Vehicle sand and dust grade identification method and vehicle

Abstract

The invention discloses a vehicle dust grade identification method and a vehicle, and is applied to the technical field of vehicle environment perception. According to the vehicle sand dust grade identification method, data of four different sources including illumination intensity, wind speed, particulate matter concentration and vehicle environment information in the vehicle running process are obtained, three feature vectors of different dimensions are generated according to the internal relation and characteristics of the data, and the feature vectors of the three dimensions are combined for identification. And multi-dimensional information can be integrated, so that the dust grade identification result is more comprehensive and accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle environment perception, in particular to a vehicle sand-dust grade identification method and vehicle. BACKGROUND

[0002] With the rapid development of the automobile industry and the continuous improvement of people's requirements for driving safety and in-vehicle environmental comfort, the vehicle environment perception system has become a key component of modern vehicles. The system obtains real-time vehicle external environment information through various sensors, provides decision basis for vehicle protection strategies, driving assistance systems, etc., to ensure safe and stable operation of the vehicle under different environmental conditions. Under normal weather conditions, such systems can accurately identify roads, obstacles and weather conditions, however, in complex meteorological environments such as sand-dust weather, the identification accuracy of existing environment perception systems decreases significantly, the response delay is obvious, and false triggering or missed triggering problems are prone to occur, which seriously affects driving safety and system reliability. Therefore, how to improve the sand-dust grade identification accuracy has become a technical problem to be solved in the current vehicle environment perception technology field. SUMMARY

[0003] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a vehicle sand-dust grade identification method and vehicle, which fuse the light intensity, wind speed, particulate matter concentration and vehicle environment information in the vehicle running process to jointly identify the sand-dust grade and improve the accuracy of sand-dust grade identification.

[0004] According to a first aspect of the present application, a vehicle sand-dust grade identification method is provided, comprising: acquiring light intensity, wind speed, particulate matter concentration and vehicle environment information in the vehicle running process; generating a spatial feature vector based on the time relationship and spatial relationship of the collected light intensity, wind speed and particulate matter concentration; generating a time feature vector based on the correlation of changes of the wind speed and the particulate matter concentration; generating a probability distribution vector of the sand-dust grade based on the vehicle environment information; and generating a sand-dust grade identification result based on the spatial feature vector, the time feature vector and the probability distribution vector.

[0005] As a possible implementation manner, the generating of the spatial feature vector based on the time relationship and spatial relationship of the collected light intensity, wind speed and particulate matter concentration comprises: taking each sensor information for collecting the light intensity, wind speed and particulate matter concentration as a node; wherein the spatial relationship between the nodes constitutes edge feature information, and a group of continuous observation values of each node within a time window constitutes a time feature sequence; generating an attribute graph structure based on the edge feature information and the time feature sequence; and generating a spatial feature vector based on feature fusion and spatial modeling of the attribute graph structure by a graph convolution network.

[0006] In order to more effectively model the correlation between multiple source sensors, the system constructs a set of sensor representation methods based on graph structure after data collection is completed. By integrating the light intensity time series, wind speed time series and particulate matter concentration time series, the static correlation and structural relationship of light intensity, wind speed and particulate matter concentration at a certain time or specific state can be analyzed, and a time dimension expression is constructed. The definition of the edge considers both physical and statistical correlations. The physical connection is set according to the vehicle-mounted arrangement structure of each sensor. The edge connection formed by this spatial relationship can better reflect the structural relationship between the data in the spatial dimension and form a more three-dimensional data description. For statistical correlation, the correlation between any two nodes in the light intensity node, wind speed node and particulate matter concentration node is calculated, and the edge connection is established through the correlation. The edges designed in this way can dynamically adapt to the signal response law of different vehicles and different scenes. After the graph is constructed, each frame of sample data window will be organized into an attribute graph structure body, which will be directly input into the model of the graph convolution network for learning the latent space coupling mode between sensors. This graph modeling method is different from the traditional vector splicing method, which emphasizes the collaborative response characteristics of sensors and is a key pre-step for solving the problem of multi-source information fusion in complex dust environments. After the construction of the attribute graph structure body, the original multi-dimensional signal data is successfully converted into a graph data structure. Then, the graph convolution network can fully utilize the connection relationship between nodes and the edge weight information to realize the associated modeling of different sensor signals, thereby extracting the spatial joint feature expression required for dust grade recognition. Therefore, based on the graph convolution network for feature fusion and spatial modeling, the graph convolution network model is obtained. Through the graph convolution network model, the data completes the conversion process from the original graph structure to the high-order spatial fusion feature, and establishes the linkage mode expression between multiple sensors in the dust process, laying a solid foundation for more accurate grade recognition.

[0007] As a possible implementation manner, the spatial relationship between the nodes constitutes edge feature information, including: based on the air flow sequence of the sensor position for collecting the light intensity, the sensor position for collecting the wind speed and the sensor position for collecting the particulate matter concentration, a first edge connection is constituted; correlation degree calculation is performed on any two nodes, when the correlation degree between any two nodes is greater than a preset threshold, a second edge connection between the two nodes is established; based on the first edge connection and the second edge connection, an adjacency matrix is generated; wherein the edge feature information further includes an edge set and an edge weight, and based on the edge information set and the time feature sequence, an attribute graph structure body is generated, including: based on the edge set, the edge weight, the adjacency matrix and the time feature sequence, an attribute graph structure body is generated.

[0008] The edge definition considers both physical and statistical correlation. Physical connections are set according to the vehicle-mounted layout of each sensor. This spatial relationship-based edge connection can better reflect the structural relationships between data in different spatial dimensions, forming a more three-dimensional data characterization. Regarding statistical correlation, correlation is calculated between any two nodes among the light intensity, wind speed, and particulate matter concentration nodes. Edge connections are established based on these correlations. This edge design can dynamically adapt to the signal response patterns of different vehicles and scenarios. After edge construction, a complete adjacency matrix is ​​formed based on the first and second edge connections. This matrix serves as the structural input for subsequent graph convolutional networks, providing a three-dimensional structure for dust level identification and increasing parameter diversity.

[0009] As one possible implementation, feature fusion and spatial modeling of the attribute graph structure based on a graph convolutional network (GCNN) are performed to generate spatial feature vectors. This includes: inputting the attribute graph structure into the GCNN; aggregating edge feature information between adjacent nodes and adjusting the aggregation strength based on edge weights to perform node feature fusion, thereby extracting spatial features characterizing the physical laws of dust propagation; wherein the edge weights include the average correlation coefficient between two nodes corresponding to an edge in the historical training set; using a preset dust level label as a supervision signal and aiming to minimize the loss function between the prediction result and the preset dust level label, the GCNN is trained; when the training process of the GCNN model satisfies a first convergence condition, the training of the GCNN model is considered complete; wherein the output of the GCNN model is a global feature vector within each time window, and the global feature vector represents the joint state of multiple sensors within the time window.

[0010] The classification of dust storm levels (e.g., light, moderate, heavy) is typically based on a comprehensive threshold of multi-dimensional features (e.g., wind speed, particulate matter concentration, visibility), but the relationship between these features and the level is often non-linear. For example, after wind speed increases to a certain critical value, the increase in particulate matter concentration may grow exponentially, leading to a jump in dust storm level. Supervised learning, by directly optimizing the loss function between the predicted results and the true labels, can automatically learn this complex non-linear mapping relationship without manually designing threshold rules. Pre-set dust storm level labels can be formulated by meteorological experts based on long-term observations and standards, which inherently contain rich domain knowledge. Through label feedback, supervised learning can force the model to learn feature representations that conform to meteorological standards. For example, graph convolutional network models will prioritize features most sensitive to level classification while ignoring irrelevant features. Furthermore, in dust storm monitoring, graph convolutional network models aggregate features from adjacent sensors through adjacency matrices to capture spatial diffusion patterns, while supervised learning, through label feedback, can optimize the weights of the adjacency matrix or learn dynamic graph structures, making spatial modeling more realistic. For example, if a label indicates a region is experiencing severe dust storms, the graph convolutional network model will strengthen the association between this region and upstream high-concentration areas, while suppressing connections with irrelevant regions. Supervised learning, by minimizing the loss function, enables the feature representations learned by the model on the training set to generalize to unseen data. Therefore, training the graph convolutional network model with preset dust storm level labels as supervisory signals significantly improves the accuracy, robustness, and real-time performance of dust storm level identification by accurately capturing nonlinear relationships, utilizing domain knowledge, optimizing spatial modeling, and enhancing generalization ability. Compared to existing model training methods, training for dust storm level identification is more closely aligned with actual meteorological conditions, adaptable to complex and ever-changing dust storm scenarios, and provides more reliable data support for dust storm early warning.

[0011] As one possible implementation, generating a time feature vector based on the correlation between the changes in wind speed and particulate matter concentration includes: generating a wind speed change sequence and a particulate matter concentration change sequence based on the time series of wind speed and the time series of particulate matter concentration within the same time window; extracting the joint dynamic features of the wind speed change sequence and the particulate matter concentration change sequence within the time window based on a temporal convolutional network, and using the joint dynamic features as the time feature vector of the relationship between wind speed and particulate matter concentration.

[0012] Temporal feature vectors provide a stable, sensitive, and time-series-logical high-dimensional dynamic representation, which can assist in making level predictions when wind speed or particulate matter changes rapidly, and complement spatial feature vectors in both space and time.

[0013] As one possible implementation, the training method of the temporal convolutional network includes: using a preset dust level label corresponding to the same window length as a supervision signal, and using wind speed and particulate matter signals obtained from testing on dusty road sections as training samples, and training the temporal convolutional network based on the cross-entropy loss function; when the training process of the temporal convolutional network model meets the second convergence condition, it is determined that the training of the temporal convolutional network model is complete; wherein, the output of the temporal convolutional network model is a temporal feature vector.

[0014] The change in dust storm levels is a dynamic process with non-uniform time dependencies; for example, changes are slow in the early stages and accelerate abruptly in the later stages. Fixed-length windows may lead to information loss or noise interference if the window is too long or too short. However, training with labels corresponding to the same window length forces the model to learn the complete mapping relationship between wind speed, particulate matter concentration, and dust storm level within each time window, thus capturing the local dynamic patterns of dust storm level evolution. Using real wind speed and particulate matter signals as training samples, combined with supervision signals using cross-entropy loss, allows the temporal convolutional network model to learn noise-robust feature representations. For example, the model can use the loss function feedback to suppress transient peaks in particulate matter concentration caused by vehicle exhaust, while reinforcing the impact of continuous changes in wind speed and particulate matter concentration on the dust storm level.

