Automobile user behavior analysis system and method based on big data

Abstract

The invention discloses an automobile user behavior analysis system and method based on big data, and relates to the technical field of computer data processing, and the method comprises the steps: obtaining and preprocessing automobile operation data, user operation logs, geographic information and external environment data in real time through a multi-source heterogeneous data collection module; generating a high-dimensional behavior feature vector by using a data fusion and feature extraction module; identifying a user behavior intention by introducing a time sequence neural network of an attention mechanism; the self-adaptive evolution tracking module dynamically monitors user behavior mode drift and updates the model; the distributed real-time computing engine provides low-delay data stream processing and model reasoning services. The system comprises a multi-source heterogeneous data acquisition module, a data fusion and feature extraction module, a behavior intention recognition and modeling module, a self-adaptive evolution tracking module and a distributed real-time calculation engine. According to the method, accurate alignment and deep fusion of multi-source data can be realized, and the behavior intention recognition accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing, and in particular to a vehicle user behavior analysis system and method based on big data. BACKGROUND

[0002] With the rapid development of intelligent networked vehicles and vehicle networking technology, the automobile industry is accelerating the transformation to a data-driven service model. User behavior data, as the core link connecting vehicles, services, and the ecosystem, has become a key support for improving user experience, optimizing product design, and implementing personalized services. Vehicle user behavior covers driving habits, vehicle usage frequency, functional preferences, environmental interactions, and other dimensions, with characteristics of high dimensionality, strong time sequence, and multi-source heterogeneity. Its effective analysis relies on efficient collection, fusion, and modeling capabilities of massive, dynamic, and unstructured data. Currently, automobile manufacturers and travel service providers generally rely on vehicle terminals, mobile applications, and cloud platforms to build a preliminary data collection system. However, the overall system is still in the data accumulation stage and has not yet formed a systematic analysis framework focusing on the nature of user behavior.

[0003] Among them, the vehicle user behavior analysis based on big data aims to integrate vehicle operation data, user operation logs, geographic information, and external environment, etc. multi-modal information, to build a quantifiable, predictable, and interpretable user behavior model, thereby supporting intelligent recommendation, risk warning, product iteration, and precise marketing, etc. high-level applications. The core of this technology direction is to achieve efficient mapping from raw data to behavior semantics, and to improve the generalization ability and real-time response performance of the model under the premise of ensuring data security and user privacy. However, existing technologies face multiple structural bottlenecks in achieving this goal.

[0004] Existing technologies generally have problems such as coarse data fusion granularity, single behavior modeling dimension, lack of dynamic evolution capability, and insufficient scalability of computing architecture. First, most systems only aggregate limited vehicle state parameters, and cannot effectively align multi-source data streams with different time stamps, sampling frequencies, and semantic levels, resulting in distorted behavior representation. Second, user behavior analysis relies on static rules or shallow statistical models, lacking modeling capabilities for long-term dependencies and context situations, making it difficult to capture behavior intentions in complex driving scenarios. Third, existing methods generally ignore the time sequence evolution characteristics of user behavior, and cannot achieve adaptive tracking and updating of individual behavior pattern drift. Finally, the analysis system mostly uses centralized batch processing architecture, which is difficult to meet the real-time analysis requirements of high concurrency and low latency, and has significant performance bottlenecks when facing millions of users. SUMMARY

[0005] The application aims to provide a big data-based automobile user behavior analysis system and method to solve the problems of coarse data fusion granularity, single behavior modeling dimension, lack of dynamic evolution capability and insufficient computing architecture scalability in the prior art.

[0006] To solve the above technical problems, the application provides the following technical solutions.

[0007] A big data-based automobile user behavior analysis system, which comprises the following components:

[0008] A multi-source heterogeneous data acquisition module, configured to acquire vehicle operation data, user operation logs, geographic information and external environment data from a vehicle terminal, a mobile application and a cloud platform in real time, and perform timestamp alignment and sampling frequency normalization preprocessing on the raw data;

[0009] A data fusion and feature extraction module, connected to the multi-source heterogeneous data acquisition module, configured to perform time-space alignment and semantic level mapping on the preprocessed multi-modal data stream, and generate a high-dimensional behavior feature vector based on a deep feature extraction network;

[0010] A behavior intention recognition and modeling module, connected to the data fusion and feature extraction module, configured to receive the high-dimensional behavior feature vector, and capture long-term dependency and context situation through a time series neural network with an attention mechanism, so as to realize quantitative analysis of user behavior intention in a complex driving scenario;

[0011] An adaptive evolution tracking module, connected to the behavior intention recognition and modeling module, configured to dynamically monitor individual behavior pattern drift based on a sliding window mechanism and an online learning algorithm, and incrementally update the user behavior model;

[0012] A distributed real-time computing engine, connected to all the above modules, configured to provide low-latency data stream processing and model inference services in a scenario of concurrent use of ten million users, and support elastic scaling of analysis tasks.

