CN111832225B - Method for constructing driving condition of automobile - Google Patents

Abstract

The invention relates to the field of automobile working condition data construction, in particular to a method for constructing automobile driving working conditions, which comprises the following steps: acquiring original GPS data of automobile driving, and preprocessing; dividing the preprocessed data into kinematic fragments; performing feature calculation on the kinematics segment to obtain feature parameters of the kinematics segment; dividing the kinematic fragments into four fragment libraries by adopting K-Means clustering; constructing a training data set; inputting the training data set into a model for training to obtain a trained long-short term memory neural network model; predicting by using the trained long and short term memory neural network model to obtain time-speed prediction curves corresponding to the four segment libraries respectively; and combining the curves of the four speed sections into a working condition curve. The method effectively identifies the implicit characteristics in the automobile driving data in the special region through the LSTM network, thereby constructing the automobile driving condition curve according with the driving characteristics.

Description

A method for constructing vehicle driving conditions

technical field

The invention relates to the field of construction of automobile operating condition data, in particular to a method for constructing automobile operating condition.

Background technique

Driving Cycle, also known as the vehicle test cycle, is the speed-time curve that describes the driving of the car. Generally, the total time is within 1800 seconds, but there is no limit standard. It can reflect the kinematic characteristics of the car on the road. An important and common basic technology, it is the basis of vehicle energy consumption/emission test methods and limit standards, and is also the main benchmark for the calibration and optimization of various performance indicators of automobiles. At present, developed countries such as Europe, the United States, Japan, etc., adopt the standards adapted to their own driving conditions for vehicle performance calibration optimization and energy consumption/emission certification.

At the beginning of this century, my country directly adopted the European NEDC driving conditions to certify the energy consumption/emissions of automobile products, but the two most important characteristics of the working conditions, the idling time ratio and the average speed, are different from the actual driving conditions of our country. The difference is huge. As the most basic basis for vehicle development and evaluation, it is becoming more and more important to carry out in-depth research and formulate test conditions that reflect the actual road driving conditions in our country. At the same time, our country has a vast territory, and the development degree, climatic conditions and traffic conditions of each city are different, which makes the characteristics of automobile driving conditions in each city significantly different. Therefore, it is more and more urgent to conduct research on the construction of urban vehicle driving conditions based on the city's own vehicle driving data. Therefore, it is necessary to construct the vehicle driving conditions of each city according to the different road conditions of each city.

Current traditional construction methods involve the use of methods such as fuzzy cluster analysis, Markov, etc. The fuzzy clustering analysis method basically finds representative segments in the kinematics segment library and combines them to form a working condition curve close to the actual situation. Whether the working condition curve is representative or not depends largely on the search for the best segment. In addition, such methods cannot effectively estimate and extract the hidden information in all the data, resulting in the selection of only a common segment that can represent all routes. The Markov method uses the state transition probability to obtain the final road driving conditions according to the influence of the current data on the future data. Although the representative working conditions of a specified duration can be randomly generated, the result is more dependent on the accuracy of the state transition probability, and The previous data that can be retained in the new state has less information, so the new forecast data can generalize the future situation less than the previous data.

Contents of the invention

In order to solve the above problems, the present invention provides a method for constructing driving conditions of a vehicle. Based on the collected driving data that fits a special region and area, the method deeply learns the hidden features in the driving data, thereby extracting effective information, predicting the working condition curve of a given interval segmentally, and finally constructing a Complete vehicle driving cycle curves of regional characteristics.

A method for constructing a vehicle driving condition, comprising the following steps:

Obtain the original GPS data of the car, and preprocess the original GPS data of the car;

The short-stroke division method is used to divide the preprocessed data into kinematic segments;

Carry out feature calculation on the kinematics segment to obtain the characteristic parameters of the kinematics segment, and use the principal component analysis method to filter irrelevant features to obtain effective feature parameters;

K-Means clustering is used to divide the kinematics fragments into four fragment libraries, namely: low-speed interval fragment library, medium-speed interval fragment library, high-number interval fragment library and extremely high-speed interval fragment library;

Construct the training data set: splice all the kinematic fragments in each fragment library to obtain four long fragments, and use the four long fragments as the training data set;

Inputting the above-mentioned training data set into the long-term short-term memory neural network model for training, and obtaining the trained long-term short-term memory neural network model model;

Use the trained long-short-term memory neural network model to predict and obtain the time-speed prediction curves corresponding to the four fragment libraries. The specific process includes: taking the last sample data of the training data set as the first input element and inputting it into the training In a good long-short-term memory neural network model, the first prediction sequence is output; the first input element is deleted, and the first prediction value is used as the second input element, and the input model obtains the second prediction sequence; and so on, a fragment library is finally obtained The predicted sequences of the four fragment libraries are respectively corresponding to the time-speed prediction curves;

After obtaining the time-velocity prediction curves corresponding to the four fragment libraries, according to the time ratios of the four fragment libraries in the entire kinematics fragment, determine the time occupied by the four types of fragment libraries in the final working condition synthesis, Merge the curves of four speed sections into one working condition curve;

The working condition curve is sent to the control device, and the control device evaluates the exhaust emission of the vehicle and evaluates the environmental protection level according to the working condition curve.

Further, the preprocessing of the raw GPS data of the car includes:

Carry out traversal search on the original GPS data of car driving from the beginning, find the first time breakpoint, divide the original GPS data into different driving segments from the first time breakpoint, and the first time breakpoint is the original time breakpoint of car driving Areas with a time interval greater than 55 seconds in the GPS data;

Judging whether there is a second time breakpoint in the obtained driving segment, if there is a second time breakpoint, a series of new speed data points are fitted by an improved polynomial fitting method based on the speed data before and after the second time breakpoint , supplementing the second time breakpoint inside the driving segment, the second time breakpoint is an area where the time interval is greater than 2 seconds and less than or equal to 55 seconds in the original GPS data of the car driving;

After the data fitting is completed, the acceleration of each time point of each driving segment is calculated, and the driving segment with abnormal acceleration is eliminated from the data according to the abnormal acceleration screening rules;

For the abnormal data of long-term idle speed greater than 180 seconds, use a sliding window with a size of 180 to slide the time and vehicle speed of each segment, and the sliding step is 1s. During the window sliding process, if all the data in the window are For idle data, filter out the first piece of data in the window; when the end of the window slides to the end of the driving segment, if all the data in the window are idle data, all the data in the window will be filtered out, and so on All driving segments are screened out to obtain preprocessed data.

