A multi-time scale driving style analysis method

Abstract

The present application relates to the technical field of intelligent auxiliary driving, and specifically provides a multi-time-scale driving style analysis method, which obtains driving time sequence segments in seconds through a multi-resolution analysis method and classifies working conditions; short-time driving style labels are marked through a driving behavior clustering method, a short-time driving style classification model is constructed and trained; a medium-long-time driving task window in minutes or hours is set, short-time driving styles are extracted using the trained short-time driving style classification model, distribution vectors are obtained according to short-time driving style sequences in the medium-long-time driving task window and mapped to a predefined medium-long-time style mode space, and a corresponding medium-long-time style mode is identified through a clustering model; a long-time style window in weeks, months, seasons or years is set, distribution vectors are obtained by counting medium-long-time style mode sequences in the long-time style window and mapping to a predefined long-time style evolution space, and long-time style evolution analysis of a driver is realized through a clustering classification model.

Description

Technical Field

[0001] This invention belongs to the field of intelligent assisted driving technology, and in particular relates to a driving style analysis method with multiple time scales. Background Technology

[0002] Driving style analysis, as a core component of traffic behavior modeling, essentially uses multi-dimensional driving data to analyze a driver's operating habits and decision-making patterns. With the rapid development of vehicle-to-everything (V2X) technology, onboard sensors can collect real-time data from various channels, such as steering wheel angle, accelerator / brake pedal travel, and vehicle dynamics parameters, providing a rich data foundation for refined driving style analysis. However, driving behavior exhibits significant differences across time scales: short-term scales (seconds) primarily reflect the intensity and frequency of specific actions, such as rapid acceleration and braking; medium- to long-term scales (minutes to hours) reflect combinations of personalized driving behaviors, such as high-frequency aggressive driving or constantly switching driving styles; and long-term scales (days to grades) reflect the stability and periodicity of driving habits, such as style differences during weekly commutes.

[0003] For driving style analysis technology, existing technologies can be mainly divided into two categories:

[0004] 1. Short-term analysis method: Based on real-time sensor data, instantaneous driving behaviors (such as rapid acceleration and sudden braking) are identified through threshold judgment or shallow machine learning models (such as SVM and decision tree).

[0005] 2. Long-term analysis method: Relying on historical statistical data (such as monthly average vehicle speed and frequency of emergency braking), statistical models or clustering algorithms (such as K-means) are used to classify driving habit types.

[0006] While existing technologies can determine driver style to some extent, they still have many problems:

[0007] First, current driving style analysis methods suffer from problems such as fragmented time scales and a lack of fusion across multiple time scales. Existing short-term methods typically ignore long-term behavioral evolution trends, while long-term methods lack fine-grained analysis of specific driving scenarios (such as overtaking and lane changing). Existing methods often process short-term and long-term data in isolation, failing to construct a full-cycle profile of driving style. They also lack the ability to model medium- and long-term driving style strategies, making it difficult to reveal the intrinsic connection between short, rapid driving behaviors and stable, long-term driving habits.

[0008] Furthermore, existing driving style analysis methods have weak dynamic modeling capabilities and insufficient feature representation. Their analysis and classification models rely on fixed time windows or static features, making it difficult to capture the temporal evolution of driving behavior. Most analysis frameworks only use time-domain statistical features (mean, variance) and frequency-domain dynamic features (such as power spectral density), without integrating deep semantic features (spatiotemporal patterns extracted through neural networks), making it difficult to characterize the nonlinear laws of driving behavior and affecting classification accuracy.

[0009] Furthermore, existing driving style analysis methods exhibit low sensitivity to scenarios and coarse-grained condition recognition. They typically fail to differentiate the impact of different driving conditions on style and lack differentiated analysis strategies for specific driving conditions (such as turning and following other vehicles), resulting in poor generalization of classification results. Even when considering the impact of driving conditions on style, most condition recognition relies on fixed rules (such as determining a turn as a steering wheel angle >30°) or simple similarity measures (Euclidean distance). In complex scenarios (such as curves with gradually changing curvature or lane changes involving multiple vehicles), the misclassification rate is high. Therefore, a single rule or simple measure is insufficient to distinguish semantically similar conditions, leading to impure subset splits, high input noise to subsequent classification models, and consequently affecting the accuracy of subsequent driving style analysis. Summary of the Invention

[0010] In view of this, the present invention aims to provide a driving style analysis method with multiple time scales. By fusing multiple time scales, it realizes hierarchical mapping and dynamic association of second-level operation actions, minute-level driving strategy preferences, and weekly / monthly / yearly habit patterns. It also integrates a rule engine and intelligent algorithms to significantly reduce the probability of misjudgment of driving conditions. Through deep feature representation and probabilistic simplex feature representation, it realizes multi-dimensional time-series data feature expression, breaks down the intrinsic connection between short and fast driving behaviors and stable long-term driving habits, and realizes full-cycle style characterization of drivers.

[0011] To achieve the above objectives, the technical solution created by this invention is implemented as follows:

[0012] This invention provides a multi-timescale driving style analysis method, comprising:

[0013] S1: The multi-channel driving time-series data stream is split into driving time-series segments of various lengths in seconds using a multi-resolution analysis method, and the driving time-series segments are classified according to working conditions to obtain multiple working condition datasets;

[0014] S2: Use driving behavior clustering to cluster driving time segments in each working condition dataset and label each driving time segment with a short-term driving style label;

[0015] A short-term driving style classification model is constructed, and the short-term driving style classification model is trained using driving time-series segments labeled with short-term driving styles; the trained short-term driving style classification model is used to extract the short-term driving style of the input driving time-series segments;

[0016] S3: Set a time window in minutes or hours as the medium-to-long-term driving task window. The driving time sequence segments within the medium-to-long-term driving task window constitute the driving task segment. Use the short-term driving style classification model to extract the short-term driving style of each driving time sequence segment in the driving task segment, obtain the short-term driving style distribution vector of each driving task segment, map the short-term driving style distribution vector to the predefined medium-to-long-term style pattern space, and use clustering method to identify the medium-to-long-term style pattern corresponding to the driving task segment.

[0017] S4: Set a time window with a week, month, quarter, or year as the long-term style window. The driving task segments within the long-term style window constitute the long-term driving style evolution segment. Statistically analyze the medium-to-long-term style patterns corresponding to the driving task segments within the long-term driving style evolution segment to obtain the medium-to-long-term style pattern sequence of the long-term driving style evolution segment. Map the medium-to-long-term style pattern sequence to a predefined long-term style evolution space and use clustering methods to perform long-term style evolution analysis on the driver.

[0018] Preferably, the process of acquiring the working condition dataset includes:

[0019] Multi-channel driving time-series data streams are collected by onboard sensors;

[0020] The multi-channel driving time-series data stream is split and classified using a multi-resolution analysis method to obtain a set of driving time-series data segments with different time lengths.

[0021] Select driving time series data segments corresponding to longitudinal driving conditions and driving time series data segments corresponding to lateral driving conditions from the set of driving time series data segments;

[0022] Based on rule-based methods, secondary classification is performed on driving time-series data segments corresponding to longitudinal driving conditions to obtain longitudinal driving condition subset datasets.

