A driver style identification method under different cycle conditions based on clustering model

By using a driver style identification method based on a clustering model, and utilizing in-vehicle T-Box data and a SOM network, nine typical driving scenarios are divided, solving the data interference problem in short-term driving style research and realizing accurate driver style analysis and guidance under different cyclic operating conditions.

CN115092165BActive Publication Date: 2024-11-19JILIN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210730724.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-24
Publication Date
2024-11-19
Estimated Expiration
2042-06-24

AI Technical Summary

Technical Problem

Existing technologies cannot effectively guide driving behavior over longer time periods when studying drivers' short-term driving styles, and abnormal environmental variables affect data analysis, leading to inaccurate driver style identification.

Method used

Based on a clustering model, data was collected through an in-vehicle T-Box to establish a driver style database. Using K-means clustering and a self-organizing map neural network (SOM) model, nine typical driving scenarios were divided, interference information was eliminated, and the style changes of drivers under different driving conditions were analyzed.

Benefits of technology

It enables more accurate identification of driver style under different cyclic operating conditions, eliminates interference from special driving behaviors, ensures that the data reflects the driver's true driving tendencies, and supports more effective driving behavior guidance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115092165B_ABST
    Figure CN115092165B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for identifying driver styles under different cycle conditions based on a clustering model, comprising the following steps: S1, establishing a driving style database: selecting a vehicle model and a route, collecting driving data, and preprocessing the data; S2, dividing typical driving scenarios: normalizing the data, setting the driving scenario type and boundary conditions, determining the optimal number of coarse clusters, and performing secondary clustering to obtain typical driving scenarios; S3, identifying the driver style under different cycle conditions: establishing driving style identification data features of different cycle conditions according to data information contained in typical driving scenarios, and eliminating and retaining data features based on the idea of ​​principal component analysis, introducing a self-organizing mapping neural network to establish a driver style identification model, inputting the filtered data features to obtain the driving behavior tendencies of different drivers under various cycle conditions, and realizing the identification of driver styles under different cycle conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of driving style identification, and more particularly to a method for driver style identification under different cycle conditions based on a clustering model. Background Art

[0002] At present, driving styles are classified into long-term driving styles and short-term driving styles according to the time period of driving style. Long-term driving style refers to the stable driving tendency shown by the driver in a long driving process, and is defined as a unified style label. Research on long-term driving style appeared earlier, and a large number of scholars have conducted extensive research. Shi-Huang Chen et al. obtained vehicle information such as vehicle speed and throttle opening through the on-board diagnostic system (OBD), analyzed the driver's behavioral tendencies through ensemble learning methods and support vector machine models (SVM), and then evaluated the driver's safety level during the entire work process. Chi-PanHwang et al. used statistical methods to process the collected multi-feature driving data, and based on this, used a decision tree model to study the driver's behavioral characteristics, and finally divided the driving style into four types. Although the research on long-term driving style has achieved good results, long-term driving style is a comprehensive evaluation of the driver's long-term driving behavior, ignoring the impact of road conditions and working conditions on the driver's style, and cannot reflect the differences in driving style under different driving scenarios. Therefore, the concept of short-term driving style was introduced. Short-term driving style is proposed based on specific driving conditions. It refers to the habitual driving tendency shown by the driver when he is in a certain driving scenario. When the driving scenario changes, the driver's behavioral tendency may also change. In recent years, research on short-term driving style has gradually emerged. Yao Y introduced variables such as road type, traffic environment and weather conditions in the study, used linear mixed models to analyze the impact of vehicle external factors on driver style, and established a prediction model to evaluate the risk level of driver behavior. Zhao Shishu extracted vehicle driving behavior data in typical scenarios of following, overtaking and lane changing, analyzed the differences in driver style in different scenarios, and finally proposed a quantitative evaluation method for short-term driving style. Overall, long-term driving style focuses on the driver's behavioral tendency throughout the driving process. It is a comprehensive evaluation of the driver's style in various driving scenarios, but it cannot reflect the differences in driver style under different conditions; short-term driving style recognizes that drivers have long-term fixed driving tendencies, and further studies the changes in driver style under specific working conditions. It can be seen that short-term driving style is a supplement and improvement to long-term driving style.

