Air pollution identification method and vehicle

Abstract

The invention relates to an air pollution identification method and a vehicle, features corresponding to abrupt change events in multi-modal data are encoded based on an attention mechanism to obtain pollution abrupt change event information, and the sudden change trend of air pollution can be accurately identified from high-frequency environmental data. According to the method, particulate matter concentration data in the multi-modal data are used as a drift term, other data in the multi-modal data are used as diffusion terms, pollution trend information is obtained through prediction, and prospective prediction and uncertainty quantification of the pollution evolution trend are achieved. A sensor map is coded based on an attention mechanism to obtain pollution spatial distribution information, and non-uniformity perception of pollution source direction and spatial distribution is realized. Therefore, through multi-modal fusion and time-space collaborative perception, the problems of recognition lag and misjudgment caused by dependence on single sensor data are effectively solved, so that a relatively complex air pollution scene can be dealt with, and the accuracy and speed of air pollution recognition are improved.

Description

Technical Field

[0001] This disclosure relates to the field of air quality control, and more particularly to an air pollution identification method and vehicle. Background Technology

[0002] Air quality control systems are an important component of vehicle control systems.

[0003] In related technologies, air quality control systems identify air pollution in the following ways: based on a particulate matter concentration sensor to detect particulate matter concentration, if the detected particulate matter concentration is higher than a particulate matter concentration threshold, it is determined that air pollution has been identified; or based on a light intensity sensor to detect light intensity, if the detected light intensity is lower than a light intensity threshold, it is determined that air pollution has been identified.

[0004] However, air pollution identification schemes in related technologies rely on single-modal data triggering, which has obvious lag and misjudgment problems when dealing with more complex air pollution scenarios. Summary of the Invention

[0005] To overcome or at least partially solve the above problems, this disclosure provides an air pollution identification method and a vehicle.

[0006] A first aspect of this disclosure provides an air pollution identification method, the method comprising: Using a preset mutation event as a trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on an attention mechanism to obtain contamination mutation event information; Using the particulate matter concentration data in the multimodal data as the drift term and the other data in the multimodal data besides the particulate matter concentration data as the diffusion term, pollution trend information is predicted. The sensor map is encoded based on an attention mechanism to obtain the spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data. Based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, the air pollution identification result is determined.

[0007] A second aspect of this disclosure provides a vehicle including a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the steps of the method described in the first aspect.

[0008] The technical solution provided in this disclosure has the following advantages: By encoding features corresponding to abrupt events in multimodal data using an attention mechanism, pollution abrupt event information is obtained, enabling rapid and accurate identification of changing trends related to sudden air pollution events from high-frequency environmental data. Using particulate matter concentration data as the drift term and other data in the multimodal data (excluding particulate matter concentration) as the diffusion term, pollution trend information is predicted, achieving forward-looking prediction and uncertainty quantification of pollution evolution trends. Encoding sensor maps using an attention mechanism yields pollution spatial distribution information, enabling perception of the non-uniformity of pollution source direction and spatial distribution. Therefore, through multimodal fusion and temporal-spatial collaborative perception, the identification lag and misjudgment problems caused by relying on single sensor data are effectively overcome, thus enabling the handling of more complex air pollution scenarios and improving the accuracy and speed of air pollution identification. Attached Figure Description

[0009] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0010] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram illustrating an application scenario of the air pollution identification method provided in the exemplary embodiments of this disclosure; Figure 2 This is a schematic diagram of the first process of an air pollution identification method provided in an exemplary embodiment of this disclosure; Figure 3 This is a schematic diagram of a second process for an air pollution identification method provided in an exemplary embodiment of this disclosure; Figure 4 This is a schematic diagram of a third process for an air pollution identification method provided in an exemplary embodiment of this disclosure; Figure 5 This is a schematic diagram of the fourth process of the air pollution identification method provided in the exemplary embodiments of this disclosure. Detailed Implementation

[0012] Air pollution identification schemes in related technologies rely on single-modal data triggering, which has obvious lag and misjudgment problems when dealing with more complex air pollution scenarios.

[0013] To overcome the above problems, this disclosure provides an air pollution identification scheme, which identifies pollution mutation events, predicts pollution trends, and models the spatial distribution of pollution based on multimodal data. The obtained pollution mutation event information, pollution trend information, and pollution spatial distribution information are fused to determine the air pollution identification result.

[0014] Among them, through multimodal, multi-scale, and time-space joint perception and prediction, it can cope with more complex air pollution scenarios and improve the accuracy and speed of air pollution identification.

[0015] The application scenarios of the embodiments of this disclosure are described below. See also... Figure 1 The application scenarios of the air pollution identification method provided in this embodiment include vehicle terminal 110, sensor set 120, and server 130.

[0016] The vehicle-mounted terminal 110 is a terminal installed on the vehicle, possessing data acquisition and processing capabilities. It can be implemented as a vehicle controller or a central processing unit in an autonomous driving system. The vehicle-mounted terminal 110 can acquire relevant information and process the acquired information. In this embodiment, the vehicle-mounted terminal 110 is the electronic device that ultimately controls the vehicle.

[0017] The sensor set 120 includes multiple sensors that can be installed in different locations on the vehicle to collect environmental perception information as comprehensively as possible.

[0018] The vehicle terminal 110 is connected to the sensor set 120 for communication, and the vehicle terminal 110 can acquire environmental perception information collected by the sensor set 120.

[0019] Server 130 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0020] The vehicle terminal 110 and the server 130 can be connected via a wired or wireless communication network to achieve data interaction.

[0021] In some exemplary embodiments, when the air pollution identification method is running on server 130, server 130 is used to provide air pollution identification services to vehicle terminal 110.

[0022] The vehicle-mounted terminal 110 collects multimodal data through itself and the sensor set 120, and sends the multimodal data to the server 130.

[0023] Server 130 uses a preset mutation event as a trigger condition, encodes the features corresponding to the mutation event in the multimodal data based on an attention mechanism to obtain pollution mutation event information; uses particulate matter concentration data in the multimodal data as a drift term, and other data in the multimodal data besides the particulate matter concentration data as a diffusion term, to predict pollution trend information; encodes sensor maps based on an attention mechanism to obtain pollution spatial distribution information, the sensor maps being constructed based on the multimodal data and the corresponding sensor location information; and determines the air pollution identification result based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information.

[0024] Server 130 can send air pollution identification results to vehicle terminal 110 to help vehicle terminal 110 deal with air pollution.

[0025] The following is combined Figure 1 The application scenarios described above are used to illustrate the air pollution identification method according to exemplary embodiments of this disclosure. It should be noted that the above application scenarios are shown only to facilitate understanding of the spirit and principles of this disclosure, and the embodiments of this disclosure are not limited in any way. Rather, the embodiments of this disclosure can be applied to any applicable scenario.

[0026] refer to Figure 2 This is a schematic diagram of a first process of an air pollution identification method provided by an exemplary embodiment of the present disclosure.

[0027] Air pollution identification methods include the following steps: Step S210: Using a preset mutation event as a trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on an attention mechanism to obtain contamination mutation event information.

[0028] In specific implementation, step S210 is used to quickly and accurately identify the changing trends related to sandstorms or sudden air pollution from high-frequency environmental data, such as signal patterns with sparsity characteristics, especially sudden drops in light intensity, sudden increases in image blur, drastic changes in particulate matter concentration, and sudden changes in wind speed.