[0023] Based on the time series similarity measurement method, the driving time series data segments corresponding to the horizontal driving conditions are classified into two levels to obtain the horizontal driving condition subset dataset;

[0024] The longitudinal driving condition subset and the lateral driving condition subset are used as the driving condition dataset.

[0025] Preferably, the multi-resolution analysis method is a multi-scale sliding window, dynamic range adjustment, or recursive segmentation and matching method.

[0026] Preferably, the rule-based method for performing secondary classification of driving time-series data segments corresponding to longitudinal driving conditions includes:

[0027] Based on threshold determination or state machine methods, secondary classification rules for longitudinal driving conditions are set, and the driving time series data segments corresponding to the longitudinal driving conditions are classified in a secondary manner according to the secondary classification rules.

[0028] Preferably, the method based on time series similarity measurement performs secondary classification of driving time-series data segments corresponding to lateral driving conditions, including:

[0029] A lateral driving condition template is set, and a time series similarity measurement method is used to match the driving time series data segments corresponding to the lateral driving condition with the lateral driving condition template to realize the identification and classification of lateral driving conditions. The time series similarity measurement method is Euclidean distance, dynamic time warping, time envelope edit distance, cosine similarity, Pearson correlation coefficient or symbolic aggregation approximation method.

[0030] Preferably, the step of clustering driving time-series segments in each working condition dataset using a driving behavior clustering method, and labeling each driving time-series segment with a short-term driving style label, includes:

[0031] A deep neural network encoder is constructed, and a driving time-series segment is input into the deep neural network encoder to extract a low-dimensional deep feature vector. The deep neural network encoder includes: a convolutional feature extraction layer, an LSTM encoding layer connected to the output of the convolutional feature extraction layer, and a feature fusion layer connected to the output of the LSTM encoding layer. The convolutional feature extraction layer is used to expand the feature dimension and compress the temporal dimension of the driving time-series segment to obtain local features; the LSTM encoding layer is used to capture the temporal features of the driving time-series segment, establish long-term temporal dependencies of local features, and obtain feature vectors; the feature fusion layer is used to reduce the dimensionality of the feature vectors to obtain the low-dimensional deep feature vector.

[0032] The low-dimensional deep feature vectors are clustered using a deep embedding clustering method. Short-term driving styles are generated based on the clustering results, and each driving time-series segment is labeled with a short-term driving style tag.

[0033] Preferably, the clustering of the low-dimensional deep feature vector using a deep embedding clustering method includes:

[0034] The K-Means algorithm is used to initialize the low-dimensional deep feature vector multiple times, and the optimal K cluster centers are selected based on the contour coefficient, where K is the number of predefined driving styles;

[0035] The soft assignment probability between each low-dimensional deep feature vector and each cluster center is calculated based on Student's t-distribution;

[0036] Construct a target loss function, and iteratively optimize the location parameters of the cluster centers by minimizing the target loss function.

[0037] Preferably, the short-term driving style distribution vector for each driving task segment is obtained, the short-term driving style distribution vector is mapped to a predefined medium-to-long-term style pattern space, and a clustering method is used to identify the medium-to-long-term style pattern corresponding to the driving task segment, including:

[0038] Calculate the weighted frequency of short-time driving style for each driving time segment in the driving task segment, and generate a short-time driving style distribution vector;

[0039] The short-term driving style distribution vector is mapped to the probability simplex space to obtain the coordinate representation of the short-term driving style distribution vector in the probability simplex space;

[0040] Based on the coordinate representation in the probabilistic simplex space, the mid-to-long-term style patterns corresponding to the driving task segments are analyzed and identified using clustering methods.

[0041] Preferably, the step of mapping the medium- to long-term style pattern sequence to a predefined long-term style evolution space and using clustering methods to perform long-term style evolution analysis on the driver includes:

[0042] A variational autoencoder model is constructed, and the variational autoencoder model is used to extract features from medium- and long-term style pattern sequences to obtain low-dimensional long-term style evolution feature vectors.

[0043] Clustering methods are performed in the long-term style evolution space to conduct long-term style evolution analysis of drivers.

[0044] Preferably, the short-term driving style labels include: aggressive, conservative, and normal.

[0045] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0046] This invention employs a multi-scale sliding window technique with 50-2000 sampling points to achieve full coverage from instantaneous operations at the second level to continuous operating conditions at the minute level, avoiding information loss caused by fixed time windows and providing excellent data samples for subsequent driving style analysis.

[0047] In the process of identifying driving conditions, this invention integrates a rule-based judgment method based on the yaw rate energy threshold with a similarity measurement algorithm to achieve high-precision, fine-grained identification of longitudinal and lateral driving conditions, significantly reducing the misjudgment rate of driving conditions in complex scenarios such as curves with gradual curvature and continuous lane changes.

[0048] This invention proposes a personalized driving behavior recognition technology based on deep embedding clustering. The 32-dimensional feature vectors extracted by the deep embedding clustering model based on the CNN-LSTM encoder have correlation coefficients between features that are generally lower than 0.1, with the vast majority in a state of extremely weak correlation or no correlation. This achieves low-redundancy feature representation, which helps the subsequent clustering model to learn the relationship between features more accurately, improves the model's generalization ability, and significantly improves the convergence speed and generalization ability of clustering and classification models.

[0049] This invention uses time windows measured in minutes or hours as driving task windows, compensating for the lack of long-term modeling in traditional analysis frameworks. It maps the mixing ratio of driving styles to a two-dimensional probabilistic simplex space, utilizing its geometric properties to transform abstract driving strategy preferences into intuitive spatial locations and evolutionary trajectories, greatly facilitating the understanding and analysis of driver behavior strategies. Furthermore, this invention constructs a long-term style panoramic analysis architecture based on VAE (Visual Architecture for Environment Optimization). By extracting six long-term features, including style separation, variability, and RSI (Responsive Stability Index), it achieves a comprehensive characterization of the evolution of driving styles on daily to yearly scales. The Style Momentum Index (RSI) can predict the changing trend of driving styles, providing a 3-6 month lead time for risk warning. Attached Figure Description

[0050] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0051] Figure 1 This is a flowchart of a multi-timescale driving style analysis method provided according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and do not constitute a limitation thereof. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0053] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined to form various implementations. Furthermore, the order of the steps or actions in the method description can be changed or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various orders in the specification and drawings are merely for the clear description of a particular embodiment and do not imply a mandatory order, unless otherwise stated that a particular order must be followed.

[0054] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0055] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0056] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0057] Please see Figure 1 In one embodiment of the present invention, a multi-timescale driving style analysis method is provided, integrating multi-resolution data processing, hierarchical feature modeling, and time series analysis technologies. This enables full-cycle analysis covering second-level operation actions, minute / hour-level driving strategy preferences, and weekly / monthly / year-level habit patterns. It can be applied to multiple fields such as personalized control of autonomous driving, optimization of intelligent driving assistance systems, driver behavior modeling, traffic safety assessment and early warning, and insurance risk modeling. This driving style analysis method includes:

[0058] S1: Multi-resolution analysis is used to split the multi-channel driving time-series data stream, obtaining driving time-series segments of various lengths in seconds. These segments are then classified according to work conditions to obtain multiple work condition datasets.