[0015] As one possible implementation, the vehicle environment information includes geographical location, current time period, temperature and humidity, vehicle speed, and wind direction. The step of generating a probability distribution vector for the dust level based on the vehicle environment information includes: inputting the geographical location, current time period, temperature and humidity, vehicle speed, and wind direction into a classification model; the classification model outputs a probability distribution vector for the dust level corresponding to the current vehicle environment information. The classification model is trained by using a preset dust level label as the target variable, and using collected environmental context data and corresponding level label data as training samples to construct a decision tree. In each iteration, the number of trees, depth, learning rate, autonomous driving regularization coefficient, and category feature processing method are adjusted.

[0016] Besides the core sensors directly involved in dust monitoring, the environmental background information of the vehicle also plays a crucial role in dust level identification. Vehicle environmental information includes geographical location, current time period, temperature and humidity, vehicle speed, and wind direction, which have a complex nonlinear relationship with the probability and intensity of dust occurrence. Therefore, generating a probability distribution vector of dust levels by combining vehicle environmental information can provide environmental dimension data for fusion, taking into account the impact of environmental factors. The probability distribution vector of dust levels can be used as the output of the environmental dimension and fused with spatial and temporal feature vectors to complete the final dust level identification. During training, dynamically adjusting the number and depth of trees can improve the balance between model complexity and generalization ability. The learning rate controls the contribution weight of each tree to the final model, and regularization is used to prevent overfitting. Through autonomous driving regularization, the risk of overfitting can be dynamically suppressed. Therefore, introducing the learning rate and regularization coefficient can optimize the convergence and stability of ensemble learning. In the dust level classification task, environmental context data may contain a large number of category features. By iteratively adjusting the category feature processing method, the optimal encoding strategy can be selected, improving the utilization efficiency of non-numerical data. Finally, by adjusting parameters based on the performance of the validation set in each iteration, the system can dynamically adapt to the dynamic nature of the data, better match the complexity and dynamism of dust storm data, and provide reliable data support for dust storm early warning.

[0017] As one possible implementation, the dust level identification result includes a multidimensional probability distribution vector, which represents the confidence level of each dust level in the current time period; the vehicle dust level identification method further includes: selecting the maximum value of the confidence level of each dust level; using the dust level corresponding to the maximum value as the real-time dust level of the current time period; and executing a preset vehicle protection strategy based on the real-time dust level; wherein each dust level corresponds to at least one vehicle protection strategy, and the vehicle protection strategy includes air conditioning mode switching, air intake system adjustment, warning information release, and driving behavior intervention.

[0018] By establishing a mapping relationship between dust levels and specific protective actions, the abstract dust levels are transformed into a series of concrete and executable vehicle protection actions, providing a unified and clear action guide for different users. This offers an intelligent, standardized, and efficient vehicle dust protection solution, thereby improving vehicle driving safety.

[0019] As one possible implementation, the vehicle dust storm level identification method further includes: acquiring historical dust storm levels; comparing the difference between the real-time dust storm level and the historical dust storm level to check for any level abrupt changes; wherein, the level abrupt change indicates that the difference is greater than a preset difference; comparing the light intensity, wind speed, and particulate matter concentration in the current time period with the light intensity, wind speed, and particulate matter concentration in historical time periods to check for any parameter abrupt changes; wherein, the parameter abrupt change indicates that at least one of the light intensity, wind speed, and particulate matter concentration changes relative to historical time periods by a value greater than a preset change threshold; when the level abrupt change and the parameter abrupt change occur simultaneously, determining that the real-time dust storm level is a high-confidence outbreak level; wherein, based on the real-time dust storm level, executing a preset vehicle protection strategy includes: when the real-time dust storm level is a high-confidence outbreak level, executing the preset vehicle protection strategy based on the real-time dust storm level.

[0020] To reduce the possibility of false triggers or alarms caused by fluctuations in edge values, the trigger thresholds for each level can be dynamically adjusted based on factors such as the vehicle's historical perception status, the current operating environment, and the trend of level fluctuations. For example, the reliability of the real-time dust level can be jointly assessed by two factors: level abrupt changes and parameter abrupt changes. When the reliability is high, a preset vehicle protection strategy is executed to prevent unnecessary vehicle strategy switching caused by misjudgments or false triggers.

[0021] According to a second aspect of this application, a vehicle dust level identification device is provided, comprising: an acquisition module for acquiring light intensity, wind speed, particulate matter concentration, and vehicle environmental information during vehicle operation; a first generation module for generating a spatial feature vector based on the temporal and spatial relationships of the acquired light intensity, wind speed, and particulate matter concentration; a second generation module for generating a temporal feature vector based on the correlation between changes in wind speed and particulate matter concentration; a third generation module for generating a probability distribution vector of dust level based on the vehicle environmental information; and a fusion module for generating a dust level identification result based on the spatial feature vector, the temporal feature vector, and the probability distribution vector.

[0022] According to a third aspect of this application, a vehicle is provided, comprising: a sensor for acquiring the light intensity, a sensor for acquiring the wind speed, and a sensor for acquiring the particulate matter concentration; and a vehicle dust level identification device, which is used to perform the vehicle dust level identification method as described in the first aspect or any implementation thereof, wherein the vehicle dust level identification device is communicatively connected to the sensor for acquiring the light intensity, the sensor for acquiring the wind speed, and the sensor for acquiring the particulate matter concentration.

[0023] According to a fourth aspect of this application, a computer device is provided, the computer device comprising: one or more processors; a memory; and one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the processor to implement the method as described in the first aspect or any implementation thereof.

[0024] According to a fifth aspect of this application, a computer-readable storage medium is provided, the storage medium storing a computer program for performing the method as described in the first aspect or any implementation thereof.

[0025] According to a sixth aspect of this application, an electronic device is provided, including a module for performing the method as described in the first aspect or any implementation thereof.

[0026] According to a seventh aspect of this application, a computer program product is provided, comprising program code for performing the method as described in the first aspect or any implementation thereof.

[0027] The vehicle dust storm level identification method and vehicle provided in this application utilize data from four different sources during vehicle operation: light intensity, wind speed, particulate matter concentration, and vehicle environmental information. Different types of feature vectors are generated for each source, enabling multi-dimensional characterization of the data. Light intensity, wind speed, and particulate matter concentration are closely correlated spatially; spatial feature vectors generated from these data better capture spatial characteristics. For dust storm formation, wind speed and particulate matter concentration show a strong correlation in their changes; temporal feature vectors reflect the temporal variation patterns of wind speed and particulate matter concentration. Vehicle environmental information plays an auxiliary role in identifying dust storm levels; quantifying this information into probability distribution vectors provides an additional dimension of reference information. Generating feature vectors based on the inherent relationships and characteristics of the data avoids interference from irrelevant information, improves feature quality, and provides a more reliable foundation for subsequent dust storm level identification. Finally, the feature vectors from the three dimensions are combined for identification. These different dimensions are complementary. First, the time dimension provides dynamic information about the dust's changes over time, while the spatial dimension shows its spatial distribution. Combining these two allows for a more comprehensive understanding of the spatiotemporal evolution of dust. Second, the time dimension reflects the overall trend of dust changes, while the vehicle environment dimension provides real-time local dust conditions. By combining the time dimension, the overall dust situation within the vehicle's time period can be understood, and by combining it with the vehicle environment dimension, the local dust level can be corrected and refined. Third, the spatial dimension provides the global state of the collaborative relationships between sensors, while the vehicle environment dimension provides the probability of dust occurrence under current external conditions. Real-time data acquired by sensors can be compared with the vehicle environment to evaluate the dust level, thus correcting the accuracy of dust level identification in special circumstances. Therefore, fusing correlated multi-dimensional feature vectors allows for a comprehensive understanding of dust information from multiple aspects, more accurately grasping the essential characteristics of dust, and thus more precisely judging the dust level. Attached Figure Description

[0028] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0029] Figure 1 This is a flowchart illustrating a vehicle dust level identification method provided in an exemplary embodiment of this application.

[0030] Figure 2This is a schematic diagram of the structure of a vehicle dust level identification device provided in an exemplary embodiment of this application.

[0031] Figure 3 This is a structural diagram of an electronic device provided in an exemplary embodiment of this application. Detailed Implementation

[0032] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0033] As a key component of intelligent vehicles and autonomous driving technologies, vehicle environmental perception systems primarily rely on single or limited types of sensors, such as cameras, millimeter-wave radar, and lidar, to identify the state of the external environment. Under normal weather conditions, such systems can accurately identify roads, obstacles, and weather conditions, and trigger corresponding vehicle protection strategies accordingly, such as automatically closing windows, switching to recirculation mode, or adjusting the powertrain operating status.

[0034] However, in addition to regular weather conditions (such as rain and fog), vehicles may also encounter dust storms during operation. Dust storms are not an isolated, well-defined physical phenomenon, but a complex weather system resulting from the coupling of multiple factors. Sensor signals during dust storms are easily confused with other weather or objects; for example, cameras and lidar struggle to distinguish between dust particles, water droplets (fog), pollutants (haze), or combustion particles (smoke) in the air. All of these factors contribute to reduced visibility, manifesting as blurred and distorted images. Dense dust clouds may be mistaken for stationary obstacles (such as walls or mountains) in radar echoes, causing the system to brake incorrectly. The concentration, distribution, and speed of dust movement are driven by wind speed and can change drastically within minutes. A light dust storm area can become a heavily dust storm area in a very short distance. This non-linear, abrupt change makes prediction based on historical data extremely difficult.

[0035] Existing perception systems relying on single or simple fusion algorithms struggle to cope with the complexities of sandstorm weather. Several reasons contribute to this: First, the inherent limitations of single sensors. For example, cameras heavily rely on lighting and visibility, rendering them largely ineffective in sandstorms. LiDAR is sensitive to aerosols, but the relationship between the returned point cloud density and intensity information and sandstorm concentration is complex, making it difficult to directly and accurately quantify sandstorm levels. Millimeter-wave radar cannot directly perceive sandstorms. Second, the challenge of multi-sensor fusion. In sandstorm weather, different sensors may produce contradictory perceptions of the same environment. For instance, a camera might fail to detect obstacles due to low visibility, while radar might detect non-existent obstacles due to sandstorm clutter. Reliable fusion under such highly uncertain conditions presents a significant challenge. Third, multi-source heterogeneous sensors can acquire information about the vehicle's external environment from different angles and levels. However, most existing systems fail to fully utilize the correlations between these sensors, simply processing and analyzing data from each sensor independently. This difficulty in data fusion further complicates sandstorm level identification. Fourth, dust storms vary significantly across different regions. For example, dust particles in desert areas are typically larger and more concentrated, accompanied by higher wind speeds; while dust storms around cities may be influenced by industrial emissions and traffic dust, resulting in a more complex particle composition and concentration distribution. Furthermore, real-time environmental conditions are constantly changing; even within the same region, dust storm characteristics may differ at different times. However, the judgment logic of existing vehicle environmental perception systems is mostly fixed and cannot be flexibly adjusted according to regional environmental differences and real-time environmental conditions. This lack of adaptive judgment logic makes it difficult for the system to achieve optimal perception performance in different environments, failing to meet the protection requirements of vehicles in various regions and real-time environments.