Further, using the short-stroke division method to divide the preprocessed data into kinematic segments includes: first judging whether the driving time of each driving segment is greater than 20s, if it is less than 20s, then rejecting the driving segment; if it is greater than 20s, then Find the kinematics segment from the driving segment according to the search rule for the kinematics segment, the search rule for the kinematics segment includes:

(1) From the starting time of the driving segment, look for the first point where the GPS vehicle speed is 0, that is, the starting point of idling speed. If the starting point of idling speed is found, record the position of the starting point of idling speed; then continue to find the first GPS point downward The point at which the vehicle speed is not 0, that is, the middle point, records the position of the middle point;

(2) Calculate the time difference from the middle point to the start point of idle speed. If the time difference is greater than 20s, move the position of the start point of idle speed down for 20s, and then judge the time difference from the middle point to the start point of idle speed until the time difference is less than 20s; find the next GPS vehicle speed as 0, that is, the idling end point of the kinematics segment, record the position of the idling end point;

(3) Screen the kinematics segment according to the kinematics segment screening rule, if the kinematics segment screening rule is met, then extract the kinematics segment from the driving segment according to the recorded positions of the idle start point and the idle end point;

The kinematic fragment screening rules include:

(1) The duration of the kinematic segment is not less than 20 seconds, that is, the time from one idle start point to the next idle start point is at least 20 seconds;

(2) The kinematics segment contains at least one acceleration state and one deceleration state. Therefore, the kinematics segment must at least have a continuous segment satisfying that the acceleration of the vehicle is greater than 0.1m/s ² and the deceleration is less than -0.1m/s ² ;

(3) The idle time of the kinematics segment shall not exceed 20 seconds.

Further, the characteristic parameters of the kinematics segment include time characteristic parameters, velocity characteristic parameters and acceleration characteristic parameters, wherein the time characteristic parameters include: running time t(s), constant speed time t _i (s), idle time t _c (s ), acceleration time t _a (s), deceleration time t _d (s); speed characteristic parameters include: average speed v _m (km/h), average driving speed v _mr (km/h), maximum speed v _max (km /h), speed standard deviation v _std (km/h); acceleration characteristic parameters include: average acceleration a _ma (m/s ² ), average deceleration a _md (m/s ² ), acceleration standard deviation a _std (m /s ² ), constant speed time ratio P _c (%), idle time ratio P _i (%), acceleration time ratio P _a (%), and deceleration time ratio P _d (%).

Further, using K-Means clustering to divide the kinematics fragments into four fragment libraries specifically includes the following steps: first randomly select 4 kinematics fragments from all kinematic fragments as the initial cluster centers; then perform the cluster assignment operation : Calculate the Euclidean distance from each kinematic segment to the four initial cluster centers, classify according to the Euclidean distance between the kinematic segment and the initial cluster center, and assign each kinematic segment to the initial segment with the closest Euclidean distance Clustering centers form 4 clusters; after obtaining 4 clusters, recalculate the clustering centers of each cluster, and perform step S42 until the composition of kinematic segments of each cluster no longer changes, and finally obtain the four clusters of kinematic segments. fragment library.

Further, the calculation method of the Euclidean distance from the kinematics segment to the cluster center includes:

为聚类中心j的第m个特征要素。Among them, d _ij is the Euclidean distance from the i-th kinematics segment to the cluster center j, and x' _im is the m-th feature element of the i-th kinematics segment,

is the mth feature element of the cluster center j.

Further, the structure of the long short-term memory neural network model includes: an input layer, an LSTM layer, a fully connected layer and an output layer.

Further, the training data set needs to be preprocessed before being input into the model. The preprocessing of the training data set includes: discarding the time dimension of long segments, and retaining the speed dimension of long segments; setting a sliding window whose length is time The size of the step, slide the window backward from the starting position on the long segment, each sliding step is 1 second, take the velocity-time sequence segment of the area covered by the window at the current moment as the velocity-time sequence at the current moment, take The initial velocity value of the area covered by the window at the next moment is used as the label of the area covered by the window at the current moment; and so on, the four long segments are preprocessed respectively to obtain the preprocessed D _L , D _M , D _H , D _EH training data set.

Beneficial effects of the present invention:

The present invention builds a convenient and fast method for driving conditions of automobiles. The method can effectively identify the hidden features in the driving data of automobiles in special regions through the long short-term memory neural LSTM network model, thereby constructing a vehicle that conforms to the driving characteristics. Vehicle driving curve.

Description of drawings

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments. The accompanying drawings are only for the purpose of illustrating preferred embodiments, and are not considered to limit the present invention.

Fig. 1 is a kind of flow chart of the method for constructing automobile running condition according to the embodiment of the present invention;

Figure 2 is a schematic diagram of the network structure of the LSTM prediction model;

Fig. 3 is a flow chart of data preprocessing;

Fig. 4 is the comparative figure before and after data processing of data processing;

Fig. 5 is a schematic diagram of two short stroke division methods;

Fig. 6 is a flow chart of kinematics segment screening;

Fig. 7 is a schematic diagram of the input data generation strategy of the LSTM model;

Fig. 8 is a schematic diagram of the kinematics segment training and prediction results of the medium-speed section;

Fig. 9 is a schematic diagram of the kinematics segment training and prediction results of the high-speed section;

Fig. 10 is a result diagram of a construction example of a vehicle driving condition curve.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

As shown in Figure 1, a method for constructing a vehicle driving condition includes but is not limited to the following steps:

Get the original GPS data of the car. The original GPS data of the car includes the acquisition time, GPS speed and GPS acceleration. Due to various external factors and the vehicle itself, some bad data values are often included, but the reliability and accuracy of the original data It is directly related to the effectiveness of subsequent construction conditions, so the original GPS data of the car is preprocessed, and the data screening principle is established according to the bad data type in the original data to filter the original data, and the original GPS data of the car is deleted Valid and reasonable data are retained to facilitate subsequent data analysis and improve the quality of working condition construction by optimizing characteristic parameters.

In one embodiment, the bad data types in the raw data and the data screening principles include:

(1) Fragment screening for time discontinuities: due to high-rise building coverage or passing through tunnels, the GPS signal may be lost when the car is driving, resulting in time discontinuity in the original GPS data of the car. In order to ensure the effectiveness of the original data on the prediction results, it is necessary to preprocess this time-interrupted situation first. The processing methods include: looking for time discontinuity points, if there is no missing value, it means that the time is continuous, and no processing is performed; if there is a missing value, it means that the time is discontinuous, if the discontinuity time is less than or equal to 55 seconds, the data will be supplemented; if discontinuous If the time is greater than 55 seconds, it will not be processed. Data with an intermittent time greater than 55 seconds will not have a great impact on the overall data, and at the same time, the idle time ratio is guaranteed.