[0003] However, the research on short-term driving style is only based on specific driving conditions and cannot effectively guide driving behavior in a longer time domain. During the natural driving process of vehicles on actual roads, they are affected by non-driver factors such as surrounding vehicle behavior and traffic scenes, and the speed and acceleration and other states will change. For example, during the starting process, due to the abnormality of the vehicle in front, the main vehicle needs to complete braking avoidance or lane change overtaking operations, and the statistical data of the entire driving process will change significantly. Abnormal environmental variables belong to interference information in the driving data statistics process. With the increase in the complexity of the driving environment, the non-driver factor components contained in the driving data gradually increase, affecting the identification of the driver's style. It can be seen that there are certain limitations in analyzing data based on the continuous driving process and obtaining data features and using them to evaluate the style of natural driving vehicle drivers.

[0004] Therefore, providing a driver style identification method under different cycle conditions based on a clustering model is an urgent problem to be solved by those skilled in the art. Summary of the invention

[0005] In view of this, the present invention provides a driver style identification method under different cycle conditions based on a clustering model; based on the cloud data collected and uploaded by the T-Box on a fixed-route heavy commercial vehicle, a driver style database under natural driving conditions on actual roads is established for subsequent driving style classification research, and the speed-acceleration data space is introduced to redistribute the continuous driving process according to the vehicle speed and acceleration information, and nine typical driving scenes are obtained by clustering twice using K-means, each of which represents a specific driving state, and the driving cycle is divided into three driving conditions: starting, cruising and braking, and the driving condition is used to replace the special driving conditions in the previous short-term driving style research. The corresponding relationship between different cycle conditions and driving scenes is established, and the statistical characteristics of driving styles under different cycle conditions are constructed based on the driving scene, eliminating the interference of special driving behaviors on vehicle data, and ensuring that the driving data can more purely reflect the driver's driving style. Finally, a self-organizing map neural network (SOM) model is established to obtain the driver's driving style under starting conditions and cruising conditions, and further analyze the changes in the driver's style as the driving conditions change.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A driver style identification method under different cycle conditions based on a clustering model comprises the following steps:

[0008] S1. Establish a driving style database: select vehicle models and routes, collect driving data, and pre-process the data;

[0009] S2. Typical driving scene division: normalize the data, set the driving scene type and boundary conditions, determine the optimal number of coarse clusters, and perform secondary clustering to obtain typical driving scenes;

[0010] S3. Identify the driver's style under different cycle conditions: Establish the driving style identification data features under different cycle conditions based on the data information contained in the typical driving scenarios, and realize the elimination and retention of data features based on the principal component analysis idea. Introduce the self-organizing map neural network to establish a driver style identification model, input the filtered data features to obtain the driving behavior tendencies of different drivers under various cycle conditions, and realize the identification of driver style under different cycle conditions.

[0011] Preferably, the step S1 specifically includes:

[0012] S11. Based on the Laida criterion method, the abnormal data is eliminated. Assuming that a set of feature data only has random errors, the mean and standard deviation of the data are calculated; the sample points of the data set are re-tested, and the data points that differ from the sample mean by three times the standard deviation are regarded as abnormal data points:

[0013]

[0014] Among them, x i represents the th measurement value in the sample space, represents the mean of all measured values, and σ represents the standard deviation of all measured values;

[0015] S12. Use the classic filtering method to filter the high-frequency signal through Fourier transform, retain the low-frequency information that conforms to the vehicle acceleration characteristics, and finally obtain the noise-reduced acceleration data through inverse Fourier transform.

[0016] Preferably, the step S2 specifically includes:

[0017] S21, use 0-1 normalization method to process velocity and acceleration data:

[0018]

[0019] in, represents the measurement value of the i-th data at vehicle state b, and Represent the maximum and minimum measured values ​​on vehicle state b, respectively. To obtain the normalized value;

[0020] S22. Obtain the average value of the acceleration of the vehicle under acceleration and braking states:

[0021]

[0022]

[0023] Among them, a i and a j For all forward and reverse acceleration measurements, n + and n - are the number of acceleration measurements in the acceleration state and the number of acceleration measurements in the braking state, and finally the average value of all positive accelerations is obtained and the mean of the reverse acceleration

[0024] S23. Based on the acceleration mean information and the acceleration characteristic value of the cluster center, the following formula is used to define the vehicle acceleration state of each data point in the corresponding driving scene:

[0025]

[0026] in, and are the speed measurement value and acceleration measurement value of the i-th sample point in driving scene b, respectively, U acc , U cru and U bra They represent the acceleration state, speed maintenance state and braking state respectively. Represents the cluster center acceleration value of scene b; when the data type is an acceleration scene, the minimum acceleration value of the data cluster is greater than When the data type is the vehicle speed maintenance scenario, the maximum acceleration value of the data cluster is less than Minimum value greater than When the data type is a braking scenario, the maximum acceleration value of the data cluster is less than

