Improved Transform intelligent network connection automobile anomaly detection method based on graph neural network

By combining graph neural network and improved Transformer model, the spatio-temporal characteristics of intelligent connected vehicle sensors are extracted and feature representations are optimized. The problem of low detection efficiency of traditional methods in complex environments is solved, and efficient and real-time abnormal detection is achieved to ensure that the vehicle is Safety.

CN119089346APending Publication Date: 2024-12-06CHONGQING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202411122071.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The prior art method used for vehicle sensor data abnormality detection in intelligent connected vehicles has problems such as strong data dependence, poor adaptability, low model training efficiency, and difficult balance of computing resources and time delays, which is difficult to implement in a complex network of vehicles environment. Quick and accurate detection.

Method used

Using graph neural network (GNN) and improved Transformer model, we use the in-vehicle sensor graph model to extract spatiotemporal features and add channel attention mechanism to optimize feature representation to improve the accuracy and efficiency of abnormal detection and realize real-time detection.

Benefits of technology

It improves the accuracy and efficiency of abnormal detection, enhances the robustness and real-timeness of the model, and can maintain high detection accuracy in complex environments, which is suitable for the actual deployment of intelligent connected vehicles, ensuring the safety of vehicle driving.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089346A_ABST
    Figure CN119089346A_ABST
Patent Text Reader

Abstract

The invention relates to an improved Transform intelligent network connection automobile anomaly detection method based on a graph neural network, and belongs to the technical field of mobile communication. Complexity and heterogeneity of data collected by an in-vehicle sensor network enable a traditional anomaly detection method to be high in false alarm rate and poor in real-time performance. In order to solve the above problem, the method achieves the quick and effective detection of the abnormal state of the vehicle through the fusion of Transform, GNN and a channel attention mechanism. The method comprises the following steps: modeling a spatial association relationship of an in-vehicle sensor network based on GNN, regarding each sensor as a node, and forming an edge of a graph by a data interaction relationship among the sensors; and the Transform is used for extracting the time sequence characteristics of the data. By adding a channel attention mechanism, the importance weight of each channel is adaptively adjusted, and the accuracy and efficiency of anomaly detection are improved. The method can be used for detecting the abnormity of the multi-dimensional data collected by the intelligent networked automobile sensor.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile communications and relates to an improved Transformer intelligent connected vehicle anomaly detection method based on a graph neural network. Background Art

[0002] With the development of Internet of Vehicles technology, vehicles have improved their perception capabilities through various sensors and communication modules to achieve intelligent control and management. However, traditional anomaly detection methods have limited processing capabilities and cannot effectively cope with the complex and changing Internet of Vehicles environment, and cannot quickly and accurately detect anomalies in vehicle sensor data. With the assistance of artificial intelligence technology, intelligent connected vehicles can not only make decisions based on the current environment, but also continuously learn and adapt to environmental changes. Anomaly detection methods based on deep learning have performed well in improving detection accuracy and efficiency, but they also face the following challenges.

[0003] Current research mainly uses traditional statistical analysis and machine learning methods for anomaly detection. These methods are highly dependent on data and have poor adaptability, and cannot achieve ideal results in complex vehicle networking environments. At the same time, when the data of a single sensor is subject to noise and interference, the detection effect is difficult to guarantee. Graph neural networks (GNNs) use graph structure models to treat sensors as nodes, which can effectively capture the spatial correlation between sensor data; at the same time, the improved Transformer model is good at processing long time series data, and through the multi-head self-attention mechanism, it can improve the ability to extract time series information.

[0004] Combining the advantages of GNN and Transformer, this paper proposes an efficient abnormal data detection method for intelligent connected vehicles, CaTransGNN. First, the spatial features of sensor data are extracted by GNN model, and then the time series features are extracted by Transformer model, the spatiotemporal feature information is integrated, and the channel attention mechanism is added to optimize the feature representation to improve the accuracy and efficiency of anomaly detection. Four main types of anomalies are considered in this scenario: instantaneous anomalies, constant anomalies, gradual drift anomalies, and deviation anomalies.

[0005] The anomaly detection method includes two core steps: offline training and online detection. In the offline training stage, a large amount of normal and abnormal data is collected and preprocessed, and spatial and temporal features are extracted using GNN and Transformer models, and CAM is combined to further optimize the features for training. In the online detection stage, the model receives sensor data in real time and performs anomaly detection to ensure vehicle driving safety.