(2) Screening rules for abnormal acceleration: In general, the acceleration time of an ordinary car from speed 0 to 100km/h is greater than 7 seconds, and the maximum acceleration of the car is 3.968m/s ² ; 7.5-8m/s ² . In the raw data of the driving process of the car, if the acceleration does not meet the range of the above-mentioned maximum acceleration and/or maximum deceleration, it is considered that this type of data is data with abnormal acceleration and deceleration. If the prediction of the situation has a great influence, the entire kinematics segment of the abnormal acceleration and deceleration segment and a connected idle segment are removed, and the motion segments before and after the abnormal segment are connected after the removal.

(3) Maximum idling duration rule: In the original GPS data of the car, there may be long-term parking or long-term traffic jams such as parking without turning off the engine, parking and turning off the engine but the data acquisition device is not turned off, etc. In these cases, the vehicle stops running, but the data acquisition The device is not turned off, and the data device is still working at this time. The data segment that keeps the GPS speed at 0 for a long time is collected, which does not conform to the actual driving conditions on the road, so it is handled according to the idling speed; in addition, the vehicle also has intermittent low-speed driving during driving. In the case that the maximum vehicle speed is less than 10km/h, this type of data belongs to glitch data, and these data points are usually modified to idle speed points. The data segment with too long idling time does not meet the requirements of the kinematics segment and should be deleted to reduce the impact of these data on the working condition curve. The data with an idling time exceeding 180 seconds is considered to be too long and the data within 180 seconds is retained. The idle speed data points are taken as the idle speed period, and the remaining part is deleted.

In one embodiment, the order of preprocessing includes but is not limited to:

As shown in Figure 3, firstly, a traversal search is performed on three raw data files collected by the same vehicle in different time periods. All the data in a file are divided into different driving segments according to the first time breakpoint (the area with a time interval greater than 55 seconds in the original GPS data of the car driving) from the beginning.

Then for all the driving segments obtained, judge whether there is a second time breakpoint inside all the driving segments obtained (if the time interval in the original GPS data of automobile travel is greater than 2 seconds and less than or equal to 55 seconds, then this section interval area is considered to be the first Two time breakpoints), the improved polynomial fitting method is used to supplement the second time breakpoint inside the driving segment, the supplementary length is the length of the time difference, and a speed data is fitted according to the speed data before and after the second time breakpoint. A series of new speed data points are filled at the second time breakpoint to supplement the second time breakpoint inside the driving segment. Since the vehicle speed of GPS is not less than 0 in actual situations, the data points obtained by fitting the polynomial function Replace all negative numbers with 0.

After the data fitting is completed, the acceleration of each time point of each driving segment is calculated, and according to the abnormal acceleration screening rules, the driving segment with abnormal acceleration and an idling segment connected to the abnormal segment are removed from the data, and the abnormal Motion clips before and after the clip are connected.

For the abnormal data of long-term idle speed greater than 180 seconds, use a sliding window with a size of 180 to slide the time and vehicle speed of each segment, and the sliding step is 1s. During the sliding process of the window, if all the data in the window are idle speed data, then filter out the first piece of data in the window; Then all the data in this window will be filtered out. By analogy, all driving segments are screened out to obtain preprocessed data.

The preprocessed data is divided into kinematic segments by short-stroke division method.

As shown in Figure 5, the kinematics segment includes an idling time period and a driving time period, and the length M of each kinematics segment does not exceed 600s, and the longest idling time period does not exceed 20s, and the idle time exceeds the delete. Each kinematics segment starts with a velocity of zero and ends with a velocity of zero, and an idle segment can be included in the interval between the start and end segments.

In an optional embodiment, in the kinematics segment, the interception of the idle time period and the driving time period can adopt the first method: the kinematics segment starts with a speed of zero and ends with a speed of zero, and the idle time period is within the driving time period Before.

In an optional embodiment, in the kinematics segment, the interception of the idle time period and the driving time period can adopt the second method: the kinematics segment starts with a speed of zero and ends with a speed of zero, and the idle time period is within the driving time period after.

In one embodiment, first establish kinematics segment screening rules:

(1) The duration of the kinematics segment is not less than 20 seconds, that is, the time from one idle start point to the next idle start point is at least 20 seconds.

(2) The kinematics segment contains at least one acceleration state and one deceleration state. Therefore, the kinematics segment must at least have a continuous segment satisfying that the acceleration of the vehicle is greater than 0.1m/s ² and the deceleration is less than -0.1m/s ² .

(3) The idle time of the kinematics segment shall not exceed 20 seconds.

As shown in Figure 6, according to the kinematics segment screening principle, the short-stroke division method is used to find the kinematics segment in the pre-processed driving segment data, which specifically includes: firstly, judging whether the driving time of each driving segment is greater than 20s, If it is less than 20s, then delete the driving segment; if it is greater than 20s, then search for the kinematics segment from the driving segment according to the search rule of the kinematics segment.

In one embodiment, the search rule for the kinematics segment includes:

(1) Find the starting point of idling (the point where the GPS vehicle speed is 0) downwards from the starting time of the driving segment. If the starting point of idling is found, record the position of the starting point of idling. Then continue to find the first point where the GPS speed is not 0, record it as the middle point, record the position of the middle point and judge the time difference from the middle point to the starting point of idling;

(2) If the time difference is greater than 20s, move the position of the starting point of the idle speed down by 20, and then judge the time difference from the middle point to the starting point of the idle speed until the time difference is less than 20s, and then look for the next idle point, which is the kinematics segment The idle speed end point (the point where the GPS vehicle speed is 0), if found, then record the position of the idle speed end point.

(3) The kinematics segment is screened according to the kinematics segment screening rule, and if the kinematics segment screening rule is satisfied, the kinematics segment is extracted from the driving segment according to the recorded positions of the idle start point and the idle end point.

According to the characteristic parameter calculation formula of the kinematic segment, the feature calculation of the kinematic segment is performed to obtain the characteristic parameter of the kinematic segment. The characteristic parameters of the kinematics segment include 16 characteristic parameters, which can be divided into three types: time characteristic parameters, velocity characteristic parameters and acceleration characteristic parameters. The time characteristic parameters include: running time t(s), constant velocity time t _i (s) , idle time t _c (s), acceleration time t _a (s), deceleration time t _d (s); speed characteristic parameters include: average speed v _m (km/h), average driving speed v _mr (km/h) , maximum speed v _max (km/h), speed standard deviation v _std (km/h); acceleration characteristic parameters include: average acceleration a _ma (m/s ² ), average deceleration a _md (m/s ² ), Acceleration standard deviation a _std (m/s ² ), constant speed time ratio P _c (%), idle time ratio P _i (%), acceleration time ratio P _a (%), deceleration time ratio P _d (%).