[0027] S24. Find the minimum number of clusters that meet the boundary conditions, define the minimum number of clusters as N, and find the optimal number of clusters K within the limited range * , suppose that the data of type b contains m sample points, and the measurement value of a certain sample point is (u bi ,a bi ), the measurement values ​​of other sample points in the same type of data are (u bj ,a bj ), then the cohesion is:

[0028]

[0029] Assume that the data of type b contains n data points, and the measurement value of a data point is (u bi ,a bi ), there are n data points outside category b, then the separation expression is:

[0030]

[0031] The expression of the silhouette coefficient SC(i) of a single data point in the entire sample space is:

[0032]

[0033] Among them, SC(i)∈[-1,1]. As B(i) increases and A(i) decreases, SC(i) gradually increases, and the clustering effect of the model improves. The mean value of the silhouette coefficient is:

[0034]

[0035] Set different K values ​​to establish clustering models, compare the silhouette coefficients corresponding to the model output results, and determine the optimal number of coarse clusters K * The value of

[0036] S25. Determine K * Afterwards, the vehicle status is divided into three driving scenarios according to the speed: low speed, medium speed, and high speed; and the vehicle status is divided into three driving scenarios according to the acceleration: acceleration, speed maintenance, and braking. Finally, nine typical driving scenarios with both vehicle speed and vehicle acceleration are obtained by combining two of them. Here, the number of clusters of the K-means model is set to 9 to further classify the cluster centers obtained by the coarse clustering, and finally nine main driving scenarios with typical significance are obtained.

[0037] Preferably, the step S3 specifically includes:

[0038] S31, using the sliding window method, by analyzing the speed change information within the time window, determine the time when the vehicle enters and leaves the cruise state, realize the division of the entire cycle condition, define the window width as WI, use W b It represents the time window starting from the bth data in the data set. The average speed of the vehicle in this window is use As the reference benchmark for the vehicle speed change at each moment in the window, define S b To characterize the accumulated value of speed change within the window:

[0039]

[0040] Among them, v b,j Represents window W b For the jth element in C, define b,N To evaluate the cumulative change of the mean speed of N adjacent sliding windows:

[0041]

[0042] Based on S bThe definition states that the cumulative value of speed changes within the same window indicates the fluctuation level of the vehicle speed during that time window, based on C b,N According to the definition, the cumulative change in the average speed of adjacent sliding windows reflects the overall change trend of the vehicle speed between adjacent windows. Define the critical value tol for speed fluctuations within the window in and the critical value tol for speed changes between windows am When S b ≤tol in and C b,N ≤tol an it is considered that the vehicle is already in the cruise state;

[0043] S32. Define the original feature data matrix D ini and the low-dimensional feature data matrix D obtained through PCA processing PC Let D ini have p features, denoted as (f1, f2, …, f p ); D PC have q principal component features, denoted as (c1, c2, …, c q ), where q < p. Use the matrix U q×p to represent the feature transformation matrix:

[0044]

[0045] such that D PC = U q×p D ini Each element of U q×p can represent the weight coefficient of a certain original data feature in the corresponding principal component feature;

[0046] S33. Suppose a total of m principal components are retained. The i-th principal component is denoted as PC i , and its variance value is Var i . Then its contribution rate is expressed as:

[0047]

[0048] Obtain the contribution rate vector of all principal components

[0049] Let be the j-th column vector in the feature transformation matrix U q×p , that is, the weight coefficient of the j-th original data feature corresponding to each principal component Then it represents the sum of the weights occupied by feature j among all principal components and can also be used as the information contribution rate of feature j to the entire data space;

[0050] Let And let Conini Represents the vector composed of the contribution rate of all original data features, then Reference The calculation process of

[0051]

[0052] Con ini Each element in is arranged from large to small, the original data features are sorted, and the critical value Cri is set. The current k data features meet

[0053]

[0054] S34, select the SOM network model to realize the classification and identification of driver styles, and set the final number of driver styles to be N d , then the number of neurons in the output layer of the SOM network is N d The k data features of the driver are normalized to form the driving style recognition model input vector corresponding to the tens of elements, and the original data features X = (x1, x2, ... x k ), then after normalization

[0055]

[0056] in, is the input vector of the final SOM;

