CN-121993591-A - On-line gear shifting method and device for heavy electric vehicle based on reinforcement chart learning frame

CN121993591ACN 121993591 ACN121993591 ACN 121993591ACN-121993591-A

Abstract

The invention discloses a heavy electric vehicle on-line gear shifting method and a device based on an enhanced graph learning frame, which firstly solve the optimal gear shifting track of a full stroke by utilizing a dynamic programming algorithm based on known driving conditions and a vehicle dynamics model in an off-line stage and obtain a time sequence optimal solution of a vehicle state quantity, and then discretizing the optimal track into a series of graph structure samples according to time steps, constructing a graph time sequence data set facing gear shifting decision by taking the optimal middle-rear axle gear position output by dynamic planning as a label, finally designing EE-GAT, coding and time sequence aggregation the graph sequence data in a sliding time window, and outputting the gear decision of the middle axle and the rear axle at the next moment. According to the invention, through coding the key physical coupling relation in the power assembly in the graph structure, the side characteristic enhanced graph attention network can learn the deep association between the running state of the vehicle and the optimal gear shifting logic from the system level, so that the model has stronger generalization and decision stability for the unseen working condition.

Inventors

HE DEQIANG
ZHAO JIAYANG
JIN ZHENZHEN
Lai Zhengqiang
CHEN YANJUN
LI XIANWANG
LI HONGWEI

Assignees

广西大学

Dates

Publication Date: 20260508
Application Date: 20260202

Claims (7)

