A Personalized Driving Behavior Identification Method Based on Deep Embedding Clustering

Abstract

This invention relates to the field of intelligent driving technology, specifically providing a personalized driving behavior identification method based on deep embedding clustering. It extracts low-dimensional deep feature vectors from driving time-series segments using a deep neural network autoencoder, and performs end-to-end joint optimization of the autoencoder and unsupervised clustering using deep embedding clustering to generate driving style clustering results and label them. The trained autoencoder parameters are frozen, a driving style classification model is constructed, and a mapping relationship between driving time-series segments and style labels is established. Semantic labels are assigned to the clustering results by extracting interpretable features. This invention, based on end-to-end joint optimization technology using deep embedding clustering, significantly improves the model's convergence speed and generalization ability, achieving driving style identification accuracy far exceeding traditional machine learning methods.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving technology, and in particular relates to a driving behavior recognition method. Background Technology

[0002] Driving behavior analysis is one of the core technologies of Intelligent Transportation Systems (ITS) and Advanced Driver Assistance Systems (ADAS). Its key lies in the accurate identification and modeling of drivers' personalized operating habits and decision-making patterns. Traditional driving style analysis methods mainly rely on the following two types of technologies:

[0003] The first type is rule-based threshold methods, which determine instantaneous driving events such as rapid acceleration and sudden braking by setting fixed thresholds for sensor readings, thereby classifying driving styles in a coarse-grained manner. However, this method relies heavily on expert experience, the set thresholds have poor universality, it cannot capture the complex and non-linear temporal patterns in driving behavior, and it is difficult to distinguish behavioral differences under different driving situations, resulting in low classification accuracy and weak generalization ability.

[0004] The second category is based on traditional machine learning methods, which extract time-domain or frequency-domain features from driving data and use shallow machine learning models for classification. These methods have significant limitations: First, feature extraction and model training are independent, and manually designed features are insufficient to fully represent the essence of driving behavior; second, shallow models have limited ability to process high-dimensional, time-series data and cannot effectively mine deep, abstract feature representations within the data; third, clustering methods divide the original high-dimensional data space, are sensitive to noise, and struggle to discover non-spherical cluster structures.

[0005] The aforementioned existing technologies generally suffer from problems such as insufficient feature representation, weak model generalization ability, and imprecise characterization of individual differences, making it difficult to achieve end-to-end, high-precision automatic identification of personalized driving behavior and failing to meet the needs of advanced applications such as personalized control of autonomous driving and driver risk profiling. Summary of the Invention

[0006] In view of this, the present invention aims to provide a personalized driving behavior identification method based on deep embedding clustering. Through personalized driving behavior clustering, unsupervised learning is performed in a suitable feature space to automatically discover different driving behavior patterns and generate corresponding personalized style labels. The classifier is trained using driving time-series segments with style labels to establish a mapping relationship between driving time-series segments and driving styles, which can accurately identify driving style categories and mine deep feature representations.

[0007] To achieve the above objectives, the technical solution created by this invention is implemented as follows:

[0008] This invention provides a personalized driving behavior identification method based on deep embedding clustering, comprising:

[0009] S1: Obtain driving time sequence segments under various operating conditions;

[0010] S2: Construct a deep neural network autoencoder, input the driving time sequence segment into the deep neural network autoencoder, and extract low-dimensional deep feature vectors;

[0011] S3: Unsupervised clustering is performed on the low-dimensional deep feature vector using a deep embedding clustering method to generate multiple driving styles, and each driving time segment is labeled with a corresponding driving style label;

[0012] S4: Freeze the intrinsic parameters of the deep neural network autoencoder, and construct a driving style classification model based on the deep neural network autoencoder with frozen intrinsic parameters. Train the driving style classification model using driving time segments labeled with driving style tags to establish a mapping relationship between driving time segments and driving styles. Input the driving time segment to be identified into the trained driving style classification model to achieve driving style classification.

[0013] Preferably, the acquisition of driving time segments under multiple operating conditions includes:

[0014] Multi-channel driving time-series data streams are collected by onboard sensors;

[0015] Multiple sliding windows of different time lengths and the working condition category to be identified for each sliding window are set. The multi-channel driving time-series data stream is divided using the sliding windows to obtain driving time-series data segments of different time lengths.