[0059] Specifically, firstly, the multi-channel driving time-series data stream collected by onboard sensors must include: steering wheel angle, steering wheel angle direction indicator, vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate, and timestamp. Additionally, it may include channels for longitudinal vehicle speed, steering angle, steering angle change rate, accelerator pedal opening, brake pedal pressure, and gear information.

[0060] Since the data is collected by different sensors, the collection frequency and time delay may be inconsistent, requiring preprocessing. First, the driving time sequence data for each channel is aligned using timestamps, and the sampling frequency is unified to a set reference frequency. In this embodiment of the invention, the reference frequency is set to 100Hz, and data continuity is ensured through resampling and interpolation processing.

[0061] Secondly, the multi-channel driving time-series data stream is split using a sliding window. Driving time-series data segments corresponding to longitudinal and lateral driving conditions are selected from the set of driving time-series data segments. Based on a rule-based method, the driving time-series data segments corresponding to longitudinal driving conditions are classified into secondary categories to obtain a longitudinal driving condition subset. Based on a time-series similarity metric, the driving time-series data segments corresponding to lateral driving conditions are classified into secondary categories to obtain a lateral driving condition subset. The longitudinal and lateral driving condition subsets are then used as the driving condition dataset.

[0062] Specifically, a driving characteristic index is set, with yaw rate as the driving characteristic index, and the average energy of yaw rate is calculated according to the following formula. for:

[0063] ;

[0064] in, Indicates the first The square of the yaw rate at the sampling time, where N represents the total number of sampling points. The average energy of the yaw rate The length of the energy calculation interval, , This indicates the length of the long-scale sliding window. The sampling interval represents the yaw rate.

[0065] Based on statistical analysis of a large amount of actual driving data, it was found that setting the energy threshold for yaw rate to 0.5 rad² / s can accurately distinguish between lateral and longitudinal driving conditions. In this embodiment of the invention, when the calculated average energy of yaw rate... If the energy level is below the set threshold of 0.5 rad² / s, the corresponding time is considered to be a longitudinal driving condition; otherwise, it is considered a lateral driving condition.

[0066] When performing secondary classification on driving time-series data segments corresponding to longitudinal driving conditions, common rule-based methods include threshold determination methods or state machine methods. Secondary classification rules for longitudinal driving conditions are set based on these methods, and the driving time-series data segments corresponding to these conditions are then classified according to these rules. Furthermore, to avoid frequent switching of driving conditions due to noise, this invention introduces a driving condition continuity constraint. Specifically, a data segment is only marked as a longitudinal driving condition if 100 consecutive sampling points (1 second) meet the longitudinal driving condition conditions. For brief fluctuations in yaw rate energy, the system performs smoothing processing to maintain the stability of the driving condition identification.

[0067] When performing secondary classification of driving time-series data segments corresponding to lateral driving conditions, selectable time series similarity measurement methods include Euclidean distance, dynamic time warping, time envelope edit distance, cosine similarity, Pearson correlation coefficient, or symbolic aggregation approximation methods. Taking dynamic time warping as an example, firstly, a lateral driving condition template is set, and then the driving time-series data segments corresponding to the lateral driving conditions are selected from the data segment set using time series similarity measurement methods. The driving time-series data segments corresponding to the lateral driving conditions are then matched with the lateral driving condition template to achieve the identification and classification of lateral driving conditions. Three typical lateral driving condition templates are set: lane-changing condition template, turning condition template, and overtaking condition template.

[0068] To improve computational efficiency, this embodiment of the invention employs a GPU-based batch Dynamic Time Warping (DTW) algorithm, parallelizing the traditional serial DTW computation and enabling simultaneous processing of matching calculations for multiple windows and templates. Specifically, the window data and template data to be matched are first transferred to GPU memory, and CUDA tensor operations are used to calculate the point-to-point distance matrix between all windows and templates. Then, the optimal alignment path is solved using dynamic programming to obtain the DTW distance. The batch size is set to 1024, which can be dynamically adjusted according to the GPU memory capacity. Furthermore, for each lateral sub-condition, the dynamic time warping distance between the template and the corresponding scale fragment set is calculated. The dynamic time warping algorithm calculates the optimal alignment path and cumulative distance for each pair of channels. Typically, the distances from multiple channels are fused (e.g., weighted summation, taking the maximum value), resulting in a comprehensive dynamic time warping distance value, i.e., the DTW distance. The formula for calculating the DTW distance is:

[0069] ;

[0070] in, A sequence of one or more channels representing a segment of mesoscale driving time-series data. This represents one or more channel sequences for a specific lateral driving condition template. This indicates that the dynamic time warping algorithm obtains the optimal planned path. Represents distance metric, Indicates the first One or more channels of a mesoscale driving time-series data segment at the sampling time. Indicates the first One or more channels of a lateral driving condition template at a sampling time.

[0071] Furthermore, a composite condition verification based on physical constraints is introduced to improve recognition accuracy. The requirements are set as follows: lane changing requires a steering wheel angle change between 15-90 degrees; turning requires 25-120 degrees; overtaking requires 20-60 degrees; the average vehicle speed for lane changing is required to be no less than 20 m / s; for turning, no less than 15 m / s; and for overtaking, no less than 25 m / s. In addition, duration constraints can be set to ensure that the duration of each condition is within a reasonable range, avoiding misidentification of transient disturbances as complete conditions. Only when the data from the sampling points of the mesoscale driving time-series data segment simultaneously meets the above physical constraints will the corresponding mesoscale driving time-series data segment be labeled as the corresponding type of lateral condition.

[0072] This allows for a two-level classification of lateral and longitudinal driving conditions. The resulting driving time-series data segments are then grouped into longitudinal and lateral driving condition sub-data sets according to their categories. Examples of longitudinal sub-data sets include following and free driving, while lateral sub-data sets include lane changing, turning, and overtaking. These multiple types of longitudinal and lateral driving condition sub-data sets are then combined into a single driving condition dataset.

[0073] Finally, to focus on second-level maneuvers, key driving events need to be labeled for the driving time-series segments in each work condition dataset. Three types of key driving events are defined: instantaneous steering events, rapid acceleration events, and emergency braking events. A rapid acceleration event is determined to exist in the driving time-series data segment when the longitudinal acceleration in the sliding window exceeds a set rapid acceleration threshold; an emergency braking event is determined to exist when the longitudinal acceleration in the sliding window is less than a set emergency braking threshold; and an instantaneous steering event is determined to exist when the speed of change of steering wheel angle in the sliding window exceeds a set speed of change of steering wheel angle threshold.

[0074] Specifically, rapid acceleration events are defined as moments when longitudinal acceleration exceeds 2.5 m / s², and the number, duration, and frequency of rapid acceleration events are recorded for each operating condition segment. Emergency braking events are defined as moments when longitudinal acceleration exceeds -2.5 m / s², and the number, duration, and frequency of emergency braking events are recorded for each operating condition segment. Instantaneous steering events are defined as moments when the absolute value of the steering wheel angular velocity exceeds 200° / s, etc.

[0075] The three key driving events mentioned above can be further flagged as indicators of dangerous driving behavior. The number of events is further calculated based on the dangerous driving behavior event indicator function. :

[0076] ;

[0077] Event frequency per unit time :

[0078] ;

[0079] in, This indicates the total duration of the segment in question.

