CN112434735B - A method, system and device for constructing dynamic driving conditions - Google Patents

Abstract

The invention discloses a method for constructing a dynamic driving condition, which comprises the following steps: acquiring speed data of a vehicle, preprocessing the speed data, and generating an input fragment X; constructing a joint learning framework based on a deep neural network and a bidirectional long-term and short-term memory network, and inputting input segments into the joint learning framework to obtain a feature space Z; soft distribution clustering of the characteristic space Z is realized by utilizing a regularization item based on relative entropy, and a clustering result is obtained after iterative updating; and classifying the input fragments according to the corresponding relation between the clustering result and the input fragments to obtain fragment libraries of various types, and selecting the input fragments from the fragment libraries of various types to form the driving working condition.

Description

A method, system and device for constructing dynamic driving conditions

technical field

The invention relates to the technical field of environmental detection, in particular to a method, system and device for constructing a dynamic driving condition.

Background technique

According to the "Technical Policy for the Prevention and Control of Motor Vehicle Pollution" organized by the Ministry of Environmental Protection, it is pointed out that further improving the environmental quality is the core to build a motor vehicle pollution prevention and control system, and promote the systematization, scientific and informatization of motor vehicle pollution prevention and control. The "Technical Policy" clarifies that the emission limits of pollutants such as carbon monoxide (CO), total hydrocarbons (THC), nitrogen oxides (NOx) and particulate matter (PM) will be gradually tightened for new production vehicles. The "2020-2025 China Motor Vehicle Industry Market Prospects and Investment Opportunities Research Report" shows that by the end of 2019, the national motor vehicle ownership reached 348 million vehicles. my country is the world's largest automobile consumption market and production country. The quantity ranks among the top in the world all the year round. With the rapid increase in the number of motor vehicles, the resulting urban traffic congestion and vehicle exhaust pollution have become increasingly serious. Motor vehicle pollutant emissions are mainly affected by the driving conditions of the vehicle, such as long idling time and high acceleration and deceleration frequency of vehicles in traffic jams, which will result in higher exhaust emissions. Driving condition construction is a method of constructing vehicle driving profiles based on typical traffic conditions, which plays an important role in the evaluation of vehicle emissions, economy and mileage.

At present, the construction methods of driving conditions are mainly divided into two categories: Markov analysis method and cluster analysis method. The Markov analysis method regards the speed and time relationship of the vehicle driving process as a random process, and combines the different model events by using the characteristic that the state at time t only depends on the state at time t-1 (that is, there is no after-effect). form the entire driving process. The cluster analysis method divides all micro-stroke segments into several categories according to their similarity, and then selects segments from each type of segment library according to certain principles to form the final operating condition curve. Compared with the Markov analysis method, the cluster analysis method can obtain different types of working conditions, which is closer to the actual road conditions and is simple and easy to implement.

Due to the different actual traffic conditions and road characteristics in various areas of the city, it will affect the driving cycle of the vehicle. At the same time, with the development of new energy vehicles, the model data is becoming more and more abundant. The traditional method uses hand-designed features to represent the spatial distribution of driving data. Speed-time distribution, and driving data is treated as static data, its inherent dynamic characteristics and temporal dependencies are often ignored, resulting in low accuracy and insufficient robustness.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the present invention provides a method, system and device for constructing a dynamic driving condition.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solutions:

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

Step 1: Obtain the speed data of the vehicle and preprocess it to generate the input segment X;

Step 2: Construct a joint learning framework based on a deep neural network and a bidirectional long-short-term memory network, input the input segment into the joint learning framework, and obtain the feature space Z;

Step 3: Use the regularization term based on relative entropy to realize the soft assignment clustering of the feature space Z, and obtain the clustering result after iterative update;

Step 4: According to the correspondence between the clustering results and the input segments, classify the input segments to obtain multiple types of segment libraries, and select input segments from the various segment libraries to form driving conditions.