[0036] Existing systems lack the ability to jointly model dynamic changes in factors such as wind speed, light intensity, and particulate matter concentration. Dust storms are complex dynamic systems where various environmental factors interact and influence each other. Existing systems fail to establish effective dynamic models to describe the relationships between these factors, resulting in an inability to accurately predict environmental trends when facing dynamic changes in dust storms. To address this issue, this application provides a vehicle dust storm level identification method, device, and vehicle. By fusing Graph Convolutional Networks (GCN), Temporal Convolutional Networks (TCN), and CatBoost (Categorical Boosting) models, it jointly models the spatial dependence, temporal trend, and environmental semantics of multi-source sensor data to jointly identify dust storm levels, achieving dynamic and high-precision determination of dust storm levels.

[0037] To address the aforementioned problems in identifying sand and dust, Figure 1 This is a schematic flowchart of a vehicle dust level identification method provided in an exemplary embodiment of this application. Figure 1 For example, firstly, acquire information on light intensity, wind speed, particulate matter concentration, and vehicle environmental conditions during vehicle operation (see...). Figure 1 (S110). Light intensity, wind speed, particulate matter concentration, and vehicle environmental information are all key factors affecting dust storm levels, and these environmental factors such as wind speed, light intensity, and particulate matter concentration interact with each other. For example, changes in wind speed affect the diffusion and concentration distribution of dust particles, while changes in light intensity may affect the sensor's accuracy in detecting particulate matter concentration. First, these influencing factors of dust storm levels are extracted to provide a data foundation for subsequent dust storm level identification. Second, based on the temporal and spatial relationships of collected light intensity, wind speed, and particulate matter concentration, a spatial feature vector is generated (see S110). Figure 1 (S120). The spatial feature vector reflects the static correlation and structural relationship between light intensity, wind speed, and particulate matter concentration at a certain moment or under specific conditions. Next, based on the correlation between changes in wind speed and particulate matter concentration, a temporal feature vector is generated (see S120). Figure 1 (S130). The time feature vector reflects the dynamic patterns and trends of wind speed and particulate matter concentration over time, capturing the short-term fluctuations and long-term trends of wind speed and particulate matter concentration, further enriching the dynamic information required for dust storm level identification. Then, based on vehicle environmental information, a probability distribution vector for dust storm levels is generated (see S130). Figure 1 (S140). Besides the core sensors directly involved in dust monitoring, the environmental background information of the vehicle also influences dust level identification. Exploring the potential relationship between vehicle environmental information and dust level, and generating supplementary probability distribution vectors, can enable the system to achieve more comprehensive judgment and response under complex conditions. Finally, based on spatial feature vectors, temporal feature vectors, and probability distribution vectors, dust level identification results are generated (see S140). Figure 1 (S150). The three types of vectors are fused to generate the final dust level determination result, taking into account the changes of various factors to improve the robustness of dust level identification.

[0038] The following text combines Figure 1 This paper provides a more detailed description of the vehicle dust level identification method provided in the embodiments of this application.

[0039] In S110, information such as light intensity, wind speed, particulate matter concentration, and vehicle environment during vehicle operation is acquired.

[0040] In some embodiments, vehicles are exposed to complex natural environments during actual operation, especially in arid and semi-arid regions, such as the Northwest region where dust storms are frequent. Therefore, constructing a multi-source sensing system is a prerequisite for dust storm level identification. Multiple heterogeneous sensors can be deployed on the vehicle to collect key environmental information related to dust storms. This key environmental information may include light intensity, wind speed, particulate matter concentration, and vehicle environmental information. Changes in light intensity affect the sensor's accuracy in detecting particulate matter concentration; under strong or weak light conditions, the sensor may not accurately reflect the true concentration of dust particles. Wind speed is one of the key factors affecting dust storm levels; sudden increases or decreases in wind speed can cause changes in the diffusion and concentration distribution of dust particles, thus affecting dust storm level identification. The size, shape, and concentration distribution of particulate matter are uneven and significantly affected by factors such as wind speed and light intensity. Environmental factors such as wind speed, light intensity, and particulate concentration interact and influence each other. For example, changes in wind speed affect the diffusion and concentration distribution of dust particles, while changes in light intensity may affect the sensor's accuracy in detecting particulate matter concentration. This interaction makes dust storm level identification more complex.

[0041] To mitigate the negative impacts of interactions and enhance the correlation between environmental factors in a dust storm environment, key environmental information is first collected for subsequent identification. Deployed sensors may include ambient light sensors (sensors that collect light intensity), wind speed sensors (sensors that collect wind speed), and air particulate matter sensors (sensors that collect particulate matter concentration). Ambient light sensors detect changes in light intensity caused by dust obstruction, wind speed sensors capture sudden wind disturbances, and particulate matter sensors collect PM2.5 and PM10 particle concentrations in the outside air to reflect the actual physical concentration of dust.

[0042] As one possible implementation, the system can acquire various sensor signals using high-frequency sampling. For example, the sampling frequency for light and wind speed signals is 10Hz to accurately capture the rapidly changing process of sand and dust entering the system. Particulate matter sensors operate at 1Hz to reflect slower trends. All raw signals are transmitted back to the environmental perception system in real time via the vehicle's Ethernet or CAN bus. A data fusion module is set up in the environmental perception system to synchronously collect timestamps and perform time alignment between various data sources. For differences in sampling frequencies between different sensors, a resampling mechanism can be used to unify them to a 5Hz time base, ensuring a consistent timing structure when subsequently input into the neural network model.

[0043] In some embodiments, in addition to the three core sensor types (ambient light sensor, wind speed sensor, and air particulate matter sensor), vehicle environmental information can also be collected. This vehicle environmental information includes GPS location information, vehicle speed, current altitude, and external temperature and humidity parameters. This vehicle environmental information will be provided as supplementary features for subsequent identification and analysis to reflect the geographical conditions and environmental factors that cause dust storms. For example, in some high-altitude areas, wind speeds may be the same, but particulate matter concentrations may vary significantly; such differences can be compensated for through geographical and meteorological data modeling.

[0044] In some embodiments, data such as light intensity, wind speed, particulate matter concentration, and vehicle environmental information during vehicle operation each have unique characteristics. To unify data representation and establish a foundation for data analysis, all data (light intensity, wind speed, particulate matter concentration, and vehicle environmental information during vehicle operation) undergo preliminary preprocessing through an edge processing unit after acquisition. For example, the system sets up a missing value handling mechanism, using a sliding window average for interpolation when a sensor experiences a brief failure. Furthermore, outliers are removed to ensure the stability and reliability of the signals relied upon for subsequent modeling. A data synchronization module is set up in the environmental perception system to perform data standardization. After data standardization, sensor data over a period of time is organized into a sample window, typically 30 time steps long, corresponding to a 6-second environmental data sequence.

[0045] One possible implementation involves incorporating a data acquisition module into the environmental perception system. This module acquires data on light intensity, wind speed, particulate matter concentration, and the vehicle's environmental conditions during operation. After preprocessing the data, it outputs a structured time-series data packet in a standardized format. Each packet contains 30 sampling points, with each frame including light intensity, wind speed, particulate matter concentration, and corresponding environmental state variables (data included in the vehicle's environmental information). These structured time-series data packets serve as input for subsequent graph modeling and temporal modeling. This highly synchronized and standardized data acquisition system ensures a stable and consistent input source for subsequent algorithm models, thereby improving recognition accuracy and system robustness.

[0046] In S120, a spatial feature vector is generated based on the temporal and spatial relationships of collected light intensity, wind speed, and particulate matter concentration.

[0047] To more effectively model the correlations between multiple sensors, the system constructs a graph-based sensor representation after data acquisition. This representation generates an attribute graph structure based on the temporal and spatial relationships of collected light intensity, wind speed, and particulate matter concentration. The attribute graph structure includes a set of nodes, edges, an adjacency matrix, and a set of time-series features. This structure not only reflects the physical connections and spatial arrangement of the sensors but also incorporates the signal coupling characteristics of multiple sensors within a time window.

[0048] In some embodiments, the process of constructing the attribute graph structure begins with node definition, using each sensor signal that collects light intensity, wind speed, and particulate matter concentration as a node; that is, defining each valid sensor signal source as a node. The spatial relationships between nodes constitute edge feature information, and a set of continuous observations within a time window for each node forms a time feature sequence. For example, when the sensors are ambient light sensors, wind speed sensors, and air particulate matter sensors, the time nodes for collecting light intensity, wind speed, and particulate matter concentration are designated as light intensity nodes, wind speed nodes, and particulate matter concentration nodes. From a temporal perspective, the temporal relationship for collecting light intensity, wind speed, and particulate matter concentration includes the time nodes for collecting light intensity, wind speed, and particulate matter concentration. Node features are the continuous observations of the sensor within the sampling window, forming a one-dimensional time series that represents the trajectory of the node's change within the window. Therefore, based on the light intensity node, wind speed node, and particulate matter concentration node, a light intensity time series, a wind speed time series, and a particulate matter concentration time series can be constructed. By integrating time series data of light intensity, wind speed, and particulate matter concentration, we can analyze the static correlations and structural relationships among these three parameters at a given moment or under specific conditions, thus constructing a time-dimensional representation.

[0049] In some embodiments, the definition of edges takes into account both physical and statistical correlation.

[0050] From a physical perspective, the locations of sensors collecting light intensity, wind speed, and particulate matter concentration have a physical relationship and order. When the vehicle is running, the airflow passes through these sensors in a specific order. Therefore, the spatial relationship between light intensity, wind speed, and particulate matter concentration can be set as the airflow sequence of the sensor locations collecting light intensity, wind speed, and particulate matter concentration. In other words, the physical connection is set according to the vehicle's onboard layout of each sensor. For example, light and wind speed sensors are often installed at the front of the vehicle, while particulate matter sensors are located in the air intake or the middle of the front compartment. There is a physical order of airflow based on the sensor locations collecting light intensity, wind speed, and particulate matter concentration. This sequential relationship forms a boundary connection (considered the first boundary connection). This spatial boundary connection better reflects the structural relationships between data in different spatial dimensions, creating a more comprehensive data profile.

[0051] Regarding statistical correlation, the system calculates the correlation between any two nodes among the light intensity, wind speed, and particulate matter concentration nodes. When the correlation between any two nodes exceeds a preset threshold, a second edge connection is established between them. For example, in the training dataset, correlation calculations are performed between different nodes, and when the Pearson correlation coefficient or mutual information value exceeds a preset threshold, the system automatically establishes an edge connection. This edge design allows for dynamic adaptation to the signal response patterns of different vehicles and scenarios.