[0080] Event Time Percentage :

[0081] ;

[0082] in, This indicates that the sampling point at that time was marked as a critical driving event.

[0083] Key driving events reflect a driver's emergency avoidance or aggressive driving behavior. Based on event statistics, the system calculates derived characteristics such as event frequency (events / s) and event duration, comprehensively characterizing the degree of aggressive driving.

[0084] Furthermore, after identifying key driving events in driving time segments belonging to lateral and longitudinal driving conditions, basic statistical features are extracted from driving time segments in specific types of driving conditions within lateral and longitudinal driving conditions. For example, assuming that a lane-changing condition includes N data sampling points, the corresponding sensor sequence is as follows: ,in, For sensor data representing different sampling points, the following basic statistical characteristics are calculated for each sensor channel in each operating condition segment:

[0085] Calculate the mean of the data This is used to characterize the average level of data within a work condition segment:

[0086] ;

[0087] in, This represents sensor data at any sampling time.

[0088] Calculate data variance Used to reflect the degree of data fluctuation:

[0089] ;

[0090] Calculate the minimum value of the data and maximum value :

[0091] ;

[0092] ;

[0093] Calculate the data skewness S to quantify the asymmetry of the data distribution (left-skewed distribution). Right-skewed distribution ):

[0094] ;

[0095] Calculate the kurtosis K of the data to characterize the sharpness (peak morphology) of the data distribution. Low peak state ):

[0096] .

[0097] Basic statistical features can characterize driving behavior from different perspectives, providing a foundation for subsequent behavioral analysis.

[0098] S2: Use driving behavior clustering to cluster driving time segments in each working condition dataset and label each driving time segment with a short-term driving style label;

[0099] A short-term driving style classification model is constructed, and the short-term driving style classification model is trained using driving time-series segments labeled with short-term driving styles; the trained short-term driving style classification model is used to extract the short-term driving style of the input driving time-series segments.

[0100] Specifically, a deep neural network encoder is first constructed to extract low-dimensional deep feature vectors from each driving time-series slice, forming a deep feature space. In this embodiment, the deep neural network encoder is a fusion framework of a convolutional neural network and a long short-term memory network, and the input to the deep neural network encoder is the driving time-series segment obtained by S1. In this embodiment, the feature dimension is 8, which represents the number of sensor channels. The number of selected feature dimensions can be adjusted according to actual needs. The time step is selected as 201, corresponding to a sampling duration of 10.05 seconds. The driving time-series segment with this time step may simultaneously cover specific types of longitudinal driving conditions and lateral driving conditions, and is marked with key driving events. The time step can also be adjusted according to actual needs.

[0101] Deep neural network encoders include:

[0102] The convolutional feature extraction layer is used to expand the feature dimension and compress the temporal dimension of the driving time sequence to obtain local features. The convolutional feature extraction layer consists of at least two sequentially connected convolutional modules; generally, two convolutional modules are sufficient. One convolutional module uses a 3×1 convolutional kernel to expand the 8-channel input to 16 channels with a stride of 1 and padding of 1, expanding the feature dimension of the driving time sequence while maintaining the temporal dimension. After batch normalization, ReLU activation, and random deactivation with a probability of 0.3, the output is fed to the other convolutional module. The second convolutional module uses a 3×1 convolutional kernel to expand the 16 channels to 32 channels with a stride of 2 and padding of 1, expanding the feature dimension of the driving time sequence while reducing the temporal dimension from 201 to 101.

[0103] An LSTM encoding layer connected to the output of the convolutional feature extraction layer is used to capture the temporal features of driving time segments, establish long-term temporal dependencies of local features, and obtain feature vectors. The input dimension of the LSTM encoding layer is 32, the hidden layer dimension is 64, and a single-layer unidirectional structure is adopted.

[0104] The feature fusion layer, connected to the output of the LSTM encoding layer, is used to reduce the dimensionality of the feature vector to obtain a low-dimensional deep feature vector. The feature fusion layer consists of two fully connected layers. The first fully connected layer maps the 64-dimensional LSTM output to a 64-dimensional intermediate representation. After ReLU activation and random deactivation with a probability of 0.3, the second fully connected layer maps it to a final 32-dimensional low-dimensional deep feature vector.

[0105] After obtaining the 32-dimensional low-dimensional deep feature vectors, an improved deep embedding clustering algorithm is used for joint optimization. Clustering methods that can be used in this process include: K-Means clustering, Hierarchical Agglomerative Clustering (HAC), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Gaussian Mixture Model (GMM), and Deep Embedded Clustering (DEC). Taking K-Means clustering as an example, the process of clustering the low-dimensional deep feature vectors using deep embedding K-Means clustering is as follows:

[0106] The K-Means algorithm is used to initialize the low-dimensional deep feature vector multiple times. Based on the silhouette coefficient, the optimal K cluster centers are selected and initialized, where K is the number of predefined driving styles. In this embodiment, K=3, corresponding to three driving styles: aggressive, conservative, and general. A multiple K-means initialization strategy is employed, performing K-means clustering five times, each time setting 10 different initial centers. The clustering quality is evaluated using the silhouette coefficient, and the three cluster centers with the highest scores are selected as the initialization for deep embedding clustering.

[0107] The soft assignment probability between each low-dimensional deep feature vector and each cluster center is calculated based on Student's t-distribution. For low-dimensional deep feature vectors... Calculate the squared Euclidean distance between it and each cluster center, and then calculate the soft assignment probability using the t-distribution kernel function. A temperature parameter of 0.5 is introduced to control the smoothness of the distribution, and the degree of freedom parameter 'a' is set to 1.0. Target distribution. The calculation is obtained by squaring and normalizing the soft assignment probability, while setting the minimum probability of each cluster to 0.01 multiplied by the batch size to prevent cluster degradation.

[0108] A target loss function is constructed, and the location parameters of cluster centers are iteratively optimized by minimizing the target loss function. The target loss function includes at least one of reconstruction loss, KL divergence loss, cluster balance loss, and feature regularization term. In this embodiment of the invention, the reconstruction loss, KL divergence loss, cluster balance loss, and feature regularization term are used simultaneously, and the target loss function is calculated by weighting the four. The reconstruction loss uses mean squared error to measure the difference between the input data and the reconstructed data, with a weight of 1.0. The KL divergence loss measures the difference between the soft assignment and the target distribution, employing an adaptive weighting strategy: in the first 30 training rounds, the weight linearly increases from 0 to 50% of the base weight; from rounds 30 to 60, the weight remains at 50%; after round 60, it is dynamically adjusted based on the average KL loss. The balance loss prevents cluster degradation by calculating the entropy of the cluster assignment, and its weight gradually increases from 0.01 as training progresses. The feature regularization term ensures that the feature standard deviation is not less than 0.1, preventing feature collapse. During training, the target distribution is updated every 20 rounds, and cluster health status analysis is performed every 10 rounds. Health status analysis includes counting the number of samples in each cluster, checking the minimum cluster size, checking for over-clustering, and calculating the silhouette coefficient. If three consecutive clusters are detected as unhealthy, a re-initialization strategy is implemented.