Specifically, in step 1, when the speed data is preprocessed, the invalid data in the speed data is removed and the missing values are filled; the micro-stroke segment is extracted from the speed data to generate a micro-stroke segment library; the micro-stroke segment library is interpolated The equal-length sequence library is obtained by processing, and the input fragment is obtained by normalizing the equal-length sequence library.

Specifically, in step 2, when the input segment is input into the joint learning framework, the joint learning framework includes an autoencoder, and the autoencoder includes an encoder and a decoder. The encoder uses a deep neural network and a bidirectional long-term memory The network processes the input fragments in turn;

The deep neural network learns short-time-scale waveforms in the input segment and extracts its local features;

The bidirectional long-short-term memory network learns the temporal connection between waveforms across time scales in the input segment, extracts its global features, and then forms the feature space Z;

The decoder uses upsampling and deconvolution to reconstruct the feature space to form a reconstructed segment X';

Pre-train the autoencoder so that the reconstructed segment X′ output by the decoder has the smallest mean square error with the input segment

其中k₀是最优聚类数，包括以下步骤：Specifically, in step 3, when the relative entropy-based regularization term is used to realize the soft assignment clustering of the feature sequence, the joint learning framework further includes a time series clustering layer for clustering the feature space, an encoder and a time series The clustering layer is iteratively updated until stable results are obtained, and finally the input fragments are clustered into multiple types of fragment libraries

where k ₀ is the optimal number of clusters, including the following steps:

Step 41: use the Euclidean distance ED to calculate the distance d _ij from each element _zi in the feature space to the cluster center c _j ;

Step 42: Normalize the distance d _ij to a probability distribution using Student’s t distribution, the probability that the feature vector _zi belongs to the jth cluster

Among them, the larger the value of q _ij is, the closer the eigenvector _zi is to the cluster center, the greater the possibility of belonging to the kth cluster, and α is the number of degrees of freedom of the student's t distribution;

Step 43: The target distribution p _ij is set to the delta distribution of data points above the confidence threshold, and the remaining values are ignored, where,

Step 44: Set the goal of iterative training to minimize the relative entropy loss between the probability distribution q _ij and the target distribution p _ij

Step 45: The total loss Loss _total =Loss _C +λLoss _ae , where λ is a proportional coefficient, and Loss _C is used as a regularization term to prevent overfitting in the encoder feature extraction process.

Specifically, selecting the optimal number of clusters according to the Davidson Bodding Index DBI includes the following steps:

Set the k value and bring it into the training codec and clustering network respectively;

Calculate the DBI value of the clustering results under each k value:

表示聚类i内特征向量到其质心的平均距离，代表聚类i中数据的分散程度，

表示聚类j内特征向量到其质心的平均距离，代表聚类j中数据的分散程度；

M_i表示聚类i的数据个数；X_is表示聚类i中的第s个数据，X_js表示聚类j中的第s个数据，c_i表示聚类i的质心，c_j表示聚类j的质心；p通常取2；Among them, k represents the number of clusters; ‖c _i -c _j || ₂ represents the Euclidean distance between the centroid of cluster i and the centroid of cluster j;

Represents the average distance from the feature vector in cluster i to its centroid, represents the degree of dispersion of data in cluster i,

Represents the average distance from the feature vector in cluster j to its centroid, and represents the dispersion degree of data in cluster j;

M _i represents the number of data in cluster i; X _is represents the s-th data in cluster i, X _js represents the s-th data in cluster j, _{ci represents the centroid of cluster i, and c j represents} cluster i The centroid of class j; p usually takes 2;

The k value when the DBI value first appears in the local minimum value is selected as the optimal number of clusters k ₀ .

Specifically, in step 3, when using the relative entropy-based regularization term to realize the soft assignment clustering of the feature sequence, the K-means algorithm is used to initialize the cluster center.

Specifically, in step 4, when input segments are selected from various types of segment libraries to form driving conditions, the feature vectors in the feature space in the clustering result have class labels, and each type of The feature vectors under the tags are sorted, and the priority of the feature vectors under various tags is determined; according to the time proportion of the various fragment libraries in the total fragment library, the number of selected fragments in the various fragment libraries is determined. The vector of priorities picks input segments to form driving conditions.