[0013] A big data-based automobile user behavior analysis method, which comprises the following specific steps:

[0014] In step S110, a data acquisition interface deployed on a vehicle terminal, a mobile application and a cloud platform is used to acquire vehicle operation data, user operation logs, geographic information and external environment data in real time, and perform timestamp alignment and sampling frequency normalization preprocessing on the raw data to generate a standardized multi-source data stream;

[0015] Step S120: Perform spatiotemporal alignment and semantic hierarchical mapping on the standardized multi-source data stream, and use a deep feature extraction network to generate high-dimensional behavioral feature vectors from the multimodal data. The deep feature extraction network includes convolutional layers and recurrent neural network layers, which are used to capture spatial local patterns and temporal series dependencies, respectively.

[0016] Step S130: Input the high-dimensional behavioral feature vector into a temporal neural network that incorporates an attention mechanism. This network captures long-term dependencies and contextual situations by calculating the hidden state weight distribution and outputs a quantitative score and classification label of the user's behavioral intent.

[0017] Step S140: Based on the sliding window mechanism and online learning algorithm, the statistical deviation of user behavior patterns relative to historical benchmarks is dynamically monitored. When significant pattern drift is detected, the incremental update process of the behavior model is triggered to maintain the model's adaptive tracking ability to individual evolutionary characteristics.

[0018] Step S150: Real-time scheduling of concurrent requests from tens of millions of users is performed through a distributed stream processing framework. The behavior analysis task is decomposed into multiple parallel subtasks and executed in the computing cluster. The allocation of computing resources is dynamically adjusted according to the system load to ensure low-latency response.

[0019] Preferably, the multi-source heterogeneous data acquisition module includes a data interface adaptation unit, a time synchronization unit, and a data quality control unit. The data interface adaptation unit supports multi-protocol data parsing from CAN bus, OBD-II, GPS receiver, inertial measurement unit, and cloud API, and implements unified data format encapsulation. The time synchronization unit uses a combination of network time protocol and hardware clock calibration to control the timestamp deviation of asynchronously arriving multi-source data streams to within 10 milliseconds. The data quality control unit uses outlier detection algorithms and data integrity verification rules to repair or remove missing values ​​and noisy data, ensuring the reliability and consistency of the input data.

[0020] Furthermore, the data fusion and feature extraction module employs a heterogeneous data fusion method based on graph neural networks. This method models vehicles, users, and environmental entities as nodes in a graph structure, and represents the spatiotemporal associations and interactions between entities through edge attributes. The graph neural network aggregates neighborhood node information through a multi-layer message passing mechanism, achieving deep fusion of cross-modal features. The deep feature extraction network consists of two one-dimensional convolutional layers and one bidirectional long short-term memory network layer connected in series. The kernel sizes of the convolutional layers are 3 and 5, respectively, with a stride of 1, used to extract local temporal patterns; the hidden layer of the bidirectional long short-term memory network has a dimension of 128, used to capture forward and backward long-term dependencies, and finally, the features are mapped to a 512-dimensional behavioral feature vector through a fully connected layer.

[0021] Furthermore, the temporal neural network in the behavioral intent recognition and modeling module employs a combined architecture of a multi-head self-attention mechanism and a gated recurrent unit. The multi-head self-attention mechanism has 8 heads, each with a dimension of 64, and assigns importance to features at different time steps by calculating the similarity weights of query-key-value pairs. The gated recurrent unit has a hidden state dimension of 256, used to fuse attention-weighted sequence features and output the probability distribution of behavioral intent. This module defines six typical behavioral intent categories: aggressive driving, economical driving, long-distance cruising, urban commuting, and function exploration and risk avoidance, achieving a classification accuracy of over 92.5%.