[0052] In some embodiments, after the edges are constructed, a complete adjacency matrix is ​​formed based on the first edge connection and the second edge connection, and then normalized as the structural input in the GCN (Graph Convolutional Network).

[0053] In some embodiments, a weight can be assigned to each edge (either in the first or second edge connection), where the weight is the average correlation coefficient between the two nodes corresponding to that edge in the historical training set. By setting edge weights, the importance of edges can be differentiated in subsequent graph convolution calculations. This graph structure allows edge weights to be updated according to changes in vehicle type and geographical region, making the entire graph more adaptable.

[0054] In some embodiments, to accommodate the potential for missing or faulty vehicle sensors in real-world deployments, the system allows for the construction of a partial graph during the graph building phase. This means that even if some nodes are missing, the graph structure remains valid, only the number of nodes is reduced. During training, the model enhances robustness by randomly discarding some nodes, thus achieving fault tolerance for sensor failures.

[0055] As one possible implementation, an attribute graph structure is generated based on edge feature information and time feature sequence.

[0056] In some embodiments, the edge feature information also includes edge sets and edge weights. Based on the edge sets, edge weights, adjacency matrices, and temporal feature sequences, an attribute graph structure is generated. That is, after graph construction, each frame of sample data window will be organized into an attribute graph structure, including a node set, edge set, adjacency matrix, edge weight tensor, and temporal feature set. The attribute graph structure will be directly input into the GCN model to learn the potential spatial coupling patterns between sensors. This graph modeling approach differs from traditional vector concatenation, emphasizing the cooperative response characteristics of sensors and serving as a crucial preliminary step in addressing the multi-source information fusion problem in complex dust environments. The completion of the attribute graph structure construction indicates that the system has successfully converted the original multidimensional signal data into a graph data structure, laying a solid data foundation for the next stage of graph convolutional modeling.

[0057] In some embodiments, based on the constructed attribute graph structure, feature fusion and spatial modeling are performed on the attribute graph structure using a graph convolutional network (GCN) to generate spatial feature vectors. For example, the attribute graph structure is input into a GCN, and features are aggregated based on edge feature information between adjacent nodes. The aggregation strength is adjusted by combining edge weights to perform node feature fusion, thereby extracting spatial features characterizing the physical laws of dust propagation. The edge weights include the average correlation coefficient between two nodes corresponding to an edge in the historical training set. In other words, the system introduces a graph convolutional network to mine the potential spatial dependencies between multi-source sensors. The GCN can fully utilize the connection relationships and edge weight information between nodes to achieve correlation modeling of signals from different sensors, thereby extracting the spatial joint feature representation required for dust level identification. Therefore, feature fusion and spatial modeling are performed based on a graph convolutional network to obtain a GCN model. The original sensor graph data is encoded into a structured high-dimensional representation using the GCN model to generate spatial feature vectors for subsequent fusion models.

[0058] As one possible implementation, the input to the GCN model is the attribute graph structure output from the previous stage, containing time-series features of multiple nodes, edge sets, and a normalized adjacency matrix. Each node has a set of continuous observations within a time window, forming a one-dimensional feature sequence. The system first converts this into a vector representation of uniform dimension through an embedding layer to ensure that subsequent graph convolutional layers can perform operations with consistent dimensionality. The embedded features are then input into the graph convolutional units, where aggregation calculations are performed using information between adjacent nodes. The core of graph convolution lies in propagating the state information of adjacent nodes through each edge, while adjusting the propagation strength based on edge weights, thereby achieving layer-by-layer feature fusion in each round of propagation.

[0059] For example, a two-layer graph convolutional network is designed, with each layer followed by a ReLU activation function and a normalization module to enhance nonlinear modeling capabilities and improve training stability. The first layer maps the original features of nodes to an intermediate space, extracting locally correlated features; the second layer further performs high-order feature fusion, enhancing cross-sensor information exchange capabilities. In each graph convolutional computation, the updated state of a node is influenced not only by its own input but also by the aggregated features of all neighboring nodes. This structure can effectively reflect the pre-driving relationship between changes in illumination and changes in particle concentration, as well as the spatial physical laws such as the accelerating effect of wind speed on the rapid propagation of dust.

[0060] As a possible implementation, when modeling the correlation between signals from different sensors, the GCN model can be trained using supervised learning. Using preset dust level labels as supervisory signals, the model is trained on a graph convolutional network with the objective of minimizing the loss function between the predicted result and the preset dust level labels. The preset dust level labels are sourced from statistics on particulate matter deposition on vehicle air filters, inference results from blurred camera images, and data validation using an external dust weather database. The graph convolutional network model is considered complete when the training process meets the first convergence condition. The output of the graph convolutional network model is a global feature vector for each time window, representing the joint state of multiple sensors within that time window. In other words, preset dust level label data collected in the vehicle operating environment is used as the training target. The label sources include statistics on particulate matter deposition on vehicle air filters, inference results from blurred camera images, and data validation using an external dust weather database, forming a reliable multi-source label system. The GCN model is based on the cross-entropy loss function, combined with KL divergence as a regularization term to suppress overfitting, and trained via backpropagation using the Adam optimizer. During training, the system employs data augmentation mechanisms, randomly perturbing edge weight coefficients and deleting some edge connections to enhance the robustness of the GCN model to graph structure variations. Furthermore, the training set includes data samples from multiple different geographical regions and seasons to ensure the GCN model has cross-regional generalization capabilities.

[0061] The classification of dust storm levels (e.g., light, moderate, heavy) is typically based on a comprehensive threshold of multi-dimensional features (e.g., wind speed, particulate matter concentration, visibility), but the relationship between these features and the level is often non-linear. For example, after wind speed increases to a certain critical value, the increase in particulate matter concentration may grow exponentially, leading to a jump in dust storm level. Supervised learning, by directly optimizing the loss function between the predicted results and the true labels, can automatically learn this complex non-linear mapping relationship without manually designing threshold rules. Pre-set dust storm level labels are derived from the collation and summarization of historical dust storm data, which may include dust storm level standards formulated by experts based on experience. Through label feedback, supervised learning can force the GCN model to learn feature representations that conform to the dust storm level standards. For example, the GCN model will prioritize features most sensitive to level classification while ignoring irrelevant features. Furthermore, in dust storm monitoring, GCN aggregates features from adjacent sensors through an adjacency matrix to capture spatial diffusion patterns, while supervised learning, through feedback from pre-set dust storm level labels, can optimize the weights of the adjacency matrix or learn dynamic graph structures, making spatial modeling more realistic. For example, if a pre-defined dust storm level label indicates that a certain area is experiencing severe dust storms, the Gaussian Cloud Computing (GCN) will strengthen the association between this area and upstream high-concentration areas, while suppressing connections with irrelevant areas. Supervised learning, by minimizing the loss function, enables the feature representations learned by the GCN model on the training set to generalize to unseen data. Therefore, training the GCN with pre-defined dust storm level labels as supervisory signals significantly improves the accuracy, robustness, and real-time performance of dust storm level identification by accurately capturing nonlinear relationships, utilizing domain knowledge, optimizing spatial modeling, and enhancing generalization ability. Compared to existing model training methods, training for dust storm level identification is more closely aligned with actual meteorological conditions, adaptable to complex and ever-changing dust storm scenarios, and provides more reliable data support for dust storm early warning.

[0062] To further enhance the efficiency and adaptability of the GCN model, the vehicle environment perception system can also build the capability for parameter fine-tuning based on vehicle platform characteristics (such as sensor model, installation location, and main control chip performance) within the vehicle manufacturer's system. Before deployment, the GCN model can be fine-tuned with small samples for the target vehicle model. Through transfer learning, the main model parameters are converged and adjusted on a small sample set, thereby improving the final recognition accuracy. The final output of the GCN model is the global feature vector (i.e., spatial feature vector) of the entire image within each time window, which represents the joint state of multiple sensors within that time period. This vector will be passed to subsequent time-series modeling for further processing. Through the GCN model, the data completes the transformation from the original graph structure to high-order spatial fusion features, establishing an expression of the linkage mode between multiple sensors in the sandstorm process, laying a solid foundation for achieving more accurate level recognition.

[0063] The introduction of the GCN model specifically for processing graph-structured data enables efficient capture of spatial relationships between nodes (such as sensors). Based on the characteristics of dust storms, their spread is influenced by factors such as terrain, wind direction, and obstacles. Sensor data from different locations exhibit spatial correlations. GCN aggregates information from adjacent nodes through an adjacency matrix, learning the spatial propagation patterns of dust. For example, sensor data near a dust source influences the feature representations of downstream sensors through the graph structure, thus more accurately reflecting the extent of dust spread. GCN's neighborhood aggregation mechanism ensures that each node not only contains its own features but also incorporates information from surrounding nodes, simultaneously capturing both local details and global trends. For instance, in the edge region of a dust storm, GCN can infer the dust level of the current area based on features from surrounding high-concentration areas. Therefore, the GCN model can provide spatial features, improving the correlation between local and global data.

[0064] In S130, a time feature vector is generated based on the correlation between changes in wind speed and particulate matter concentration.

[0065] In some embodiments, the trends in wind speed and particulate matter concentration typically exhibit a strong temporal correlation in dust storm level identification. Especially as a dust storm approaches, wind speed increases rapidly within a short period, followed by a rapid increase in particulate matter concentration. This temporal characteristic with a causal relationship is a crucial basis for determining dust storm intensity. Therefore, the system introduces a Temporal Convolutional Network (TCN) to model wind speed and particulate matter data, capturing short-term fluctuations and long-term trends to further enrich the dynamic information required for dust storm level identification. The established TCN model can output a temporal feature vector based on the correlation between wind speed and particulate matter concentration changes.

[0066] One possible implementation involves generating wind speed change sequences and particulate matter concentration change sequences based on the time series of wind speed and particulate matter concentration within the same time window. Specifically, the input data for the TCN model consists of time series from wind speed and air particle sensors within a fixed window length. For example, each time series is 30 time steps long, corresponding to a sampling duration of 6 seconds and a sampling frequency of 5Hz. To avoid training bias caused by different units of measurement, wind speed and particulate matter concentration are preprocessed and normalized to ensure they are distributed within a uniform numerical range. Each sample data point contains two dimensions: a wind speed change sequence and a particulate matter concentration change sequence. These two sets of sequences are input to the TCN model in a dual-channel manner. The TCN model extracts the joint dynamic features of the wind speed change sequence and particulate matter concentration change sequence within the time window, using these joint dynamic features as the time feature vectors for wind speed and particulate matter concentration.