[0029] Step S220: Using the particulate matter concentration data in the multimodal data as the drift term and the other data in the multimodal data besides the particulate matter concentration data as the diffusion term, pollution trend information is predicted.

[0030] In specific implementation, step S220 is used to model the pollution evolution, mainly focusing on modeling the temporal changes of airborne particulate matter such as PM2.5 and PM10. It also incorporates external disturbance factors such as wind speed, wind direction, and vehicle operating status to construct a dynamic evolution process of pollution concentration over time. Step S220 aims to address the problem of slow response of time-series models in related technologies to complex external environmental disturbances, and can more realistically depict the dynamic characteristics of pollution diffusion, accumulation, and dilution.

[0031] Step S230: Encode the sensor map based on the attention mechanism to obtain the spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data.

[0032] In specific implementation, step S230, after completing the detection of pollution mutation events in step S210 and the prediction of pollution concentration change trends in step S220, further models the spatial correlation between sensors around the vehicle body to improve the completeness and accuracy of overall environmental perception. Because air pollution, especially dust pollution, exhibits strong spatial non-uniformity—for example, wind-driven particulate matter rapidly intrudes into the front or side of the vehicle from one side—it is necessary to jointly perceive and model using multiple spatially distributed sensors. Therefore, this stage models the spatial structure and information flow relationships between sensors to construct a dynamic, learnable sensor map, enabling more refined extraction of spatial distribution features and directional identification of pollution sources.

[0033] Step S240: Based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, determine the air pollution identification result.

[0034] In specific implementation, step S240 is used to jointly process the core output results of the three sub-modules after completing the identification of sudden events, prediction of pollution dynamic trends and spatial distribution modeling, to generate a set of pollution state vectors in a unified format, obtain air pollution identification results, and further support the automatic decision-making of subsequent control strategies.

[0035] It should be noted that the above steps S210 to S240 are a simplified description of the air pollution identification method provided in the embodiments of this application. The air pollution identification method provided in the embodiments of this application will be described in more detail below with some examples.

[0036] Next, we will introduce the sources of multimodal data.

[0037] In some exemplary embodiments, the multimodal data includes at least: image data, wind data (including wind speed and wind direction data), particulate matter concentration data, ambient light data, and dynamic state data.

[0038] Combination Figure 1 In some exemplary embodiments, server 130 acquires multimodal data from vehicle terminal 110. A portion of the multimodal data (e.g., the aforementioned image data, wind data (including wind speed and direction data), particulate matter concentration data, and ambient light data) is sensor data collected from the environment by sensor set 120, while another portion is dynamic state data collected from the vehicle by vehicle terminal 110. Optionally, the multimodal data may also include vehicle configuration data (including vehicle hardware and software configuration) pre-stored in vehicle terminal 110.

[0039] In practical implementation, the collaborative operation of multiple sensor modules in the vehicle system establishes a multimodal data acquisition system for sudden environmental events such as sandstorms and air pollution. Under normal vehicle operation, the system continuously collects high-frequency heterogeneous data, including images, wind speed, particulate matter concentration, and ambient light.

[0040] As an example, image data includes: raw video frames captured by a camera (e.g., a front-facing camera); wind data includes: instantaneous velocity vector values ​​collected by a wind speed / direction sensor; particulate matter concentration data includes: PM2.5 and PM10 concentration values ​​collected by a particulate matter sensor; ambient light data includes: Lux (lux, a unit of illuminance) values ​​detected by a light sensor; and dynamic state data includes: dynamic state parameters such as vehicle speed, gear, and vehicle posture.

[0041] In some exemplary embodiments, the sampling frequency of each sensor in the sensor set 120 can be configured with reference to vehicle-related regulations, or it can be freely configured by users (including vehicle manufacturers and drivers).

[0042] As an example, the camera is configured to capture images at a frequency of 15 frames per second, and the sampling frequency of the wind speed / direction sensor and particulate matter sensor is not less than 1 Hz (Hertz).

[0043] Among these features, the high sampling frequency ensures a millisecond-level response capability to sudden changes.

[0044] Through the above exemplary embodiments, this disclosure obtains multimodal data for sudden environmental events such as sandstorms and air pollution. Before processing the multimodal data in subsequent steps, this disclosure also preprocesses and fuses the multimodal data, specifically: In some exemplary embodiments, after determining the multimodal data, the method further includes: The multimodal data is preprocessed to obtain preprocessed multimodal data; Based on linear interpolation and nearest neighbor resampling, the preprocessed multimodal data is aligned and fused in the time dimension to obtain multimodal time series data.

[0045] In specific implementation, the multimodal data is preprocessed to obtain preprocessed multimodal data, including: Image data can be encoded using a lightweight ResNet network (residual neural network), but is not limited to lightweight ResNet; other encoding methods that achieve the same or similar functions are also included. The original image is converted into a fixed-length feature vector sequence while preserving inter-frame information for subsequent temporal modeling. Ambient lighting data is smoothed using a sliding window to eliminate abnormal fluctuations caused by momentary occlusion or noise. Particulate matter concentration data is normalized to eliminate drift during initial sensor startup, and joint compensation based on ambient temperature and humidity information is performed to calibrate data deviations under different external climates. Wind speed and direction data undergo coordinate transformation to a vector representation in the vehicle's coordinate system, ensuring a stable reference direction under different steering and speed conditions.

[0046] In specific implementation, based on linear interpolation and nearest neighbor resampling, the preprocessed multimodal data is aligned and fused along the time dimension to obtain multimodal time-series data, including: Data from different modalities are fused according to a strict timestamp alignment mechanism. Because image data and other sensor data have different sampling frequencies, the system employs linear interpolation and nearest-neighbor resampling to achieve cross-modal data synchronization. Furthermore, to address unique issues in image data such as occlusion and blurring, an image quality assessment module is introduced to detect blur in each frame, identify areas of reduced field of view caused by sand and dust obstruction, and assign data confidence labels for weight adjustment in subsequent feature extraction. Finally, all raw data is organized into a time-series sample set in a unified format. Each sample contains synchronized image encoding vectors, wind speed vectors, PM concentration values, light levels, and vehicle dynamics state variables, constituting the system's original input features.

[0047] After processing, the data is cached in the vehicle's local feature cache module for real-time access by subsequent model processing modules. To adapt to different vehicle models and platform hardware and software environments, the data standardization module adopts a configurable parameter structure. Automakers can adjust the priority and weight factors of each sensor in the fusion process based on vehicle positioning, powertrain type, vehicle height, and intake system structure. For example, in SUVs (sport / suburban utility vehicles), the front wind speed and PM sensors, due to their high exposure, may detect sudden sandstorm intrusion earlier, so the system automatically increases their initial weight in the fusion process. In contrast, in urban sedans with a more enclosed air intake structure, the system relies more on image and lighting changes for sudden situation detection.

[0048] Through the above exemplary embodiments, meticulous data acquisition, organization, and standardized processing provide a high-quality, multi-channel, and alignable feature input foundation for subsequent fusion recognition, ensuring unified support for the algorithm's global environmental perception and local emergency response capabilities. After deployment, the system can generalize data acquisition capabilities under complex and variable weather conditions, significantly improving the basic accuracy and data availability of environmental anomaly detection.