[0057] Take a random number between (0, 1) as the connection weight between the input layer neurons and the output layer neurons to initialize the value, let W be the initialized connection weight matrix, each input neuron corresponds to N d There are k×N output neurons in W. d elements, set is the input layer connection weight corresponding to the i-th neuron node in the output layer. It is known that it contains k elements, each element represents the weight coefficient of the input layer node corresponding to each output layer node;

[0058] Select an input feature vector And calculate the connection weights The output value after Calculate Y using Euclidean distance i The similarity of the weight coefficients corresponding to the neuron nodes in different output layers is calculated, and the output layer node with the highest similarity, that is, the smallest Euclidean distance, is defined as the winning node corresponding to the feature vector. The connection weight of the winning node is updated. The expression is as follows:

[0059] W i (t+1)=W i (t)+η(t)(Y i -Wi (t))

[0060] Among them, η(t) is the learning rate at time t, which decays over time and is set to t tol is the maximum number of iterations set, t is the current number of iterations, a is the initial learning rate, between 0 and 1; W i (t+1) is the (t+1)th connection weight value, W i (t) is the value of the t-th connection weight;

[0061] Continue to input the next driving behavior state feature vector into the SOM network and implement weight update. tol After several times, the network training is completed and the winning nodes of the output layer corresponding to different input vectors are determined. Each output layer node represents a comprehensive driving behavior tendency.

[0062] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses a method for identifying driver styles under different cycle conditions based on a clustering model; based on the cloud data collected and uploaded by the T-Box on a fixed-route heavy commercial vehicle, a driver style database under natural driving conditions on actual roads is established for subsequent driving style classification research, and the speed-acceleration data space is introduced to redistribute the continuous driving process according to the vehicle speed and acceleration information, and nine typical driving scenes are obtained by clustering twice using K-means, each of which represents a specific driving state, and the driving cycle is divided into three driving conditions: starting, cruising and braking, and the driving condition is used to replace the special driving conditions in the previous short-term driving style research. The corresponding relationship between different cycle conditions and driving scenes is established, and the statistical characteristics of driving styles under different cycle conditions are constructed based on the driving scene, eliminating the interference of special driving behaviors on vehicle data, and ensuring that the driving data can more purely reflect the driver's driving style. Finally, a self-organizing map neural network (SOM) model is established to obtain the driver's driving style under starting conditions and cruising conditions, and further analyze the changes in the driver's style as the driving conditions change. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0064] Figure 1 The accompanying drawing is a schematic diagram of the driver style change structure under different cycle conditions provided by the present invention.

[0065] Figure 2 The accompanying drawing is a schematic diagram of the division of typical driving scenarios provided by the present invention.

[0066] Figure 3 The accompanying drawing is a schematic diagram of driver style classification under different cycle conditions provided by the present invention. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] The embodiment of the present invention discloses a method for identifying driver style under different cycle conditions based on a clustering model, comprising the following steps:

[0069] S1. Establish a driving style database: select vehicle models and routes, collect driving data, and pre-process the data;

[0070] S2. Typical driving scene division: normalize the data, set the driving scene type and boundary conditions, determine the optimal number of coarse clusters, and perform secondary clustering to obtain typical driving scenes;

[0071] S3. Identify the driver's style under different cycle conditions: Establish the driving style identification data features under different cycle conditions based on the data information contained in the typical driving scenarios, and realize the elimination and retention of data features based on the principal component analysis idea. Introduce the self-organizing map neural network to establish a driver style identification model, input the filtered data features to obtain the driving behavior tendencies of different drivers under various cycle conditions, and realize the identification of driver style under different cycle conditions.

[0072] To further optimize the above technical solution, step S1 specifically includes:

[0073] S11. Based on the Laida criterion method, the abnormal data is eliminated. Assuming that a set of feature data only has random errors, the mean and standard deviation of the data are calculated; the sample points of the data set are re-tested, and the data points that differ from the sample mean by three times the standard deviation are regarded as abnormal data points:

[0074]

[0075] Among them, x i represents the th measurement value in the sample space, represents the mean of all measured values, and σ represents the standard deviation of all measured values;

[0076] S12. Use the classic filtering method to filter the high-frequency signal through Fourier transform, retain the low-frequency information that conforms to the vehicle acceleration characteristics, and finally obtain the noise-reduced acceleration data through inverse Fourier transform.