[0016] By setting driving characteristic indicators, driving time-series data segments corresponding to longitudinal driving conditions are selected from the data segment set based on the driving characteristic indicators, thereby realizing the identification and classification of longitudinal driving conditions.

[0017] A lateral driving condition template is set, and the driving time series data segments in the data segment set are matched with the lateral driving condition template using a time series similarity measurement method. The driving time series data segments corresponding to the lateral driving conditions are then selected to achieve the identification and classification of lateral driving conditions.

[0018] Preferably, the deep neural network autoencoder is a fusion framework of convolutional neural network and long short-term memory network.

[0019] Preferably, the deep neural network autoencoder includes:

[0020] Convolutional feature extraction layer is used to expand the feature dimension and compress the time dimension of driving time segments to obtain local features;

[0021] The 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.

[0022] A feature fusion layer connected to the output of the LSTM encoding layer is used to reduce the dimensionality of the feature vector to obtain the low-dimensional deep feature vector.

[0023] Preferably, the convolutional feature extraction layer includes at least two sequentially connected convolutional modules, wherein at least one convolutional module expands the feature dimension while maintaining the time dimension, and at least another convolutional module expands the feature dimension while compressing the time dimension.

[0024] Preferably, the unsupervised clustering of the low-dimensional deep feature vector using a deep embedding clustering method includes:

[0025] 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;

[0026] The soft assignment probability between each low-dimensional deep feature vector and each cluster center is calculated based on Student's t-distribution;

[0027] Construct a target loss function, and iteratively optimize the location parameters of the cluster centers by minimizing the target loss function.

[0028] Preferably, the target loss function includes at least one of reconstruction loss, KL divergence loss, clustering balance loss, and feature regularization term.

[0029] Preferably, the driving styles include: aggressive, conservative, and normal.

[0030] Preferably, the driving style classification model includes a deep neural network autoencoder with frozen network intrinsics at the front end and a fully connected neural network at the back end. The input layer dimension of the fully connected neural network 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 driving style types.

[0031] Preferred options also include:

[0032] Calculate the time-domain characteristics of driving time segments, wherein the time-domain characteristics include at least one of mean, standard deviation, zero-crossing rate, peak-to-peak value, and absolute integral;

[0033] The frequency domain characteristics of each driving time segment are calculated based on Fourier transform, and the frequency domain characteristics include at least one of power spectral density, spectral centroid, spectral bandwidth and spectral flatness.

[0034] Temporal and frequency domain features are used as quantitative data identifiers for driving time-series segments.

[0035] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0036] This invention addresses the weakness of existing driving behavior recognition technologies in dynamic modeling by proposing a personalized driving behavior recognition technology based on deep embedding clustering. The 32-dimensional feature vectors extracted by the deep embedding clustering model using a CNN-LSTM encoder exhibit correlation coefficients generally below 0.1, with the vast majority being in a state of extremely weak or no correlation. This achieves low-redundancy feature representation, which helps subsequent clustering models learn the relationships between features more accurately, improving model generalization ability. It significantly enhances the convergence speed and generalization ability of both clustering and classification models, resulting in clearer delineation of driving style boundaries and recognition accuracy far exceeding that of traditional machine learning methods relying on handcrafted features.

[0037] Furthermore, this invention utilizes deep embedding clustering, resulting in clusters with high intra-cluster density and large inter-cluster distance. This ensures clear boundaries for each driving style category and significantly improves cluster purity compared to existing technologies, laying the foundation for accurate classification of personalized driving behaviors.

[0038] This invention also features a unique multi-dimensional feature fusion interpretation system. In addition to deep features, it simultaneously extracts time-domain and frequency-domain features and associates abstract clustering results with specific driving behavior indicators through feature importance analysis. This gives style labels such as "aggressive" and "robust" clear physical meaning and indicators, greatly improving the credibility and usability of the model results and providing a convincing decision-making basis for applications such as safety warnings and driving scores. Attached Figure Description

[0039] 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:

[0040] Figure 1 This is a schematic diagram of a driving behavior recognition method provided according to an embodiment of the present invention. Detailed Implementation

[0041] 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.

[0042] 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.

[0043] 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.

[0044] 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.