[0109] After iteratively optimizing the deep embedding clustering model using the objective loss function, hard assignment is performed on the driving time sequence segments to obtain their respective cluster categories, i.e., short-term driving style labels.

[0110] A short-term driving style classification model is trained using the aforementioned subset of driving behavior labeled with short-term driving style tags, thereby achieving personalized driving behavior style identification. First, a short-term driving style classification model is constructed. Optional models include: decision trees, random forests, support vector machines (SVM), K-nearest neighbors (KNN), backpropagation neural networks, and ensemble learning models (such as adaptive boosting AdaBoost and gradient boosting). In this embodiment, the short-term driving style classification model uses a backpropagation neural network. The model is trained using driving time-series segments labeled with short-term driving style tags to establish a mapping relationship between driving time-series segments and short-term driving styles. A lightweight two-layer fully connected neural network is used as the classifier. The input layer dimension of the classifier is consistent with the dimension of the low-dimensional deep feature vector, and the number of neurons in the output layer is consistent with the number of short-term driving style types. The classifier's input layer receives a 32-dimensional deep feature vector, the first hidden layer contains 8 neurons using the ReLU activation function, and the output layer contains 3 neurons, corresponding to three driving styles: aggressive, conservative, and general. This simplified network structure significantly reduces computational complexity and the risk of overfitting while maintaining classification accuracy.

[0111] During the classification model training process, the AdamW optimizer was selected, with a learning rate of 0.001 and a weight decay coefficient of 0.001, effectively preventing overfitting. Cross-entropy loss was used as the loss function, suitable for multi-class classification tasks. The training dataset was divided chronologically, with the first 80% used as the training set and the last 20% as the test set, ensuring the continuity of the time-series data. An early stopping mechanism was implemented, setting patience to 20 and a minimum improvement threshold of 0.001. Training was stopped when the validation loss failed to improve beyond the threshold for 20 consecutive epochs to avoid overfitting. The network parameters were initialized using the Kaiming normal distribution method, suitable for the ReLU activation function, with bias initialized to 0. During training, the model parameters with the highest validation accuracy were saved, and a checkpoint was saved every 50 epochs for model recovery and analysis.

[0112] Once the classifier is trained, the driving time sequence segment to be identified can be input into the trained driving style classification model to label the driving time sequence segment with short-term driving style tags.

[0113] Furthermore, to quantify short-term driving style labels and extract or learn more comprehensive and discriminative behavioral features, the following steps are included before inputting the driving time-series segments into the deep neural network encoder:

[0114] The temporal characteristics of driving time segments are calculated, and the temporal statistical characteristics include at least one of the following: mean, standard deviation, zero-crossing rate, peak-to-peak value, and absolute integral. In this invention, the temporal statistical characteristics are calculated as follows:

[0115] ;

[0116] in, The characteristic time-domain mean of a driving time sequence segment. This represents the characteristic time-domain standard deviation of a driving time sequence segment. The zero-crossing rate of a characteristic segment representing a driving time sequence. The characteristic peak-to-peak value of a driving time sequence segment. For symbolic functions, The feature absolute integral represents the driving time sequence segment.

[0117] The frequency domain characteristics of each driving time segment are calculated based on Fourier transform. These frequency domain characteristics include at least one of power spectral density, spectral centroid, spectral bandwidth, and spectral flatness. In this invention, the frequency domain statistical characteristics are calculated as follows:

[0118] ;

[0119] in, This represents the power spectral density at the k-th frequency point. It is the complex result obtained at the k-th frequency point after the characteristic signal undergoes a Fast Fourier Transform (FFT). Indicates the sampling frequency. Indicates the centroid of the spectrum. This represents the frequency value corresponding to the k-th frequency point. Indicates the spectrum bandwidth. Spectral flatness This represents an exponential function.

[0120] Using time-domain and frequency-domain features as quantized data identifiers for driving time-series segments, label quantization is achieved.

[0121] S3: Set a time window in minutes or hours as the medium-to-long-term driving task window. The driving time sequence segments within the medium-to-long-term driving task window constitute the driving task segment. Use the short-term driving style classification model to extract the short-term driving style of each driving time sequence segment in the driving task segment, obtain the short-term driving style distribution vector of each driving task segment, map the short-term driving style distribution vector to the predefined medium-to-long-term style pattern space, and use clustering method to identify the medium-to-long-term style pattern corresponding to the driving task segment.

[0122] In the process of identifying medium- and long-term driving styles, a time window measured in minutes or hours is first set as the driving task window, which generally corresponds to a complete driving trip. In this embodiment of the invention, based on actual driving patterns, a 15-minute time window and a 5-minute sliding step are used for medium- and long-term driving style analysis. The original multi-channel driving time-series data stream is divided using the driving task window to obtain driving task segments, which contain a large number of driving time-series segments. The information of each driving time-series segment in the driving task segment is extracted using a short-term driving style classification model, including the driving condition type, timestamp, and short-term driving style label (conservative, general, and aggressive, with conservative denoted as 0, general as 1, and aggressive as 2). In addition, the driver ID can also be recorded to avoid mixed analysis of different drivers.

[0123] After short-term driving style labels are labeled in the driving task segments, the short-term driving style label sequence corresponding to the driving task segments is obtained by sorting by time, and the driving style distribution vector is obtained based on the short-term driving style label sequence. In this embodiment of the invention, the driving style distribution vector is generated by calculating the weighted frequency of the short-term driving style labels of the driving task segments. Specifically, it is first necessary to set the weights of each driving condition. According to different analysis preferences, the following weighting strategies can be selected:

[0124] Based on a safety risk strategy, where lane change weight is 1.0, turning weight is 0.8, and overtaking weight is 1.5, this strategy is applicable to insurance assessment and fleet management.

[0125] Based on the operational complexity strategy, lane changing has a weight of 1.2, turning has a weight of 0.8, and overtaking has a weight of 1.8, which is suitable for skills assessment;

[0126] Based on the frequency inverse ratio strategy, higher weights are given to low-frequency, high-risk behaviors.

[0127] Based on the accident correlation strategy, the weights are adjusted according to historical accident data;

[0128] The equal weighting strategy assigns the same weight to all operating conditions and is suitable for generalized analysis.

[0129] For each driving task window, calculate the weighted frequency of the corresponding driving task segment. Since driving conditions vary, the appropriate driving style differs depending on the conditions. Therefore, different weights are assigned to different short-term driving style labels based on different conditions. The weighted frequency calculation formula is: the weighted sum of short-term driving style labels divided by the total weight, ensuring that the sum of the three frequencies equals 1.

[0130] To generate the short-term driving style distribution vector, the calculated weighted frequencies of the three short-term driving styles are mapped to a two-dimensional probability simplex space. The mapping employs a transformation from barycentric coordinates to Cartesian coordinates, ensuring the geometric properties of the probability distribution: the three vertices correspond to pure conservative driving (0,0), normal driving (1,0), and aggressive driving, respectively. Any interior point represents a mixture of the three styles, and the distance to each vertex reflects the weight of the corresponding style. Through sliding window analysis, each driver generates a three-dimensional probability vector for each time window, representing the distribution of driving style within that time period. For a complete trip, multiple such driving style distribution vectors are generated, forming a temporal feature sequence. These feature vectors not only contain static distribution information of driving styles but also implicitly contain dynamic change patterns of styles.