[0077] To further optimize the above technical solution, step S2 specifically includes:

[0078] S21, use 0-1 normalization method to process velocity and acceleration data:

[0079]

[0080] in, represents the measurement value of the i-th data at vehicle state b, and Represent the maximum and minimum measured values ​​on vehicle state b, respectively. To obtain the normalized value;

[0081] S22. Obtain the average value of the acceleration of the vehicle under acceleration and braking states:

[0082]

[0083]

[0084] Among them, a i and a j For all forward and reverse acceleration measurements, n + and n - are the number of acceleration measurements in the acceleration state and the number of acceleration measurements in the braking state, and finally the average value of all positive accelerations is obtained and the mean of the reverse acceleration

[0085] S23. Based on the acceleration mean information and the acceleration characteristic value of the cluster center, the following formula is used to define the vehicle acceleration state of each data point in the corresponding driving scene:

[0086]

[0087] in, and are the speed measurement value and acceleration measurement value of the i-th sample point in driving scene b, respectively, U acc , U cru and U bra They represent the acceleration state, speed maintenance state and braking state respectively. Represents the cluster center acceleration value of scene b; when the data type is an acceleration scene, the minimum acceleration value of the data cluster is greater than When the data type is the vehicle speed maintenance scenario, the maximum acceleration value of the data cluster is less than Minimum value greater than When the data type is a braking scenario, the maximum acceleration value of the data cluster is less than

[0088] Table 1. Acceleration state boundary conditions

[0089]

[0090] S24. Find the minimum number of clusters that meet the boundary conditions, define the minimum number of clusters as N, and find the optimal number of clusters K within the limited range * , suppose that the data of type b contains m sample points, and the measurement value of a certain sample point is (u bi ,a bi ), the measurement values ​​of other sample points in the same type of data are (u bj ,a bj ), then the cohesion is:

[0091]

[0092] Assume that the data of type b contains n data points, and the measurement value of a data point is (u bi ,a bi ), there are n data points outside category b, then the separation expression is:

[0093]

[0094] The expression of the silhouette coefficient SC(i) of a single data point in the entire sample space is:

[0095]

[0096] Among them, SC(i)∈[-1,1]. As B(i) increases and A(i) decreases, SC(i) gradually increases, and the clustering effect of the model improves. The mean value of the silhouette coefficient is:

[0097]

[0098] Set different K values ​​to establish clustering models, compare the silhouette coefficients corresponding to the model output results, and determine the optimal number of coarse clusters K * The value of

[0099] S25. Determine K *Afterwards, the vehicle status is divided into three driving scenarios according to the speed: low speed, medium speed, and high speed; and the vehicle status is divided into three driving scenarios according to the acceleration: acceleration, speed maintenance, and braking. Finally, nine typical driving scenarios with both vehicle speed and vehicle acceleration are obtained by combining two of them. Here, the number of clusters of the K-means model is set to 9 to further classify the cluster centers obtained by the coarse clustering, and finally nine main driving scenarios with typical significance are obtained.

[0100] To further optimize the above technical solution, step S3 specifically includes:

[0101] S31, using the sliding window method, by analyzing the speed change information within the time window, determine the time when the vehicle enters and leaves the cruise state, realize the division of the entire cycle condition, define the window width as WI, use W b It represents the time window starting from the bth data in the data set. The average speed of the vehicle in this window is use As the reference benchmark for the vehicle speed change at each moment in the window, define S b To characterize the accumulated value of speed change within the window:

[0102]

[0103] Among them, v b,j Represents window W b For the jth element in C, define b,N To evaluate the cumulative change of the mean speed of N adjacent sliding windows:

[0104]

[0105] Based on S b The definition of C is that the accumulated value of speed change in the same window indicates the fluctuation level of vehicle speed in the time window. b,N The definition of , the cumulative change of the mean speed of adjacent sliding windows reflects the overall change trend of the speed between adjacent windows, and the critical value of speed fluctuation within the window is defined as tol in and the critical value of vehicle speed change between windows tol am , when S b ≤tol in And C b,N ≤tol am When the vehicle is considered to be in the cruising state;

[0106] S32, define the original feature data matrix D ini And the low-dimensional feature data matrix D obtained after PCA processing PC , let D ini It has p features, represented by (f1,f2,…,fp ); D PC It has q principal component features, denoted as (c1, c2, …, c q ), where q < p, and the matrix U q×p is used to represent the feature transformation matrix:

[0107]

[0108] such that D PC = U q×p D ini , and each element of U q×p can represent the weight coefficient of a certain original data feature in the corresponding principal component feature;

[0109] S33. Suppose a total of m principal components are retained, and the i-th principal component is denoted as PC i , and its variance value is Var i , then its contribution rate is denoted as:

[0110]