Specifically, two methods of relative error and joint distribution of speed and acceleration are used to evaluate the constructed driving conditions.

A system for constructing dynamic driving conditions, comprising:

A data acquisition module, which acquires the speed data of the vehicle and performs preprocessing to generate an input segment X;

an encoding module, which constructs a joint learning framework based on a deep neural network and a bidirectional long-short-term memory network, inputs the input segment into the joint learning framework, and obtains the feature space Z;

Clustering module, which utilizes the regularization term based on relative entropy to realize soft assignment clustering of feature space Z, and obtains the clustering result after iterative update;

The driving condition building module classifies the input segments according to the corresponding relationship between the clustering results and the input segments to obtain various types of segment libraries, and selects input segments from various segment libraries to form driving conditions.

A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the construction method when the processor executes the computer program.

Compared with the prior art, the beneficial technical effects of the present invention are:

Different from the traditional working condition construction method, the present invention adopts an unsupervised joint feature learning and clustering framework, utilizes the continuity of driving data and considers the time dependence of dynamic data, and does not use any hand-designed feature expression in the driving condition construction process. And select segments, and can achieve higher accuracy and robustness model construction on real driving data.

Description of drawings

Fig. 1 is the schematic flow chart of the working condition construction method of the present invention;

2 is a schematic structural diagram of a joint learning framework of the present invention;

Fig. 3 is the result visualization diagram of clustering of the present invention;

Fig. 4 is a working condition model diagram constructed by the present invention;

5 is a visualization diagram of the driving speed of the test vehicle of the present invention;

FIG. 6 is a visualization diagram of the estimated CO emission status of the present invention;

Figure 7 shows the calculated coefficients for each pollutant.

Detailed ways

A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

As shown in Figures 1 and 2, a method for constructing a dynamic driving condition includes the following steps:

S1: Obtain the speed data of the vehicle and perform preprocessing to generate the input segment X.

Specifically, in step 1, when the speed data is preprocessed, the invalid data in the speed data is removed and the missing values are filled; the micro-stroke segment is extracted from the speed data to generate a micro-stroke segment library; the micro-stroke segment library is interpolated The equal-length sequence library is obtained by processing, and the input fragment is obtained by normalizing the equal-length sequence library.

When interpolating the micro-stroke segment library, two methods of cubic spline interpolation and linear interpolation are used to combine one column of sequences into two columns of input.

The micro-stroke segment consists of an idle speed segment (the speed is kept at 0) and a kinematic segment (the speed is always greater than 0) that does not exceed 180s. The extracted micro-stroke segment starts from the idle speed state and ends at the next idle speed state.

S2: Construct a joint learning framework based on a deep neural network and a bidirectional long-short-term memory network, and input the input segment into the joint learning framework to obtain the feature space z.

Specifically, in step 2, when the input segment is input into the joint learning framework, the joint learning framework includes an autoencoder, and the autoencoder includes an encoder and a decoder. The encoder uses a deep neural network and a bidirectional long short-term memory. The network processes the input fragments in turn;

The deep neural network learns short-time-scale waveforms in the input segment and extracts its local features;

The bidirectional long short-term memory network learns the temporal connection between waveforms across time scales in the input segment, extracts its global features, and then forms the feature space Z;

The decoder uses upsampling and deconvolution to reconstruct the feature space to form a reconstructed segment X';

Pre-train the autoencoder so that the reconstructed segment X′ output by the decoder has the smallest mean square error with the input segment

The purpose of this step is to solve the time dependence of dynamic data and realize dimension reduction of nonlinear time dimension.

Specifically, in step 3, when using the relative entropy-based regularization term to realize the soft assignment clustering of the feature sequence, the K-means algorithm is used to initialize the cluster center.

S3: Use the regularization term based on relative entropy to realize the soft assignment clustering of the feature space Z, and obtain the clustering result after iterative update.