[0131] Furthermore, based on the coordinate representation in the probabilistic simplex space, the driving style distribution vector is mapped to a predefined medium-to-long-term style pattern space, and cluster analysis is used to identify the medium-to-long-term style pattern corresponding to the driving task segment. The cluster analysis process based on the driving style distribution vector is as follows:

[0132] Seven initial cluster centers are pre-defined on the probabilistic simplex to fully utilize its geometric structure: Conservative: located at the conservative vertex (0,0), representing a purely conservative driving style; Moderate: located at the intermediate vertex (1,0), representing a purely moderate driving style; Aggressive: located at the aggressive vertex... Represents a purely aggressive driving style; Cons-Mod: Located at the midpoint between the Cons-Mod and Cons-Mod sides. General-Radical Hybrid (Mod-Agg): Located at the midpoint between the general and radical edges. Aggressive-Cons hybrid: located at the midpoint between the aggressive and conservative sides. Comprehensive Balanced Type (Center): Located at the centroid of the triangle This represents a balanced distribution of the three styles.

[0133] Then, the K-means clustering algorithm is executed using the preset initial cluster centers. Parameters include: number of clusters k=7, maximum number of iterations 300, and the use of fixed initial cluster centers to ensure reproducibility. During clustering, the algorithm iteratively updates the cluster center positions and sample assignments: it calculates the Euclidean distance from each sample point to each cluster center and assigns the sample to the nearest cluster center; based on the assignment results, the cluster center position is updated to the mean of all samples within that cluster. This iteration continues until convergence or the maximum number of iterations is reached.

[0134] After clustering, clustering quality was assessed and the results were interpreted. Two indicators were used to evaluate clustering quality: the silhouette score, which measures the compactness and separation of clusters, with a value ranging from -1 to 1, where a value closer to 1 indicates better clustering; and the Calinski-Harabasz index, which assesses the ratio of inter-cluster dispersion to intra-cluster dispersion, where a higher value indicates better clustering. Experimental results show that the clustering method based on pre-defined cluster centers can effectively identify different driving style combinations, with silhouette scores typically above 0.4, indicating a clear cluster structure.

[0135] After clustering, each cluster represents a typical medium- to long-term driving style pattern. For example, the "Conservative-dominant" type employs conservative driving for over 60% of the time, suitable for safety-critical scenarios such as long-haul freight. The "Moderate-dominant" type primarily uses moderate-intensity driving, representing the standard driving mode for daily commuting. The "Aggressive-dominant" type is dominated by aggressive driving and may require safety intervention. The "Conservative-Moderate Hybrid," "Moderate-Aggressive Hybrid," and "Aggressive-Conservative Hybrid" types represent combinations of different styles, reflecting the driver's ability to adjust strategies according to road conditions. The "Balanced" type shows a balanced distribution of the three styles, indicating good driver adaptability.

[0136] To further describe the medium- to long-term style patterns, more descriptive medium- to long-term features can be obtained through statistical analysis. The specific implementation method is as follows:

[0137] First, the basic style score features are extracted.

[0138] The basic style score directly reflects the proportion of each driving style in the total driving time. The aggressiveness score is calculated by dividing the total aggressive driving time by the total driving time, with a value ranging from [0,1]. Similarly, the conservativeness score and the generality score are calculated. The sum of these three scores is 1, constituting a complete description of the style distribution. In implementation, the overall style score is obtained by accumulating the weighted frequencies within each time window and considering the window duration weight. For example, if a driver's average weighted frequency of aggressive driving is 0.3 across multiple 15-minute windows, then their aggressiveness score is approximately 0.3.

[0139] Secondly, dynamic features of the style are extracted.

[0140] The sensitivity index is calculated by counting the number of style category changes between adjacent time windows. In practice, it compares the cluster labels of two consecutive time windows; if a change occurs, the count is incremented by 1. The sensitivity index equals the number of changes divided by the total number of windows, reflecting the driver's responsiveness to environmental changes and further categorizing drivers as sensitive.

[0141] The style stability score assesses the consistency of driving style. It is calculated by subtracting the coefficient of variation of the style vector from 1, where the style vector is composed of the time proportions of the three styles. The closer the stability score is to 1, the more stable the driving style.

[0142] Style fluctuation frequency is defined as the number of style changes per unit of time, usually measured in times per hour. This indicator intuitively reflects the frequency of changes in driving behavior.

[0143] Finally, the composite features are extracted.

[0144] Style Mix The degree of style mixing is quantified by Shannon entropy:

[0145] ;

[0146] in, The time percentages for the three driving styles are normalized to the [0,1] interval using a base-3 logarithm. A mixing degree of 0 indicates a purely single style, while 1 indicates a completely uniform distribution of the three styles.

[0147] Aggressive action density measures the intensity of aggressive driving behavior and can be calculated by distance or time. When calculated by distance, it equals the number of aggressive behavior events divided by the total mileage (events / km); when calculated by time, it equals the number of aggressive behavior events divided by the total driving time (events / hour). This indicator provides an important reference for risk assessment and can further classify drivers as truly aggressive.

[0148] S4: Set a time window with weekly, monthly, quarterly, or yearly units as a long-term style window. The driving task segments within the long-term style window constitute long-term driving style evolution segments. Statistically analyze the medium-to-long-term style patterns corresponding to the driving task segments within the long-term driving style evolution segments to obtain the medium-to-long-term style pattern sequence of the long-term driving style evolution segments. Map the medium-to-long-term style pattern sequence to a predefined long-term style evolution space and use clustering methods to perform long-term style evolution analysis on the driver.

[0149] Specifically, to conduct long-term style evolution analysis of drivers and study their driving habits and changes in those habits, a long-term style window needs to be established, measured in weeks, months, quarters, or years. Segments are obtained from the multi-channel driving time-series data stream using this long-term style window; these are called long-term driving style evolution segments. The S3 method is used to obtain the medium-to-long-term style patterns of the driving task segments contained within these long-term driving style evolution segments. These medium-to-long-term style patterns are quantified using the weighted frequencies of three driving styles.

[0150] The mid-to-long-term style patterns corresponding to the driving task segments within the long-term driving style evolution segment are subjected to temporal sequencing to obtain the mid-to-long-term style pattern sequence of the long-term driving style evolution segment, including:

[0151] The system loads the original dataset containing driver IDs, timestamps, two-dimensional coordinate features, and weighted frequencies of three types of driving behaviors: file-based, general, and aggressive. During data preprocessing, the system employs a time parsing strategy, supporting automatic recognition of various date and time formats. It uniformly converts time fields to a standard format and extracts month and year information to construct monthly time series.