[0045] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0046] Please see Figure 1 In one embodiment of the present invention, a driving behavior recognition method is provided, which aims to automatically discover and define typical driving styles from driving data without human intervention. This embodiment of the present invention uses natural driving data of vehicles on urban roads as an example for study. The driving behavior recognition method includes:

[0047] S1: Obtain driving time sequence segments under various operating conditions.

[0048] Specifically, S101: First, raw driving data needs to be acquired. This raw driving data is collected from onboard sensors using a multi-channel driving time-series data stream. This data stream must include the following channels: steering wheel angle, steering wheel angle direction indicator, vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate, and timestamp. That is, the locomotive system collects these types of driving data through corresponding sensors, and these channels correspond to the feature dimensions of the driving time-series segment. In addition, the multi-channel driving time-series data stream will also include channels for longitudinal vehicle speed, steering angle, steering angle change rate, accelerator pedal opening, brake pedal pressure, and gear information.

[0049] Since the data is acquired from different sensors, the acquisition frequency and time delay may be inconsistent, necessitating preprocessing of the raw data. First, the driving time-series data for each channel is aligned using timestamps, and the sampling frequency of each channel's driving time-series data is unified to a set reference frequency. The sampling frequency unification method includes: ensuring the time consistency of driving time-series data for each channel through timestamp alignment, and setting a reference frequency. In this embodiment, the reference frequency is set to 100Hz. By resampling channel data with different sampling rates and interpolating missing data, data continuity is ensured, unifying the driving time-series data for each channel to 100Hz.

[0050] S102: Set multiple sliding windows of different time lengths and corresponding driving behavior categories for each sliding window. Use the sliding windows to divide the multi-channel driving time-series data stream and obtain multiple sets of driving time-series data segments of different time lengths. Multiple time scales and sliding windows of different time lengths are set according to different analysis needs, and each sliding window of different time length has a corresponding driving behavior category.

[0051] In this embodiment of the invention, three time scales are defined, corresponding to three sliding windows of different time lengths, including:

[0052] Short-scale sliding window, with a window length of 50 to 200 sampling points (window length corresponds to 0.5 to 2 seconds, sampling frequency 100Hz);

[0053] Mesoscale sliding window, with a window length of 200 to 500 sampling points (window length corresponds to 2 to 5 seconds);

[0054] A long-scale sliding window, with a window length of 500 to 2000 sampling points, corresponding to 5 to 20 seconds.

[0055] The driving behavior categories corresponding to short-scale sliding windows include: momentary driving behaviors such as instantaneous steering, rapid acceleration, and sudden braking.

[0056] The driving behavior categories corresponding to the medium-scale sliding window include typical driving actions such as lane changing, turning, and overtaking.

[0057] The driving behavior categories corresponding to the long-scale sliding window include steady-state conditions such as following other vehicles and free driving.

[0058] For each time scale, during the data stream partitioning process, the window movement step size for all sliding windows is set to 20% of the window length, achieving an 80% window overlap between adjacent sliding windows. This high overlap design ensures sufficient sampling of driving behavior and avoids missing key driving events. During window generation, the system simultaneously records metadata information for each window, including: window start index, end index, data segment identifier, and average vehicle speed within the window. This metadata provides crucial contextual information for subsequent condition identification and feature extraction.

[0059] The multi-channel driving time-series data stream is divided and classified using sliding windows of the three time lengths mentioned above. Each window division yields a driving time-series data segment of a corresponding time length, thus obtaining a set of driving time-series data segments of three different time lengths.

[0060] S103: Set driving characteristic indicators, and select driving time-series data segments corresponding to longitudinal driving conditions from the data segment set based on the driving characteristic indicators, and classify the longitudinal driving conditions. The yaw rate is selected as the driving characteristic indicator. The longitudinal driving conditions are identified by calculating the average energy of the yaw rate of the long-scale driving time-series data segments obtained by dividing with a long-scale sliding window, which serves as the basis for dividing the longitudinal and lateral driving conditions.

[0061] Specifically, the average energy of the yaw rate for:

[0062] ;

[0063] 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.