[0067] For the TCN model, the temporal convolutional network structure employs a multi-level dilated convolution approach. Each TCN layer uses a one-dimensional causal convolution, ensuring that the TCN model relies only on historical information rather than future values ​​at each time point, thus aligning with the temporal characteristics of actual vehicle inference. To enhance the receptive field (the region of the input image visible from a point on the feature map) and improve the modeling ability for long-term dependencies, an incremental dilation factor design is introduced, allowing features spanning a wider range of time spans to be extracted. For example, the first layer uses a dilation factor of 1, the second layer 2, the third layer 4, and so on, doubling with each layer. This allows for the coverage of longer temporal dependencies without significantly increasing the network depth.

[0068] Furthermore, in some embodiments, to prevent gradient explosion or vanishing, the system can add residual connections after each convolutional layer and use the ReLU activation function. The residual structure improves the model's stability during deep training, especially suitable for signals with frequent fluctuations, such as particulate matter concentration. The output features, after being stacked through multiple convolutional layers, are then processed by global average pooling to obtain a fixed-length vector representing the joint dynamic features of wind speed and particulate matter within that time window.

[0069] As one possible implementation, the TCN model can be trained using a supervised approach, with the objective consistent with the GCN model: classification based on a pre-defined dust level label corresponding to each sample window. The training method for the temporal convolutional network (TCN) includes using the pre-defined dust level labels corresponding to the same window length as the supervision signal, and wind speed and particulate matter signals obtained from tests on dusty road sections as training samples. The TCN is trained based on a cross-entropy loss function. The training is considered complete when the second convergence condition is met. The output of the TCN is a temporal feature vector. The loss function uses cross-entropy, and class weight coefficients are introduced to prevent the model from biasing towards predicting low-level dust. The training data comes from wind speed and particulate matter signals measured by automakers on typical dusty road sections, including data from normal weather, light, moderate, and severe dust storms. The system specifically constructs samples containing extreme events such as sudden wind speed changes and delayed particulate matter rise to enhance the model's response to actual sudden weather events. The optimizer uses Adam (Adaptive Moment Estimation) with an initial learning rate of 0.001. During training, a dynamic learning rate decay strategy is employed, automatically adjusting the learning rate based on the validation set accuracy. To prevent overfitting, Dropout (a training mechanism in deep learning to combat overfitting) is introduced with a dropout rate of 0.3, primarily applied to intermediate convolutional layers. The model training cycle is limited to under 100 epochs, employing an early stopping mechanism: training terminates if the validation set accuracy fails to improve for five consecutive epochs to conserve computational resources.

[0070] The change in dust storm levels is a dynamic process with non-uniform temporal dependencies; for example, changes are slow in the early stages and accelerate abruptly in the later stages. Fixed-length windows may lead to information loss or noise interference due to excessively long or short windows. However, training with labels corresponding to the same window length forces the model to learn the complete mapping relationship between wind speed, particulate matter concentration, and dust storm level within each time window, capturing the local dynamic patterns of dust storm level evolution. The TCN model expands its receptive field through dilated convolution, enabling it to cover long time spans exponentially. The cross-entropy loss function penalizes the difference between predicted and true labels, forcing the TCN model to learn the most relevant feature combination for dust storm level at each time step. For example, during a prolonged dust storm, the TCN model can simultaneously capture the impact of short-term wind speed fluctuations and long-term particulate matter concentration accumulation on the level through dilated convolution, while the cross-entropy loss optimizes the convolution kernel weights, making the TCN model more focused on temporal patterns sensitive to level classification, thereby improving the long-term modeling capability of the TCN model. Using real wind speed and particulate matter signals as training samples, combined with the supervision signal of cross-entropy loss, the temporal convolutional network model can learn noise-robust feature representations. For example, the model uses a loss function feedback to suppress transient peaks in particulate matter concentration caused by vehicle exhaust, while amplifying the impact of sustained wind speed and the coordinated changes in particulate matter concentration on the severity level. Furthermore, the TCN model can optimize the joint modeling of temporal and spatial features. If a dusty road section includes multiple sensors, the TCN can learn joint spatial-temporal patterns through multivariate time-series inputs. Cross-entropy loss optimizes the time window length, allowing the model to simultaneously consider spatial correlations and temporal evolution within each window. This joint modeling capability significantly outperforms existing methods that rely solely on either spatial or temporal dimensions.

[0071] The TCN model efficiently processes time-series data through dilated convolution and causal convolution, capturing temporal features and addressing long-distance dependencies. Since changes in dust storm levels are long-term processes (e.g., a dust storm can last for several hours from formation to dissipation), TCN's dilated convolution expands the receptive field, enabling it to capture dependencies over extended periods. For example, TCN can learn complete temporal patterns of continuously increasing wind speed, rapidly rising particulate matter concentration, and sudden drops in visibility without stacking multiple layers of RNNs (Recurrent Neural Networks). Furthermore, in real-time dust storm monitoring scenarios, the TCN model can quickly process time-series data from multiple sensors, demonstrating high efficiency. Moreover, TCN mitigates information loss issues in deep networks through residual connections, ensuring that long-term dependencies are not weakened. For instance, during a dust storm, TCN can continuously track trends in wind speed and particulate matter concentration, avoiding information attenuation caused by network depth.

[0072] In some embodiments, at the automotive engineering implementation level, the TCN model is deployed in the vehicle's edge computing unit, enabling rapid level prediction when wind speed or particulate matter changes quickly, thus complementing the GCN model in both space and time. The Temporal Convolutional Network completes the modeling task of the dynamic behavior of wind speed and particulate matter (TCN model), providing a stable, sensitive, and temporally logical high-dimensional dynamic representation, which is an important supporting part of the system's overall recognition capability.

[0073] In S140, a probability distribution vector of the dust level is generated based on vehicle environmental information.

[0074] In some embodiments, besides the core sensors directly involved in dust monitoring, the environmental background information of the vehicle also plays a crucial role in dust level identification. Vehicle environmental information includes geographical location, current time period, temperature and humidity, vehicle speed, and wind direction, which have complex nonlinear relationships with the probability and intensity of dust occurrence. For example, in a specific time period or area, increased wind speed makes dust propagation more likely; while under high humidity conditions, the activity of dust particles is suppressed. Therefore, a classification model is introduced to fully explore the potential relationship between these environmental context variables and dust levels. Geographical location, current time period, temperature and humidity, vehicle speed, and wind direction are input into the classification model. The model outputs a probability distribution vector of the dust level corresponding to the current vehicle environmental information. This probability distribution vector is used as a supplementary discrimination result. The classification model uses a preset dust level label as the target variable and the collected environmental context data and corresponding level labeled data as training samples to construct a decision tree. In each iteration, the number of trees, depth, learning rate, autonomous driving regularization coefficient, and category feature processing method are adjusted.

[0075] For classification models, the CatBoost model can be used. For example, when the CatBoost model is applied to vehicle environmental information processing, the input of the CatBoost model includes a set of structured environmental features, which mainly include the following categories: geographic coordinates (latitude and longitude), current hour (0 to 23), whether it is daytime (calculated by sunrise and sunset times), current external temperature and humidity, vehicle speed, wind direction (such as northeast, southwest, etc.), and historical dust records from weather stations. All structured environmental features are acquired in real time through onboard sensors and the navigation system, and the latest set of feature data is extracted within each recognition window to form a complete environmental description vector.

[0076] Because the input features of the CatBoost model include a large number of categorical fields, such as wind direction and whether it is daytime, unlike traditional tree models, CatBoost employs a unique target encoding method when handling categorical variables. Through sequential processing and information-free encoding, it can improve accuracy while preserving the model's generalization ability. Furthermore, CatBoost possesses a natural feature cross-modeling capability, automatically learning higher-order combinations of variables. For example, in the combination of "afternoon + northwest wind + high temperature," complex relationships such as moderate dust storms are more likely to emerge, improving overall judgment ability.

[0077] As one possible implementation, the CatBoost model employs a multi-class classification task for training. The target variable is a predefined dust storm level label defined by the system, divided into five levels, from light to heavy, representing different degrees of environmental pollution and visual occlusion. During training, the system uses environmental context data and corresponding level-labeled data collected from real vehicles by automakers, covering multiple typical dust storm weather areas and multiple seasonal scenarios to ensure broad representativeness of the training samples. The CatBoost model automatically constructs thousands of decision trees during training and optimizes the overall loss function in each iteration to improve classification performance. To address the issue of uneven distribution of sample classes, the CatBoost model can introduce class balancing weights to prevent predictions from favoring samples with lighter class levels.

[0078] Understandably, the five preset dust storm level labels in the example above can be customized. For example, the five levels could correspond to floating dust, blowing sand, dust storm, severe dust storm, and extremely severe dust storm, respectively. These five levels could also correspond to Level 1, Level 2, Level 3, Level 4, and Level 5. Furthermore, other levels, such as four or six, can also be set.

[0079] The hyperparameter tuning phase of the CatBoost model primarily focuses on the following key hyperparameters: the number of trees, tree depth, learning rate, L2 regularization coefficient, and categorical feature processing method. For example, a combination of grid search and Bayesian optimization can be used to evaluate the performance of different parameter combinations on a development dataset. The final parameters that can be adopted are: a learning rate of 0.05, a tree depth of 8, 1000 iterations, a symmetric tree construction method, and cross-statistical encoding for categorical features.

[0080] During training, dynamically adjusting the number and depth of trees can improve the balance between model complexity and generalization ability. The learning rate controls the contribution weight of each tree to the final model, and regularization is used to prevent overfitting. Autonomous driving regularization can dynamically suppress the risk of overfitting. Therefore, introducing the learning rate and regularization coefficient can optimize the convergence and stability of ensemble learning. In the dust storm classification task, environmental context data may contain a large number of category features. By iteratively adjusting the category feature processing method, the optimal encoding strategy can be selected, improving the utilization efficiency of non-numerical data. Finally, by adjusting parameters based on the validation set performance in each iteration, the system can dynamically adapt to the dynamic nature of the data, better matching the complexity and dynamism of dust storm data, and providing reliable data support for dust storm early warning.

[0081] After training, the CatBoost model demonstrates high accuracy and stability on the validation set, particularly in static conditions or scenarios lacking wind speed signals, where it can still infer dust level trends from environmental variables, significantly enhancing the system's fault tolerance. During deployment, the CatBoost model can be exported as a general prediction file, supporting import into the vehicle-mounted inference framework using the ONNX (Open Neural Network Exchange) format. Due to its lightweight structure, the inference time is typically less than 10 milliseconds, making it suitable for real-time judgment requirements in edge deployment conditions. The final output of CatBoost is a set of five-dimensional vectors (other dimensions can be set independently), representing the probability distribution of five dust levels (other level classification methods can be set independently) under the current environmental characteristics. This set of probability vectors serves as the environmental dimension output, and together with the spatial feature vector output by GCN and the temporal feature vector output by TCN, it is fed into the fusion model to complete the final dust level identification.