1. The heavy electric vehicle on-line gear shifting method based on the reinforcement chart learning frame is characterized by comprising the following steps of: S1, solving an optimal gear shifting track of a full stroke by using a dynamic programming algorithm based on a known driving condition and a vehicle dynamics model, and obtaining a time sequence optimal solution of a vehicle state quantity; S2, discretizing the vehicle state quantity into a series of graph structure samples according to time steps, and constructing graph time sequence data oriented to gear shifting decision by taking the optimal middle-rear axle gear output by dynamic programming as a label; s3, processing the graph time sequence data by utilizing a sliding window to obtain a graph data set, introducing an edge feature enhanced graph attention network, coding and time sequence aggregation the graph data set, outputting a gear decision of a middle bridge and a rear bridge at the next moment, and training the network by minimizing the difference between the minimum gear and the dynamic planning optimal gear; s4, a map data set is built by utilizing the vehicle state obtained by sampling the test working condition, and is input into a trained model, and a gear instruction is obtained by fast forward reasoning.
2. The on-line shifting method for the heavy electric vehicle based on the reinforcement map learning framework of claim 1, wherein in step S1, the method specifically comprises the following steps: S101, dispersing the driving working condition into a plurality of stages according to time, and taking step length At each moment in time Defining discrete time state quantities as: Wherein, the Is the longitudinal speed of the vehicle; is the state of charge of the battery; 、 Respectively representing the current gear values of the middle axle and the rear axle; s102, selecting a control quantity of dynamic programming as an absolute gear at the next moment, enumerating all possible combinations of rear axle gears at the next moment, wherein the control quantity is defined as: Wherein, the And Respectively represent Gear values of the middle axle and the rear axle at moment; s103, under the above definition, the discrete-time state equation of the system can be expressed as: Wherein, the Representation of The amount of time state of the moment in time, Representation of A control amount of time; s104, expanding each component of the discrete-time state quantity, which can be expressed as: Wherein, the Is that The longitudinal acceleration of the vehicle at moment is determined by the resultant force and resultant resistance of wheel ends; the battery current is calculated by a battery power balance equation; representing battery capacity; 、 Representing a gear update result; Representation of The amount of time state of the moment in time, Representation of A control amount of time; s105, constructing a stage cost function: Wherein, the Indicating the battery power consumption of the current step, A shift penalty representing energy equivalence; S106, giving initial state x0 and terminal time Under the condition of (1) minimizing cumulative cost To be from the moment The minimum cumulative cost to endpoint, a recurrence relation can be obtained according to the Bellman optimality principle: Wherein, the Representing a control feasible region; Representation of Time state quantity of moment; Representation of A control amount of time; s107, the optimal control law obtained by optimization is as follows: Wherein, the Representation of Cost function of time of day.
3. The on-line shifting method for the heavy electric vehicle based on the reinforcement map learning framework of claim 1, wherein in the step S2, the method specifically comprises the following steps: S201, the moment of time Expressed as a pattern book, can be expressed as: Wherein, the Representing a node set; Representing a set of edges; representing a node feature matrix; representing an edge feature matrix; s202, enabling the vehicle to be at the moment The running state information of (2) is expressed as 6 nodes, and each node constructs a 3-dimensional feature vector to form a node feature matrix: Wherein, the A feature vector representing the current node; s203, forming an edge feature matrix for the structural edge feature, which can be expressed as: Wherein, the Representing slave nodes Pointing node Is used for the correlation strength of the (a).
4. The method for on-line shifting of a heavy electric vehicle based on an enhanced graph learning framework of claim 3, wherein the nodes comprise a center bridge motor node, a rear axle motor node, a center bridge gearbox node, a rear axle gearbox node, a battery node and an environmental node.
5. The on-line shifting method and device for heavy electric vehicles based on the reinforcement map learning framework of claim 1, wherein in step S3, the method specifically comprises the following steps: S301, encoding the constructed pattern book by using a drawing meaning force layer with edge characteristics, wherein the drawing meaning force operation can be expressed as the following formula: Wherein, the Representing nodes In the first place Hidden features of layers, when When the number of the codes is =1, ; Representing a trainable linear transformation matrix; Is a node Is a neighborhood of (a); is the attention weight; s302, attention weight is defined as: Wherein, the Representing nodes Pointing node Is a side feature of (2); transpose of the attention parameter vector; Representing a splicing operation; Representing a trainable linear transformation matrix; Is a node Is a neighborhood of (a); S303, after the schematic force operation, introducing an edge feature enhancement function, which can be expressed as: Wherein, the Representing a layer 2 feed forward network; Representing nodes Hidden features at layer i; Representing nodes Pointing node Is a side feature of (2); s304, obtaining a graph-level embedded vector through global average pooling, wherein the graph-level embedded vector is expressed as: Wherein, the Representing the final state matrix of all nodes; s305, dynamically modeling the graph embedding sequence by adopting a gating circulation unit, wherein the graph embedding sequence can be expressed as follows: Wherein, the For the length of the timing window, Indicating time of day Is a timing hidden state of (a); S306, based on time sequence state The gear classifiers of the middle axle and the rear axle are respectively constructed and expressed as follows: Wherein, the 、、、 Is a trainable parameter; S307, the final output shift decision may be expressed as: And S308, obtaining loss and carrying out back propagation by comparing with a dynamic planning optimal gear shifting sequence, and iteratively updating the model.
6. The on-line shifting method for the heavy electric vehicle based on the reinforcement map learning framework of claim 1, wherein in the step S4, the method specifically comprises the following steps: s401, under a test working condition, acquiring a vehicle state quantity in a preset sampling period, and constructing graph time sequence data; S402, inputting the graph time sequence data into a trained enhanced graph attention network model to obtain a target gear at the current moment; S403, issuing a target gear instruction to a vehicle control unit to execute gear shifting, updating the running state of the vehicle and entering the next sampling period, and repeating the steps S401 to S402.
7. The heavy electric commercial vehicle on-line gear shifting device based on the enhanced graph learning framework is characterized by comprising a data acquisition module, a data processing module, a graph data construction module, a model reasoning module, a parameter updating module and a gear shifting output module; The data acquisition module is used for acquiring running state data of the heavy electric commercial vehicle in real time; The data processing module is used for preprocessing the collected running state data to obtain vehicle state data; The graph data construction module is used for forming graph data which comprises a node characteristic matrix and an edge characteristic matrix and is used for representing the energy flow direction and the coupling logic inside the power system; The model reasoning module is used for inputting the constructed graph data into a gear shifting decision model based on the combination of a graph attention network with a time sequence model, extracting high-dimensional association characteristics among power assembly components through a multi-layer graph attention encoder, modeling dynamic evolution relations of vehicle states of different time steps on a time sequence dimension, and obtaining an optimal gear instruction and a corresponding torque distribution strategy of a middle axle and a rear axle through online reasoning under the condition of giving the running state of the current vehicle; The parameter updating module is used for updating trainable parameters in the graphic neural network and the time sequence model based on historical operation data, simulation comparison results or offline calibration data so as to adapt to state distribution changes under different vehicle configurations, different road working conditions and different driving behaviors; The gear shifting output module is used for receiving the gear instruction and the torque distribution result output by the model reasoning module, checking the safety and the feasibility of the gear shifting instruction and transmitting the gear shifting instruction to the AMT transmission actuating mechanism and the torque control unit through the communication interface of the vehicle controller.

