113177258B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110351401.1A CN113177258B (en) 2021-03-31 2021-03-31 Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishment method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110351401.1A CN113177258B (en) 2021-03-31 2021-03-31 Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishment method

Publications (2)

Publication Number Publication Date
CN113177258A true CN113177258A (en) 2021-07-27
CN113177258B CN113177258B (en) 2024-02-20

Family

ID=76922994

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110351401.1A Active CN113177258B (en) 2021-03-31 2021-03-31 Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishment method

Country Status (1)

Country Link
CN (1) CN113177258B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104881563A (en) * 2015-01-28 2015-09-02 柳州师范高等专科学校 Chaotic characteristic analysis and non-linear prediction method of run-off
US20190114557A1 (en) * 2017-10-18 2019-04-18 Nxgen Partners Ip, Llc Unified nonlinear modeling appoach for machine learning and artificial intelligence (attractor assisted ai)
CN111260124A (en) * 2020-01-11 2020-06-09 大连理工大学 Chaos time sequence prediction method based on attention mechanism deep learning

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104881563A (en) * 2015-01-28 2015-09-02 柳州师范高等专科学校 Chaotic characteristic analysis and non-linear prediction method of run-off
US20190114557A1 (en) * 2017-10-18 2019-04-18 Nxgen Partners Ip, Llc Unified nonlinear modeling appoach for machine learning and artificial intelligence (attractor assisted ai)
CN111260124A (en) * 2020-01-11 2020-06-09 大连理工大学 Chaos time sequence prediction method based on attention mechanism deep learning

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
丁明 等: "基于多变量相空间重构和RBF神经网络的光伏功率预测方法", 电子测量与仪器学报, no. 08 *
付学良 等: "基于混沌―支持向量机的大气污染物浓度预测模型", 电子世界, no. 04 *
刘林刚 等: "基于支持向量机的航空发动机振动预测模型研究", 微计算机信息, no. 16 *
王军栋 等: "基于相空间重构的RBF江水浊度预报研究", 系统仿真学报, no. 024 *
蔡英凤 等: "基于注意力机制的车辆行为预测", 江苏大学学报(自然科学版), no. 02 *

Also Published As

Publication number Publication date
CN113177258B (en) 2024-02-20

Similar Documents

Publication Publication Date Title
Sathyamoorthy et al. Terrapn: Unstructured terrain navigation using online self-supervised learning
EP2591443B1 (en) Method for assisting vehicle guidance over terrain
Acosta et al. Teaching a vehicle to autonomously drift: A data-based approach using neural networks
CN112762957B (en) Multi-sensor fusion-based environment modeling and path planning method
CN110687907B (en) Intelligent automobile decoupling prediction controller based on model dynamic reconstruction and control method thereof
Wang et al. From agile ground to aerial navigation: Learning from learned hallucination
KR102425657B1 (en) Mobile robot capable of evaluating self-training based traversability
Kullberg et al. Online joint state inference and learning of partially unknown state-space models
CN115303289A (en) Vehicle dynamics model based on depth Gaussian, training method, intelligent vehicle trajectory tracking control method and terminal equipment
Visca et al. Conv1D energy-aware path planner for mobile robots in unstructured environments
CN113341994B (en) Intelligent automobile path tracking control method based on optimal control of piecewise affine system
Sadhukhan et al. Terrain estimation using internal sensors
Kruger et al. Estimating wheel slip of a planetary exploration rover via unsupervised machine learning
Masmoudi et al. Autonomous car-following approach based on real-time video frames processing
Sebastian et al. Neural network based heterogeneous sensor fusion for robot motion planning
CN117376920A (en) Intelligent network connection automobile network attack detection, safety state estimation and control method
CN110926470B (en) AGV navigation control method and system
Hou et al. Crash mitigation controller for unavoidable T-bone collisions using reinforcement learning
Joo et al. Real time traversability analysis to enhance rough terrain navigation for an 6× 6 autonomous vehicle
Liang et al. Vehicle side-slip angle estimation with deep neural network and sensor data fusion
CN113177258A (en) Automatic driving automobile deep neural network prediction model system based on phase space reconstruction and establishing method
Mohamed-Ahmed et al. Non-linear control based on state estimation for the convoy of autonomous vehicles
Dallas et al. Neural network based terramechanics modeling and estimation for deformable terrains
Liu et al. Slip estimation for autonomous tracked vehicles via machine learning
Gräber et al. Application of Neural Networks to External Parameter Estimation for Nonlinear Vehicle Models

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