[0111] Obtain the contribution rate vector of all principal components

[0112] Suppose is the j-th column vector in the feature transformation matrix U q×p , that is, the weight coefficient of the j-th original data feature corresponding to each principal component, then it represents the sum of weights occupied by feature j in all principal components, and can also be used as the information contribution rate of feature j to the entire data space;

[0113] Let and let Con ini represent the vector composed of the contribution rates of all original data features, then Referring to the calculation process, we can obtain

[0114]

[0115] Sort each element in Con ini from largest to smallest, sort the original data features, set the critical value Cri, and the current k data features satisfy

[0116]

[0117] S34. Select the SOM network model to achieve the classification and identification of driver styles. SOM is a neural network layer with only a feedforward structure. Generally, it only has an input layer and a competitive layer. Therefore, the SOM network structure is determined by the input layer and the output layer network structures.

[0118] The number of neurons in the input layer is determined by the number of features of the input data. One neuron in the input layer corresponds to one feature of the input data. Therefore, the number of neurons in the input layer is equal to the number of input features, which is the k value determined in S33.

[0119] The competition layer network can have a variety of network structures. Since the number of competition layer neurons determines the number of clusters, and the number of clusters is the number of driving styles obtained in the end, the number of output layer neurons can be determined by determining the number of driving styles. In engineering applications, the number of driving styles is generally only 3 to 5, that is, the number of competition layer neurons is about 3 to 5. The number of neurons is relatively small, so a two-dimensional planar network structure can be selected. Different neurons in a two-dimensional planar structure are generally connected to each other through geometric structures such as rectangles or hexagons. Suppose the final number of driver styles is N. d The number of neurons in the output layer of the SOM network is also N d indivual.

[0120] Each driver's driving data contains multiple data features, and k data features are finally selected by S33 process; the k data features of each driver are normalized to form the input vector of the driving style recognition model corresponding to the driver. Assume that the original data feature X = (x1, x2, ... x k ), then after normalization

[0121]

[0122] This is the input vector of the final SOM.

[0123] Take a random number between (0, 1) to initialize the connection weight between the input layer neurons and the output layer neurons. Let W be the initialized connection weight matrix, and each input neuron corresponds to N d There are k×N output neurons in W. d elements. is the input layer connection weight corresponding to the i-th neuron node in the output layer, which contains k elements, and each element represents the weight coefficient of each output layer node corresponding to the input layer node.

[0124] Select an input feature vector And calculate the connection weights The output value after Calculate Y using Euclidean distance iThe similarity of the weight coefficients corresponding to the neuron nodes in the different output layers is calculated, and the output layer node with the highest similarity (the smallest Euclidean distance) is defined as the winning node corresponding to the feature vector, and the relevant connection weight values ​​are further updated based on the winning node. The general SOM network weight update requires setting the neighborhood radius σ, and all nodes within σ participate in the weight update, and the closer to the winning node, the larger the update amplitude, and the farther away from the winning node, the smaller the update amplitude. However, since the number of neurons in the output layer is small in this method, the weight update can be simplified to only update the connection weights of the winning node.

[0125] Update the weight of the winning node, the expression is as follows:

[0126] W i (t+1)=W i (t)+η(t)(Y i -W i (t))

[0127] Among them, η(t) is the learning rate at time t, which decays over time and is generally set to t tol is the maximum number of iterations set, t is the current number of iterations, a is the initial learning rate, generally between 0 and 1; W i (t+1) is the (t+1)th connection weight value, W i (t) is the tth connection weight value; it can be seen that in each update process, the connection weight value corresponding to the winning node continues to approach the input feature.

[0128] Continue to input the next driving behavior state feature vector into the SOM network and implement weight update. tol After the network training is completed, the winning nodes of the output layer corresponding to different input vectors are determined. Each input vector corresponds to a set of driving state features, different driving state features correspond to specific input layer neuron nodes, and each output layer neuron node has a corresponding set of input layer neuron nodes. From the training process of the connection weights between the input layer network nodes and the output layer network nodes, it can be seen that the value of the connection weight of each winning node contains a set of state information of the input data features. These state information also intuitively reflects the style characteristics of the driver. Therefore, it is believed that each output layer node represents a comprehensive driving behavior tendency.