[0082] CatBoost is an optimized version of Gradient Boosting Decision Tree (GBDT), natively supporting categorical feature processing and reducing overfitting risk through ranking boosting. The CatBoost model not only provides independent level prediction paths but also offers rich environmental contextual semantic information for multimodal fusion, enabling the overall system to achieve more comprehensive judgment and response under complex conditions. Dust monitoring data may contain categorical features (such as sensor type, geographic location encoding, and weather type labels). CatBoost eliminates the need for manual one-hot encoding, automatically encoding through target statistics and Bayesian smoothing, avoiding the dimensionality explosion problem caused by high cardinality categorical features. For example, sensor data from different geographic locations can automatically learn their correlation with dust levels through CatBoost. Dust data may contain noise (such as sensor malfunctions or local interference). CatBoost's ranking boosting technique calculates gradients by randomly arranging training samples, avoiding the gradient estimation bias of traditional GBDT, thus performing more robustly on noisy data. For example, at low dust levels, CatBoost can reduce misjudgments caused by abnormal readings from individual sensors. CatBoost's symmetric tree structure supports efficient prediction, making it suitable for real-time dust monitoring systems. Its built-in regularization methods (such as L2 regularization and feature penalty) further simplify the parameter tuning process and reduce the model deployment cost.

[0083] In S150, the dust storm level identification result is generated based on the spatial feature vector, the temporal feature vector, and the probability distribution vector.

[0084] In some embodiments, after completing the modeling of multi-sensor spatial dependencies using a graph convolutional network, the extraction of dynamic trends in wind speed and particulate matter using a temporal convolutional network, and the nonlinear inference of environmental context variables using CatBoost, spatial feature vectors, temporal feature vectors, and probability distribution vectors are obtained. Finally, the spatial feature vectors, temporal feature vectors, and probability distribution vectors need to be effectively fused to generate the final dust storm level determination result. Since each model mines features from different dimensions and there is a complementary relationship between them, designing a reasonable fusion strategy is key to improving recognition accuracy and system stability.

[0085] GCN captures spatial correlations, TCN handles temporal dependencies, and CatBoost optimizes category features; the combination of these three technologies comprehensively covers the key dimensions of dust storm level identification. Dust storms are influenced by various factors, such as meteorological conditions, topography, and human activities, resulting in complex and diverse dust storm scenarios. Information from a single dimension is often insufficient to fully and accurately describe these complex scenarios. However, by combining information from the three dimensions of time, space, and vehicle environment, the model can comprehensively analyze and judge dust storms from multiple perspectives, better adapting to different dust storm scenarios. Furthermore, because the multi-dimensional fusion model integrates information from multiple aspects, it is more adaptable to data. Even if the feature data of one dimension is abnormal or missing, the features of other dimensions can still provide useful information, ensuring that the model can make relatively accurate judgments.

[0086] For example, by analyzing wind speed and particulate matter concentration data collected by sensors at different spatial locations over a period of time, the direction and speed of dust propagation can be determined. If sensors in a certain area detect an increase in wind speed and particulate matter concentration earlier, and neighboring areas show similar changes later in the same direction, it can be inferred that the dust is propagating from the former area to the latter. Furthermore, combining temporal information allows us to understand the duration and stage of dust development at different spatial locations. For instance, some areas may be affected by dust earlier and for a longer period, while others may only be briefly affected. This temporal and spatial complementarity helps to more accurately determine the source, propagation path, and affected area of ​​dust, thus more precisely identifying the dust level. During a prolonged dust storm, although the dust concentration may appear high from a macroscopic temporal perspective, vehicles may pass through localized areas, such as areas shaded by buildings or green belts, where the dust concentration may differ from the surrounding environment. For example, when a vehicle drives behind a building, the wind speed decreases due to the building's obstruction, causing dust to settle. At this time, the concentration of particulate matter around the vehicle may decrease, and visibility may improve. By combining time-based information, we can understand the overall dust situation during the vehicle's time period. Then, by combining vehicle environmental information, we can correct and refine the local dust level. If the time dimension indicates that the current period is a strong dust storm, but the vehicle's environmental sensors detect a low dust concentration in a local area, we can comprehensively consider that the dust level in that local area may be relatively low, but still within the scope of dust storm influence. Furthermore, the particulate matter concentration and visibility data detected by vehicle sensors may contain errors. Combining vehicle environmental information can provide data support for updating and correcting spatial dimension information. Therefore, by integrating information from the three dimensions of time, space, and vehicle environment, the model can comprehensively analyze and judge dust from multiple perspectives, better adapting to different dust storm scenarios. For example, in cities, with dense buildings and complex terrain, the diffusion and deposition of dust differs significantly from open areas. By integrating information from three dimensions, the model can take into account the impact of buildings on wind speed and dust dispersion, as well as the differences in the environment in which vehicles travel on different streets, and more accurately identify the dust levels in different areas of the city.

[0087] As one possible implementation, the vehicle environment perception system constructs a fusion model to achieve the final fusion. The input to the fusion model comes from three parts. The first part is the graph-level spatial feature vector (spatial feature vector) output by the GCN model, representing the global state of the cooperative relationship between sensors in the current time window. The second part is the temporal dynamic feature representation (spatial feature vector) output by the TCN model, reflecting the historical trends of wind speed and particulate matter. The third part is the environmental level prediction probability distribution (probability distribution vector) output by CatBoost, used to express the probability of dust storms occurring under the current external conditions. These three types of outputs represent the results of spatial modeling, temporal modeling, and environmental modeling, respectively. Since they differ in dimension and format, they require unified conversion processing.

[0088] As a possible approach to the result transformation, the three types of features can first be aligned and encoded. The feature vectors of GCN and TCN are mapped to a unified dimension through a linear transformation layer, and then standardized using Batch Normalization to eliminate the impact of scale differences on the fusion effect. The output of CatBoost is a five-dimensional probability vector, directly representing the confidence level of each dust level. After being encoded by a fully connected network layer, it is concatenated with the first two types of features. The overall feature vector structure after fusion is clear, containing information on graph space state, temporal behavior, and external environmental conditions.

[0089] In some embodiments, the fusion method can be modeled using a lightweight neural network structure. The system uses a three-layer Multilayer Perceptron (MLP) as the fusion discriminator. The first layer is the input concatenation layer, the second layer is the feature interaction layer, employing ReLU activation and Dropout to improve generalization ability, and the third layer is the output layer, generating a five-dimensional dust level prediction distribution. The entire fusion structure is end-to-end trainable, and parameters can be backpropagated using a unified loss function to ensure that the outputs of each model continuously converge towards the optimal solution during the fusion process.

[0090] In some embodiments, to improve the discrimination quality during the fusion stage, a weighted guidance mechanism can be added to dynamically adjust the influence ratio of the outputs of different models (GCN model, TCN model, and CatBoost model) on the final judgment during the training phase. In each training iteration, the confidence score and historical stability performance of each sub-model on the current sample are calculated, and different fusion weights are assigned. For example, under conditions of drastic wind speed changes but relatively stable lighting, the system will increase the weight of the TCN model output and correspondingly reduce the participation of the GCN model and CatBoost model, achieving an adaptive fusion strategy that better matches the scene characteristics.

[0091] In some embodiments, the training of the fusion model can employ a joint optimization approach, connecting the fusion model with the GCN, TCN, and CatBoost models into a unified structure for end-to-end training. The loss function is weighted cross-entropy, where each dust level is weighted according to its frequency of occurrence to address class imbalance in the training set. A fusion stability regularization term is also introduced to encourage a certain degree of predictive consistency among the outputs of different models, improving the overall stability and reliability of the judgment. The training data is the same as that of the aforementioned models, derived from real-world datasets from automotive manufacturers, with labels validated through multi-source comparisons to ensure reliable quality. After training, the fusion model demonstrates significantly better accuracy and robustness than individual models on the validation set. Under typical sudden dust storm conditions, the fusion model can integrate multi-dimensional features such as light shading, particulate matter rise, and sudden wind speed changes to achieve accurate level classification, with a significant improvement in accuracy. Furthermore, when a model input experiences fluctuations or malfunctions, the fusion mechanism possesses fault tolerance, relying on the remaining models to maintain reasonable judgments.

[0092] The fusion model, as the final judgment link in the dust storm level identification system, is responsible for outputting the vehicle protection strategy trigger signal. Its output not only includes the five-level dust storm level classification results but also provides a confidence score, which the upper-level control system uses to evaluate the reliability and select whether to trigger the protection logic. The design of this model fully demonstrates the technical advantages of multimodal and heterogeneous data fusion, providing strong decision support for the vehicle-mounted environmental perception system.

[0093] In some embodiments, a fusion model outputs dust level identification results, which include a multidimensional probability distribution vector representing the confidence level of each dust level in the current time period. The maximum confidence level for each dust level is selected, and the dust level corresponding to the maximum value is taken as the real-time dust level for the current time period. After obtaining the real-time dust level, a correlation between the dust level and vehicle control strategy can be established. Based on the real-time dust level, a preset vehicle protection strategy is executed, thereby achieving intelligent protection strategy triggering. By establishing a mapping relationship between dust levels and specific protection actions, the abstract dust level is transformed into a series of specific and executable vehicle protection actions, providing a unified and clear action guide for different users. This provides an intelligent, standardized, and efficient vehicle dust protection solution, improving vehicle driving safety.

[0094] As one possible approach, each dust storm level corresponds to at least one vehicle protection strategy, which includes air conditioning mode switching, intake system adjustment, warning information dissemination, and driving behavior intervention.

[0095] As one possible implementation, let's take a five-level sandstorm level as an example. First, the maximum value is selected from the five-dimensional probability distribution vector output by the fusion model to obtain the most probable sandstorm level at the current moment. In most standard cases, the level determination directly uses the level value corresponding to the highest probability as the real-time sandstorm level. However, to reduce the possibility of false triggers or false alarms caused by fluctuations in edge values, the trigger threshold for each level can be dynamically adjusted based on factors such as the vehicle's historical perception status, the current operating environment, and the level fluctuation trend.

[0096] As a method for dynamically adjusting the level trigger threshold, historical dust storm levels are first obtained. A set of historical cached data is collected. While predicting the level in the current window, the system records the prediction results of the previous N windows and the evolution of the confidence level for each level. The difference between the real-time dust storm level and the historical dust storm level is compared to check for any level abrupt changes. A level abrupt change indicates that the difference is greater than a preset difference. Simultaneously, the light intensity, wind speed, and particulate matter concentration of the current time period are compared with those of historical time periods to check for any parameter abrupt changes. Parameter abrupt changes indicate that at least one of the light intensity, wind speed, and particulate matter concentration has changed more than a preset change threshold relative to the historical time period. When both level abrupt changes and parameter abrupt changes occur simultaneously, the real-time dust storm level is determined to be a high-confidence outbreak level. When the real-time dust storm level is a high-confidence outbreak level, a preset vehicle protection strategy is executed based on the real-time dust storm level.