[0152] To address the uneven temporal distribution of real-world driving data, this invention innovatively proposes an adaptive temporal sampling technique. This technique first assesses the temporal density of data for each month, defined as the number of data points divided by the total number of hours in that month. When the data density is greater than 0.05, the system uses a 4-hour sampling frequency to retain more detailed information; when the density is between 0.01 and 0.05, a 12-hour sampling frequency is used to balance information retention and computational efficiency; when the density is less than 0.01, daily sampling is used to avoid over-interpolation. In the interpolation process, the system first performs temporal alignment of the original data using a resampling method. For missing values, the system intelligently selects an interpolation strategy based on the number of available data points: when there are sufficient data points, polynomial or spline interpolation is prioritized to better preserve data trends; otherwise, linear interpolation is used to ensure stability. For boundary missing values ​​that still exist after interpolation, a combination of forward and backward padding is used. To ensure the consistency of subsequent deep learning model inputs, this invention designs an intelligent sequence length normalization mechanism. The system first calculates the statistical characteristics of all sequence lengths, including the minimum, maximum, and median. Then, it dynamically determines the target length based on the median: for short sequences with a median of 30 or less, the original length is maintained to avoid information loss; for medium-length sequences with a median between 30 and 100, the target length is set to no more than 80; for long sequences with a median exceeding 100, the target length is limited to 120. In practice, for sequences requiring downsampling, a uniform interval sampling strategy is used to ensure the preservation of the overall trend of the sequence; for sequences requiring upsampling, linear interpolation is used to generate a smooth transition between adjacent points.

[0153] A variational autoencoder (VAE) model is constructed to extract features from medium- to long-term style pattern sequences, yielding low-dimensional long-term style evolution feature vectors. Specifically, the VAE model employs a multi-layer fully connected network structure, dynamically adjusting the network capacity based on the length of the input medium- to long-term style pattern sequence: short sequences use a hidden layer configuration of [128, 64, 32] mapped to a 16-dimensional latent space; medium-length sequences use a configuration of [256, 128, 64] mapped to a 20-dimensional space; and long sequences use a configuration of [512, 256, 128] mapped to a 24-dimensional space. Each hidden layer is followed by a ReLU activation function, layer normalization, and Dropout regularization with a probability of 0.2. The model is trained by minimizing a weighted combination of reconstruction loss and KL divergence, where the weight coefficient of KL divergence linearly increases from 0 to 0.5 in the first 100 training epochs, achieving a smooth transition from reconstruction to regularization. The training process employs the Adam optimizer, setting an initial learning rate and utilizing the ReduceLROnPlateau scheduler for adaptive learning rate decay. To determine the optimal number of clusters, the system performs a grid search within the range of 3 to 8, calculating multiple evaluation metrics for each candidate cluster number, including silhouette coefficient, Calinski-Harabasz index, Davies-Bouldin index, and inertia. Considering practical application requirements, when the silhouette coefficient of the target cluster number (5) differs from the optimal value by less than 0.05, 5 is preferentially selected as the final cluster number to balance model complexity and interpretability. Subsequently, K-means clustering is performed in the extracted latent feature space to analyze the distribution of different medium- to long-term styles within each cluster, calculating the information entropy of style diversity and identifying the dominant style and its temporal stability. Furthermore, seasonal patterns are identified through statistical monthly distribution, and the temporal evolution trend of trajectories is determined through linear fitting. Finally, the above analysis results are used to generate descriptive cluster names, giving the abstract mathematical clusters intuitive meanings, such as long-term stable clusters and split clusters. Long-term stable clusters reflect stable long-term driving styles, while split clusters indicate significant fluctuations in driving styles.

[0154] Furthermore, long-term style classification is performed based on transfer learning. Specifically, the variational autoencoder (VAE) is used as a feature extractor. Some parameters of the pre-trained VAE are frozen, serving as a fixed feature extractor to avoid disrupting the learned unsupervised feature representation on limited labeled data. For the input sequence, a low-dimensional latent feature vector, i.e., a low-dimensional long-term style evolution feature vector, is extracted using the frozen VAE model and then input into a specially designed lightweight classification network. The long-term classification network employs a three-layer fully connected structure: the input layer receives the latent features output by the VAE, with dimensions ranging from 16 to 24 depending on the sequence length; the first hidden layer contains 64 neurons, using the ReLU activation function and Dropout with a 0.3 dimensionality; the second hidden layer contains 32 neurons, configured with the same activation and regularization; the output layer has the same number of neurons as the number of clusters, using the Softmax activation function to output the probability distribution of each category. Network training uses the cross-entropy loss function, updates parameters using the Adam optimizer, sets the initial learning rate to 32, the batch size to 32, the maximum number of training epochs to 100, and configures an early stopping mechanism to prevent overfitting. This transfer learning-based approach fully leverages the feature representation capabilities of unsupervised pre-training, achieving good classification results even with limited labeled samples.

[0155] By treating medium- to long-term style pattern sequences as univariate time series analysis, the long-term evolution characteristics of driver style are obtained, describing the long-term features of style on scales ranging from weeks to years. For medium- to long-term style pattern sequences... Multi-dimensional long-term feature quantization extraction is performed, specifically including obtaining the following quantized features:

[0156] Style Separation It quantifies the distinguishability between different driving style categories, and its calculation formula is as follows:

[0157] ;

[0158] in, This represents the difference between the mean scores of two different driving style categories. The mean score refers to a quantitative indicator of driving style, such as the aggressiveness score or the conservativeness score. and These represent the probabilities of these two driving styles. A higher value indicates a more significant difference between the two styles, which is beneficial for subsequent style identification and early warning.

[0159] Dynamic variability characterization The stability of driving style is characterized by calculating the coefficient of variation of the medium- to long-term style pattern sequence. The calculation formula is as follows:

[0160] ;

[0161] in, This represents the standard deviation of a medium- to long-term style pattern sequence. The mean of the medium- to long-term style pattern sequence is used. In actual calculations, the standard deviation and mean of the features are calculated separately for each long-term wind window. A value exceeding 0.5 indicates significant fluctuations in driving behavior during that period, which may require close monitoring.

[0162] Style Momentum Indicator Borrowing the concept of the Relative Strength Index (RSI) from financial technical analysis, it is used to identify trends in driving style. Its calculation formula is as follows:

[0163] ;

[0164] in, This represents the average increase in style intensity over the past six months. This represents the average decrease in style intensity over the past 6 months. When RSI is greater than 70, it indicates that the trend of aggressive driving is strengthening; when RSI is less than 30, it indicates that the trend of conservative driving is dominant.

[0165] Long-term trend capture A 12-period weighted moving average is used to identify the annual evolution trend of driving style. The calculation formula is as follows:

[0166] ;

[0167] in, This represents the weight value; the more recent the date of the driving data, the greater the weight. This represents driving data such as speed, acceleration, frequency of emergency braking events, and frequency of rapid acceleration events.

[0168] Pattern complexity entropy value High entropy values ​​indicate random driving behavior, while low entropy values ​​reflect regular operations.

[0169] ;

[0170] in, Permutation entropy is used to quantify temporal regularity by calculating the diversity of local permutation patterns in a sequence. The entropy of the samples is used to characterize complexity by evaluating the self-similarity of the sequences. High entropy values ​​indicate strong randomness in driving behavior, while low entropy values ​​reflect regular operating patterns.

[0171] Periodicity intensity The energy of the one-year periodic component is extracted by continuous wavelet transform to quantify the significance of seasonal driving patterns. The calculation formula is as follows:

[0172] ;

[0173] in, The wavelet coefficients represent the local energy at time and frequency f. This index can quantify the intensity of seasonal variations in driving behavior, providing important information for understanding long-term driving patterns.