[0006] Although the anomaly detection method based on deep learning has broad application prospects in the field of intelligent transportation, the following problems still need to be solved. First, model training requires a large amount of labeled data, which is costly to obtain, and the data quality directly affects the detection effect. Second, in complex environments, there may be problems with low model training efficiency. Finally, when the model is deployed in actual intelligent connected vehicles, it is necessary to consider the balance between computing resources and latency.

[0007] In summary, the present invention proposes an improved Transformer intelligent connected vehicle data anomaly detection method CaTransGNN based on graph neural network, which overcomes the shortcomings of the prior art, can effectively improve the detection accuracy and real-time performance, and has important application prospects. Summary of the invention

[0008] In view of this, the purpose of the present invention is to provide an improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] In the first aspect, the present invention can solve the problem of abnormal detection of vehicle sensor data, and the solution is divided into the following steps:

[0011] S1: Building an in-vehicle sensor graph model

[0012] S2: Data collection and preprocessing

[0013] S3: Data spatiotemporal feature extraction

[0014] S4: Model training and testing

[0015] S5: Data anomaly classification and analysis feedback

[0016] In the second aspect, in S1, the embodiment of the present invention proposes a method for constructing a vehicle sensor graph model. According to the location and function of the vehicle sensors, a graph structure model is constructed. Assume that there are N sensors in the vehicle, each sensor is represented by a node V i , the data interaction between sensors is represented by the edge E of the graph ij The entire graph structure G = (V, E) can effectively capture the complex correlations between sensors, help extract spatial features, and lay the foundation for subsequent anomaly detection. This step effectively establishes a relationship network between sensors, making subsequent feature extraction and anomaly detection more accurate.

[0017] In a third aspect, in S2, an embodiment of the present invention proposes a method for data collection and preprocessing. A data acquisition system is used to collect historical data of sensors in the vehicle, including camera image frames and inertial measurement unit data. Suppose the data collected by sensor i at time t is x i (t). Preprocess and annotate the collected data to generate a high-quality training dataset D = {x i (t), y i (t)}, where y i (t) is the annotation information of sensor data. Through denoising and data cleaning, the integrity and consistency of the data are ensured.

[0018] In a fourth aspect, an embodiment of the present invention proposes a method for extracting spatiotemporal features in S3. An improved Transformer model is used to extract time series features. The input data can accurately capture the complex dependencies in the time series data through a multi-head self-attention mechanism. In addition, the time characteristics of the input data are retained during the training process through position encoding based on the shortest path, thereby improving the accuracy and effect of time feature extraction. The Transformer model can process long time series data and improve the time series sensitivity of anomaly detection. At the same time, the GNN model is used to extract spatial features between sensors. Each node V i The initial features are Through the message passing mechanism, in each iteration, node V i From its neighbor node set The GNN model receives information from the sensor and updates its own features. Through iterative updates, the information of each node and its neighboring nodes is integrated to effectively capture the potential spatial correlation in the data. The advantage of this method is that the GNN model can process graph structure data, capture the spatial correlation between sensors, and improve the spatial feature extraction capability of anomaly detection.

[0019] In a fifth aspect, in S4, an embodiment of the present invention proposes a method for training an anomaly detection model. The time series features T extracted by Transformer are i And the spatial features S extracted by GNN i Fusion is performed to obtain the fusion feature Z i The fusion features are reduced in dimension through the fully connected layer, and the feature representation is optimized by combining the channel attention mechanism to generate the final anomaly detection result. The anomaly detection model includes offline training and online monitoring. During the training process, a comprehensive loss function is used for optimization to improve the accuracy and stability of the anomaly detection model. In the online detection phase, based on the trained model, sensor data x(t) is received in real time for inference and anomaly detection. Offline training can ensure the detection accuracy of the model in various situations, while online detection can respond to anomaly detection results in real time, thereby improving the safety of vehicle operation.

[0020] In a sixth aspect, in S5, an embodiment of the present invention proposes a method for classifying and analyzing abnormal data. The detected abnormal data is classified and analyzed. The four attributes of instantaneous abnormality, constant abnormality, gradual drift abnormality and deviation abnormality are combined, and the abnormality is divided into these four categories using a classification algorithm, and different processing methods and early warning mechanisms are used to improve the effectiveness and accuracy of abnormality processing. The detection results are fed back to the in-vehicle control system in a timely manner to adjust the safety strategy.