[0022] Preferably, the sliding window length of the adaptive evolution tracking module is 30 days, and the window sliding step size is 1 day. The online learning algorithm adopts an incremental learning strategy based on stochastic gradient descent, with an initial learning rate of 0.01, which decays exponentially with the number of iterations. Mode drift detection is based on Kullback-Leibler divergence to calculate the difference between the current window behavior distribution and the historical baseline distribution. When the divergence value exceeds the threshold of 0.15, the model update process is triggered. The update process only trains on the latest window data for 3 epochs, and the adjustment of model parameters is limited to within 10% to balance adaptability and stability.

[0023] Furthermore, the distributed real-time computing engine is built on the Apache Flink framework and integrates a model service interface. The computing cluster consists of at least 32 nodes, each equipped with a 16-core CPU and 64GB of memory, supporting exactly-Once semantics for data stream processing. An elastic scaling mechanism dynamically adjusts task parallelism by monitoring queue backlog length and CPU utilization metrics, with the parallelism adjustment range from 8 to 256. The real-time inference service encapsulates the trained behavioral model as a gRPC microservice, achieving an average response latency of less than 50 milliseconds and a system throughput of 20,000 queries per second.

[0024] Furthermore, the deep feature extraction network in step S120 is trained end-to-end using the following feature reconstruction loss function:

[0025]

[0026] in To input multimodal data, To reconstruct the output, The KL divergence regularization term is the term of the potential spatial distribution. The value is set to 0.01. This loss function ensures that the extracted behavioral feature vectors have a smooth underlying representation while preserving the main information of the original data.

[0027] Preferably, the attention mechanism temporal neural network in step S130 calculates the time step using the following formula. Context vector:

[0028]

[0029] in, For the encoder's hidden state, attention weights Score normalized by the softmax function:

[0030]

[0031] in, The score function uses an additive model to represent the hidden state of the decoder at the previous time step. This mechanism allows the model to dynamically focus on the historical fragments most relevant to the current behavioral intent.

[0032] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0033] This invention achieves accurate alignment and deep fusion of multi-source data streams with asynchronous timestamps, different sampling frequencies, and different semantic levels by using a multi-source heterogeneous data acquisition module and a data fusion method based on graph neural networks, thereby reducing the distortion rate of behavioral representation.

[0034] This invention, through a behavioral intent recognition and modeling module, employs a combination architecture of multi-head self-attention mechanism and gated loop unit to effectively capture long-term dependencies and complex contextual situations, thereby improving the accuracy of capturing behavioral intent in driving scenarios.

[0035] The adaptive evolution tracking module of this invention achieves real-time monitoring and incremental model updates of individual behavior pattern drift through a sliding window and an online learning algorithm, enabling the system to continuously track the temporal evolution characteristics of user behavior and shortening the model adaptive update response time to within 2 hours.

[0036] This invention's distributed real-time computing engine is based on the Apache Flink framework and integrates an elastic scaling mechanism. It supports high-concurrency, low-latency analysis needs for tens of millions of users, with a system throughput of 20,000 queries per second and an average response latency of less than 50 milliseconds. This significantly improves the effectiveness of automotive user behavior analysis in key scenarios such as smart cockpits, active safety, and user operations. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall technical architecture of the big data-based automotive user behavior analysis system proposed in this invention.

[0038] Figure 2This is a schematic diagram of the core principle framework of multi-source heterogeneous data fusion and behavioral intent recognition in this invention. Detailed Implementation

[0039] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present 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 merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0040] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0041] In the embodiments of the present invention, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of the present invention shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on the present invention; the term "multiple" in the present invention refers to two or more (including two).

[0042] Example 1

[0043] In real-world operating scenarios of intelligent connected vehicles, a big data-based vehicle user behavior analysis system collects real-time vehicle operation data, user operation logs, geographic information, and external environmental data from in-vehicle terminals, mobile applications, and cloud platforms via a multi-source heterogeneous data acquisition module. The in-vehicle terminal continuously acquires vehicle operating parameters such as engine speed, vehicle speed, throttle opening, and brake pressure through the Controller Area Network (CAN) bus interface, with a sampling frequency of 100 Hz. The mobile application records the vehicle's position trajectory through a GPS receiver and collects acceleration and angular velocity data through an inertial measurement unit (IMU), with a sampling frequency of 50 Hz. The cloud platform acquires external environmental data such as traffic conditions, weather conditions, and road grades through an application programming interface (API), updating once per minute. The time synchronization unit uses a combination of network time protocols and hardware clock calibration to control the timestamp deviation of asynchronously arriving multi-source data streams to within 10 milliseconds. The data quality control unit identifies and removes sensor readings exceeding the normal range using a 3σ criterion-based outlier detection algorithm and performs linear interpolation compensation for missing GPS coordinates using data integrity verification rules, ensuring the reliability and consistency of the input data.