In one embodiment, the formula for calculating the characteristic parameters of the kinematic segment includes:

(1) Running time t(s): Since the sampling frequency is 1Hz, the calculation method of running time is:

t=n

Among them, n is the number of collected driving data.

(2) Idle time t _i : t _i is the number of zeros in the kinematics segment.

(3) Acceleration time t _a : t _a is the total number of points whose acceleration is greater than 0.1m/s ² .

(4) Deceleration time t _d : t _d is the total number of points whose acceleration is less than -0.1m/s ² .

(5) The calculation method of constant velocity time t _c is: t _c =tt _i -t _a -t _d

(6) The calculation method of the average velocity v _m is:

In formula (6), q is the total number of data points of a kinematic segment, and v _p represents the velocity value of data point p.

(7) The calculation method of the average driving speed v _mr is:

where q is the total number of data points for a kinematic segment.

(8) The calculation method of the maximum speed v _max is:

v _max =max{v _p ,p=1,2,3...q}

(9) The calculation method of the speed standard deviation v _std is:

(10) The calculation method of the average acceleration a _ma is:

(11) The calculation method of the average deceleration _amd is:

Among them, a _p represents the acceleration value corresponding to the data point p measured by GPS.

(12) The calculation method of acceleration standard deviation a _std is:

(13) The calculation method of the idle time ratio P _i is: P _i =t _i /t

(14) The calculation method of constant speed time ratio P _c is: P _c =t _c /t

(15) The calculation method of the acceleration time ratio P _a is: P _a =t _a /t

(16) The calculation method of the deceleration time ratio P _d is: P _d =t _d /t

Further, in a preferred embodiment, after obtaining the above-mentioned features of the kinematic segment, a principal component analysis method is used to filter irrelevant features to obtain feature parameters with a relatively large relationship.

Since the use of the principal component analysis method to filter irrelevant features is not the innovation of the present invention, nor is it the focus of the present invention, so this specification will not go into details. The specific process of using the principal component analysis method to filter irrelevant features Reference can be made to the corresponding principal component analysis steps in the prior art "Research on the Construction Method of Vehicle Driving Conditions on Urban Roads Based on K-Means Cluster Analysis".

According to the characteristics of kinematic fragments, K-Means clustering is used to divide kinematic fragments into four fragment libraries, namely: low-speed interval fragment library L, medium-speed interval fragment library M, high-speed interval fragment library H and extremely high-speed interval Fragment library EH. Among them, the maximum speed of the kinematic segments in the low-speed segment library is not more than 60km/h, the maximum speed of the kinematic segment in the medium-speed segment library is not more than 80km/h, and the maximum speed of the kinematic segment in the high-speed segment library is not more than 100km/h, the maximum speed of the kinematics fragments in the very high-speed section fragment library shall not exceed 130km/h.

In one embodiment, the K-Means clustering algorithm is used to divide the kinematic segments, and the Euclidean distance is used as the distance metric of the category distance, which specifically includes:

(1) First randomly select 4 kinematic segments from all kinematic segments as initial cluster centers;

(2) Perform cluster assignment operation: calculate the Euclidean distance from each kinematic segment to the four initial cluster centers, classify according to the Euclidean distance between the kinematic segment and the initial cluster center, and assign each kinematic segment To the initial cluster center with the closest Euclidean distance, 4 clusters are formed;

(3) After obtaining 4 clusters, recalculate the new cluster center of each cluster, and perform step (2) until the composition of kinematic segments of each cluster does not change, and finally obtain four segments of kinematic segments library.

The calculation method of the new cluster center in each cluster includes:

C _i ＝mean(x ₁ ,x ₂ ,x ₃ ,…,x _n )

Among them, C _i represents the new cluster center in cluster i, x is the kinematics segment assigned to cluster i, n is the number of kinematics segments belonging to cluster i, and mean is the mean value operation.

Further, in one embodiment, in the K-means clustering algorithm, the criterion function for measuring clustering quality is the sum of squares of errors, and the smaller the sum of squares of errors, the higher the quality of clustering. The criterion function of the K-means clustering algorithm is:

Wherein, SSE represents the sum of squared errors (Sumofthe Squared Error, SSE), and k represents k clustering centers, that is, the final clustering result has k classes, k=4 in this specification, x is the data point that needs to be clustered, C _i is the cluster center, and dist is the Euclidean metric.

Further, in one embodiment, the K-means clustering algorithm uses the Euclidean metric to calculate the degree of similarity, and the Euclidean distance between the kinematic segment and the cluster center is:

为聚类中心j的第m个特征要素。Among them, d _ij is the Euclidean distance from the i-th kinematics segment to the cluster center j, and x' _im is the m-th feature element of the i-th kinematics segment,

is the mth feature element of the cluster center j.

It should be noted that, in addition to using the Euclidean metric to calculate the similarity, alternatively, other methods can be used to measure the similarity in this specification, for example: Euclidean square distance, Minkowski distance, Chebi Schiff distance, Block distance, etc., may be any achievable similarity measurement method in the prior art.

Further, in one embodiment, if the minimum Euclidean distances from kinematics segments to various centers are equal and the degree of similarity cannot be judged, the closeness criterion is used for judgment, and the formula of the closeness criterion is as follows:

表示任一需要进行贴近度判定的运动学片段i的速度-时间向量，

为聚类中心m的速度-时间向量，

表示运动学片段i与聚类中心m的贴近度值。in,

Represents the velocity-time vector of any kinematics segment i that needs to be judged by closeness,

is the velocity-time vector of the cluster center m,

Indicates the closeness value between the kinematics segment i and the cluster center m.

Calculate the closeness value of the kinematics segment and all cluster centers according to the formula of the above closeness criterion, find the maximum value from the obtained closeness value, and divide the kinematics segment into the cluster corresponding to the maximum closeness value. The cluster center, that is, the cluster center corresponding to the maximum value of closeness, is the category to which the kinematic segment belongs.

All the kinematic fragments in each fragment library are spliced separately to obtain four long fragments, and the four long fragments are used as the training data set.

The training data set construction method includes: for each fragment library, splicing all the kinematic fragments s in each fragment library to obtain the speed-time sequence Si (ie long fragment), and for the four fragment libraries, correspondingly obtain S _L , _SM , _SH and S _EH four long fragments.

The above training data set is input into the long-term short-term memory neural network model (LSTM) for training, and the long-term short-term memory neural network model (LSTM) continuously learns the change characteristics of the speed-time series contained in the fragment library of different categories during the training process, Four corresponding time-speed sequence prediction models are trained to obtain a trained long-short-term memory neural network model. It should be noted that the time-speed sequence prediction model in this specification is the same model as the long-short-term memory neural network model.