[0021] The beneficial effects of the present invention are:

[0022] (1) Efficient anomaly detection

[0023] Combining the time series processing capabilities of Transformer and the spatial feature extraction advantages of GNN, the temporal and spatial characteristics of sensor data are fully captured, thereby improving the accuracy and efficiency of anomaly detection.

[0024] Adaptively adjust the importance weight of each channel to highlight key features, suppress interference information, and further improve the performance of the model.

[0025] (2) Improved real-time performance

[0026] The model can receive sensor data in real time and perform anomaly detection to ensure vehicle driving safety and respond to abnormal situations in a timely manner.

[0027] The model is lightweight in design, which reduces the demand for computing resources and latency, and is more suitable for deployment in intelligent connected vehicles.

[0028] (3) High robustness

[0029] GNN can effectively capture the spatial correlation between sensor data, enhance the robustness of the model, and maintain high detection accuracy even in the presence of noise and interference in the data.

[0030] Automatically learn the connection relationship between sensor nodes to avoid the limitations of node selection in traditional methods and further improve the robustness of the model.

[0031] (4) Widely used

[0032] It can be used to detect anomalies in multi-dimensional data collected by smart connected car sensors, such as images, IMU data, etc.

[0033] It can be applied to different intelligent transportation and unmanned driving scenarios, such as autonomous driving, vehicle-road collaboration, etc.

[0034] (5) Security assurance

[0035] The detected anomalies are classified and analyzed, such as instantaneous anomalies, constant anomalies, gradual drift anomalies and deviation anomalies, and corresponding processing methods and early warning mechanisms are adopted to ensure vehicle driving safety.

[0036] The detection results are promptly fed back to the in-vehicle control system to adjust the safety strategy, such as triggering an alarm or taking preventive measures.

[0037] The present invention provides an efficient, real-time and robust intelligent connected vehicle anomaly detection method, which can effectively improve the safety and data processing efficiency of intelligent transportation and has important application value.

[0038] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0040] Figure 1 This is the architecture diagram of the improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network;

[0041] Figure 2 Schematic diagram of sensor node dependency capture and update based on GNN;

[0042] Figure 3 Execute the flowchart for the GNN-based improved Transformer method. DETAILED DESCRIPTION

[0043] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0044] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0045] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0046] Figure 1 FIG. 1 shows a model architecture diagram of an anomaly detection system involved in an embodiment of the present invention. Figure 1 As shown in the figure, the improved Transformer model CaTransGNN based on graph neural network includes four parts. First, the dependency relationship of multiple sensors in the car is constructed by the wireless sensor network. Second, spatiotemporal feature sampling is performed. Third, the spatial features are further enhanced through position encoding based on the shortest path. Finally, CaTransGNN is trained to achieve the purpose of abnormal detection of in-vehicle sensor data. Among them, CaTransGNN contains a three-layer structure. The first layer is the channel attention layer, which is used to suppress useless features and enhance useful features. The second layer is the GNN layer, which is used to capture rich spatial features. The third layer is the Transformer layer, which captures time information with its powerful time series processing capabilities.

[0047] 1. Transformer model design based on combining channel attention mechanism and GNN

[0048] The Transformer model design based on the combination of channel attention mechanism and graph neural network (GNN) proposed in this paper aims to improve the spatiotemporal feature extraction capability of intelligent connected vehicle data anomaly detection. The core idea is to combine the time series feature extraction capability of Transformer with the spatial feature extraction capability of graph neural network, and optimize feature fusion by channel attention mechanism (CAM), and finally achieve efficient and accurate anomaly detection. This method mainly includes the following steps:

[0049] 1) Data input

[0050] The in-vehicle sensor data includes various types, such as camera image frames, inertial measurement unit data, etc. These data are transmitted through the node V i and edge E ij The form is constructed as a graph structure G = (V, E), each node V i represents a sensor, each edge E ij Represents the data interaction between sensors.