[0044] See Figure 1 This system includes a multi-source heterogeneous data acquisition module, a data fusion and feature extraction module, a behavior intent recognition and modeling module, an adaptive evolution tracking module, and a distributed real-time computing engine.

[0045] The data fusion and feature extraction module performs spatiotemporal alignment and semantic hierarchical mapping on the preprocessed multimodal data streams. The spatiotemporal alignment process first resamples the data streams with different sampling frequencies to 10 Hz, and then precisely matches vehicle operation data, user operation data, and environmental data based on timestamps. The semantic hierarchical mapping labels continuous operation sequences with throttle opening greater than 70% and brake pressure less than 10% as aggressive driving semantic segments, and driving segments with stable speeds of 60-80 km / h and throttle opening maintained at 30%-50% as economic driving semantic segments. See also... Figure 2 This module employs a heterogeneous data fusion method based on graph neural networks, modeling vehicles, users, and environmental entities as nodes in a graph structure. Edge attributes represent the spatiotemporal relationships and interactions between entities. The graph neural network aggregates neighborhood node information through a three-layer message passing mechanism, with hidden dimensions of 64, 128, and 256 for each layer, achieving deep fusion of cross-modal features. The deep feature extraction network consists of two one-dimensional convolutional layers and one bidirectional long short-term memory (LSTM) network layer connected in series. The convolutional layers have kernel sizes of 3 and 5, a stride of 1, and identical padding. The ReLU activation function is used to extract local temporal patterns. The LSTM network has a hidden dimension of 128 and a dropout rate of 0.2, used to capture long-term dependencies in both forward and backward directions. Finally, a fully connected layer maps the features to a 512-dimensional behavioral feature vector.

[0046] The deep feature extraction network employs a feature reconstruction loss function for end-to-end training. This function is defined as the mean squared error between the input multimodal data and the reconstructed output, plus a KL divergence regularization term for the latent space distribution. During training, the Adam optimizer is used with a batch size of 256, an initial learning rate of 0.001, and a training epoch of 100. This loss function ensures that the extracted behavioral feature vectors retain the main information of the original data while possessing a smooth latent representation, with the reconstruction error controlled within 0.05.

[0047] The behavioral intent recognition and modeling module receives a 512-dimensional high-dimensional behavioral feature vector and captures long-term dependencies and contextual situations through a temporal neural network incorporating an attention mechanism. The temporal neural network employs a combination architecture of a multi-head self-attention mechanism and a gated recurrent unit. The multi-head self-attention mechanism has 8 heads, each with a dimension of 64, and assigns importance to features at different time steps by calculating the similarity weights of query-key-value pairs. The attention mechanism calculates the context vector for each time step using a specific formula; this vector is a weighted sum of the encoder's hidden states, with weights determined by the softmax function normalization score. The gated recurrent unit has a hidden state dimension of 256 and a dropout rate of 0.3, used to fuse the attention-weighted sequence features and output the probability distribution of the behavioral intent. This module defines six typical behavioral intent categories: aggressive driving, economical driving, long-distance cruising, urban commuting, and functional exploration and risk avoidance. The judgment threshold for each category is set to 0.75, achieving a classification accuracy of over 92.5%.

[0048] The adaptive evolution tracking module dynamically monitors individual behavior pattern drift based on a sliding window mechanism and an online learning algorithm. The sliding window length is 30 days, with a sliding step size of 1 day, processing approximately 10,000 behavior records per day. The online learning algorithm employs an incremental learning strategy based on stochastic gradient descent, with an initial learning rate of 0.01 and a decay rate of 0.95, which decreases exponentially with each iteration. Pattern drift detection calculates the difference between the current window behavior distribution and the historical baseline distribution based on Kullback-Leibler divergence, which is obtained from the data of the previous 30 windows. When the divergence value exceeds a threshold of 0.15, a model update process is triggered. The update process only trains the latest window data for 3 epochs, with model parameter adjustments limited to within 10% to balance adaptability and stability. The model adaptive update response time is controlled within 2 hours to ensure the system can continuously track the temporal evolution characteristics of user behavior.