其中k₀是最优聚类数，包括以下步骤：Specifically, in step 3, when the relative entropy-based regularization term is used to realize the soft assignment clustering of the feature sequence, the joint learning framework further includes a time series clustering layer for clustering the feature space, an encoder and a time series The clustering layer is iteratively updated until stable results are obtained, and finally the input fragments are clustered into multiple types of fragment libraries

where k ₀ is the optimal number of clusters, including the following steps:

Step 41: use the Euclidean distance ED to calculate the distance d _ij from each element _zi in the feature space to the cluster center c _j ;

Step 42: Normalize the distance d _ij to a probability distribution using Student’s t distribution, the probability that the feature vector _zi belongs to the jth cluster

Among them, the larger the value of q _ij is, the closer the eigenvector _zi is to the cluster center, the greater the possibility of belonging to the kth cluster, and α is the number of degrees of freedom of the student's t distribution;

Step 43: The target distribution p _ij is set to the delta distribution of data points above the confidence threshold, and the remaining values are ignored, where,

Step 44: Set the goal of iterative training to minimize the relative entropy loss between the probability distribution q _ij and the target distribution p _ij

Step 45: The total loss Loss _t4tal = Loss _C + λLoss _ae , where λ is a proportional coefficient, and Loss _C is used as a regularization term to prevent the encoder feature extraction process from overfitting. The autoencoder has been pre-trained, so fine-tuning can be done. In this embodiment, the proportional coefficient may take a constant of 0.01 at this time.

The joint learning framework is trained using velocity data, that is, the encoder and time series clustering layers are iteratively updated until a stable result is obtained. The process is as follows.

enter:

A set of velocity values for a given vehicle micro-travel segment, that is, the input segment X, the training sample size n, the number of pre-training iterations iteration0, the number of optimization iterations iteration1, and the number of clusters k.

Output: Encoder-decoder network θ after training; cluster center C.

The specific process is as follows:

1: Initialize learning parameters θ, learning rate η, momentum v;

2: for the i-th random selection of n training samples x(1≤i≤iteration0)do{

3: encoder network output Z;

4: The decoder network outputs X';

5: Calculate the loss function Loss _ae (i) according to the formula;

6: Update the weight parameter θ ⁱ ←θ ^i-1 -αv ^j-2 -η ^i-1 ΔLoss _ae ;}

7: end for

8: Initialize the cluster center C, initialize the learning rate η, the exponential decay rate β ₁ estimated by the first-order moment, the exponential decay rate β ₂ estimated by the second-order moment, the constant ∈, the first-order momentum term m ₀ , the second-order momentum term v ₀

9: for the i-th random selection of n training samples x(1≤i≤iteration1)do{

10: calculate the kl divergence Lossc according to the formula;

11: Calculate the total loss function Losstotal according to the formula;

12: First-order momentum term correction value

13: Second-order momentum term correction value

14: Update the codec network weight parameters

15: Update cluster centers

16: end for

17: Return the encoder-decoder network θ that has been trained; the cluster center C.

The above process uses the form of pseudocode to describe the training process of the joint learning framework, where return represents the output value; for A do{B} means that each element in A is iterated once, and the content in B is executed once, and endfor means End the loop.

Specifically, selecting the optimal number of clusters according to the Davidson Bodding Index DBI includes the following steps:

Set the k value and bring it into the training codec and clustering network respectively;

Calculate the DBI value of the clustering results under each k value:

表示聚类i内特征向量到其质心的平均距离，代表聚类i中数据的分散程度，

表示聚类j内特征向量到其质心的平均距离，代表聚类j中数据的分散程度；

M_i表示聚类i的数据个数；X_is表示聚类i中的第s个数据，X_js表示聚类j中的第s个数据，c_i表示聚类i的质心，c_j表示聚类j的质心；p通常取2；Among them, k represents the number of clusters; ‖c _i -c _j || ₂ represents the Euclidean distance between the centroid of cluster i and the centroid of cluster j;

Represents the average distance from the feature vector in cluster i to its centroid, represents the degree of dispersion of data in cluster i,

Represents the average distance from the feature vector in cluster j to its centroid, and represents the dispersion degree of data in cluster j;

M _i represents the number of data in cluster i; X _is represents the s-th data in cluster i, X _js represents the s-th data in cluster j, _{ci represents the centroid of cluster i, and c j represents} cluster i The centroid of class j; p usually takes 2;

The k value when the DBI value first appears in the local minimum value is selected as the optimal number of clusters k ₀ .