The long-short-term memory neural network model can complete the reconstruction of time series according to the spatiotemporal correlation of different categories of fragment libraries, rely on model training to identify and strengthen the temporal correlation characteristics, and predict different categories of kinematic fragments with representative characteristics. Since the network structure of the long-short-term memory neural network model is definite, it needs to be input as a vector of a specific dimension, and the length of the data in the training data set is different, which does not meet this condition. Therefore, the training data set needs to be pre-prepared before inputting the model. Processing, the training data set is processed into supervised learning data that can be used by the time-velocity series prediction model.

Preprocessing the training dataset includes:

1. Discard the time dimension of long fragments and retain the speed dimension of long fragments;

2. Set a sliding window, the length of the window is the size of the time step, slide the window backward from the starting position on the long segment, each sliding step is 1 second, take the speed of the area covered by the window at the current moment - The time series segment is used as the velocity-time series at the current moment, and the initial velocity value of the area covered by the window at the next moment is taken as the label of the area covered by the window at the current moment;

取该覆盖区域的速度-时间序列片段作为t₀时刻的速度-时间序列s₀，取t_n+1时刻的速度值

作为s₀的标签，构成t₀时刻的数据片段

以此类推，滑动窗口继续向后滑动，产生t₁时刻的数据片段

直到生成t_L-n时刻的数据片段

共计得到L-n个数据片段，根据L-n个数据片段构建出长片段有预处理后的训练数据集为D＝{d_m|0<m<L-n}。其中，L表示长片段S的数据长度，d_m表示第m个数据片段，s₁表示t₁时刻的速度-时间序列，

表示t_n+2时刻的速度值，

表示t_L时刻的速度值。Specifically, the length of the sliding window W is n, and the sliding window W slides on the speed-time sequence S (a long segment) of length L, starting at time t ₀ , and the coverage area of the sliding window at time t ₀ is

Take the speed-time series segment of the coverage area as the speed-time series s ₀ at time t ₀ , and take the speed value at time t _n+1

As the label of s ₀ , it constitutes the data segment at time t ₀

By analogy, the sliding window continues to slide backwards to generate data fragments at time t ₁

Until the data fragment at time t _Ln is generated

A total of Ln data fragments are obtained, and a long-segment preprocessed training data set is constructed based on the Ln data fragments as D={d _m |0<m<Ln}. Among them, L represents the data length of the long segment S, d _m represents the mth data segment, s ₁ represents the speed-time sequence at time t ₁ ,

Indicates the velocity value at time t _n+2 ,

Indicates the velocity value at time t _L.

3. By analogy, the four long fragments are preprocessed respectively, and four long fragment preprocessed D _L , D _M , D _H , D _EH training data sets are obtained.

The invention utilizes the LSTM cycle neural network as the prediction core to generate the driving condition curve of the automobile, and constructs the time-speed sequence prediction model. The time-velocity sequence prediction model is shown in Figure 2. The model includes an input layer, an LSTM layer, a fully connected layer and an output layer. The input layer is followed by a layer of LSTM layer, and then through two layers of fully connected layers. The last layer is connected to the output layer. Since the input layer of the model is the kinematic fragment library of four categories after K-Means clustering, the input data is potentially periodic, which is helpful for training. The LSTM layer contains 3 hidden layers. Use As an activation function, tanh is used to learn potential temporal relationships and abstract features in the input sequence; the fully connected layer reduces the output dimension of the LSTM layer and maps the abstract features learned by the LSTM layer, and the output dimension of the fully connected layer is 1 ; The final output layer outputs the predicted speed data results, and the predicted results are sequentially spliced, and finally the vehicle driving condition curve is obtained.

After preprocessing the input training data set, input the preprocessed training data set D _L , D _M , D _H , D _EH into the long short-term memory neural network model to train the time-speed sequence prediction model, time-speed The training of the sequence prediction model includes: for the LSTM layer part, the input dimension of the first hidden layer is (1,n), the input dimension of the second hidden layer is (1,6), and the third hidden layer The input dimension is (1,8), and the flow direction of the input data in the LSTM layer is: first pass through the LSTM gate structure to obtain the processed intermediate output, and then pass the intermediate output through the tanh activation function as the output result of this layer. Finally, the output of the LSTM layer is input to a fully connected layer. The input dimension of the fully connected layer is 4, and the output dimension is 1. The output is the speed prediction value of the speed point immediately after the current input vector. The number of training iterations of the model is 100 rounds, and the training process is divided into forward propagation process and error back propagation process: Forward propagation uses time-velocity sequence prediction model The input is the velocity vector data in the preprocessed training set X D _i , i=(L, M, H, EH) is used as input, after the above-mentioned LSTM layer and fully connected layer, the weights of all layers are randomly deactivated with a probability of 0.2, and the output is obtained, and the mean square error is used as the loss function. Get the error, perform error backpropagation, and use the Adam function as the optimization function of each layer to update the weight of each layer. The learning rate of the optimization function is set to 0.001, and the batch size of network training is set to 1, respectively. Train all the input data in the four fragment libraries D _L , D _M , D _H , D _EH in the training set X, and get four corresponding time-speed sequence prediction models M _L , M _M , M _H , M _EH , the internal structure of the four models is the same, but the training data sets are different, so the internal weights obtained are different, corresponding to the fragments that can predict the four parts of the low, medium, high and extremely high parts of the vehicle operating conditions.

Preferably, the basic gate structure in the adopted LSTM is:

Forget gate: f _t = σ(W _f ×[h _t-1 ,x _t ]+b _f )

Input gate: i _t = σ(W _i ×[h _t-1 , x _t ]+ _bi )

Output gate: o _t = σ(W _o ×[h _t-1 , x _t ]+b _o )

h _t ＝o _t ×tanh(c _t )

代表当前输入的单元状态，x_t为细胞输入，h_t为细胞输出，W，b分别为算法中各个门的权重和偏移量。具体为： f_t是遗忘向量，决定上一时刻中哪些信息从单元状态中遗忘，W_f为遗忘门中的权重，h_t-1为上一个门结构的输出，x_t为t时刻的输入，b_f为遗忘门中的偏置； i_t是输入门向量，决定单元状态中被保留的信息，W_i是输入门中的权重，b_i为输入门中的偏置；o_t为输出门的输出向量，W_o是输出门中的权重，b_o是输出门中的偏置，

是细胞单元状态更新值，W_c是细胞单元网络中的权重，b_c是细胞单元网络中的偏置，c_t-1是上一个时刻(t-1)的细胞状态。Among them, the symbol "." represents the multiplication of elements at the corresponding positions of the two vectors; the symbol "σ" represents the sigmoid function, which is used as the activation function; c represents the cell state,

Represents the current input cell state, x _t is the cell input, h _t is the cell output, W, b are the weight and offset of each gate in the algorithm, respectively. Specifically: f _t is the forgetting vector, which determines which information is forgotten from the unit state at the previous moment, W _f is the weight in the forgetting gate, h _t-1 is the output of the previous gate structure, and x _t is the input at time t , b _f is the bias in the forget gate; _it is the input gate vector, which determines the information retained in the cell state, W _i is the weight in the input gate, b _i is the bias in the input gate; o _t is the output The output vector of the gate, W _o is the weight in the output gate, b _o is the bias in the output gate,

is the update value of the cell state, W _c is the weight in the cell network, b _c is the bias in the cell network, c _t-1 is the cell state at the last moment (t-1).