[0097] For example, if the dust storm level suddenly and rapidly escalates from level two to level four, accompanied by a sharp increase in wind speed and particulate matter concentration, the current state is considered a high-confidence sudden dust storm. In this case, the confidence threshold for triggering protection strategies can be lowered to enable a rapid response to sudden environmental changes. Conversely, if the current level is level four, but historically it has remained stable around level two, and there are no significant fluctuations in sunlight and wind speed, the trigger threshold can be appropriately raised to avoid misjudgments that could lead to unnecessary vehicle strategy switching.

[0098] As one possible implementation method, vehicle protection strategies mainly include four categories: air conditioning mode switching, air intake system adjustment, warning information dissemination, and driving behavior intervention. When the sandstorm level is greater than or equal to level three, the system will automatically switch the air conditioning to the in-vehicle recirculation mode to prevent external sandstorms from entering the vehicle. When the sandstorm level is greater than or equal to level four, the system will simultaneously close the external air intake passage for the engine and activate preset air filtration protection logic; some models support electric filter closure to prevent particulate matter from entering the engine compartment. In extreme sandstorm conditions with a sandstorm level greater than or equal to level five, the system will issue warning information through the instrument panel or central control screen, prompting the driver to reduce speed, turn on low beam headlights, and assess whether to adjust the route or suggest pausing driving based on the vehicle's current condition.

[0099] It is understandable that the vehicle protection strategies corresponding to each sandstorm level mentioned above can be adjusted independently, and are not limited to the vehicle protection strategies in the examples above.

[0100] In some embodiments, considering the differences in hardware capabilities and sensor accuracy among different vehicle models, independent trigger parameter sets can be established for each vehicle configuration. For example, high-end models have more complex intake regulation systems and multi-level filter control strategies, allowing for early intervention of some protective actions when the sandstorm level reaches level three; while in some lower-spec models, the system can set the threshold to sandstorm level four before triggering actions, avoiding resource waste. In addition, some vehicle models allow users to manually adjust the threshold strategy through the vehicle's infotainment system settings interface, enabling flexible switching between intelligent system judgment and user-controlled operation.

[0101] In other embodiments, a critical state maintenance mechanism can be set up so that when the dust level just drops below the threshold, the protection action is not immediately canceled, but delayed by more than N seconds (N can be set by the user) before executing the recovery command. This is to avoid frequent strategy switching from affecting passenger experience and component lifespan. For example, when the dust level drops from level four to level three but high wind speeds still exist, the system will maintain the air circulation and filter closure strategy for a period of time and continuously assess whether the decline is stable.

[0102] Understandably, with the introduction of dynamic threshold determination, the final strategy trigger signal is jointly determined by the dust level identification result output by the fusion model and the dynamic threshold determination result.

[0103] In some embodiments, the action signals of the vehicle protection strategy are transmitted to the vehicle controller in the form of standardized control commands, and future strategy updates and adjustments are supported by OTA (Over The Air) remote configuration.

[0104] In some embodiments, local logs can be set to record the results of sandstorm level identification and the execution of vehicle protection strategies in real time, and upload them to the vehicle manufacturer's cloud data platform regularly to ensure that all strategy triggers have high traceability and recording capabilities, and can also be used for subsequent model optimization.

[0105] In some embodiments, during long-term vehicle operation, a single static model often struggles to maintain high accuracy in the face of constantly changing geographical environments, climate seasons, and differences in vehicle sensor configurations. Therefore, to improve the usability of each model in dust storm level identification, a feedback learning mechanism can be set up to construct a closed-loop model update system based on vehicle-side data uploads, centered on vehicle manufacturer cloud-based model management, and aimed at continuous iteration, ensuring that each model can operate stably and continuously optimize over the long term.

[0106] As a possible implementation of model optimization, after each sandstorm level identification and corresponding strategy trigger, the vehicle can package and store key data from the identification process to form feedback data. Feedback data includes the input raw sensor sequence data, the output level and confidence score of the fusion model, the type of vehicle protection strategy triggered, and auxiliary information such as the vehicle's current geographical location, time information, and driving status. Simultaneously, based on the characteristics of the vehicle's deployment area, ground environmental monitoring data can be linked to supplement the labels. For example, when the vehicle is near a weather station, the sandstorm level for that period can be automatically retrieved via API as a reference label; when the vehicle has an onboard forward-facing camera, it can also analyze indicators such as image clarity and color saturation, combined with particulate matter concentration changes, to verify and correct the original labels, forming a credible feedback sample. After being cached locally, the feedback data is uploaded to the vehicle manufacturer's cloud platform via OTA at set intervals (e.g., weekly or daily after the vehicle is turned off). The upload process is encrypted to ensure vehicle user privacy and data security. After receiving the feedback data, the cloud system automatically enters the sample screening and quality assessment process, evaluating indicators such as the representativeness, confidence consistency, and geographical coverage of the samples, eliminating noisy samples and incomplete data, and constructing a high-quality incremental training set.

[0107] After the sample set is prepared, model update strategies can be designed based on the differences in model structure. GCN and TCN models primarily rely on sequence and structured graph data. During updates, they employ an incremental training mechanism, not re-initializing the original model parameters, but fine-tuning the training based on existing model weights and new samples. To ensure the original model's performance on key data distributions does not degrade during training, knowledge distillation and regularization constraints are introduced to prevent new data from over-covering old knowledge. CatBoost models, on the other hand, use a rollback retraining approach, concatenating new samples with some historical representative samples to form a new training set, performing lightweight reconstruction to ensure the model remains sensitive to the current combination of environmental variables.

[0108] In some embodiments, the cloud platform can also maintain a multi-model, multi-region model version library, with each model bound to a specific model ID and region code. When the misjudgment rate reported by vehicles in a certain region remains high, the system will automatically trigger a local model optimization process, generate a dedicated version, and distribute it to vehicles in the corresponding region via OTA, achieving region-customized recognition capabilities. For example, in the arid Northwest region, the system will strengthen the linkage modeling of sudden changes in high-altitude wind speed and sunlight shading, while in the North China Plain, it will pay more attention to the humidity interference factors of dust transport paths, improving the targeting of recognition in different regions.

[0109] In some embodiments, to ensure the availability of the updated model, the updated model undergoes multiple rounds of simulation verification and hardware / software compatibility testing before deployment to ensure that its efficiency and accuracy on the vehicle side meet safety policy requirements. After the model is deployed, a verification and rollback preparation will be performed before local deployment on the vehicle. If the new model cannot adapt to the current hardware or performs worse than the old version, the system will automatically revert to the previous stable version to ensure the stability and safety of vehicle operation. Furthermore, the deployment effect of each model, regional false alarm rate, policy trigger count, and other key indicators can be continuously monitored through a cloud platform to guide subsequent data collection tasks and model optimization directions, enabling dynamic learning and optimization of the model.

[0110] Figure 2 This is a schematic diagram of the structure of a vehicle dust level identification device provided in an exemplary embodiment of this application, as shown below. Figure 2 As shown, the vehicle dust level identification device 2 includes: an acquisition module 21 for acquiring light intensity, wind speed, particulate matter concentration, and vehicle environmental information during vehicle operation; a first generation module 22 for generating a spatial feature vector based on the temporal and spatial relationships of the acquired light intensity, wind speed, and particulate matter concentration; a second generation module 23 for generating a temporal feature vector based on the correlation between changes in wind speed and particulate matter concentration; a third generation module 24 for generating a probability distribution vector of dust level based on vehicle environmental information; and a fusion module 25 for generating a dust level identification result based on the spatial feature vector, temporal feature vector, and probability distribution vector.

[0111] As one possible implementation, the first generation module 22 can be configured as follows: each sensor signal that collects light intensity, wind speed and particulate matter concentration is taken as a node; wherein the spatial relationship between nodes constitutes edge feature information, and a set of continuous observations of each node within a time window constitutes a time feature sequence; based on the edge feature information and the time feature sequence, an attribute graph structure is generated; based on the graph convolutional network, feature fusion and spatial modeling are performed on the attribute graph structure to generate a spatial feature vector.

[0112] As one possible implementation, the first generation module 22 can be configured to: form a first edge connection based on the airflow sequence of the sensor positions for collecting light intensity, wind speed, and particulate matter concentration; calculate the correlation between any two nodes, and establish a second edge connection between the two nodes when the correlation between any two nodes is greater than a preset threshold; generate an adjacency matrix based on the first and second edge connections; wherein the edge feature information also includes edge set and edge weight, and the first generation module 22 can also be configured to: generate an attribute graph structure based on the edge set, edge weight, adjacency matrix, and time feature sequence.

[0113] As one possible implementation, the first generation module 22 can be configured to: input the attribute graph structure into the graph convolutional network, aggregate based on the edge feature information between adjacent nodes, and adjust the aggregation strength in combination with the edge weights to perform node feature fusion to extract spatial features representing the physical laws of dust propagation; wherein, the edge weights include the average correlation coefficient between two nodes corresponding to the edge in the historical training set; using a preset dust level label as a supervision signal, and aiming to minimize the loss function between the prediction result and the preset dust level label, the graph convolutional network is trained; wherein, the sources of the preset dust level label include the statistics of particulate matter deposition in vehicle air filters, the inference results of camera image blurring, and data verification combined with an external dust weather database; when the graph convolutional network model training process meets the first convergence condition, the graph convolutional network model training is determined to be completed; wherein, the output of the graph convolutional network model is a global feature vector within each time window, and the global feature vector is used to represent the joint state of multiple sensors within the time window.

[0114] As one possible implementation, the second generation module 23 can be configured to: generate a wind speed change sequence and a particulate matter concentration change sequence based on the time series of wind speed and the time series of particulate matter concentration within the same time window; extract the joint dynamic features of the wind speed change sequence and the particulate matter concentration change sequence within the time window based on a temporal convolutional network, and use the joint dynamic features as the time feature vector of wind speed and particulate matter concentration.

[0115] As one possible implementation, the training method of the temporal convolutional network model can be configured as follows: using preset dust level labels corresponding to the same window length as supervision signals, and using wind speed and particulate matter signals obtained from tests on dusty road sections as training samples, the temporal convolutional network is trained based on the cross-entropy loss function; when the training process of the temporal convolutional network model meets the second convergence condition, the training of the temporal convolutional network model is determined to be complete; wherein, the output of the temporal convolutional network model is a temporal feature vector.

[0116] As one possible implementation, vehicle environmental information includes geographical location, current time period, temperature and humidity, vehicle speed, and wind direction. The third generation module 24 can be configured to input geographical location, current time period, temperature and humidity, vehicle speed, and wind direction into a classification model, and the classification model outputs a probability distribution vector of the dust level corresponding to the current vehicle environmental information. The classification model is trained by using a preset dust level label as the target variable, and using the collected environmental context data and corresponding level labeled data as training samples to construct a decision tree. In each iteration, the number of trees, depth, learning rate, autonomous driving regularization coefficient, and category feature processing method are adjusted.