[0051] 2) Temporal feature extraction and position encoding based on the shortest path

[0052] The improved Transformer model is used to extract time series features. First, the input sequence X = {x i (t)} is embedded, and the position encoding based on the shortest path retains the time information and introduces spatial features. In the traditional Transformer model, position encoding encodes the position of the input sequence through sine and cosine functions, but this has certain limitations in capturing graph structure information. In order to solve this problem, the present invention introduces a position encoding method based on the shortest path. Consider a sensor graph structure G = (V, E), where V is a node set (representing sensors) and E is an edge set (representing data interaction between sensors). The present invention calculates each pair of nodes (V i , V j ) between the shortest path distance d(V i , V j ). If node V i and V j If there is an edge between them, then d(V i , V j )=1; otherwise, d(V i , V j )=min(d(V i , V k )+d(V k , V j )), where V k is any intermediate node. Secondly, the present invention uses the shortest path distance to generate the position code. Assume that the shortest path distance matrix of graph G is represented by D, where D ij =d(V i , V j ). For each V i Generate positional encoding PEi:

[0053]

[0054] Where N is the total number of nodes, f is a mapping function, and f(d) = W d ·d+b d , where W d and b d is a trainable parameter.

[0055] Assume that the position encoding dimension is d model , encode the shortest path distance element by element to get the final position code PE i :

[0056]

[0057] Among them, 0∈[1,d model ] is the dimension index of the positional encoding, W oj and b oj is the linear transformation parameter on dimension o. Again, the positional encoding is added to the Transformer input, and the input data x at time slot t is i , and encode its position PE i Add to the input embedding layer to get the new input embedding layer H (0) :

[0058] H (0) =Embedding(x i )+PE i (3) In each layer of the multi-head self-attention mechanism, the self-attention score is calculated:

[0059] Q=H (l) W Q , K=H (l) W K , V=H( l) W V (4)

[0060]

[0061] For the multi-head self-attention mechanism, multiple independent attention heads are used:

[0062] MultiHead(Q,K,V)=Concat(head1,...,head h )W o (6)

[0063] head i =Attention(Q i , K i , V i ) (7)

[0064] The updated hidden state obtained is:

[0065] H (l+1) =LayerNorm(H (l) +MultiHead(Q,K,V)) (8)

[0066] By calculating layer by layer, we can get the time series feature T i .

[0067] 3) Spatial feature extraction

[0068] Initially each V i The characteristics are In GNN, node V i The spatial characteristics of can be calculated as follows:

[0069]

[0070] Where m is the normalization coefficient, Represents node V i The set of neighbor nodes, W (l) is the weight matrix of the lth layer, and σ is the activation function, such as ReLU.

[0071] After multiple layers of iterative updates, node V is extracted i The spatial characteristic is S i .

[0072] 4) Feature fusion and channel attention mechanism

[0073] The time series features T extracted by Transformer i And the spatial features S extracted by GNN i To perform the fusion:

[0074] Z i =Concat(T i , S i ) (10)

[0075] Feature fusion and dimensionality reduction through fully connected layers:

[0076] Z′ i =σ(W f Z i +b f ) (11)

[0077] The channel attention mechanism is introduced to optimize feature representation and calculate the channel attention weight:

[0078] A i =σ(W c Z′ i +b c ) (12)

[0079] The fused spatiotemporal features can be expressed as:

[0080]

[0081] 5) Model training and online monitoring

[0082] In the model training stage, the comprehensive loss function is used to optimize the model parameters. The loss function can be expressed as:

[0083]

[0084] in, is the temporal feature loss, is the spatial feature loss, is the fusion feature loss, αβ, γ are weight coefficients.

[0085] During the online detection phase, the trained model is deployed to receive sensor data x(t) in real time, perform inference and anomaly detection, and ensure the safety of vehicle driving and the timeliness of detection results.

[0086] The present invention firstly effectively extracts time series features through the Transformer model, then captures the spatial relationship between sensors based on GNN, and finally optimizes feature fusion by the channel attention mechanism, so that the model can realize efficient abnormal data detection in a complex Internet of Vehicles environment.