[0064] Furthermore, an energy setting threshold for judging longitudinal driving conditions is determined. Based on statistical analysis of a large amount of actual driving data, it was found that an energy setting threshold of 0.5 rad² / s for yaw rate can accurately distinguish between longitudinal and lateral driving conditions. In this embodiment of the invention, when the calculated average energy of the 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.

[0065] Furthermore, longitudinal driving conditions are categorized based on following distance, including:

[0066] The following distance threshold is set at 50 meters, and the average energy of the yaw rate is... If the energy level is below the set threshold of 0.5 rad² / s and the actual following distance is less than 50 meters, it is classified as a following-vehicle driving condition. Under this condition, the driver mainly controls the longitudinal speed to maintain a safe distance from the vehicle in front.

[0067] When the average energy of the yaw rate If the energy level is below the set threshold and the actual following distance is greater than 50 meters or there is no vehicle in front, it is considered a free driving condition. Under this condition, the driver can freely choose the driving speed, which is mainly characterized by stable longitudinal driving.

[0068] Furthermore, additional constraints need to be set to ensure the continuity of the operating conditions and prevent noise from causing frequent switching between operating conditions. Specifically, a data segment is only marked as a longitudinal operating condition if 100 consecutive sampling points (1 second) meet the longitudinal operating condition conditions. For brief fluctuations in yaw rate energy, the system will perform smoothing processing to maintain the stability of the operating condition identification.

[0069] S104: Set a lateral driving condition template, and use the time series similarity measurement method to match the driving time series data segments with the lateral driving condition template. Select the driving time series data segments corresponding to the lateral driving condition from the data segment set, thereby realizing the identification and classification of lateral driving conditions.

[0070] This invention employs time series similarity metrics to filter driving time series data segments corresponding to lateral driving conditions from a set of data segments. Optional time series similarity metrics include Euclidean distance, Dynamic Time Warping (DTW), Time Warp Edit Distance (TWED), cosine similarity, Pearson correlation coefficient, and Symbolic Aggregate Approximation (SAX). In this embodiment, Dynamic Time Warping is specifically chosen. To improve computational efficiency, a GPU-based batch parallel DTW algorithm is used, parallelizing the traditional serial DTW computation and simultaneously processing the matching calculations of multiple windows and templates. Specifically, the window data to be matched (i.e., the mesoscale driving time series data segments obtained by mesoscale sliding window partitioning) and the lateral driving condition template data are first transferred to GPU memory. CUDA tensor operations are used to calculate the point-to-point distance matrix between all windows and templates. Then, a dynamic programming method is used to solve for the optimal alignment path, calculating the DTW distance between the mesoscale driving time series data segments obtained by mesoscale sliding window partitioning and each lateral driving condition template. The batch size is set to 1024, which can be dynamically adjusted according to the GPU memory capacity. Then, for each horizontal sub-case, the dynamic time warping distance (DTW) between the template and the corresponding scale fragment set is calculated. The DTW algorithm calculates the optimal alignment path and cumulative distance for each pair of channels. Typically, the distances of multiple channels are fused (e.g., weighted summation, taking the maximum value) to obtain a comprehensive DTW distance value, which is calculated as follows:

[0071] ;

[0072] 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.

[0073] Furthermore, a dynamic threshold is set for each lateral driving condition template. The setting of the dynamic threshold needs to take into account the noise level of each template. The specific dynamic threshold calculation process is as follows:

[0074] First, the templates are Z-score standardized, and the median of the absolute value of the first-order difference of the standardized templates is used as an estimate of the noise level. Then, the noise level, template duration, and configuration sensitivity coefficient are multiplied to obtain the dynamic threshold. The dynamic threshold calculated by this method can adapt to the characteristics of different templates. Considering that left turns and right turns exhibit mirror symmetry in steering wheel angle, this invention generates a corresponding mirror template for each basic template. The mirror template is obtained by inverting the angle value of the original template. This design allows the system to simultaneously recognize symmetrical driving actions without defining a separate template for each direction. The dynamic thresholds corresponding to lane-changing, turning, and overtaking scenarios can be calculated using the above method. For any mesoscale driving time-series data segment, its DTW distance with the lateral driving scenario template is calculated. If the calculated DTW distance is less than the dynamic threshold set by the corresponding lateral driving scenario template, the mesoscale driving time-series data segment is determined to be a lateral scenario and belongs to the driving scenario represented by that lateral driving scenario template.