[0049] Step S210: Using a preset mutation event as a trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on an attention mechanism to obtain contamination mutation event information.

[0050] In specific implementation, step S210 is used to quickly and accurately identify the changing trends related to sandstorms or sudden air pollution from high-frequency environmental data, such as signal patterns with sparsity characteristics, especially sudden drops in light intensity, sudden increases in image blur, drastic changes in particulate matter concentration, and sudden changes in wind speed.

[0051] The fully connected Transformer structure commonly used in related technologies suffers from low computational efficiency and the attention mechanism is easily diluted by non-critical features when processing such sparse signals. Therefore, this disclosure performs key attention and feature focusing processing on the input data sequence through step S210 to improve the sensitivity and accuracy of identifying sudden environmental changes.

[0052] In practice, step S210 can be implemented using a sparse Transformer network, but it is not limited to this one method. Therefore, in the following exemplary embodiments, a sparse Transformer network will be used as an example for explanation.

[0053] The input data consists of standardized multimodal samples from the previous stage, including image feature vector sequences, wind speed vector time series, illumination variation curves, and PM concentration rate vectors over time. The system uses fixed time windows as processing units, for example, constructing time series tensors from the most recent 10 seconds of data and inputting them into a sparse Transformer network.

[0054] In terms of network structure, a local sparse attention strategy is adopted, activating the attention mechanism only in regions where predefined key events (i.e., mutation events) occur. The event regions are defined based on the following criteria: a sudden increase in image blur index, wind speed exceeding 12 m / s, a rate of decrease in Lux value exceeding a set threshold, and an abnormal rate of increase in PM concentration within a short period. These events serve as Trigger Tokens, gaining higher participation in the attention mechanism.

[0055] In some exemplary embodiments, while encoding features corresponding to the mutation event in multimodal data based on an attention mechanism using a preset mutation event as a triggering condition, the method further includes: An attention mask is generated for the features in the multimodal data that do not correspond to the mutation event.

[0056] In practice, in addition to activating the attention mechanism in the regions where predefined key events (i.e. mutation events) occur, non-event regions are weakened or ignored within the model through pruning strategies, which further avoids overlearning of stable segments and improves learning efficiency.

[0057] In some exemplary embodiments, learning the features corresponding to the mutation event in the multimodal data based on the attention mechanism includes: Based on the spatial sparse attention mechanism, the image features corresponding to the preset mutation event in the multimodal data are learned; Based on the temporal sparse attention mechanism, the temporal features corresponding to the preset mutation events in the multimodal data are learned.

[0058] In its implementation, the Transformer network employs a hierarchical attention module to model image sequences and temporal sensor data separately before cross-fusion. The image portion uses spatial sparse attention, extracting features only in areas with blurred edges or sudden drops in brightness within image frames, significantly reducing computational cost and enhancing response to critical regions. The temporal sensor data portion uses temporal sparse attention, modeling only time periods where changes exceed historical averages, avoiding overlearning of stable segments. This design improves overall computational efficiency while significantly enhancing the model's responsiveness to sudden environmental changes, making it particularly suitable for the rapid identification of sandstorm weather.

[0059] The output of the sparse Transformer network is a set of weighted vectors representing abrupt events. Each vector contains the event category (e.g., sudden drop in light, PM surge, image blurring), the time of occurrence, the intensity of the change, and its correlation with the multimodal input. These vectors are further mapped to a sudden environmental score vector, which contains a comprehensive score of the degree of anomaly within the entire processing window. The model also outputs the position index of the abrupt change point in the input sequence, used to mark the key modeling areas for subsequent models within a specific time period. The standardized abrupt change score serves as an important input signal for pollution status assessment and enters the fusion decision module.

[0060] To adapt to the differences in sensor layout and intake system response time across various vehicle models, the attention gating mechanism within the sparse Transformer network supports soft parameter adjustment. For example, for vehicles equipped with high-sensitivity image modules, the weight of the image channel in attention fusion can be increased; while in off-road vehicles with more open air ducts and faster sensor responses, the attention trigger weight guided by wind speed changes can be increased. Through this structural adaptability, the module can be uniformly deployed across different platforms and possesses a certain degree of model transfer capability.

[0061] During deployment, the system sets trigger thresholds and alarm levels, and uses mutation point scores to determine whether to proceed to the next stage of pollution state modeling and control strategy planning. This module, as the core of the entire solution's emergency identification, directly impacts the activation timing and priority settings of subsequent air system control and power distribution modules, making it a crucial link in ensuring accurate identification and timely control.

[0062] Step S220: Using the particulate matter concentration data in the multimodal data as the drift term and the other data in the multimodal data besides the particulate matter concentration data as the diffusion term, pollution trend information is predicted.

[0063] In some exemplary embodiments, step S220 specifically includes: A stochastic differential equation is constructed using a neural network based on the drift term and the diffusion term. Based on the stochastic differential equation, the pollution trend information is predicted.

[0064] In specific implementation, step S220 is used to achieve pollution evolution modeling, mainly focusing on modeling the time-varying processes of particulate matter in the air such as PM2.5 and PM10, and combining external disturbance factors such as wind speed, wind direction, and vehicle operating status to construct the dynamic evolution process of pollution concentration over time. Step S220 aims to solve the problem of slow response of time series models in related technologies to complex external environmental disturbances, and can more realistically depict the dynamic characteristics of pollution diffusion, accumulation, and dilution.

[0065] In specific implementation, step S220 can be achieved through a pre-built and trained pollution trend prediction model, but is not limited to this method. The pollution trend prediction model is built based on the Stochastic Differential Equation (SDE) model framework. In the following exemplary embodiments, the pollution trend prediction model is used as an example for illustration: The pollution trend prediction model primarily receives historical particulate matter concentration data (PM2.5 / PM10) within a continuous time window, along with synchronized wind speed vector sequences, wind direction angle changes, vehicle speed, vehicle acceleration, and ambient light levels and image blur indices. All of this data is high-frequency time-series data and has undergone uniform alignment during the preprocessing stage. The pollution trend prediction model continuously receives the latest input using a sliding window approach, predicting pollution trends over several future time slices (e.g., the next 30 seconds to 1 minute). This trend not only encompasses the direction of concentration changes (increasing, decreasing, or remaining stable) but also outputs a probability distribution to describe the interval uncertainty of pollution changes.

[0066] The pollution trend prediction model is structured as a system-level dynamic pollution prediction network composed of multiple stochastic differential equations. Each particulate matter index (e.g., PM2.5) is treated as a state variable, and its change is modeled as a stochastic process including drift and diffusion terms. The drift term simulates the natural variation trend of pollutant concentration under normal environmental evolution, while the diffusion term introduces uncertain external disturbances such as wind speed and direction changes. In vehicle-mounted deployment, the system implements this complex process through a neural SDE modeling framework, employing a parameterized neural network structure to model and learn the drift and diffusion functions, thereby avoiding the complex differential equation construction required for manual modeling and ensuring compatibility with multimodal inputs.