[0049] The distributed real-time computing engine is built on the Apache Flink framework and integrates a model service interface. The computing cluster consists of 32 nodes, each equipped with a 16-core CPU and 64GB of memory, running CentOS 7.9 and using OpenJDK 11 as the Java runtime environment. Data stream processing supports exactly-once semantics, with a checkpoint interval of 30 seconds, and RocksDB is used as the state backend. The elastic scaling mechanism dynamically adjusts task parallelism by monitoring queue backlog length and CPU utilization. When the queue backlog exceeds 10,000 entries and CPU utilization remains above 80% for 5 minutes, the system automatically increases the parallelism from the initial value of 64 to a maximum of 256. When the queue backlog falls below 1,000 entries and CPU utilization remains below 30% for 10 minutes, the system gradually reduces the parallelism to a minimum of 8. The real-time inference service encapsulates the trained behavioral model as a gRPC microservice, using protocol buffer version 3.19 and service port 50051. The average response latency is less than 50 milliseconds, the system throughput reaches 20,000 queries per second, and the service availability reaches 99.95%.

[0050] In the specific implementation process, the system executes a big data-based automotive user behavior analysis method, which includes the following steps: Step S110, real-time acquisition of vehicle operation data, user operation logs, geographic information, and external environment data is achieved through data acquisition interfaces deployed on in-vehicle terminals, mobile applications, and cloud platforms. The in-vehicle terminal data acquisition interface supports the Controller Area Network (CAN) bus protocol and the vehicle diagnostic system protocol, collecting 100 data points per second. The mobile application data acquisition interface calls Android location services and sensor services, collecting 50 data points per second. The cloud platform data acquisition interface obtains external environment data through a RESTful application programming interface, updating once per minute. The raw data undergoes timestamp alignment and sampling frequency normalization preprocessing, with a timestamp alignment accuracy of 10 milliseconds and a sampling frequency uniformly set to 10 Hz, generating a standardized multi-source data stream.

[0051] Step S120 involves spatiotemporal alignment and semantic hierarchical mapping of the standardized multi-source data streams. The spatiotemporal alignment process first establishes a unified time reference system, then uses Lagrange interpolation to unify data streams with different sampling frequencies to 10 Hz. Semantic hierarchical mapping is based on a predefined rule base, marking operation sequences with throttle opening greater than 80% and lateral acceleration exceeding 0.5g within 5 consecutive seconds as aggressive driving semantics, and marking driving segments with speed fluctuations less than 10% and fuel consumption below average within 10 consecutive minutes as economical driving semantics. A deep feature extraction network is used to generate high-dimensional behavioral feature vectors from the multimodal data. The deep feature extraction network includes convolutional layers and recurrent neural network layers to capture spatial local patterns and temporal series dependencies, respectively. The convolutional layers use 32 and 64 filters, with ReLU activation function and batch normalization layer momentum set to 0.99. The recurrent neural network layer uses a bidirectional long short-term memory network with 128 hidden units and a sequence length of 60 time steps, corresponding to a 6-second behavioral segment.

[0052] Step S130: The high-dimensional behavioral feature vector is input into a temporal neural network incorporating an attention mechanism. This network captures long-term dependencies and contextual information by calculating the hidden state weight distribution. During the attention weight calculation, an additive model is used for the scoring function, with a parameter matrix dimension of 256×256 and a bias vector dimension of 256. During the context vector calculation, the encoder hidden state dimension is 256, and the sequence length is set to 60. The network outputs a quantitative score and classification label for the user's behavioral intent. The quantitative score ranges from 0 to 1, and the classification label corresponds to 6 behavioral intent categories. For each time step, the network outputs the probability distribution of the 6 categories, and the maximum probability value is taken as the final classification result. When the maximum probability value exceeds 0.75, it is determined to belong to that category.

[0053] Step S140: Based on the sliding window mechanism and online learning algorithm, the statistical deviation of user behavior patterns relative to historical baselines is dynamically monitored. The sliding window contains 30 days of behavior data, approximately 10,000 records per day, for a total window capacity of 300,000 records. The online learning algorithm uses a stochastic gradient descent optimizer with a momentum parameter set to 0.9 and a weight decay coefficient of 0.0001. Pattern drift detection calculates the Kullback-Leibler divergence between the current window behavior distribution and the historical baseline distribution. The baseline distribution is obtained from the data of the previous 30 windows and includes the frequency and duration statistics for each behavior category. When a significant pattern drift is detected, i.e., the Kullback-Leibler divergence value exceeds the threshold of 0.15, the incremental update process of the behavior model is triggered. The update process uses the latest window data, with a batch size of 128, a learning rate that decays exponentially from 0.01 to 0.95, and is trained for 3 epochs to ensure the model's adaptive tracking ability to individual evolutionary characteristics.