[0174] Through the aforementioned quantitative features, this invention can comprehensively depict the long-term evolution of driving styles, accurately identify current driving style categories, and predict future trends, providing a solid technical foundation for applications such as personalized driving assistance, risk assessment, and safety warnings.

[0175] In summary, the above description is merely a preferred embodiment of this specification and is not intended to limit the scope of protection of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of protection of this specification.

[0176] The systems, apparatuses, modules, or units described in one or more of the above embodiments may be implemented by a computer chip or entity, or by a product having a certain function. A typical implementation device is a computer. Specifically, a computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0177] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0178] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

Claims

1. A multi-timescale driving style analysis method, characterized in that, include: S1: The multi-channel driving time-series data stream is split into driving time-series segments of various lengths in seconds using a multi-resolution analysis method, and the driving time-series segments are classified according to working conditions to obtain multiple working condition datasets; S2: Use driving behavior clustering to cluster driving time segments in each working condition dataset and label each driving time segment with a short-term driving style label; A short-term driving style classification model is constructed, and the short-term driving style classification model is trained using driving time-series segments labeled with short-term driving styles; the trained short-term driving style classification model is used to extract the short-term driving style of the input driving time-series segments; S3: Set a time window in minutes or hours as the medium-to-long-term driving task window. The driving time sequence segments within the medium-to-long-term driving task window constitute the driving task segment. Use the short-term driving style classification model to extract the short-term driving style of each driving time sequence segment in the driving task segment, obtain the short-term driving style distribution vector of each driving task segment, map the short-term driving style distribution vector to the predefined medium-to-long-term style pattern space, and use clustering method to identify the medium-to-long-term style pattern corresponding to the driving task segment. S4: Set a time window with a week, month, quarter, or year as the long-term style window. The driving task segments within the long-term style window constitute the long-term driving style evolution segment. Statistically analyze the medium-to-long-term style patterns corresponding to the driving task segments within the long-term driving style evolution segment to obtain the medium-to-long-term style pattern sequence of the long-term driving style evolution segment. Map the medium-to-long-term style pattern sequence to a predefined long-term style evolution space and use clustering methods to perform long-term style evolution analysis on the driver.

2. The multi-timescale driving style analysis method according to claim 1, characterized in that, The process of acquiring the working condition dataset includes: Multi-channel driving time-series data streams are collected by onboard sensors; The multi-channel driving time-series data stream is split and classified using a multi-resolution analysis method to obtain a set of driving time-series data segments with different time lengths. Select driving time series data segments corresponding to longitudinal driving conditions and driving time series data segments corresponding to lateral driving conditions from the set of driving time series data segments; Based on rule-based methods, secondary classification is performed on driving time-series data segments corresponding to longitudinal driving conditions to obtain longitudinal driving condition subset datasets. Based on the time series similarity measurement method, the driving time series data segments corresponding to the horizontal driving conditions are classified into two levels to obtain the horizontal driving condition subset dataset; The longitudinal driving condition subset and the lateral driving condition subset are used as the driving condition dataset.

3. The multi-timescale driving style analysis method according to claim 2, characterized in that, The multi-resolution analysis method is a multi-scale sliding window, dynamic range adjustment, or recursive segmentation and matching method.

4. The multi-timescale driving style analysis method according to claim 2, characterized in that, The rule-based method performs secondary classification of driving time-series data segments corresponding to longitudinal driving conditions, including: Based on threshold determination or state machine methods, secondary classification rules for longitudinal driving conditions are set, and the driving time series data segments corresponding to the longitudinal driving conditions are classified in a secondary manner according to the secondary classification rules.

5. The multi-timescale driving style analysis method according to claim 2, characterized in that, The time-series similarity measurement method is used to perform secondary classification of driving time-series data segments corresponding to lateral driving conditions, including: A lateral driving condition template is set, and a time series similarity measurement method is used to match the driving time series data segments corresponding to the lateral driving condition with the lateral driving condition template to realize the identification and classification of lateral driving conditions. The time series similarity measurement method is Euclidean distance, dynamic time warping, time envelope edit distance, cosine similarity, Pearson correlation coefficient or symbolic aggregation approximation method.

6. The multi-timescale driving style analysis method according to claim 1, characterized in that, The method involves clustering driving behavior segments in each work condition dataset using a driving behavior clustering approach, and labeling each driving time segment with a short-term driving style label, including: A deep neural network encoder is constructed, and a driving time-series segment is input into the deep neural network encoder to extract a low-dimensional deep feature vector. The deep neural network encoder includes: a convolutional feature extraction layer, an LSTM encoding layer connected to the output of the convolutional feature extraction layer, and a feature fusion layer connected to the output of the LSTM encoding layer. The convolutional feature extraction layer is used to expand the feature dimension and compress the temporal dimension of the driving time-series segment to obtain local features; the LSTM encoding layer is used to capture the temporal features of the driving time-series segment, establish long-term temporal dependencies of local features, and obtain feature vectors; the feature fusion layer is used to reduce the dimensionality of the feature vectors to obtain the low-dimensional deep feature vector. The low-dimensional deep feature vectors are clustered using a deep embedding clustering method. Short-term driving styles are generated based on the clustering results, and each driving time-series segment is labeled with a short-term driving style tag.

7. The multi-timescale driving style analysis method according to claim 6, characterized in that, The clustering of the low-dimensional deep feature vectors using a deep embedding clustering method includes: The K-Means algorithm is used to initialize the low-dimensional deep feature vector multiple times, and the optimal K cluster centers are selected based on the contour coefficient, where K is the number of predefined driving styles; The soft assignment probability between each low-dimensional deep feature vector and each cluster center is calculated based on Student's t-distribution; Construct a target loss function, and iteratively optimize the location parameters of the cluster centers by minimizing the target loss function.

8. The multi-timescale driving style analysis method according to claim 1, characterized in that, Obtain the short-term driving style distribution vector for each driving task segment, map the short-term driving style distribution vector to a predefined medium-to-long-term style pattern space, and use clustering methods to identify the medium-to-long-term style pattern corresponding to the driving task segment, including: Calculate the weighted frequency of short-time driving style for each driving time segment in the driving task segment, and generate a short-time driving style distribution vector; The short-term driving style distribution vector is mapped to the probability simplex space to obtain the coordinate representation of the short-term driving style distribution vector in the probability simplex space; Based on the coordinate representation in the probabilistic simplex space, the mid-to-long-term style patterns corresponding to the driving task segments are analyzed and identified using clustering methods.

9. The multi-timescale driving style analysis method according to claim 1, characterized in that, The process of mapping medium- to long-term style pattern sequences to a predefined long-term style evolution space and using clustering methods to perform long-term style evolution analysis on drivers includes: A variational autoencoder model is constructed, and the variational autoencoder model is used to extract features from medium- and long-term style pattern sequences to obtain low-dimensional long-term style evolution feature vectors. Clustering methods are performed in the long-term style evolution space to conduct long-term style evolution analysis of drivers.

10. The multi-timescale driving style analysis method according to claim 1, characterized in that, The short-term driving style labels include: aggressive, conservative, and normal.

Images