[0067] The training process of pollution trend prediction models relies on a large amount of pollution evolution data under real-world driving conditions. Automakers collect extensive historical pollution data from real vehicles under different regions, weather conditions, and vehicle speeds, combining this data with meteorological records of dust storm levels to establish a multi-dimensional sample set. The training objective is to minimize the deviation between the model's predicted future PM concentrations and actual observations, while maximizing the confidence coverage of the uncertainty interval. Through continuous learning and model updates, the pollution trend prediction model can continuously adapt to the changing characteristics of typical pollution patterns in various regions, achieving high-precision trend prediction capabilities.

[0068] The pollution trend prediction model outputs a set of predicted pollution concentrations over a continuous time period, along with their corresponding confidence intervals. These predictions, together with the current pollution level, constitute an estimate of the future pollution state and serve as crucial input to the downstream control strategy generation module. If the pollution trend prediction model predicts a rapid increase in pollution levels, even if current values ​​are not yet exceeded, the model can proactively activate air quality protection mode to prevent passengers from being exposed to adverse conditions. Furthermore, the pollution trend prediction model also provides a pollution growth rate indicator, which serves as a basis for triggering some active control actions, such as determining when to activate the positive pressure system or close the air intake grille.

[0069] The parameters of the pollution trend prediction model are highly configurable to adapt to different platform hardware capabilities and vehicle pollution response characteristics. For some vehicle models, differentiated initial model states and learning rates can be set based on geographical region. For example, in northern regions frequently experiencing sandstorms, the model can incorporate stronger sensitivity to diffusion term disturbances; while for urban vehicles operating in humid climates, the sensitivity to sudden increases in particulate matter can be reduced to prevent false triggering of the control system.

[0070] The pollution trend prediction model provides information flow support for this disclosure based on predictions, enhancing its forward-looking judgment and environmental change prediction capabilities. The pollution trend information output by the pollution trend prediction model not only provides startup logic support for the air quality control system but also provides an environmental assessment basis for torque distribution in the dynamic response system. It is one of the core supporting mechanisms for achieving forward perception of the control closed loop in the entire scheme.

[0071] Step S230: Encode the sensor map based on the attention mechanism to obtain the spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data.

[0072] In some exemplary embodiments, the encoding of the sensor map based on the attention mechanism to obtain pollution spatial distribution information, wherein the sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data, including: Using the sensors corresponding to the multimodal data as nodes, the spatial adjacency and information coupling strength between the sensors as edges, and the physical distance between the sensors as the weight of the edges, the sensor map is constructed. Using historical coupling frequency, correlation between the multimodal data, and sensitivity to disturbance propagation between the nodes as attention weights, the features of the nodes are updated by weighted aggregation based on neighbor node information to obtain environmental state information and pollution distribution map; The spatial distribution information of pollution includes the environmental status information and the pollution distribution map.

[0073] In specific implementation, step S230 is used to further model the spatial correlation between sensors around the vehicle body after the detection of pollution mutation events and the prediction of pollution concentration change trends in this disclosure are completed, in order to improve the completeness and accuracy of overall environmental perception. Because air pollution, especially dust pollution, has a strong spatial non-uniformity, for example, wind-driven particulate matter rapidly intrudes into the front or side of the vehicle from one side, it is necessary to jointly perceive and model using multiple spatially distributed sensors. Therefore, this stage models the spatial structure and information flow relationships between sensors, thereby constructing a dynamic and learnable sensor map to achieve more refined spatial distribution feature extraction and pollution source directionality identification.

[0074] In specific implementation, step S230 can be achieved through a pre-constructed and trained pollution spatial distribution modeling model. This model is based on a graph neural network (GNN) framework, but is not limited to this approach. The following exemplary embodiments use a pollution spatial distribution modeling model as an example for illustration: The input data for the pollution spatial distribution modeling model includes the outputs of various sensors distributed throughout the vehicle, primarily covering the front wind speed sensor, the left and right PM2.5 / PM10 sensors, the ambient light sensor, the blur analysis results from the front-facing camera, and the vehicle's own heading and lateral speed information. This data is synchronized in time and mapped entirely to the vehicle coordinate system, facilitating the establishment of spatial relationships. The pollution spatial distribution modeling model uses these sensors as nodes in the graph. Each node contains the sensor's current observation value, rate of change, and historical state summary vector. The edges between nodes represent the spatial adjacency and information coupling strength between sensors. The initial edge weights are set based on the physical distance of the sensors within the vehicle's layout, while empirical weights are also introduced; for example, the front wind speed sensor and the front PM concentration sensor have a high correlation in pollution propagation.

[0075] The pollution spatial distribution modeling model is structured as a graph attention network architecture with an attention mechanism built upon a graph neural network. Each node updates its own features through weighted aggregation of information from its neighbors. The attention weights are jointly determined by historical coupling frequency, the relevance of the current observation, and the sensitivity to perturbation propagation between nodes. The pollution spatial distribution modeling model allows for dynamic changes in edge weights. During a pollution event, it adjusts the connection strength between nodes in real time based on factors such as wind direction and vehicle heading changes, thereby simulating the actual propagation path of pollution on the vehicle's surface. For example, if the current wind direction is towards the left front of the vehicle, the pollution spatial distribution modeling model will strengthen the edge weight between the left front PM sensor and the front wind speed sensor, while weakening the aggregation weight of the right rear sensor.

[0076] The training phase of the pollution spatial distribution model uses historical real-vehicle data collected by automakers to construct graph-structured samples. The training objective is to reconstruct the spatial distribution map of pollution and the direction of pollution sources, or to predict the changing trend of sensor values ​​in a certain area over a future period. The model achieves self-learning by maximizing spatial prediction accuracy and consistency with pollution levels. At the same time, it uses data labeled by automotive engineering experts for supervised correction to ensure that the model has engineering interpretability and fault tolerance in actual deployment.

[0077] The output of the pollution spatial distribution modeling model is an environmental state embedding vector that integrates spatial topological information. The vector dimension is dynamically determined based on the number of nodes and feature dimensions. Simultaneously, a pollution distribution map is output to identify the main intrusion directions of pollution, high-pollution areas, and potential risk paths. This output not only supports subsequent control strategy decisions but also provides a basis for system visualization, enabling the system to provide real-time indications of pollution direction through instrument displays, such as marking "increased pollutant concentration and high wind speed to the left front" in the vehicle's central control system.

[0078] The pollution spatial distribution model is highly customizable to improve vehicle adaptability. Different vehicle models can define their own graph template structures due to differences in sensor layout. For example, in an off-road vehicle, the edge weight of the high-position PM sensor on the right can be set to a high value to adapt to the rapid response of the right-side inhalation path in sandy environments; while in an urban SUV, the sensor near the lower front air intake can be used as a key aggregation node to enhance forward prediction capabilities. The system automatically matches the graph structure template with the sensor identification code to achieve adaptive deployment across multiple vehicle models without retraining the core model.

[0079] The pollution spatial distribution model disclosed herein greatly enhances the ability to model the spatial distribution of pollution, enabling environmental perception to move beyond single-point detection and acquire regional correlation and dynamic change tracking capabilities. As a bridge between sparse feature identification and time-series trend prediction, the pollution spatial distribution model effectively incorporates spatial information into the system's pollution assessment framework, significantly improving the accuracy of pollution source direction identification and providing solid data support for the subsequent refinement of control strategies.

[0080] Step S240: Based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, determine the air pollution identification result.