[0075] 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.

[0076] S2: Construct a deep neural network autoencoder, input the driving time sequence segment into the deep neural network autoencoder, and extract low-dimensional deep feature vectors.

[0077] Specifically, a deep neural network autoencoder is first constructed to extract low-dimensional deep feature vectors, forming a deep feature space. In this embodiment, the deep neural network autoencoder is a fusion framework of a convolutional neural network and a long short-term memory network. The input of the deep neural network autoencoder is the driving time sequence segment obtained by S1, and the time length of the input driving time sequence segment is standardized to ensure that the feature dimension and time step are the same. 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 sequence 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.

[0078] Deep neural network autoencoders include:

[0079] 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.

[0080] 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.

[0081] 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.

[0082] S3: Unsupervised clustering is performed on the low-dimensional deep feature vector using a deep embedding clustering method to generate multiple driving styles, and each driving time sequence segment is labeled with a corresponding driving style label.

[0083] Specifically, after obtaining the 32-dimensional low-dimensional deep feature vector, an improved deep embedding clustering algorithm is used for joint optimization. The K-Means algorithm is used to initialize the low-dimensional deep feature vector multiple times, selecting the optimal K cluster centers based on the silhouette coefficient, and initializing K cluster centers, where K is the number of predefined driving styles. In this embodiment, K=3, corresponding to three driving styles: aggressive, robust, and general. A multiple K-means initialization strategy is adopted, performing K-means clustering five times, setting 10 different initial centers each time. 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.

[0084] 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.

[0085] 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.

[0086] After iteratively optimizing the deep embedding clustering model using the objective loss function, hard assignment is performed on driving time-series segments to obtain their respective cluster categories, i.e., driving style labels. The specific clustering results can be labeled as category 0, category 1, and category 2, and the cluster center for each category is determined to represent the driving style under different driving scenarios.

[0087] S4: Freeze the intrinsic parameters of the deep neural network autoencoder, and construct a driving style classification model based on the deep neural network autoencoder with frozen intrinsic parameters. Train the driving style classification model using driving time segments labeled with driving style tags to establish a mapping relationship between driving time segments and driving styles. Input the driving time segment to be identified into the trained driving style classification model to achieve driving style classification.

[0088] Specifically, 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. This simplified network structure significantly reduces computational complexity and the risk of overfitting while maintaining classification accuracy.

[0089] 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 stopped when the validation loss failed to improve beyond the threshold for 20 consecutive rounds 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 rounds for model recovery and analysis. After the classifier training was complete, the driving time-series segments to be identified could be input into the trained driving style classification model to achieve driving behavior classification.

[0090] As an optional embodiment, it also includes S5: extracting statistical features or physical experience features, defining the clustering result labels through human-understandable features, and assigning interpretable semantic labels to the unsupervised clustering results.

[0091] Specifically, unsupervised clustering results are given interpretable semantic labels through statistical methods, domain knowledge, or visualization. Statistical analysis is performed on each cluster obtained from deep unsupervised learning to extract statistical or physical empirical features. The clustering results are defined using human-understandable features, transforming the technical output of unsupervised clustering into domain-understandable quantitative labels.

[0092] The temporal characteristics of driving time segments are calculated, including at least one of the following: mean, standard deviation, zero-crossing rate, peak-to-peak value, and absolute integral. In this invention, the temporal characteristic calculation includes:

[0093] ;

[0094] 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.

[0095] 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:

[0096] ;

[0097] 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.

[0098] Temporal and frequency domain features are used as domain-understandable quantitative data identifiers for driving time-series segments.

[0099] This invention innovatively integrates deep embedding clustering (DEC) with representation learning techniques to build a personalized driving style recognition system capable of accurately identifying driving style categories and mining deep feature representations. It overcomes the problems of weak dynamic modeling capabilities and insufficient feature representation in existing classification schemes.

[0100] 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.

[0101] 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.

[0102] 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.