The trained long-short-term memory neural network model is used for prediction, and the time-speed prediction curves corresponding to the four fragment libraries are obtained.

For a segment library, the speed at a certain time in the future is predicted according to the kinematics data in the segment library, and the driving curve of the vehicle operating condition is obtained.

作为输入，预测出t_L+1时刻的速度值

进而由

预测出t_L+2时刻的速度值

知道预测出t_L+M的速度值

取

速度值作为该区间的抽取出的能够代表该区间的所有速度变化曲线特征的时间-速度预测曲线P_i。以此种方式，最后得到低速区间片段库L、中速区间片段库M、高数区间片段库H和极高速区间片段库EH分别对应的时间-速度预测曲线为P_L，P_M，P_H，P_EH。The prediction process includes: the model M _i obtained above, i=(L, M, H, EH), and the last data in the corresponding training data D _i in the training set X

As input, predict the speed value at time t _L+1

And then by

Predict the speed value at time t _L+2

Knowing the predicted speed value of t _L+M

Pick

The speed value is taken as the time-speed prediction curve P _i extracted from the interval and can represent the characteristics of all speed change curves in the interval. In this way, the time-speed prediction curves respectively corresponding to the low-speed segment library L, the medium-speed segment library M, the high-speed segment library H and the extremely high-speed segment library EH are finally obtained as P _L , _PM , P _H , P _EH .

After obtaining the time-velocity prediction curves corresponding to the four fragment libraries, according to the time ratios of the four fragment libraries in the entire kinematics fragment, determine the time occupied by the four types of fragment libraries in the final working condition synthesis, The time-speed prediction curves _PL , _PM , _PH , and P _EHy of the four speed segments are sequentially spliced into a working condition curve to obtain the final vehicle driving condition curve P.

As can be seen from the technical solution provided by the vehicle driving condition curve construction method of the present invention, due to the complicated speed changes of the vehicle during driving, the time-speed curve generated by the running of the vehicle can contain all the information of the vehicle running: acceleration, Maximum acceleration, maximum deceleration and average acceleration and deceleration, average speed and idle speed range, there are no strict rules to follow, uncertainty, instability, nonlinearity in practice and space, data volume comparison big. The structure of long short-term memory (LSTM) recurrent neural network is complex, and its output information is not only related to the input, but also related to the memory content of the cell unit and the last output result. At the same time, the LSTM model makes up for typical machine learning models such as recurrent neural network Due to the problems of gradient dispersion and gradient explosion in structures such as Recurrent Neural Networks (RNN), and insufficient long-term memory capacity, compared with traditional fitting methods, recurrent neural networks can truly and effectively utilize long-distance timing information. The vehicle operating condition curve is exactly A typical time series data, so using the LSTM model to predict the operating curve can get better results. In the LSTM model, the new cell state and the previous state are an accumulation process, which can nonlinearly fit the driving kinematics curve segment of the car, effectively consider the timing of the input data, and realize the encoding and decoding of time series .

In order to make this description clearer and more complete, further description will be given in combination with specific data.

In one embodiment, the original data collected by the vehicle-mounted GPS device in the 2019 "China Graduate Mathematical Modeling Competition" question D is selected for analysis. The data is the data collected by the actual road driving of a light-duty vehicle in a certain city, and the sampling frequency is 1Hz. The data includes Acquisition time (seconds) and GPS vehicle speed.

Preprocess the raw data collected by the vehicle-mounted GPS device, eliminate bad data values and invalid data, and retain valid data.

The raw data collected by the vehicle GPS device includes 3 files. Perform corresponding processing according to the preprocessing method of the above raw data, and obtain the final preprocessing results, as shown in Table 1 below: the number of motion segment records after processing in file 1 is 186255, and the number of motion segment records after processing in file 2 is 149032 , the number of motion segment records in file 3 after processing is 170808.

Table 1 The number of records before and after data preprocessing of each file

In order to better display the preprocessing results, some processed data fragments in file 2 according to the three screening principles are selected for comparison with the original data fragments, as shown in Figure 4. The above three figures are original data fragments, and the three original The data fragments have problems of data loss, abnormal acceleration/deceleration, and abnormal fragments with excessive idle time. The following three figures are respectively preprocessing the original data abnormal fragments with data loss, abnormal acceleration/deceleration, and excessive idle time. From the final results, it can be clearly seen that the speed-time curve of the pre-processed car is closer to the operating condition curve of the actual car.

The preprocessed data is divided into short strokes to obtain kinematic segments, and the screening rules for kinematic segments are established:

(1) The duration of the kinematics segment is not less than 20 seconds, that is, the time from one idle start point to the next idle start point is at least 20 seconds.

(2) The kinematics segment contains at least one acceleration state and one deceleration state. Therefore, the kinematics segment must at least have a continuous segment satisfying that the acceleration of the vehicle is greater than 0.1m/s ² and the deceleration is less than -0.1m/s ² .

(3) The idle time of the kinematics segment shall not exceed 20 seconds.

Apply the short-stroke division method, as shown in Figure 6, use the short-stroke division method to find the kinematics segment in the driving segment data obtained by preprocessing

According to the characteristic parameter calculation formula of the kinematic segment, the feature calculation of the kinematic segment is performed to obtain the characteristic parameter of the kinematic segment.

An example of the calculated characteristic parameter values of some kinematic segments is shown in Table 2 below.

Table 2 Kinematic segment eigenvalues

According to the characteristics of the kinematic fragments, K-Means clustering is used to divide the kinematic fragments into four fragment libraries. The original 386 kinematic segments are divided into four segment libraries through the feature matrix clustering of the segments, namely the low-speed interval segment library L, the medium-speed interval segment library M, the high-speed interval segment library H, and the extremely high-speed interval segment library EH . In order to further analyze the driving characteristics of the vehicle represented by the kinematic fragments in each type of speed interval fragment library, select the characteristic parameters related to speed and acceleration from the 16 characteristic parameters to perform corresponding calculations on the four classes, and obtain various fragment libraries The comprehensive eigenvalues of all kinematic segments in , see Table 3.