[0081] In practical implementation, after the system completes the identification of mutation events, prediction of pollution dynamic trends, and spatial distribution modeling, the core outputs of these three sub-modules need to be jointly processed to generate a pollution state vector in a unified format to support the automatic decision-making of subsequent control strategies. To achieve effective integration of multi-source model information, this step designs an attention-guided multi-model fusion mechanism, which jointly inputs the temporal mutation features output by the sparse Transformer network, the trend prediction information output by the SDE model, and the spatial distribution encoding output by the GNN model to generate a high-dimensional pollution state vector. This vector uniformly describes the pollution intensity, changing trend, intrusion path, key affected areas, and system response priority in the current and short term.

[0082] The input in this stage includes three main types of features. The first type is the abrupt event vector output by the sparse Transformer, containing the abrupt event type (e.g., sudden drop in light intensity, image blurring, sudden increase in PM levels), the abrupt event time point, and weighting coefficients. The second type is the future pollution concentration change curve predicted by the SDE model, which is further encoded as the trend direction (increasing, decreasing, fluctuating), growth rate index, and prediction confidence interval. The third type is the spatial pollution distribution map and sensor embedding vector output by the GNN model, representing the distribution of pollution sources, dominant intrusion direction, and regional risk level. All input vectors are synchronized on the time axis and uniformly remapped in space to ensure consistency and comparability of the three types of information.

[0083] In some exemplary embodiments, determining the air pollution identification result based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information includes: Based on the correlation between the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, the weight information corresponding to the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information is determined by using an attention mechanism. The pollution mutation event information, pollution trend information, and pollution spatial distribution information are weighted and fused based on the weight information to obtain the air pollution identification result.

[0084] In practical implementation, the core of the fusion mechanism is a multi-channel attention-guided fusion structure. The system first performs dimensional alignment and normalization on the outputs of the three types of models. Then, a multi-head attention layer calculates the degree of mutual influence between the outputs of different models. The sparse Transformer output has strong constraints in the time dimension, effectively guiding the attention mechanism to focus on points of abrupt change; the SDE output has foresight in trend judgment, improving the ability to trigger policy decisions in advance; and the GNN output has interpretability and local difference discrimination capabilities in spatial recognition, guiding control strategies to precisely target contaminated intrusion areas. The fusion structure allows different models to have different dominant weights under different contamination scenarios, thereby achieving a dynamically weighted result fusion strategy.

[0085] The fusion output is a set of fixed-structure pollution state vectors containing six key fields. First, there's the current pollution level label, categorized into four levels—Normal, Moderate, Heavy, and Sandstorm—based on the fusion results, directly driving the selection of control strategy templates. Second, there's the pollution trend vector, indicating whether pollution is predicted to worsen within a specific future timeframe and providing a numerical rate indicator. Third, there's a list of high-risk spatial areas, informing the air system of the main directions of pollution intrusion. Fourth, there's a strategy trigger priority indicator, such as whether closing the grille or activating positive pressure inside the vehicle should be prioritized. Fifth, there's a predicted control stability value, indicating whether the control system can stably maintain the target environmental level under the current pollution conditions. Finally, there's a model fusion confidence value, indicating the reliability of the current state assessment, used to decide whether to delay execution or wait for subsequent data confirmation.

[0086] The pollution state vector plays a role in the system similar to a "strategy decision-making center," with all downstream control modules calculating and optimizing their response actions based on this vector. To adapt to the control system architecture and user comfort settings of different vehicle models, this vector structure supports soft template adaptive matching. For example, in high-end vehicles, the system can automatically switch the internal air circulation duct and cabin air conditioning volume according to the pollution level, while in commercial vehicles, it tends to maintain high-efficiency particulate filtration and control the power output mode.

[0087] The model fusion process does not involve hard-coded logical judgments; it is entirely automated through attention weights and the learning structure, thereby enhancing the system's environmental adaptability and platform portability. In actual deployment, the system can continuously learn from feedback to optimize the fusion weight distribution, achieving dynamic adjustments to regional pollution characteristics and seasonal climate changes.

[0088] This step serves as the information integration hub for multi-model collaborative identification and prediction, providing structured, stable, and interpretable decision-making basis for the subsequent air system and off-road torque control module, laying the most critical cognitive decision-making foundation in the entire identification and control closed loop.

[0089] In some exemplary embodiments, the pollution mutation event information is determined by a pollution mutation event identification model, the pollution trend information is determined by a pollution trend prediction model, and the pollution spatial distribution information is determined by a pollution spatial distribution modeling model. The training methods for the pollution mutation event identification model, the pollution trend prediction model, and the pollution spatial distribution model include: The pollution mutation event identification model, the pollution trend prediction model, and the pollution spatial distribution model are trained in parallel. In the parallel training process, feature verification is performed among the pollution mutation event identification model, the pollution trend prediction model, and the pollution spatial distribution model, and the deviation of pseudo-labels in the parallel training process is corrected through cross-checking.

[0090] In practice, steps S200, S220, and S230 can be implemented using a model, specifically: By using a pre-trained pollution mutation event identification model, with a preset mutation event as the trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on the attention mechanism to obtain pollution mutation event information. By using a pre-trained pollution trend prediction model, the particulate matter concentration data in the multimodal data is used as the drift term, and the other data in the multimodal data besides the particulate matter concentration data is used as the diffusion term, pollution trend information is predicted. By using a pre-trained pollution spatial distribution modeling model, the sensor map is encoded based on an attention mechanism to obtain pollution spatial distribution information. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data.

[0091] The training methods for the pollution mutation event identification model, the pollution trend prediction model, and the pollution spatial distribution model will be introduced below, specifically: In practical implementation, to ensure that the multiple recognition sub-modules (Sparse Transformer, SDE pollution evolution model, and graph neural network) in this disclosure possess high robustness, strong generalization, and engineering deployability under sandstorm and air pollution scenarios, this step designs a systematic multi-model joint training process and constructs a semi-automated annotation system specifically for car manufacturers to support efficient learning and model iteration of large-scale multimodal data. This mechanism combines real-vehicle data collection, rule-guided pseudo-label generation, expert manual verification, and active model learning to form a stable, efficient, and automotive-grade-compatible data training closed loop, ensuring the model's recognition accuracy and the reliability of control strategy generation under complex real-world conditions.

[0092] The training data comes from raw data logs collected by automakers using test vehicles deployed across multiple models, regions, and climate conditions. The data includes forward-facing camera video streams, PM2.5 / PM10 particulate matter concentration time series, wind speed and direction sensor data, light intensity changes, image clarity indicators, and synchronous information such as vehicle driving status, torque distribution, and control command execution records. All data logs are stored in a standardized format with precise timestamps and synchronization identifiers to support unified training tasks across models. Each vehicle can automatically upload data to a central server on an hourly basis. Environmental segmentation indexing is performed through a unique vehicle identifier and environmental label management system, facilitating categorization and organization by labels such as sandstorm, severe pollution, light pollution, and normal environment.

[0093] To construct a high-quality labeled dataset, the system designed a rule-guided semi-automatic label generation mechanism. First, by setting pollution mutation rules (such as PM concentration rising above a certain threshold within 5 seconds, wind speed suddenly exceeding 12 m / s, or rapid increase in image blur), candidate pollution segments are automatically identified and initially labeled. The system automatically defines the start and end times of the mutation, the initial pollution level, and the estimated pollution direction. Subsequently, engineering experts review the data, focusing on verifying the match between the automatic labels and the actual images and sensor trends, correcting the label levels and pollution types, and finally generating high-confidence training samples suitable for multi-model training.