Description

On-line gear shifting method and device for heavy electric vehicle based on reinforcement chart learning frame Technical Field The invention relates to the technical field of bearing residual service life prediction, in particular to an on-line gear shifting method and device for a heavy electric vehicle based on an enhancement chart learning frame. Background In road transportation systems, heavy commercial vehicles take up an extremely important role as the main equipment for logistics transportation. At present, the carbon emission of the heavy commercial vehicle accounts for 54.3% of the total emission of the automobile industry, and the motorized permeability of the heavy commercial vehicle is only 0.7%. The method means that the heavy commercial vehicle is pushed to change from traditional fuel power to electric drive, so that the method has remarkable energy saving and emission reduction potential, and is a key link for realizing carbon neutralization targets in the traffic field. In order to achieve both traction performance and energy consumption performance, most heavy-duty electric commercial vehicles adopt a multi-gear electric drive axle or a multi-gear automatic transmission so as to ensure that the motor can keep high-efficiency running in a wider vehicle speed range. However, currently applied regular shift strategies have difficulty continuously maintaining good energy consumption performance. Although the optimization method can obtain globally optimal gear shifting and torque distribution through dynamic programming and other means, the optimization method depends on a preset working condition, has high calculation cost and cannot meet the real-time requirement of on-line control. Disclosure of Invention Aiming at the problems, the invention provides the on-line shifting method and the device for the heavy electric vehicle based on the enhanced graph learning framework, and the side characteristic enhanced graph attention network can learn the deep association between the vehicle running state and the optimal shifting logic from the system level through encoding the key physical coupling relation in the power assembly in the graph structure, so that the model has stronger generalization and decision stability to the unseen working condition. In order to achieve the above object, the present invention adopts the following technical scheme: according to one aspect of the invention, an on-line shifting method and device for a heavy electric vehicle based on an enhanced graph learning frame are provided, and the method comprises the following steps: S1, solving an optimal gear shifting track of a full stroke by using a dynamic programming algorithm based on a known driving condition and a vehicle dynamics model, and obtaining a time sequence optimal solution of a vehicle state quantity; S2, discretizing the vehicle state quantity into a series of graph structure samples according to time steps, and constructing graph time sequence data oriented to gear shifting decision by taking the optimal middle-rear axle gear output by dynamic programming as a label; s3, processing the graph time sequence data by utilizing a sliding window to obtain a graph data set, introducing an edge feature enhanced graph attention network, coding and time sequence aggregation the graph data set, outputting a gear decision of a middle bridge and a rear bridge at the next moment, and training the network by minimizing the difference between the minimum gear and the dynamic planning optimal gear; s4, a map data set is built by utilizing the vehicle state obtained by sampling the test working condition, and is input into a trained model, and a gear instruction is obtained by fast forward reasoning. Preferably, in step S1, the method specifically includes the following steps: S101, dispersing the driving working condition into a plurality of stages according to time, and taking step length At each moment in timeDefining discrete time state quantities as: Wherein, the Is the longitudinal speed of the vehicle; is the state of charge of the battery; 、 Respectively representing the current gear values of the middle axle and the rear axle; s102, selecting a control quantity of dynamic programming as an absolute gear at the next moment, enumerating all possible combinations of rear axle gears at the next moment, wherein the control quantity is defined as: Wherein, the AndRespectively representGear values of the middle axle and the rear axle at moment; s103, under the above definition, the discrete-time state equation of the system can be expressed as: Wherein, the Representation ofThe amount of time state of the moment in time,Representation ofA control amount of time; s104, expanding each component of the discrete-time state quantity, which can be expressed as: Wherein, the Is thatThe longitudinal acceleration of the vehicle at moment is determined by the resultant force and resultant resistance of wheel ends; the battery