[0054] Step S150 involves real-time scheduling of concurrent requests from tens of millions of users using a distributed stream processing framework. This framework is built on Apache Flink, with the job manager configured for high availability and ZooKeeper used for leader election. The initial number of task managers is 32, each configured with 4 task slots. Behavioral analysis tasks are decomposed into multiple parallel subtasks executed in the computing cluster. The parallelism of the data source operator is set to 64, the mapping operator to 128, and the aggregation operator to 64. Computing resource allocation is dynamically adjusted based on system load, with monitoring metrics including records per second, processing latency, CPU utilization, and memory utilization. When processing latency exceeds 100 milliseconds and CPU utilization remains above 85%, the number of task managers is automatically increased to 48. When processing latency is below 20 milliseconds and CPU utilization remains below 40%, the number of task managers is gradually reduced to 24 to ensure low-latency response.

[0055] Example 2

[0056] In the operation scenario of car-sharing service platforms, a big data-based car user behavior analysis system performs refined behavior analysis on short-term rental users. A multi-source heterogeneous data acquisition module collects data from the car-sharing vehicle's onboard terminal, user mobile application, and operation management platform. The onboard terminal acquires vehicle status data, including engine operating parameters, door lock status, and seatbelt usage, via the Controller Area Network (CAN) bus interface, at a sampling frequency of 50 Hz. The user mobile application records driving trajectories via GPS and user interface operation sequences via touchscreen interaction, at a sampling frequency of 30 Hz. The operation management platform acquires business data such as basic user information, rental history, and payment behavior via an application programming interface (API), updating once per minute. The time synchronization unit uses a combination of hardware clock calibration and software timestamp correction to control the timestamp deviation of the multi-source data streams to within 15 milliseconds. The data quality control unit identifies abnormal driving behavior segments using an anomaly detection model based on the isolated forest algorithm and compensates for missing sensor readings using forward padding based on data integrity verification rules, ensuring the reliability and consistency of the input data.

[0057] The data fusion and feature extraction module performs spatiotemporal alignment and semantic hierarchical mapping on the preprocessed multimodal data stream. The spatiotemporal alignment process first establishes an analysis time window based on rental orders, then precisely matches vehicle status data, user interaction data, and business data. Semantic hierarchical mapping marks combinations of operations such as rapid acceleration, sudden braking, and sharp turns as high-risk driving semantic segments, and behaviors such as prolonged idling and frequent door opening and closing as abnormal usage semantic segments. This module employs a heterogeneous data fusion method based on graph neural networks, modeling vehicles, users, and rental orders as nodes in a graph structure, using edge attributes to represent associated features such as rental relationships, usage frequency, and driving habits. The graph neural network aggregates neighborhood node information through a two-layer message passing mechanism, with hidden dimensions of 128 and 256 in each layer, achieving deep fusion of cross-modal features. The deep feature extraction network consists of three one-dimensional convolutional layers and one bidirectional gated recurrent unit layer connected in series. The kernel sizes of the convolutional layers are 3, 5, and 3, with a stride of 1, and the activation function is ELU, used to extract local temporal patterns. The bidirectional gated recurrent unit has a hidden layer dimension of 256 and a dropout rate of 0.3. It is used to capture long-term dependencies in both forward and backward directions. Finally, the features are mapped to a 1024-dimensional behavioral feature vector through a fully connected layer.

[0058] The behavioral intent recognition and modeling module receives a 1024-dimensional high-dimensional behavioral feature vector and captures long-term dependencies and contextual situations through a temporal neural network with an attention mechanism. The temporal neural network employs a combination architecture of a multi-head self-attention mechanism and a gated recurrent unit. The multi-head self-attention mechanism has 12 heads, each with a dimension of 64, and assigns importance to features at different time steps by calculating the similarity weights of query-key-value pairs. The gated recurrent unit has a hidden state dimension of 512 and a dropout rate of 0.4, used to fuse attention-weighted sequence features and output the probability distribution of user behavioral intent. This module expands the behavioral intent categories for the car-sharing scenario, adding four special modes: short-term rental adaptation, unfamiliarity with functions, vehicle exploration, and rule testing. Combined with the original six typical modes, a total of ten behavioral intent categories are defined, with a judgment threshold of 0.7 for each category, achieving a classification accuracy of 90.8%.