[0087] 2. Sensor node capture and iterative update based on Gumbel-Softmax

[0088] The present invention uses the Gumbel-Softmax strategy to capture the connection relationship between sensor nodes and perform iterative updates. This method overcomes the problem of non-differentiability in the traditional discrete data sampling process, allowing the present invention to be optimized directly in the training of deep neural networks. Since the sampling of discrete data from categorical distribution cannot be directly applied to the back propagation in deep neural networks, the Gumbel-Softmax distribution is introduced, which is a continuous distribution that can approximate categorical distribution samples. Figure 2 As shown, it is assumed that the present invention has three sensor nodes N1, N2, and N3, and their dependencies are not yet clear. The core idea of ​​the present invention is to use the Gumbel-Softmax sampling strategy to generate a random classification vector for determining whether a directed connection can be established between two nodes. For example, for nodes N1 and N2, if the probability value P between them is 1,2 is relatively high, which means that node N1 is likely to point to node N2 and vice versa.

[0089] The present invention considers a graph learning strategy with M nodes. For any nodes i and j, define a binary connection control variable z i,j Represents the connection relationship between nodes. Assume undirected probability And there is in Represents the probability that there is information flow from node i to node j in the graph.

[0090] According to the Gumbel-Softmax strategy, the connection strategy between two nodes is sampled from the following formula, where z i,j ∈{0, 1} 2 :

[0091]

[0092] Among them, g0 and g1 are independently and identically distributed samples from the standard Gumbel distribution, which can be obtained by the inverse transformation sampling method, that is, sampling from the uniform distribution u~Uniform(0,1) and calculating: g=-log(-logu).

[0093] Since the above argmax is not differentiable, the present invention uses the Gumbel-Softmax sampling method:

[0094]

[0095] Among them, τ is a parameter that controls the smoothness of the Gumbel-Softmax distribution. When τ approaches 0, the Gumbel-Softmax distribution becomes the same as the one-hot classification distribution.

[0096] Secondly, in each training iteration, the feature representation of the sensor node is updated according to the current connection relationship. For example, the feature representation of node i is Update according to the message passing mechanism of GNN:

[0097]

[0098] in, represents the set of neighbor nodes of node i, W(k) is the weight matrix of the kth layer, and σ is the activation function.

[0099] Since the introduction of Gumbel-Softmax sampling makes the sampling process of node connections differentiable, the model can be trained using a standard gradient descent optimization algorithm. This method not only significantly reduces the computational complexity, but also automatically learns the global topological structure of the entire graph, avoiding the limitations of node selection in traditional methods. In this way, the present invention can capture the relationship between sensor nodes in the vehicle in real time and efficiently, providing more accurate and comprehensive data support for anomaly detection.

[0100] 3. System Process

[0101] Figure 3 The execution flow chart of the Transformer based on collaborative GNN, the specific steps are as follows:

[0102] S301-304: initialization phase;

[0103] S301: sensor network construction;

[0104] S302: Graph structure construction;

[0105] S303: Initialize graph structure parameters;

[0106] S304: Initialize the data collection system and check whether the graph structure and node features are constructed. If so, execute S305. If not, execute S301-304 again.

[0107] S305-308: Data collection and preprocessing stage;

[0108] S305: Sensor data integration and storage;

[0109] S306: data preprocessing;

[0110] S307: Data annotation;

[0111] S308: Data cleaning and denoising;

[0112] S309-312: Sensor node capture and iterative update based on Gumbel-Softmax;

[0113] S309: Initialize Gumbel-Softmax parameters;

[0114] S310: Gumbel-Softmax sampling;

[0115] S311: connection relationship update;

[0116] S312: sensor feature update;

[0117] S313-316: Spatiotemporal feature extraction and fusion;

[0118] S313: spatial feature extraction (GNN);

[0119] S314: Temporal feature extraction (Transformer and CAM);

[0120] S315: Feature fusion and dimensionality reduction;

[0121] S316: Model training and optimization, judging whether the model has converged according to the loss function, if the model has converged, online monitoring is performed, if not, it is iterated again and S316 is executed;

[0122] S317-319: Anomaly detection and classification;

[0123] S317: Online anomaly detection. If there is abnormal data, execute S318. If there is no abnormality, continue to execute S317.