The smaller the DBI value, the smaller the intra-class distance, and the larger the inter-class distance, that is, the better the clustering effect.

S4: According to the corresponding relationship between the clustering result and the input segment, the input segment is classified to obtain various types of segment libraries, and the input segments are selected from the various segment libraries to form the driving condition.

Specifically, in step 4, when input segments are selected from various types of segment libraries to form driving conditions, the feature vectors in the feature space in the clustering result have class labels, and each type of The feature vectors under the tags are sorted, and the priority of the feature vectors under various tags is determined; according to the time proportion of the various fragment libraries in the total fragment library, the number of selected fragments in the various fragment libraries is determined. The vector of priorities picks input segments to form driving conditions.

Each feature vector corresponds to an input segment, and the feature vector and the input segment are matched by their respective serial numbers; after the clustering operation, each feature vector has a class label. Since each feature vector corresponds to an input segment, Therefore, the input fragments can be divided into various fragment libraries through class labels; the feature vectors are sorted according to certain rules, which is essentially determining the priority of the input fragments in the various fragment libraries.

The number of fragments that should be selected for each type of fragment library is determined by the time proportion of input fragments in each type of fragment library to all input fragments; Priority is determined.

Specifically, two methods of relative error RE and speed-acceleration joint distribution SAPD are used to evaluate the constructed driving conditions.

The present invention uses the COPERT model to estimate single-vehicle emissions, specifically:

Using COPERT Ⅲ emission model to calculate the exhaust emission factor of a single vehicle

Ef _jw =(a _w +c _w v _j +e _w v _j ² )/(1+b _w v _j +d _w v _j ² );

where v _j is the average speed of the j-th type of vehicle driving cycle, and a _w , b _w , c _w , and d _w are the calculation coefficients of the w-th type of pollutant, as shown in Figure 7 .

Estimated emissions of major vehicle pollutants E=Ef _jw ×len×f, major pollutants such as carbon dioxide CO, hydrocarbons HC, nitrogen oxides NOx, where len represents the length of the driving distance, f represents the traffic flow, and when estimating single-vehicle emissions , f takes 1.

Combined with the vehicle's GPS data, it estimates and visualizes the emissions of a single vehicle to provide recommendations for urban road planning.

Different from the traditional working condition construction method, the present invention adopts an unsupervised joint feature learning and clustering framework, utilizes the continuity of driving data and considers the time dependence of dynamic data, and does not use any hand-designed feature expression in the driving condition construction process. And select segments, and can achieve higher accuracy and robustness model construction on real driving data.

Using the OBD data of light vehicles in Fuzhou City, including speed data and GPS data, the method of the present invention is verified, and the advanced nature of the method of the present invention is demonstrated from the clustering effect, and the constructed working condition model is further displayed, and the working condition model is demonstrated. an application case.

FIG. 3 shows the clustering result of this embodiment. The coupling degree between classes is low, the aggregation degree within the class is high, and the clustering effect is good.

FIG. 4 is a working condition model constructed in this embodiment, and the set driving condition period is 1200-1300s.

Figures 5 and 6 are visual representations of estimating the pollution emissions of single vehicles in Fuzhou test vehicles, of which Figure 5 is a visual diagram of the driving speed of the test vehicles; Figure 6 is a visual diagram of the estimated CO emissions.

It can be seen from Figure 5 and Figure 6 that the pollution emissions of vehicles are proportional to the speed; the high-speed driving sections and some intersections have larger emissions. It is possible to improve the efficiency of traffic at intersections with high emissions to reduce emissions.