[0094] For the three core models, a decoupled joint training strategy is adopted during the training process. The Sparse Transformer aims to identify mutation events, with training labels including pollution type, multimodal triggering points, and their weight distribution. The SDE model aims to predict pollution concentration changes, employing a self-supervised training method that does not require explicit labels, relying solely on fitting historical and future concentration curves. The Graph Neural Network aims to reconstruct the spatial distribution of pollution, with training labels including pollution source direction, key high-risk nodes, and pollution gradient maps. During the training phase, the training tasks of the three models are run in parallel on a unified data platform. A feature verification mechanism between models ensures input consistency and output compatibility. Furthermore, cross-checking can further correct pseudo-label biases, improving the overall collaborative performance of the models.

[0095] To enhance the system's continuous learning capabilities, an active model learning module was constructed. After deployment, the system automatically feeds back polluted segments with low confidence levels, marking them as "low-confidence samples" in the training data pool. These samples undergo priority review by experts, are further processed, and then reintroduced into the training process. Simultaneously, to address generalization issues such as regional and vehicle type differences, the training platform incorporates a regional distribution strategy and a vehicle-specific parameter generation mechanism. This allows the model to automatically adjust its parameter templates based on the deployment region (e.g., high-incidence sandstorms in North China, and humid, low-pollution conditions in South China) and vehicle type (e.g., SUVs, sedans, and commercial vehicles), generating a lightweight model version with customized features. This improves the system's efficiency and adaptability in automotive-grade deployments.

[0096] The entire training and annotation process ultimately forms a complete multi-model collaborative training architecture: from data collection, pseudo-label generation, manual review, model training to feedback after model deployment, and finally to model fine-tuning, forming a closed loop. Automakers can periodically update model weight files through OTA (Over-the-Air Technology) mechanisms, enabling in-service vehicles to continuously adapt to external disturbances such as climate change, environmental policy adjustments, and hardware aging, further improving the overall system's engineering availability and stability.

[0097] This module serves as the underlying support mechanism for the implementation of the multi-model algorithm disclosed herein. It ensures that the pollution identification and linkage control system can operate stably in a real vehicle environment for a long time, and builds a sustainable and self-evolving intelligent pollution protection capability for the whole vehicle platform. It is a key guarantee for the large-scale promotion and deployment of the entire technical solution.

[0098] In the above exemplary embodiments, this disclosure provides an air pollution identification method that, through multimodal, multi-scale, and time-space joint perception and prediction, can cope with more complex air pollution scenarios, improving the accuracy and speed of air pollution identification. However, this disclosure is not limited thereto. In some exemplary embodiments, after obtaining air pollution identification, this disclosure also provides a scheme for air system regulation based on the air pollution identification result, specifically: refer to Figure 3 This is a schematic diagram of a second process for an air pollution identification method provided in an exemplary embodiment of the present disclosure.

[0099] Air pollution identification methods include the following steps: Step S310: Using a preset mutation event as a trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on the attention mechanism to obtain contamination mutation event information; Step S320: Using the particulate matter concentration data in the multimodal data as the drift term and the other data in the multimodal data besides the particulate matter concentration data as the diffusion term, pollution trend information is predicted; Step S330: Encode the sensor map based on the attention mechanism to obtain the spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data. Step S340: Based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, determine the air pollution identification result.

[0100] It should be noted that the specific implementation of steps S310 to S340 is similar to the specific implementation of steps S210 to S240 in the above embodiments of this disclosure, and will not be repeated here.

[0101] Step S350: Based on the pollution level information, determine the corresponding air system control strategy template; adjust the air system control strategy template based on the pollution trend information and the pollution spatial distribution information to obtain the air system control strategy; control the air system based on the air system control strategy.

[0102] The air pollution identification results include: pollution level information, pollution trend information, and pollution spatial distribution information.

[0103] In practical implementation, after the pollution state vector is constructed, this disclosure enters the crucial execution decision-making stage. The goal of this step is to generate a set of control commands that can be directly applied to the vehicle's air management hardware system based on the aforementioned fused key indicators such as pollution level, trend, and spatial distribution. This ensures stable air quality in the passenger compartment and minimizes the risk of pollutant intrusion. To achieve tight coupling between pollution identification and the air control system, this disclosure constructs a strategy rule engine and command mapping logic. By matching the pollution state vector with predefined control strategy templates, it automatically generates multi-channel control signals, covering control components such as the air intake grille, electric air conditioning door, cabin positive pressure system, and internal circulation channel switching.

[0104] The input to the control strategy is the pollution state vector output from the previous stage. This vector contains key fields such as pollution level (Normal, Moderate, Heavy, Sandstorm), pollution trend (rising / falling), pollution intrusion direction (front / side / multi-point), strategy trigger priority, and control stability prediction value. The system first enters the corresponding strategy template branch based on the pollution level and then refines the adjustments based on trend and spatial distribution information. For example, when the pollution level is determined to be "Sandstorm," the system enters the highest-level response path. However, if the pollution trend shows signs of decline, the response intensity can be appropriately reduced to avoid energy waste. If the pollution intrusion direction is concentrated in the side area, the air conditioning recirculation logic and window control status are adjusted accordingly.

[0105] This step employs a strategy graph structure, with each pollution level corresponding to a set of basic control actions. The system quickly matches the response strategy by looking up a table. For example, at the "Moderate" pollution level, the default control actions are to close the front grille, maintain an internal / external air circulation ratio of 20:80, and start the PM filter module at low speed. At the "Heavy" level, the system further closes all external circulation channels, starts the cabin pressurization fan to establish a positive pressure differential, starts the internal circulation mode, and activates the high-speed operation mode of the PM2.5 filter in the air conditioning system. At the "Sandstorm" level, all external channels are closed, the positive pressure system operates at full load, and if the vehicle supports remote dust removal, the A-pillar external circulation purification component is activated simultaneously to achieve rapid environmental isolation.

[0106] The commands are generated in a standardized structured command set. Each command includes the controlled object, target state, duration, and whether it can be intervened by the user. For example, a control command might include: "intake_grill: closed, duration: 180s, override: false", indicating that the system will forcibly close the grille within the next 180 seconds, and the user is not allowed to cancel this control state via physical buttons or the central control system. Each command includes a unique ID (Identity document) and an execution feedback mechanism to ensure that the actual control action is consistent with the strategy settings. The system can monitor the execution status in real time and make secondary adjustments.

[0107] To ensure the accuracy and stability of command generation, this step performs redundancy checks before generating each set of control commands. A higher-level control strategy is only triggered if two consecutive identifications of the pollution level are consistent, avoiding false triggers caused by a single misidentification. Trend and spatial distribution indicators must meet a minimum duration threshold before taking effect, avoiding short-term judgment errors caused by sudden wind speed changes or image occlusion. The system also incorporates a strategy stability prediction module. The control stability index output during the fusion phase is used to evaluate the system's feedback capability after the control strategy is implemented. If the predicted value is lower than a set threshold, some control actions will be delayed or a gradual adjustment strategy will be adopted.