[0059] The adaptive evolution tracking module dynamically monitors individual behavior pattern drift based on a sliding window mechanism and an online learning algorithm. The sliding window length is 14 days, with a sliding step size of 1 day, processing approximately 5000 behavior records per day. The online learning algorithm employs an Adam-based incremental learning strategy with an initial learning rate of 0.005 and a decay rate of 0.9, which decreases exponentially with each iteration. Pattern drift detection calculates the difference between the current window behavior distribution and the historical baseline distribution based on Jensen-Shannon divergence, obtained from the data of the previous 14 windows. When the divergence value exceeds a threshold of 0.2, a model update process is triggered. The update process trains only the latest window data for 2 epochs, with model parameter adjustments limited to within 15% to balance adaptability and stability. The model adaptive update response time is controlled within 1 hour to ensure the system can quickly adapt to rapid changes in shared car user behavior patterns.

[0060] The distributed real-time computing engine is built on the Apache Flink framework and integrates a model service interface. The computing cluster consists of 16 nodes, each equipped with a 32-core CPU and 128GB of memory, running Ubuntu 20.04 on the operating system and Oracle JDK 11 in Java. Data stream processing supports exactly-once semantics, with a checkpoint interval of 60 seconds, and the state backend uses a file system. An elastic scaling mechanism dynamically adjusts task parallelism by monitoring queue backlog length and memory utilization. When the queue backlog exceeds 5000 entries and memory utilization remains above 85% for 5 minutes, the system automatically increases the parallelism from the initial value of 32 to a maximum of 128. When the queue backlog falls below 500 entries and memory utilization remains below 40% for 10 minutes, the system gradually reduces the parallelism to a minimum of 16. The real-time inference service encapsulates the trained behavioral model as a gRPC microservice, using protocol buffer version 3.15 and port 50052. The average response latency is less than 30 milliseconds, the system throughput reaches 15,000 queries per second, and the service availability reaches 99.9%.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A car user behavior analysis system based on big data, characterized in that, It includes the following components: The multi-source heterogeneous data acquisition module is used to collect vehicle operation data, user operation logs, geographic information and external environment data in real time from vehicle terminals, mobile applications and cloud platforms, and to perform timestamp alignment and sampling frequency normalization preprocessing on the raw data. The data fusion and feature extraction module is connected to the multi-source heterogeneous data acquisition module. It is used to perform spatiotemporal alignment and semantic hierarchical mapping on the preprocessed multimodal data stream, and generate high-dimensional behavioral feature vectors based on a deep feature extraction network. The behavior intention recognition and modeling module is connected to the data fusion and feature extraction module. It is used to receive high-dimensional behavioral feature vectors and capture long-term dependencies and contextual situations through a temporal neural network with an attention mechanism. The adaptive evolution tracking module, connected to the behavior intent recognition and modeling module, is used to dynamically monitor the drift of individual behavior patterns based on a sliding window mechanism and online learning algorithms, and to incrementally update the user behavior model. The distributed real-time computing engine connects to all the aforementioned modules and is used to provide low-latency data stream processing and model inference services in scenarios with tens of millions of concurrent users, while also supporting elastic scaling of analysis tasks.

2. The big data-based automotive user behavior analysis system according to claim 1, characterized in that: The multi-source heterogeneous data acquisition module includes a data interface adapter unit, a time synchronization unit, and a data quality control unit; the data interface adapter unit supports multi-protocol data parsing from CAN bus, OBD-II, GPS receiver, inertial measurement unit, and cloud API; The time synchronization unit uses a combination of network time protocol and hardware clock calibration to control the timestamp deviation of asynchronous multi-source data streams to within 10 milliseconds; the data quality control unit repairs or removes missing values ​​and noisy data through outlier detection algorithms and data integrity verification rules.

3. The car user behavior analysis system based on big data according to claim 1, characterized in that: The data fusion and feature extraction module adopts a heterogeneous data fusion method based on graph neural networks, which models vehicle, user, and environmental entities as nodes in a graph structure and represents the spatiotemporal association and interaction relationship between entities through edge attributes; the graph neural network aggregates neighborhood node information through a multi-layer message passing mechanism; The deep feature extraction network consists of two one-dimensional convolutional layers and one bidirectional long short-term memory network layer connected in series. The kernel sizes of the convolutional layers are 3 and 5 respectively, with a stride of 1. The hidden layer dimension of the bidirectional long short-term memory network is 128. Finally, the features are mapped to a 512-dimensional behavioral feature vector through a fully connected layer.