It can be seen from Table 3 that the medium-speed segment library M (the first category) includes 203 kinematic segments, the average speed is 12.15km/h, and the maximum speed does not exceed 106.74km/h; the low-speed segment library L (the first category) Class II) includes 39 kinematic segments with an average speed of 6.02km/h and a maximum speed of no more than 149.88 km/h; the very high-speed interval segment library EH (type III) includes 10 kinematic segments with an average speed of 37.54 km/h, the maximum speed does not exceed 76.60km/h; the high-speed section segment library H (the fourth category) includes 134 kinematic segments, the average speed is 23.88km/h, and the maximum speed does not exceed 139.61km/h.

Table 3 Comprehensive characteristic parameter values of various fragment libraries

Characteristic Parameters the first sort second category third category fourth category total number of fragments 203 39 10 134 v_m 12.15 6.02 37.54 23.88 v_max 106.74 149.88 76.60 139.61 v_std 9.657 14.786 14.027 14.672 a_ma 0.5970 0.4356 0.3387 0.4856 a_md -0.7155 -0.5479 -0.4549 -0.6179 a_std 0.6023 0.5509 0.4236 0.5810

All the kinematic fragments in each fragment library are spliced separately to obtain four long fragments, and the four long fragments are used as the training data set.

The input dimensions of a typical LSTM model include: sample, timestep, and feature of the previous network. The input of LSTM must be a "sequence", which coincides with our requirements: through input Velocity sequence to predict the velocity at the next moment.

(2) Preprocessing of the training data set: Since the obtained long segment is a time-velocity sequence, the time dimension in the long segment is discarded, then the actual dimension of the long segment after discarding the time dimension is dim=(s, 1), s is the length of the segment, but when performing model training, the data needs to be processed into supervised learning data that the model can use.

The way to deal with it is: for the speed sequence of long segments, set a sliding window, the length of the window is the size of the time step, slide the window backward from the initial speed, each time the sliding step is 1 second, and take the bottom of the window A value is used as the processing value of the current window, as the label h _t of the segment x _t covered in the window, and so on, to obtain the input data set 5 with a length of s-timestep, the input data generation strategy of the long-term short-term memory neural network model As shown in Figure 7.

(3) After preprocessing the training data set, input the preprocessed training data set into the long short-term memory neural network model (LSTM) for training:

The data set X is input to the model for training. The hyperparameters used are: the batch size (batch-size) is equal to 5, and the time step is equal to 300 seconds. The specific structure of the model is: a layer of LSTM, which contains 3 hidden layers. Use tanh as the activation function, use a fully connected layer to process the output of all LSTM layers, the output dimension of the fully connected layer is 1, and predict the speed at the next moment (second) for the sequence speed data of the current input model.

(4) Use the model to predict fragments:

The input of the model is a sequence of time periods corresponding to a length of len, and the corresponding velocity sequence Y _len is predicted, and the last sample data of the training data set is used as a seed input to generate the first predicted value, and delete the first element of the input at the same time , and take the first predicted value as the input of the next element, the model outputs the second predicted value, and so on to get all the predicted sequences Y. After predicting the four fragment libraries respectively, the four fragment libraries are obtained respectively The corresponding time-speed prediction curves, and then the time-speed prediction curves corresponding to the four fragment libraries are spliced together to obtain the entire complete working condition data. Table 4 shows the loss function related data during the model training process.

Table 4. Loss function related data in the model training process

Use the trained long-short-term memory neural network model to make predictions, and finally obtain the time-speed prediction curves corresponding to the four kinematics segment libraries. Figure 8 shows the schematic diagram of the training and prediction results of the medium-speed segment kinematics segments, where symbols The "+" data point part (the "True" curve in Figure 8) is the training data, and the solid curve (the "Train" curve in Figure 8) is the prediction result on the training data set after model training, showing the prediction result The degree of coincidence with the training data is relatively high, and the operating condition curve measured by the part of the data point of the symbol "▽" (the "Predict" curve in Figure 8) is shown in Figure 9. Among them, the symbol "+" data point is the training data, and the solid curve is the prediction result on the training data set after model training, showing that the prediction result has a high degree of coincidence with the training data, and the working condition measured by the symbol "▽" data point From Figure 8-9, it can be seen that the prediction results of the LSTM model have a good coincidence degree, and the construction of the working condition curve can obtain more accurate results.

After obtaining the time-velocity prediction curves corresponding to the four kinematic fragment libraries, according to the time ratio of each category in the entire kinematic fragment, determine the time occupied by each type of fragment in the final working condition synthesis, according to the LSTM model For the prediction results obtained from training, the curves of the four speed segments are merged into one operating condition curve, as shown in Figure 10.

After the working condition curve is obtained, the working condition curve is sent to the control device, and the control device controls the tested vehicle to simulate driving according to the working condition curve, and obtains the fuel consumption data and exhaust emission data of the tested vehicle during driving, and according to the obtained Fuel consumption data and exhaust emission data are used for vehicle emission assessment and environmental protection level assessment.

Those skilled in the art will understand that unless otherwise stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the words "comprising", "comprising", and "having" used in the description of the present invention refer to the existence of said features, integers, steps, operations, elements and/or components, but do not exclude the existence or Add one or more other characteristics, integers, steps, operations, elements, components and/or groups thereof. It should be understood that the term "and/or" used herein includes any and all combinations of one or more of the associated listed items.

It should be noted that those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be implemented by instructing related hardware through computer programs, and the programs can be stored in a computer-readable memory In the medium, when the program is executed, it may include the processes of the above-mentioned method embodiments. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM), and the like. The above descriptions are only the specific implementation methods of the present application. It should be pointed out that those skilled in the art can understand that various changes and modifications can be made to these embodiments without departing from the principle and spirit of the present invention. , alternatives and modifications, the scope of the present invention is defined by the appended claims and their equivalents.