[0129] The driver style corresponding to the different input features of the classification is evaluated and identified. The driving state features that can be used for driver style classification are mainly composed of information such as vehicle speed, acceleration, pedal opening, etc., and it is believed that the larger the mean of each feature, the higher the driver's operating intensity; the larger the variance of each feature, the greater the driver's operating volatility, both of which represent a more aggressive driving style. Taking vehicle speed information as an example, when the driver's average driving speed is higher and the speed changes more dramatically, it means that the driver's operation is more intense and the driving style is more aggressive.

[0130] The above shows that the driving state characteristics of each input model can be evaluated for aggressiveness and conservatism based on the actual physical meaning, and a set of input feature vectors contains multiple driving state characteristics, each of which has its own conservative and aggressive characteristics. It is believed that each feature has the same contribution rate to the driver's aggressiveness. The greater the proportion of aggressive driving state in the k input features, the higher the driver's aggressiveness level. A driver whose all characteristics are aggressive is defined as an aggressive driver, and a driver whose all characteristics are conservative is defined as a conservative driver. The transitional driving style can be calculated based on the total number of driving styles N. d Personalized division is carried out, and finally the driver's style is classified and identified.

[0131] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0132] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A driver style identification method under different cycle conditions based on a clustering model, characterized in that: The following steps are involved: S1. Establish a driving style database: Select vehicle models and routes, collect driving data, and establish a driver style database under natural driving conditions on actual roads based on the cloud data collected and uploaded by the T-Box on heavy commercial vehicles on fixed routes; and pre-process the data; S2. Typical driving scene division: normalize the data, set the driving scene type and boundary conditions, determine the optimal number of coarse clusters, and perform secondary clustering to obtain typical driving scenes; S3. Identify the driver's style under different cycle conditions: establish the driving style identification data features of different cycle conditions based on the data information contained in the typical driving scene, and realize the elimination and retention of data features based on the principal component analysis idea, introduce the self-organizing map neural network to establish the driver style identification model, input the filtered data features to obtain the driving behavior tendencies of different drivers under various cycle conditions, and realize the identification of the driver's style under different cycle conditions; The step S2 specifically includes: S21, use 0-1 normalization method to process velocity and acceleration data: in, represents the measurement value of the i-th data at vehicle state b, and Represent the maximum and minimum measured values ​​on vehicle state b, respectively. To obtain the normalized value; S22. Obtain the average value of the acceleration of the vehicle under acceleration and braking states: Among them, a i and a j For all forward and reverse acceleration measurements, n + and n - are the number of acceleration measurements in the acceleration state and the number of acceleration measurements in the braking state, and finally the average value of all positive accelerations is obtained and the mean of the reverse acceleration S23. Based on the acceleration mean information and the acceleration characteristic value of the cluster center, the following formula is used to define the vehicle acceleration state of each data point in the corresponding driving scene: in, and are the speed measurement value and acceleration measurement value of the i-th sample point in driving scene b, respectively, U acc , U cru and U bra They represent acceleration state, speed keeping state and braking state respectively. Represents the cluster center acceleration value of scene b; when the data type is an acceleration scene, the minimum acceleration value of the data cluster is greater than When the data type is the vehicle speed maintenance scenario, the maximum acceleration value of the data cluster is less than Minimum value greater than When the data type is a braking scenario, the maximum acceleration value of the data cluster is less than S24. Find the minimum number of clusters that meet the boundary conditions, define the minimum number of clusters as N, and find the optimal number of coarse clusters K within the limited range * , suppose that the data of type b contains m sample points, and the measurement value of a certain sample point is (u bi ,a bi ), the measurement values ​​of other sample points in the same type of data are (u bj ,a bj ), then the cohesion is: Assume that the data of type b contains n data points, and the measurement value of a data point is (u bi ,a bi ), there are n data points outside category b, then the separation expression is: The expression of the silhouette coefficient SC(i) of a single data point in the entire sample space is: Among them, SC(i)∈[-1,1]. As B(i) increases and A(i) decreases, SC(i) gradually increases, and the clustering effect of the model improves. The mean value of the silhouette coefficient is: Set different K values ​​to establish clustering models, compare the silhouette coefficients corresponding to the model output results, and determine the optimal number of coarse clusters K * The value of S25. Determine K * Afterwards, the vehicle status is divided into three driving scenarios according to the vehicle speed: low speed, medium speed, and high speed; and the vehicle status is divided into three driving scenarios according to the acceleration: acceleration, speed maintenance, and braking. Finally, nine typical driving scenarios with both vehicle speed