A system for constructing dynamic driving conditions, comprising:

A data acquisition module, which acquires the speed data of the vehicle and performs preprocessing to generate an input segment X;

an encoding module, which constructs a joint learning framework based on a deep neural network and a bidirectional long-short-term memory network, inputs the input segment into the joint learning framework, and obtains the feature space Z;

Clustering module, which utilizes the regularization term based on relative entropy to realize soft assignment clustering of feature space Z, and obtains the clustering result after iterative update;

The driving condition building module classifies the input segments according to the corresponding relationship between the clustering results and the input segments to obtain various types of segment libraries, and selects input segments from various segment libraries to form driving conditions.

A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the construction method when the processor executes the computer program.

In the present invention, the speed data and the GPS data are derived from the vehicle driving data of the vehicle-mounted diagnostic system.

It will be apparent to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, but that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. The embodiments are therefore to be regarded in all respects as illustrative and not restrictive, the scope of the invention being defined by the appended claims rather than the foregoing description, and are therefore intended to fall within the scope of the appended claims. All changes that come within the meaning and range of equivalents are embraced within the invention, and any reference signs in the claims shall not be construed as limiting the scope of the claims involved.

In addition, it should be understood that although this specification is described in terms of embodiments, not every embodiment only includes an independent technical solution, and this description in the specification is only for the sake of clarity, and those skilled in the art should take the specification as a whole, The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims (8)

1. A dynamic running condition construction method comprises the following steps:

the method comprises the following steps: acquiring speed data of a vehicle, preprocessing the speed data, and generating an input fragment X;

step two: constructing a joint learning framework based on a deep neural network and a bidirectional long-term and short-term memory network, and inputting input segments into the joint learning framework to obtain a feature space Z;

step three: soft distribution clustering of the characteristic space Z is realized by utilizing a regularization item based on relative entropy, and a clustering result is obtained after iterative updating;

step four: classifying the input fragments according to the corresponding relation between the clustering result and the input fragments to obtain fragment libraries of various types, and selecting the input fragments from the fragment libraries to form a driving working condition;

in the first step, when the speed data is preprocessed, invalid data in the speed data is removed and missing values are filled; extracting micro-stroke fragments from the speed data to generate a micro-stroke fragment library; carrying out interpolation processing on the micro-stroke fragment library to obtain an equal-length sequence library, and carrying out normalization processing on the equal-length sequence library to obtain an input fragment;

step two, when the input segment is input into the joint learning framework, the joint learning framework comprises a self-encoder, the self-encoder comprises an encoder and a decoder, and the encoder sequentially processes the input segment through a deep neural network and a bidirectional long-time and short-time memory network;

the deep neural network learns the waveform of the short time scale in the input segment and extracts the local features of the waveform;

the bidirectional long-short time memory network learns the time connection between the waveforms across the time scale in the input segment, extracts the global characteristics of the input segment and further forms the characteristic space Z;

the decoder reconstructs the characteristic space by adopting up-sampling and deconvolution to form a reconstructed segment X';

pre-training the self-encoder to make the reconstructed segment X' output from the decoder and the input segment have the minimum mean square error

2. The dynamic running condition construction method according to claim 1, characterized in that: in the third step, when the soft distribution clustering of the feature sequences is realized by utilizing the regularization item based on the relative entropy, the joint learning framework further comprises a time sequence clustering layer for clustering the feature space, an encoder and the time sequence clustering layer are updated in an iterative manner until a stable result is obtained, and finally the input segments are clustered into a plurality of types of segment libraries

Wherein k is ₀ Is an optimal clustering number and comprises the following steps:

step 41: computing elements z of a feature space using Euclidean distances ED _i To the center of the cluster c _j Distance d of _ij ；

Step 42: distance d using student t distribution _ij Normalized to probability distribution, feature vector z _i Probability of belonging to jth cluster

Wherein q is _ij The larger the value, the feature vector z _i The closer to the clustering center, the higher the possibility of belonging to the jth cluster, and alpha is the degree of freedom of student t distribution;

step 43: target distributionp _ij Set to a delta distribution of data points above a confidence threshold and ignore the remaining values, wherein,

step 44: setting the target of iterative training to minimize probability distribution q _ij With the target distribution p _ij Relative entropy loss therebetween

Wherein n is the number of micro-stroke segments;

step 45: total Loss _total ＝Loss _C +λLoss _ae Wherein λ is a proportionality coefficient, Loss _C As a regularization term, the encoder feature extraction process is prevented from overfitting.