Claims (6)

1. A method for building automobile driving conditions, is characterized in that, comprises the following steps: Obtain the original GPS data of the car, and preprocess the original GPS data of the car; the preprocessing of the original GPS data of the car includes: Traversing and searching the original GPS data of the car from the beginning, looking for the first time breakpoint, and dividing the original GPS data into different driving segments from the first time breakpoint; Judging whether there is a second time breakpoint in the obtained driving segment, if there is a second time breakpoint, a series of new speed data points are fitted by an improved polynomial fitting method based on the speed data before and after the second time breakpoint , to supplement the second time breakpoint inside the driving segment; After the data fitting is completed, the acceleration of each time point of each driving segment is calculated, and the driving segment with abnormal acceleration is eliminated from the data according to the abnormal acceleration screening rules; For the abnormal data of long-term idle speed greater than 180 seconds, use a sliding window with a size of 180 to slide the time and vehicle speed of each segment, and the sliding step is 1s. During the window sliding process, if all the data in the window are For idle data, filter out the first piece of data in the window; when the end of the window slides to the end of the driving segment, if all the data in the window are idle data, all the data in the window will be filtered out, and so on Screen out data for all driving segments to obtain preprocessed data; Use the short-stroke division method to divide the preprocessed data into kinematic segments, including: first judge whether the driving time of each driving segment is greater than 20s, if it is less than 20s, then remove the driving segment; if it is greater than 20s, then according to The search rule for the kinematics segment is to find the kinematics segment from the driving segment; The search rules for the kinematics segment include: (1) From the starting time of the driving segment, look for the first point where the GPS vehicle speed is 0, that is, the starting point of idling speed. If the starting point of idling speed is found, record the position of the starting point of idling speed; then continue to find the first GPS point downward The point at which the vehicle speed is not 0, that is, the middle point, records the position of the middle point; (2) Calculate the time difference from the middle point to the start point of idle speed. If the time difference is greater than 20s, move the position of the start point of idle speed down for 20s, and then judge the time difference from the middle point to the start point of idle speed until the time difference is less than 20s; find the next GPS vehicle speed of 0, that is, the idling end point of the kinematics segment, record the position of the idling end point; (3) Screen the kinematics segment according to the kinematics segment screening rule, if the kinematics segment screening rule is met, then extract the kinematics segment from the driving segment according to the recorded positions of the idle start point and the idle end point; The kinematic fragment screening rules include: (1) The duration of the kinematic segment is not less than 20 seconds, that is, the time from one idle start point to the next idle start point is at least 20 seconds; (2) The kinematics segment contains at least one acceleration state and one deceleration state. Therefore, the kinematics segment must at least have a continuous segment satisfying that the acceleration of the vehicle is greater than 0.1m/s ² and the deceleration is less than -0.1m/s ² ; (3) The idling duration of the kinematics segment shall not exceed 20 seconds; Carry out feature calculation on the kinematics segment to obtain the characteristic parameters of the kinematics segment, and use the principal component analysis method to filter irrelevant features to obtain effective feature parameters; K-Means clustering is used to divide the kinematics fragments into four fragment libraries, namely: low-speed interval fragment library, medium-speed interval fragment library, high-number interval fragment library and extremely high-speed interval fragment library; Construct the training data set: splice all the kinematic fragments in each fragment library to obtain four long fragments, and use the four long fragments as the training data set; Inputting the above-mentioned training data set into the long-short-term memory neural network model for training, and obtaining the trained long-short-term memory neural network model; Use the trained long-short-term memory neural network model to predict and obtain the time-speed prediction curves corresponding to the four fragment libraries. The specific process includes: taking the last sample data of the training data set as the first input element and inputting it into the training In a good long-short-term memory neural network model, the first prediction sequence is output; the first input element is deleted, and the first prediction value is used as the second input element, and the input model obtains the second prediction sequence; and so on, a fragment library is finally obtained The predicted sequences of the four fragment libraries are respectively corresponding to the time-speed prediction curves; After obtaining the time-velocity prediction curves corresponding to the four fragment libraries, according to the time ratios of the four fragment libraries in the entire kinematics fragment, determine the time occupied by the four types of fragment libraries in the final working condition synthesis, Merge the curves of four speed sections into one working condition curve; The working condition curve is sent to the control device, and the control device evaluates the exhaust emission of the vehicle and evaluates the environmental protection level according to the working condition curve. 2. A kind of method of building automobile running condition according to claim 1, it is characterized in that, the characteristic parameter of kinematics segment comprises time characteristic parameter, speed characteristic parameter and acceleration characteristic parameter, wherein, time characteristic parameter comprises: running Time t(s), constant speed time t _i (s), idle time t _c (s), acceleration time t _a (s), deceleration time t _d (s); speed characteristic parameters include: average speed v _m (km/ h), average driving speed v _mr (km/h), maximum speed v _max (km/h), speed standard deviation v _std (km/h); acceleration characteristic parameters include: average acceleration a _ma (m/s ² ) , average deceleration a _md (m/s ² ), acceleration standard deviation a _std (m/s ² ), constant speed time ratio P _c (%), idle time ratio P _i (%), acceleration time ratio P _a (% ), deceleration time ratio P _d (%). 3. A kind of method of building automobile running condition according to claim 1, is characterized in that, adopts K-Means clustering to divide kinematics segment into four segment libraries and specifically comprises the following steps: S41. First randomly select 4 kinematic segments from all kinematic segments as initial clustering centers; S42. Perform cluster assignment operation: calculate the Euclidean distance from each kinematic segment to the four initial cluster centers, classify according to the Euclidean distance between the kinematic segment and the initial cluster center, and assign each kinematic segment to The initial cluster center with the closest Euclidean distance forms 4 clusters; S43. After obtaining 4 clusters, recalculate the clustering center of each cluster, and perform step S42 until the composition of kinematic segments of each cluster does not change, and finally obtain four segment libraries of kinematic segments. 4. A method of constructing a vehicle driving condition according to claim 3, wherein the calculation method of the Euclidean distance from the kinematics segment to the cluster center comprises:

Among them, d _ij is the Euclidean distance from the i-th kinematics segment to the cluster center j, and x′ _im is the m-th feature element of the i-th kinematics segment,

is the mth feature element of the cluster center j. 5. A kind of method of constructing automobile running condition according to claim 1, is characterized in that, the structure of described long short-term memory neural network model comprises: input layer, LSTM layer, fully connected layer and output layer. 6. A kind of method of building automobile running conditions according to claim 1, is characterized in that, training data set needs to carry out pretreatment earlier before inputting model, and the preprocessing to training data set comprises: Discard the time dimension of long fragments and retain the speed dimension of long fragments; Set a sliding window, the length of the window is the size of the time step, slide the window backward from the starting position on the long segment, each sliding step is 1 second, and take the speed-time series of the area covered by the window at the current moment The segment is used as the velocity-time series at the current moment, and the initial velocity value of the area covered by the window at the next moment is taken as the label of the area covered by the window at the current moment; By analogy, the four long fragments are preprocessed respectively, and four long fragment preprocessed D _L , D _M , D _H , D _EH training data sets are obtained.

Images