4. The big data-based automotive user behavior analysis system according to claim 1, characterized in that: The temporal neural network in the behavioral intent recognition and modeling module adopts a combination architecture of multi-head self-attention mechanism and gated recurrent unit. The multi-head self-attention mechanism has 8 heads, each with a dimension of 64, and calculates the similarity weight of query-key value pairs. The hidden state dimension of the gated recurrent unit is 256, which is used to fuse attention-weighted sequence features and output the probability distribution of behavioral intent. The behavioral intent categories defined by this module include 6 typical modes: aggressive driving, economical driving, long-distance cruising, urban commuting, function exploration, and risk avoidance.

5. The big data-based automotive user behavior analysis system according to claim 1, characterized in that: The adaptive evolution tracking module has a sliding window length of 30 days and a window sliding step size of 1 day. The online learning algorithm employs an incremental learning strategy based on stochastic gradient descent, with an initial learning rate of 0.01 that decays exponentially with each iteration. Mode drift detection is based on Kullback-Leibler divergence to calculate the difference between the current window behavior distribution and the historical baseline distribution. When the divergence value exceeds a threshold of 0.15, the model update process is triggered. The update process only trains the latest window data for 3 epochs, and the adjustment of model parameters is limited to within 10%.

6. The car user behavior analysis system based on big data according to claim 1, characterized in that: The distributed real-time computing engine is built on the Apache Flink framework and integrates a model service interface. The computing cluster consists of at least 32 nodes, each equipped with a 16-core CPU and 64GB of memory, supporting exactly-Once semantics for data stream processing. The elastic scaling mechanism dynamically adjusts the parallelism of tasks by monitoring queue backlog length and CPU utilization metrics, with the parallelism adjustment range being 8 to 256. The real-time inference service encapsulates the trained behavioral model as a gRPC microservice, with an average response latency of less than 50 milliseconds.

7. A method for analyzing car user behavior based on big data, applied to the car user behavior analysis system based on big data as described in any one of claims 1-6, characterized in that, Includes the following steps: Step S110: Through the data acquisition interfaces deployed on the vehicle terminal, mobile application and cloud platform, real-time vehicle operation data, user operation logs, geographic information and external environment data are acquired, and the raw data is preprocessed by timestamp alignment and sampling frequency normalization to generate a standardized multi-source data stream. Step S120: Perform spatiotemporal alignment and semantic hierarchical mapping on the standardized multi-source data stream, and use a deep feature extraction network to generate high-dimensional behavioral feature vectors from the multimodal data. The deep feature extraction network includes convolutional layers and recurrent neural network layers, which are used to capture spatial local patterns and temporal series dependencies, respectively. Step S130: Input the high-dimensional behavioral feature vector into a temporal neural network that incorporates an attention mechanism. This network captures long-term dependencies and contextual situations by calculating the hidden state weight distribution and outputs a quantitative score and classification label of the user's behavioral intent. Step S140: Based on the sliding window mechanism and online learning algorithm, dynamically monitor the statistical deviation of user behavior patterns relative to historical benchmarks. When a significant pattern drift is detected, trigger the incremental update process of the behavior model. Step S150: Real-time scheduling of concurrent requests from tens of millions of users is performed through a distributed stream processing framework. The behavior analysis task is decomposed into multiple parallel subtasks to be executed in the computing cluster, and the allocation of computing resources is dynamically adjusted according to the system load.

8. The method for analyzing car user behavior based on big data according to claim 7, characterized in that: The deep feature extraction network in step S120 is trained end-to-end using a feature reconstruction loss function: in, To input multimodal data, To reconstruct the output, The KL divergence regularization term is the term of the potential spatial distribution. The value is 0.

01.

9. The method for analyzing car user behavior based on big data according to claim 7, characterized in that: The attention mechanism temporal neural network in step S130 calculates the time step using the following formula. Context vector: in, For the encoder's hidden state, attention weights Score normalized by the softmax function: in, The hidden state of the decoder at the previous time step is represented by an additive model for the scoring function.

10. The method for analyzing car user behavior based on big data according to claim 7, characterized in that: The mode drift detection in step S140 is based on the Kullback-Leibler divergence to calculate the difference between the current window behavior distribution and the historical baseline distribution. When the divergence value exceeds the threshold of 0.15, the model update process is triggered. The update process uses the latest window data, with a batch size of 128, and the learning rate starts from 0.01 and decays exponentially at a decay rate of 0.95, and is trained for 3 epochs.

Images