3. The dynamic running condition construction method according to claim 2, characterized in that: selecting an optimal clustering number according to the davison burger index DBI, comprising the following steps:

setting a k value, and respectively substituting the k value into the training coding and decoding and the clustering network;

calculating the DBI value of the clustering result under each k value:

wherein k represents a cluster number; II c _i -c _j || ₂ Representing the Euclidean distance between the centroid of the cluster i and the centroid of the cluster j;

represents the average distance of the feature vector in the cluster i to the centroid thereof, represents the dispersion degree of the data in the cluster i,

represents the average distance of the feature vector in the cluster j to the centroid thereof, and represents the clusterThe degree of dispersion of data in class j;

M _i representing the number of data of the cluster i; m _j Representing the number of data of the cluster j; x _is Representing the s-th data in cluster i, X _js Representing the s-th data in the cluster j, c _i Representing the centroid of cluster i, c _j Represents the centroid of cluster j; p is 2;

selecting the k value of the DBI value when the local minimum value appears for the first time as the optimal clustering number k ₀ 。

4. The dynamic running condition construction method according to claim 1, characterized in that: in the third step, when soft distribution clustering of the characteristic sequences is realized by utilizing the regularization item based on the relative entropy, a K-means algorithm is used for initializing a clustering center.

5. The dynamic running condition construction method according to claim 1, characterized in that: step four, when the input segments are selected from the segment libraries to form a driving working condition, the feature vectors in the feature space in the clustering result have class labels, the feature vectors under the various labels are sequenced according to the ratio of the class distance to the class distance, and the priority of the feature vectors under the various labels is determined; and determining the number of the selected fragments in each fragment library according to the time ratio of each fragment library to the total fragment library, and selecting the input fragments according to the priority of the feature vectors under each label to form a driving working condition.

6. The dynamic running condition construction method according to claim 1, characterized in that: and evaluating the constructed running condition by using two methods of relative error and speed acceleration combined distribution.

7. A dynamic driving condition construction system, characterized by comprising:

the data acquisition module acquires and preprocesses speed data of a vehicle to generate an input fragment X;

the coding module is used for constructing a joint learning framework based on a deep neural network and a bidirectional long-time memory network, and inputting the input segments into the joint learning framework to obtain a feature space Z;

the clustering module is used for realizing soft distribution clustering of the characteristic space Z by utilizing a regularization item based on relative entropy and obtaining a clustering result after iterative updating;

the driving condition construction module is used for classifying the input fragments according to the corresponding relation between the clustering result and the input fragments to obtain fragment libraries of various types, and selecting the input fragments from the fragment libraries of various types to form the driving condition;

when the speed data is preprocessed, removing invalid data in the speed data and filling missing values; extracting micro-stroke fragments from the speed data to generate a micro-stroke fragment library; carrying out interpolation processing on the micro-stroke fragment library to obtain an equal-length sequence library, and carrying out normalization processing on the equal-length sequence library to obtain an input fragment;

when an input segment is input into a joint learning framework, the joint learning framework comprises a self-encoder, the self-encoder comprises an encoder and a decoder, and the encoder sequentially processes the input segment through a deep neural network and a bidirectional long-term and short-term memory network;

the deep neural network learns the waveform of the short time scale in the input segment and extracts the local features of the waveform;

the bidirectional long-short time memory network learns the time connection between the waveforms across the time scale in the input segment, extracts the global characteristics of the input segment and further forms the characteristic space Z;

the decoder reconstructs the characteristic space by adopting up-sampling and deconvolution to form a reconstructed segment X';

pre-training the self-encoder to make the reconstructed segment X' output from the decoder and the input segment have the minimum mean square error

8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the building method according to any one of claims 1-6 when executing the computer program.

Images