[0108] Regarding vehicle model adaptation, this step supports the automatic generation of strategy templates based on hardware configuration files defined by the automaker. For example, some models are not equipped with a positive cabin pressure system, so when the pollution level reaches Heavy, the PM filter module's operating power will be increased to the maximum by default, and the external air circulation damper will be locked. In luxury models, the system can link seat ventilation, ambient lighting prompts, and the central control information interface to form a more complete air pollution response experience.

[0109] After being generated, all control commands are sent to the corresponding ECU (Engine Control Unit) module via CAN (Controller Area Network) bus or Ethernet. Execution nodes include the body controller, air conditioning system controller, and the core control unit of the air quality management system. After successful command transmission, the system enters feedback monitoring mode, recording the execution results and feedback values ​​from the passenger compartment air quality sensors in real time. If the control action fails to be completed as scheduled or the air quality continues to deteriorate, the system will automatically enter the next level of control state or switch to the backup control module.

[0110] Through the exemplary embodiments described above, this step, acting as a bridge between the identification results and the physical control system, is a crucial link in achieving real-time air protection response in the entire solution. By rapidly parsing the pollution state vector and accurately mapping it to hardware control commands, the system can dynamically manage the in-vehicle air environment in complex external environments, effectively improving occupant safety and comfort. Furthermore, it provides a foundation of environmental parameters for subsequent powertrain strategy linkages.

[0111] In the above exemplary embodiments, this disclosure provides an air pollution identification method and a scheme for air system regulation based on the air pollution identification result. However, this disclosure is not limited thereto. In some exemplary embodiments, after obtaining air pollution identification, this disclosure also provides a scheme for regulating a vehicle drive system based on the air pollution identification result, specifically: refer to Figure 4 This is a schematic diagram of a third process for an air pollution identification method provided by an exemplary embodiment of the present disclosure.

[0112] Air pollution identification methods include the following steps: Step S410: Using a preset mutation event as a trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on an attention mechanism to obtain contamination mutation event information; Step S420: Using the particulate matter concentration data in the multimodal data as the drift term and the other data in the multimodal data besides the particulate matter concentration data as the diffusion term, pollution trend information is predicted; Step S430: Encode the sensor map based on the attention mechanism to obtain the spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data. Step S440: Based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, determine the air pollution identification result.

[0113] It should be noted that the specific implementation of steps S410 to S440 is similar to the specific implementation of steps S210 to S240 in the above embodiments of this disclosure, and will not be repeated here.

[0114] Step S450: Determine vehicle operating information; based on the air pollution identification result and the vehicle operating information, generate a vehicle drive system control strategy; based on the vehicle drive system control strategy, control the vehicle drive system.

[0115] In some exemplary embodiments, generating a vehicle drive system control strategy based on the air pollution identification result and the vehicle operation information includes: Based on the air pollution identification results and the vehicle operation information, a corresponding vehicle drive system control strategy template is determined. Based on the air pollution identification results and the vehicle operation information, the vehicle drive system control strategy template is adjusted to obtain the vehicle drive system control strategy.

[0116] In practice, once the vehicle identifies a sandstorm or severe air pollution level based on the air pollution identification results, it enters a control phase linked to the powertrain. The core of this step lies in coupling the environmental identification results from the air pollution identification with the vehicle's drive system. This enables adaptive adjustment of the power strategy in dust-proof mode, ensuring stable power delivery in complex environments while minimizing the entry of sand and dust into the engine system or its impact on power response. It not only considers the direct impact of pollution levels on the drive strategy but also integrates multiple variables such as the vehicle's current state, torque demand, and intake path design to achieve adaptive optimization of the drive system in polluted environments.

[0117] The input to step S450 includes the pollution level field, trend field, spatial distribution field, and control strategy priority field from the pollution state vector. It also includes the vehicle's current operating parameters, such as vehicle speed, engine speed, current drive mode (two-wheel drive / four-wheel drive), road surface adhesion coefficient estimate, vehicle longitudinal slope, and wheel speed difference.

[0118] In some exemplary embodiments, the system determines whether the conditions for switching to "dustproof off-road mode" are met based on the above information and dynamically corrects the existing drive strategy. When the pollution level is Heavy or higher, and the wind direction is highly consistent with the air intake direction, the drive system protection strategy intervention process will be triggered.

[0119] First, the system assesses the compatibility between the current drive mode and the intake path. For vehicles with an open-type front air intake structure, when wind speed is high and particulate matter concentration is high, the system will automatically reduce the engine torque limit, limiting the instantaneous high negative pressure intake when the accelerator is pressed deeply, thus reducing the risk of sand and dust entering the intake manifold through the filter system. Simultaneously, in this state, the system will prioritize torque distribution to the rear axle, adjusting the drive ratio to front:rear = 30:70 or 40:60 to reduce the probability of front axle slippage on low-traction surfaces, thereby reducing ground dust stirred up by wheel slippage.

[0120] For vehicles with four-wheel drive capabilities, the system will automatically switch to the four-wheel drive lock mode and adjust the central differential lock strategy to ensure stable power distribution between the front and rear axles. If an active torque distribution system is equipped, the control logic will further offset the driving torque according to the direction of the pollution source. For example, when the pollution direction is concentrated on the left side of the vehicle, the system will reduce the torque of the left wheels and increase the output on the right side, controlling the vehicle to operate in an asymmetric torque mode to avoid high-concentration pollution areas as much as possible.

[0121] In an extreme sand and dust environment, when a pollution level of Sandstorm is identified and accompanied by a serious decline in visibility, the system will activate a multi-level linkage protection mechanism. First, it will limit the maximum speed of the vehicle and force it to enter the protection speed limit mode to avoid the strong suction effect caused by high-speed driving. Second, in配合发动机管理系统,适当延迟进气门开启时机,减少峰值负压的产生;再次,在具备电子涡轮技术的车型中,系统将调节涡轮响应曲线,在污染高强度阶段降低涡轮介入频率与介入速率,确保发动机进气压力维持在污染过滤系统负载能力范围内。

[0122] The adjustment of the power strategy is not only reflected in the longitudinal control dimension but also in the lateral and undulating terrain adaptation capabilities. When the system integrates camera and IMU (Inertial Measurement Unit) information and identifies an undulating sandy or wind-eroded terrain ahead, it will increase the intervention level of the traction control system to avoid secondary pollution entrainment caused by wheel spin and jamming. At the same time, the chassis domain controller will adjust the damping characteristics of the shock absorbers and the chassis height to ensure the best ground clearance in a sand and dust environment, blocking the upward path of fine surface particles into the engine compartment or the gaps at the bottom of the cockpit.

[0123] After the vehicle completes the drive strategy switch, the system will enter the "dust-proof off-road mode" state. The instrument interface will display the mode activation indicator and will also provide real-time feedback on torque distribution ratios, intake status, and pollution level-related information to assist the driver in understanding the reasons for the current strategy. During the maintenance of the mode, the system continuously monitors the pollution level and trend. When the pollution level drops below Moderate and the duration exceeds the preset threshold, the system will automatically解除防尘越野模式,恢复原始驱动分配策略。

[0124] Due to differences in power structures among different vehicle platforms, this linkage module has a high degree of configurability. For pure electric platforms, the system will mainly control the power response curve of the electric drive axle and the filter air duct configuration of the heat exchange system, setting power protection upper limits and air duct pressure drop thresholds respectively for front-wheel drive and rear-wheel drive platforms. For internal combustion engine or range-extended platforms, it will control more engine management parameters related to intake and torque distribution paths to ensure stable operation of each platform in a polluted environment.