[0124] S318: Abnormal attribute analysis and in-vehicle control system adjustment;

[0125] S319: Continue to detect anomalies and the algorithm ends.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network, characterized by: The method comprises the following steps: S1: Based on the Graph Neural Network (GNN), a graph model of in-vehicle sensors is constructed. Each sensor is regarded as a node, and the data interaction relationship between sensors is used to form the edges of the graph, thereby establishing the spatial association relationship between sensors. S2: Collect historical data from in-vehicle sensors, perform data preprocessing and annotation, and generate training data sets, including camera image frames and inertial measurement unit data; S3: An improved Transformer model is used to extract the time series features of sensor data. By introducing a multi-head self-attention mechanism and position encoding based on the shortest path, the complex dependencies in the time series data are accurately captured. S4: The GNN model is used to extract the spatial features of sensor data. Based on the topological structure of the graph, the potential spatial correlation and dependency in the sensor data are captured through the message passing mechanism. S5: Perform spatiotemporal feature fusion and data dimension reduction through the fully connected layer, and adaptively adjust the importance weight of each channel in combination with the channel attention mechanism to optimize feature representation; S6: Use the training data set to train the model offline, and use the collected normal and abnormal data to perform online detection on the model to improve the detection accuracy; S7: Classify and analyze the detected abnormal data, including instantaneous abnormalities, constant abnormalities, gradual drift abnormalities and deviation abnormalities, and adopt different processing methods and early warning mechanisms to ensure vehicle driving safety; S8: Feedback the detection results to the in-vehicle control system in a timely manner to make corresponding safety strategy adjustments. If necessary, trigger an alarm or take preventive measures to ensure safe driving of the vehicle.

2. According to claim 1, the improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network is characterized in that: In S1, the construction of the sensor graph model adopts a graph structure model based on a graph neural network (GNN) to assist in the anomaly detection of intelligent connected vehicles. The in-vehicle control units (CUs) are responsible for building and maintaining the entire sensor graph model and allocating corresponding data processing tasks. Each sensor transmits and receives data through the in-vehicle control unit, forming a dynamic interactive topology structure that reflects the complex association relationship between sensors.

3. The improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network according to claim 1 is characterized in that: In S2, the data acquisition system collects historical data from the in-vehicle camera and inertial measurement unit (IMU) to form a training data set, and pre-processes and labels the data to improve the reliability and accuracy of subsequent model training.

4. According to claim 1, the improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network is characterized in that: In S3, an improved Transformer model is used to extract the time series features of sensor data. By introducing a multi-head self-attention mechanism and a position encoding based on the shortest path, the complex dependencies in the time series data are accurately captured, thereby improving the time series sensitivity of anomaly detection.

5. According to claim 1, the improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network is characterized in that: In S4, a GNN model is used to extract the spatial features of sensor data. Based on the topological structure of the graph and through a message passing mechanism, the potential spatial correlation and dependency in the sensor data are captured to improve the spatial feature extraction capability of anomaly detection.

6. The improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network according to claim 1 is characterized by: In S5, spatiotemporal feature fusion and dimensionality reduction are performed through a fully connected layer, and the importance weight of each channel is adaptively adjusted in combination with a channel attention mechanism to optimize feature representation and improve the overall performance of the model.

7. The improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network according to claim 1 is characterized by: In S6, the collected normal and abnormal data are used to perform offline training and online detection on the model to improve the detection accuracy. In the offline training stage, the preprocessed historical data set is used for training to optimize the objective function, and the feature extraction process is further optimized by combining the channel attention mechanism. In the online detection stage, the vehicle sensor data is received in real time, and the trained model is deployed for real-time inference and anomaly detection.

8. The improved Transformer intelligent connected vehicle anomaly detection method based on graph neural network according to claim 1 is characterized by: In S7, the detected abnormal data is classified and analyzed, including instantaneous abnormalities, constant abnormalities, gradual drift abnormalities and deviation abnormalities, and different processing methods and early warning mechanisms are used to ensure the safety of vehicle driving. The detection results are fed back to the in-vehicle control system in a timely manner to make corresponding safety strategy adjustments. If necessary, an alarm is triggered or preventive measures are taken to ensure the safe driving of the vehicle.

Citation Information

Cited By

  • Vehicle-road cloud integrated complex time sequence real-time anomaly detection method and system

    CN120354889A

  • New energy automobile pressure sensor production visual management system and method

    CN120540239A

  • Vehicle network attack detection method and device, vehicle and medium

    CN121239462A

  • Automatic driving risk deduction method based on anomaly detection and graph neural network

    CN121743709A