[0103] 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 personalized driving behavior recognition method based on deep embedding clustering, characterized in that, The method comprises the following steps: S1: acquiring driving time sequence segments of multiple working conditions; The acquisition of the driving time sequence segments of multiple working conditions comprises: collecting a multi-channel driving time sequence data stream by a vehicle-mounted sensor; setting multiple different time length sliding windows and working condition categories for each sliding window, dividing the multi-channel driving time sequence data stream by using the sliding windows to obtain driving time sequence data segments of different time lengths; setting driving feature indicators, and screening driving time sequence data segments corresponding to longitudinal driving working conditions from the data segment set according to the driving feature indicators to realize the identification and classification of longitudinal driving working conditions; setting a lateral driving working condition template, and matching the driving time sequence data segments in the data segment set with the lateral driving working condition template by using a time sequence similarity measurement method to screen driving time sequence data segments corresponding to lateral driving working conditions, thereby realizing the identification and classification of lateral driving working conditions; S2: constructing a deep neural network autoencoder, inputting the driving time sequence segments into the deep neural network autoencoder, and extracting low-dimensional deep feature vectors; the deep neural network autoencoder is a convolutional neural network and long short-term memory network fusion framework; the deep neural network autoencoder comprises: a convolutional feature extraction layer for expanding the feature dimension and compressing the time dimension of the driving time sequence segment to obtain local features; an LSTM encoding layer connected to the output end of the convolutional feature extraction layer, for capturing the time sequence features of the driving time sequence segment, establishing the long-term time sequence dependence of the local features, and obtaining a feature vector; a feature fusion layer connected to the output end of the LSTM encoding layer, for dimension reduction of the feature vector to obtain the low-dimensional deep feature vector; the convolutional feature extraction layer comprises at least two convolutional modules connected in sequence, wherein at least one convolutional module is used for feature dimension expansion and time dimension maintenance, and at least another convolutional module is used for feature dimension expansion and time dimension compression; S3: unsupervised clustering of the low-dimensional deep feature vectors by a deep embedding clustering method, clustering to generate multiple driving styles, and labeling each driving time sequence segment with a corresponding driving style label; S4: freezing the network parameters of the deep neural network autoencoder, constructing a driving style classification model based on the deep neural network autoencoder with frozen network parameters, training the driving style classification model by using driving time sequence segments labeled with driving style labels, and establishing a mapping relationship between the driving time sequence segments and the driving styles; inputting a driving time sequence segment to be identified into the trained driving style classification model to realize driving style classification. 2.The personalized driving behavior recognition method based on deep embedding clustering according to claim 1, characterized in that, The unsupervised clustering of the low-dimensional deep feature vectors by the deep embedding clustering method comprises: using a K-Means algorithm to initialize the low-dimensional deep feature vectors multiple times, and selecting the optimal K cluster centers according to the silhouette coefficient, wherein K is the number of predefined driving styles; calculating the soft assignment probability of each low-dimensional deep feature vector and each cluster center based on Student's t-distribution; A target loss function is constructed, and positions of the cluster centers are iteratively optimized by minimizing the target loss function. 3.The personalized driving behavior recognition method based on deep embedding clustering according to claim 2, characterized in that, The target loss function comprises a reconstruction loss, a KL divergence loss, and at least one of a cluster balance loss and a feature regularization term. 4.The personalized driving behavior recognition method based on deep embedding clustering according to claim 1, characterized in that, The driving style comprises an aggressive type, a conservative type, and a general type. 5.The personalized driving behavior recognition method based on deep embedding clustering according to claim 1, characterized in that, The driving style classification model comprises a front-end deep neural network autoencoder after a frozen network parameter and a rear-end fully connected neural network, an input layer dimension of the fully connected neural network is consistent with a dimension of the low-dimensional deep feature vector, and a neuron number of an output layer is consistent with a type number of the driving style. 6.The personalized driving behavior recognition method based on deep embedding clustering according to claim 1, wherein, Further comprising: A time domain feature of a driving time sequence segment is calculated, and the time domain feature comprises at least one of a mean value, a standard deviation, a zero-crossing rate, a peak-to-peak value, and an absolute integral; A frequency domain feature of each driving time sequence segment is calculated based on a Fourier transform, and the frequency domain feature comprises at least one of a power spectral density, a spectral centroid, a spectral bandwidth, and a spectral flatness; The time domain feature and the frequency domain feature are taken as quantitative data identification of the driving time sequence segment.

Images