[0117] As one possible implementation, the dust level identification result includes a multidimensional probability distribution vector, which represents the confidence level of each dust level in the current time period. The vehicle dust level identification device 2 can be configured to: select the maximum value of the confidence level of each dust level; use the dust level corresponding to the maximum value as the real-time dust level of the current time period; and execute a preset vehicle protection strategy based on the real-time dust level. Each dust level corresponds to at least one vehicle protection strategy, which includes air conditioning mode switching, air intake system adjustment, warning information release, and driving behavior intervention.

[0118] As one possible implementation, the vehicle dust level identification device 2 can also be configured to: acquire historical dust levels; compare the difference between the real-time dust level and the historical dust level to check for any level abrupt changes; wherein, a level abrupt change indicates that the difference is greater than a preset difference; compare the light intensity, wind speed, and particulate matter concentration of the current time period with the light intensity, wind speed, and particulate matter concentration of historical time periods to check for any parameter abrupt changes; wherein, a parameter abrupt change indicates that at least one of the light intensity, wind speed, and particulate matter concentration changes relative to historical time periods by a value greater than a preset change threshold; when both level abrupt changes and parameter abrupt changes occur simultaneously, determine that the real-time dust level is a high-confidence outbreak level; wherein, based on the real-time dust level, execute a preset vehicle protection strategy, including: when the real-time dust level is a high-confidence outbreak level, execute the preset vehicle protection strategy based on the real-time dust level.

[0119] In some embodiments, the vehicle may include: a sensor for collecting light intensity, a sensor for collecting wind speed, and a sensor for collecting particulate matter concentration; as provided in the vehicle dust level identification device of this application, the vehicle dust level identification device is communicatively connected to the sensor for collecting light intensity, the sensor for collecting wind speed, and the sensor for collecting particulate matter concentration. By utilizing existing vehicle-mounted sensors in conjunction with the vehicle dust level identification device, no new sensors need to be installed. By leveraging the spatiotemporal correlation between multi-source heterogeneous sensors, dynamic changes in wind speed, light intensity, and particulate matter concentration are jointly modeled, and the judgment logic is adjusted according to regional environmental differences and real-time environmental conditions, thereby improving the reliability and accuracy of dust level identification.

[0120] An electronic device includes: a processor; a memory for storing processor-executable instructions; and a processor for executing a page layout method as described in the embodiments provided in this application.

[0121] Below, for reference Figure 3 This application describes an electronic device according to embodiments thereof. The electronic device may be either or both of a first device and a second device, or a standalone device independent of them, which may communicate with the first device and the second device to receive acquired input signals from them.

[0122] Figure 3 A block diagram of an electronic device according to an embodiment of this application is illustrated.

[0123] like Figure 3 As shown, the electronic device 30 includes one or more processors 31 and memory 32.

[0124] The processor 31 may be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device 30 to perform desired functions.

[0125] The memory 32 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 31 may execute the program instructions to implement the page layout methods of the various embodiments of this application described above and / or other desired functions. Various contents such as input signals, signal components, and noise components may also be stored in the computer-readable storage medium.

[0126] In one example, the electronic device 30 may also include an input device 33 and an output device 34, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0127] When the electronic device is a standalone device, the input device 33 can be a communication network connector for receiving the collected input signals from the first device and the second device.

[0128] In addition, the input device 33 may also include, for example, a keyboard, a mouse, etc.

[0129] The output device 34 can output various information to the outside, including determined distance information, direction information, etc. The output device 34 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0130] Of course, for the sake of simplicity, Figure 3 Only some of the components of the electronic device 30 relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device 30 may include any other suitable components depending on the specific application.

[0131] The methods in this application can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer programs or instructions that, when loaded and executed on a computer, perform, in whole or in part, the processes or functions described in this application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, network equipment, user equipment, core network equipment, OAM (Operational Information Management), or other programmable devices.

[0132] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0133] The computer program or instructions may be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions may be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium may be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; or an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both volatile and non-volatile types of storage media.

[0134] Furthermore, embodiments of this application may also be storage media storing a computer program, which is executed by a processor through steps of a page layout method described in any of the above embodiments of this specification: For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0135] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0136] The steps in the methods of the various embodiments of this application can be adjusted, combined, or deleted according to actual needs, and the technical features described in each embodiment can be replaced or combined. The apparatuses in the various embodiments of this application can be combined, divided, or deleted according to actual needs.

[0137] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0138] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software unit executed by a processor, or a combination of both. The software unit can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0139] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only 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 one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0140] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the dust level of a vehicle, characterized in that, include: Acquire information on light intensity, wind speed, particulate matter concentration, and vehicle environment during vehicle operation; Based on the temporal and spatial relationships of the collected light intensity, wind speed, and particulate matter concentration, a spatial feature vector is generated. Based on the correlation between the changes in wind speed and particulate matter concentration, a time feature vector is generated; Based on the vehicle environmental information, a probability distribution vector of the dust level is generated; Based on the spatial feature vector, the temporal feature vector, and the probability distribution vector, a dust storm level identification result is generated.

2. The vehicle dust level identification method according to claim 1, characterized in that, The process of generating a spatial feature vector based on the temporal and spatial relationships of the collected light intensity, wind speed, and particulate matter concentration includes: Each sensor signal that collects the light intensity, the wind speed, and the particulate matter concentration is considered as a node; wherein, the spatial relationship between nodes constitutes edge feature information, and a set of continuous observations of each node within a time window constitutes a time feature sequence; Based on the edge feature information and the time feature sequence, an attribute graph structure is generated; The attribute graph structure is fused and spatially modeled using a graph convolutional network to generate spatial feature vectors.

3. The vehicle dust level identification method according to claim 2, characterized in that, The spatial relationships between the nodes constitute edge feature information, including: The first side connection is formed based on the airflow sequence of the sensor locations for collecting the light intensity, the wind speed, and the particulate matter concentration. Calculate the correlation between any two nodes. When the correlation between any two nodes is greater than a preset threshold, establish a second edge connection between the two nodes. Based on the first edge connection and the second edge connection, generate an adjacency matrix; The edge feature information further includes an edge set and edge weights. The step of generating an attribute graph structure based on the edge information set and the time feature sequence includes: An attribute graph structure is generated based on the edge set, the edge weights, the adjacency matrix, and the time feature sequence.

4. The vehicle dust level identification method according to claim 3, characterized in that, Based on a graph convolutional network, feature fusion and spatial modeling are performed on the attribute graph structure to generate spatial feature vectors, including: The attribute graph structure is input into a graph convolutional network, and the edge feature information between adjacent nodes is aggregated. The aggregation strength is adjusted by combining the edge weights, and node features are fused to extract spatial features that characterize the physical laws of dust propagation. The edge weights include the average correlation coefficient between two nodes corresponding to an edge in the historical training set. Using a preset dust level label as a supervision signal, and aiming to minimize the loss function between the prediction result and the preset dust level label, the graph convolutional network is trained. When the training process of the graph convolutional network model meets the first convergence condition, the training of the graph convolutional network model is determined to be complete; wherein, the output of the graph convolutional network model is a global feature vector within each time window, and the global feature vector is used to represent the joint state of multiple sensors within the time window.

5. The vehicle dust level identification method according to claim 1, characterized in that, The generation of a time feature vector based on the correlation between the changes in wind speed and particulate matter concentration includes: Based on the time series of wind speed and particulate matter concentration within the same time window, wind speed change series and particulate matter concentration change series are generated. The joint dynamic features of the wind speed change sequence and the particulate matter concentration change sequence within the time window are extracted based on the temporal convolutional network, and the joint dynamic features are used as the temporal feature vectors of the wind speed and the particulate matter concentration.

6. The vehicle dust level identification method according to claim 5, characterized in that, The training methods for the temporal convolutional network include: The temporal convolutional network is trained using a preset dust level label corresponding to the same window length as a supervision signal and wind speed and particulate matter signals obtained from tests on dusty road sections as training samples, based on the cross-entropy loss function. When the training process of the temporal convolutional network model meets the second convergence condition, the training of the temporal convolutional network model is determined to be complete; wherein, the output of the temporal convolutional network model is a temporal feature vector.

7. The vehicle dust level identification method according to claim 1, characterized in that, The vehicle environmental information includes geographical location, current time period, temperature and humidity, vehicle speed, and wind direction. The step of generating a probability distribution vector for dust levels based on the vehicle environmental information includes: The geographical location, the current time period, the temperature and humidity, the vehicle speed, and the wind direction are input into the classification model, and the classification model outputs the probability distribution vector of the dust level corresponding to the current vehicle environmental information. The training method for the classification model is as follows: Using preset dust level labels as target variables and collected environmental context data and corresponding level labeled data as training samples, a decision tree is constructed. In each iteration, the number of trees, depth, learning rate, autonomous driving regularization coefficient, and category feature processing method are adjusted.

8. The vehicle dust level identification method according to claim 1, characterized in that, The dust storm level identification result includes a multidimensional probability distribution vector, which represents the confidence level of each dust storm level in the current time period; The vehicle dust level identification method also includes: The maximum confidence level was selected for each dust storm level; The dust level corresponding to the maximum value is taken as the real-time dust level for the current time period; Based on the real-time dust level, a preset vehicle protection strategy is executed; wherein, each dust level corresponds to at least one vehicle protection strategy, and the vehicle protection strategy includes air conditioning mode switching, air intake system adjustment, warning information release, and driving behavior intervention.

9. The vehicle dust level identification method according to claim 8, characterized in that, The vehicle dust level identification method also includes: Obtain historical dust storm levels; The difference between the real-time dust storm level and the historical dust storm level is compared to check for any sudden changes in the level; wherein, a sudden change in the level indicates that the difference is greater than a preset difference. The light intensity, wind speed, and particulate matter concentration in the current time period are compared with those in historical time periods to check for any parameter abrupt changes; wherein, the parameter abrupt change indicates that at least one of the light intensity, wind speed, and particulate matter concentration has a change value greater than a preset change threshold relative to the historical time period; When the level change and the parameter change occur simultaneously, the real-time dust storm level is determined to be a high-confidence outbreak level. Based on the real-time dust level, a preset vehicle protection strategy is implemented, including: When the real-time dust storm level is a high-confidence sudden occurrence level, a preset vehicle protection strategy is executed based on the real-time dust storm level.

10. A vehicle, characterized in that, include: Sensors that collect light intensity, sensors that collect wind speed, and sensors that collect particulate matter concentration; A vehicle dust level identification device is used to perform the vehicle dust level identification method as described in any one of claims 1-9, wherein the vehicle dust level identification device is communicatively connected to a sensor for collecting light intensity, a sensor for collecting wind speed, and a sensor for collecting particulate matter concentration.

Images