and vehicle acceleration status are obtained by combining two of them. Here, the number of clusters of the K-means model is set to 9 to further classify the cluster centers obtained by the coarse clustering, and finally nine main driving scenarios with typical significance are obtained. The step S3 specifically includes: S31, using the sliding window method, by analyzing the speed change information within the time window, determine the time when the vehicle enters and leaves the cruise state, realize the division of the entire cycle condition, define the window width as WI, use W b It represents the time window starting from the bth data in the data set. The average speed of the vehicle in this window is use As the reference benchmark for the vehicle speed change at each moment in the window, define S b To characterize the accumulated value of speed change within the window: Among them, v b,j Represents window W b For the jth element in C, define b,N To evaluate the cumulative change of the mean speed of N adjacent sliding windows: Based on S b The definition of C is that the accumulated value of speed change in the same window indicates the fluctuation level of vehicle speed in the time window. b,N The definition of , the cumulative change of the mean speed of adjacent sliding windows reflects the overall change trend of the speed between adjacent windows, and the critical value of speed fluctuation within the window is defined as tol in and the critical value of vehicle speed change between windows tol am , when S b ≤tol in And C b,N ≤tol am When the vehicle is considered to be in the cruising state; S32. Define the original feature data matrix D ini and the low-dimensional feature data matrix D obtained through PCA processing PC . Assume that D ini has p features, denoted as (f1, f2, …, f p ); D PC has q principal component features, denoted as (c1, c2, …, c q ), where q < p. Use the matrix U q×p to represent the feature transformation matrix: Make D PC =U q×p D ini , U q×p Each element of can represent the weight coefficient of a certain original data feature in the corresponding principal component feature; S33, suppose that a total of m principal components are retained, and the i-th principal component is represented by PC i , whose variance is Var i , then its contribution rate It is expressed as: Get all principal component contribution rate vectors set up is the feature transformation matrix U q×p The j-th column vector in , that is, the j-th original data feature corresponds to the weight coefficient of each principal component, It represents the weight and weight of feature j in all principal components, and can also be used as the information contribution rate of feature j to the entire data space; make And let Con ini Represents the vector composed of the contribution rate of all original data features, then Reference The calculation process can be obtained: Con ini Arrange each element from large to small, sort the original data features, set the critical value Cri, and the current k data features meet: S34, select the SOM network model to realize the classification and identification of driver styles, and set the final number of driver styles to be N d , then the number of neurons in the output layer of the SOM network is N d The k data features of the driver are normalized to form the driving style recognition model input vector corresponding to the neuron, and the original data features X = (x1, x2, ... x k ), then after normalization: in, is the input vector of the final SOM; Take a random number between (0, 1) as the connection weight between the input layer neurons and the output layer neurons to initialize the value, let W be the initialized connection weight matrix, each input neuron corresponds to N d There are k×N output neurons in W. d elements, set is the input layer connection weight corresponding to the i-th neuron node in the output layer. It is known that it contains k elements, each element represents the weight coefficient of the input layer node corresponding to each output layer node; Select an input feature vector And calculate the connection weights The output value after Calculate Y using Euclidean distance i The similarity of the weight coefficients corresponding to the neuron nodes in different output layers is calculated, and the output layer node with the highest similarity, that is, the smallest Euclidean distance, is defined as the winning node corresponding to the feature vector. The connection weight of the winning node is updated. The expression is as follows: W i (t+1)=W i (t)+η(t)(Y i -W i (t)) Among them, η(t) is the learning rate at time t, which decays over time and is set to t tol is the maximum number of iterations set, t is the current number of iterations, a is the initial learning rate, between 0 and 1; W i (t+1) is the (t+1)th connection weight value, W i (t) is the value of the t-th connection weight; Continue to input the next driving behavior state feature vector into the SOM network and implement weight update. tol After several times, the network training is completed and the winning nodes of the output layer corresponding to different input vectors are determined. Each output layer node represents a comprehensive driving behavior tendency.

2. The driver style identification method under different cycle conditions based on clustering model according to claim 1 is characterized in that: The step S1 specifically includes: S11. Based on the Laida criterion method, the abnormal data is eliminated. Assuming that a set of feature data only has random errors, the mean and standard deviation of the data are calculated; the sample points of the data set are re-tested, and the data points that differ from the sample mean by three times the standard deviation are regarded as abnormal data points: Among them, x i represents the th measurement value in the sample space, represents the mean of all measured values, and σ represents the standard deviation of all measured values; S12. Use the classic filtering method to filter the high-frequency signal through Fourier transform, retain the low-frequency information that conforms to the vehicle acceleration characteristics, and finally obtain the noise-reduced acceleration data through inverse Fourier transform.