[0125] Through the above exemplary embodiments, as a bridge between the pollution identification method and the vehicle power strategy, cross-domain linkage from cognition to response is completed, realizing the adaptive adjustment of the vehicle's power under extreme pollution conditions, and constructing a complete closed loop for the entire identification and protection system.

[0126] In the above exemplary embodiments, this disclosure provides an air pollution identification method and a scheme for separately regulating the air system or the vehicle drive system based on the air pollution identification result. However, this disclosure is not limited thereto. In some exemplary embodiments, after obtaining air pollution identification, this disclosure can jointly regulate the air system and the vehicle drive system based on the air pollution identification result, specifically: refer to Figure 5 This is a schematic diagram of the fourth process of the air pollution identification method provided by the exemplary embodiments of this disclosure.

[0127] Air pollution identification methods include the following steps: Step S510: Using a preset mutation event as a trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on an attention mechanism to obtain contamination mutation event information; Step S520: Using the particulate matter concentration data in the multimodal data as the drift term and the other data in the multimodal data besides the particulate matter concentration data as the diffusion term, pollution trend information is predicted; Step S530: Encode the sensor map based on the attention mechanism to obtain the spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data. Step S540: Based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, determine the air pollution identification result.

[0128] It should be noted that the specific implementation of steps S510 to S540 is similar to the specific implementation of steps S210 to S240 in the above embodiments of this disclosure, and will not be repeated here.

[0129] Step S550: Based on the pollution level information, determine the corresponding air system control strategy template; adjust the air system control strategy template based on the pollution trend information and the pollution spatial distribution information to obtain the air system control strategy; control the air system based on the air system control strategy.

[0130] It should be noted that the specific implementation of step S550 is similar to the specific implementation of step S350 in the above embodiments of this disclosure, and will not be described in detail here.

[0131] Step S560: Determine vehicle operating information; based on the air pollution identification result and the vehicle operating information, generate a vehicle drive system control strategy; based on the vehicle drive system control strategy, control the vehicle drive system.

[0132] In practice, after the vehicle identifies a sandstorm or severe air pollution level and generates air system control commands, the system enters a control phase linked to the powertrain. The core of this step lies in coupling the environmental identification results from the air pollution identification with the vehicle's drive system. This enables adaptive adjustment of the power strategy in dust-proof mode, ensuring stable power delivery in complex environments while minimizing the entry of sand and dust into the engine system or its impact on power response. This module not only considers the direct impact of pollution levels on the drive strategy but also integrates multiple variables such as the vehicle's current state, torque demand, and intake path design to achieve adaptive optimization of the drive system under polluted conditions.

[0133] It should be noted that the specific implementation of step S560 is similar to the specific implementation of step S450 in the above embodiments of this disclosure, and will not be described in detail here.

[0134] This disclosure also provides a vehicle including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the air pollution identification method provided in any of the above embodiments.

Claims

1. A method for identifying air pollution, characterized in that, The method includes: Using a preset mutation event as a trigger condition, the features corresponding to the mutation event in the multimodal data are encoded based on an attention mechanism to obtain contamination mutation event information; Using the particulate matter concentration data in the multimodal data as the drift term and the other data in the multimodal data besides the particulate matter concentration data as the diffusion term, pollution trend information is predicted. The sensor map is encoded based on an attention mechanism to obtain the spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the sensor location information corresponding to the multimodal data. Based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, the air pollution identification result is determined.

2. The method according to claim 1, characterized in that, The air pollution identification results include: pollution level information, pollution trend information, and pollution spatial distribution information; After determining the air pollution identification result, the method further includes: Based on the pollution level information, a corresponding air system control strategy template is determined; The air system control strategy template is adjusted based on the pollution trend information and the pollution spatial distribution information to obtain the air system control strategy. The air system is regulated based on the aforementioned air system regulation strategy.

3. The method according to any one of claims 1 or 2, characterized in that, After determining the air pollution identification result, the method further includes: Determine vehicle operating information, and generate a vehicle drive system control strategy based on the air pollution identification results and the vehicle operating information; The vehicle drive system is controlled based on the aforementioned vehicle drive system control strategy.

4. The method according to claim 3, characterized in that, The step of generating a vehicle drive system control strategy based on the air pollution identification results and the vehicle operation information includes: Based on the air pollution identification results and the vehicle operation information, a corresponding vehicle drive system control strategy template is determined. Based on the air pollution identification results and the vehicle operation information, the vehicle drive system control strategy template is adjusted to obtain the vehicle drive system control strategy.

5. The method according to claim 1, characterized in that, The process of determining air pollution identification results based on the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information includes: Based on the correlation between the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information, the weight information corresponding to the pollution mutation event information, the pollution trend information, and the pollution spatial distribution information is determined by using an attention mechanism. The pollution mutation event information, pollution trend information, and pollution spatial distribution information are weighted and fused based on the weight information to obtain the air pollution identification result.

6. The method according to claim 1, characterized in that, The learning of features corresponding to the mutation event in the multimodal data based on the attention mechanism includes: Based on the spatial sparse attention mechanism, the image features corresponding to the preset mutation event in the multimodal data are learned; Based on the temporal sparse attention mechanism, the temporal features corresponding to the preset mutation events in the multimodal data are learned.

7. The method according to claim 1, characterized in that, The method of predicting pollution trend information by using particulate matter concentration data from the multimodal data as the drift term and other data from the multimodal data besides the particulate matter concentration data as the diffusion term includes: A stochastic differential equation is constructed using a neural network based on the drift term and the diffusion term. Based on the stochastic differential equation, the pollution trend information is predicted.

8. The method according to claim 1, characterized in that, The sensor map is encoded based on an attention mechanism to obtain spatial distribution information of pollution. The sensor map is constructed based on the multimodal data and the corresponding sensor location information, and includes: Using the sensors corresponding to the multimodal data as nodes, the spatial adjacency and information coupling strength between the sensors as edges, and the physical distance between the sensors as the weight of the edges, the sensor map is constructed. Using historical coupling frequency, correlation between the multimodal data, and sensitivity to disturbance propagation between the nodes as attention weights, the features of the nodes are updated by weighted aggregation based on neighbor node information to obtain environmental state information and pollution distribution map; The spatial distribution information of pollution includes the environmental status information and the pollution distribution map.

9. The method according to claim 1, characterized in that, The pollution mutation event information is determined by a pollution mutation event identification model, the pollution trend information is determined by a pollution trend prediction model, and the pollution spatial distribution information is determined by a pollution spatial distribution modeling model. The training methods for the pollution mutation event identification model, the pollution trend prediction model, and the pollution spatial distribution model include: The pollution mutation event identification model, the pollution trend prediction model, and the pollution spatial distribution model are trained in parallel. In the parallel training process, feature verification is performed among the pollution mutation event identification model, the pollution trend prediction model, and the pollution spatial distribution model, and the deviation of pseudo-labels in the parallel training process is corrected through cross-checking.

10. A vehicle, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 9.

Images