Vehicle-mounted air quality identification method and vehicle

Abstract

The invention relates to a vehicle-mounted air quality recognition method and a vehicle, and belongs to the technical field of intelligent environment monitoring, and the method comprises the steps: obtaining image data, carbon dioxide concentration time sequence data and environment auxiliary data under different vehicle-mounted operation parameters, carrying out pollution level label labeling, and taking the labeled data as a multi-source training data set; a vehicle-mounted air quality recognition model is constructed, the vehicle-mounted air quality recognition model is trained based on the multi-source training data set, the vehicle-mounted air quality recognition model is used for conducting feature recognition on the multi-source training data set, and smoke particle size feature vectors and carbon dioxide concentration time sequence feature vectors are extracted; outputting a pollution level prediction result based on the feature vector and the environment auxiliary data; and acquiring real-time vehicle-mounted multi-source data, inputting the real-time vehicle-mounted multi-source data into the trained vehicle-mounted air quality recognition model, acquiring a pollution level prediction result, determining a corresponding control instruction, and controlling an air purification actuator to execute the control instruction.

Description

Technical Field

[0001] This application relates to the technical field of intelligent environmental monitoring, and in particular to a vehicle-mounted air quality identification method and vehicle. Background Technology

[0002] With the increasing intelligence of automobiles, in-vehicle air quality monitoring and purification systems have become an important component of modern vehicle environmental comfort and health protection. Common in-vehicle air quality monitoring solutions mainly rely on data from single or limited types of sensors, such as direct measurement and threshold judgment using carbon dioxide or particulate matter concentration sensors to trigger alarms when specific pollutants exceed limits. In recent years, with the development of multimodal data fusion technology, technical solutions have emerged that combine image processing with multiple physical sensors to enhance environmental perception and anomaly detection capabilities.

[0003] However, existing general environmental detection solutions based on multimodal fusion typically aim to identify specific abnormal targets and provide macro-level risk warnings in open spaces by fusing image and sensor data. These methods are not designed for the specific needs of air quality identification in the enclosed space of an in-vehicle environment. They lack deep joint modeling of the intrinsic correlation between the visual characteristics of pollution sources and the temporal changes in gas concentration, and fail to utilize the in-vehicle spatial topology to model the diffusion paths and distribution relationships of pollutants. Consequently, they cannot achieve precise identification of pollution types, accurate prediction of concentration evolution trends, or precise location of localized pollution areas within the vehicle. This results in insufficient recognition accuracy and coarse response strategies when facing dynamic, complex, and spatially constrained in-vehicle environments, making it difficult to meet the actual needs of passengers for real-time, accurate, and personalized air quality control in in-vehicle scenarios. Summary of the Invention

[0004] This application addresses, to at least some extent, one of the technical problems in the related art.

[0005] To achieve the above objectives, in a first aspect, this application provides a method for identifying in-vehicle air quality, comprising: Real-time vehicle-mounted multi-source data is acquired and input into a trained vehicle-mounted air quality recognition model to obtain pollution level prediction results; the real-time vehicle-mounted multi-source data includes image data, carbon dioxide concentration time-series data, and environmental auxiliary data. Determine the air purification control command corresponding to the pollution level prediction result, and control the air purification actuator to execute the air purification control command; The vehicle-mounted air quality identification model is used to identify features of the real-time vehicle-mounted multi-source data, extract smoke particle size feature vectors and carbon dioxide concentration time-series feature vectors, predict pollution levels based on the smoke particle size feature vectors, the carbon dioxide concentration time-series feature vectors and the environmental auxiliary data, and output pollution level prediction results.

[0006] The technical solution achieves high-precision identification and intelligent purification control of in-vehicle air pollution by introducing multimodal data fusion and deep neural network modeling. Specifically, through multi-scale image feature extraction, temporal feature compression of carbon dioxide concentration, and multi-source heterogeneous feature fusion and spectral modeling, visual, gas, and environmental parameter information are effectively integrated, significantly improving the accuracy of pollution level judgment and environmental adaptability. Simultaneously, control commands are dynamically generated based on pollution level prediction results and linked with air purification actuators to manage in-vehicle air quality, achieving closed-loop control from perception and decision-making to execution. This overcomes the problems of single perception dimension, delayed response, and lack of personalized adaptation in traditional methods, achieving real-time monitoring and precise treatment of in-vehicle air quality.

[0007] In some embodiments of this application, the vehicle air quality identification method further includes: Image data, carbon dioxide concentration time series data, and environmental auxiliary data under different vehicle operating parameters were acquired and labeled with pollution level. The labeled image data, carbon dioxide concentration time series data, and environmental auxiliary data were used as a multi-source training dataset. A vehicle air quality recognition model is constructed, and the vehicle air quality recognition model is trained based on the multi-source training dataset to obtain the trained vehicle air quality recognition model.

[0008] In this technical solution, multi-source data under different vehicle operating parameters is systematically collected and labeled to construct a multi-source training dataset that closely reflects the complex working conditions inside a real vehicle. This effectively solves the problems of scarce domain data and insufficient model generalization, providing a solid data foundation for the model. During model training, an end-to-end joint optimization mechanism drives the collaborative learning of modules such as image recognition, temporal analysis, and graph inference. This enables the model to automatically uncover deep correlations between heterogeneous features from multiple sources, achieving unified optimization of feature fusion and representation. Consequently, the overall accuracy, robustness, and environmental adaptability of the model are significantly improved.

[0009] In some embodiments of this application, the vehicle air quality recognition model includes: The image data is subjected to multi-scale feature extraction based on the smoke particle size multi-scale feature extraction branch to obtain the smoke particle size feature vector. The carbon dioxide concentration time-series data is compressed and features are extracted based on the carbon dioxide concentration time-series feature compression branch to obtain the carbon dioxide concentration time-series feature vector. Based on the multi-model heterogeneous feature fusion branch, feature alignment and weighted fusion are performed on the smoke particle size feature vector and the carbon dioxide concentration time series feature vector to generate a multi-modal fusion feature vector; Based on the air pollution level map branch, the multimodal fusion feature vector and the environmental auxiliary data are used for map inference to obtain the pollution level prediction result.

[0010] In this technical solution, a multi-branch vehicle-mounted air quality recognition model is constructed to achieve in-depth mining and effective fusion of multi-source pollution information. The model captures fine-grained visual features from images through a multi-scale feature extraction branch for smoke particle size, extracts dynamic change patterns of gas concentration through a carbon dioxide concentration temporal feature compression branch, and achieves cross-modal feature alignment and weighted fusion using a multi-model heterogeneous feature fusion branch. Finally, a structured pollution state map is constructed through an air pollution level map branch for inference and judgment. This modular and structured design allows the system to fully utilize the complementary information between images, temporal concentrations, and environmental parameters, significantly improving the accuracy and robustness of pollution level recognition. It also provides a reliable basis for subsequent precise purification control, effectively overcoming the limitations of traditional single-modal sensing methods.

[0011] In some embodiments of this application, the multi-model heterogeneous feature fusion branch includes: Based on the feature alignment module, the smoke particle size feature vector and the carbon dioxide concentration time series feature vector are projected to the same feature dimension through a fully connected layer to obtain the aligned feature vector. The attention weighting module receives the aligned feature vector, calculates the corresponding weight coefficients through a neural network, and performs a weighted summation on the aligned feature vector based on the weight coefficients to obtain a multimodal fusion feature vector.

[0012] In this technical solution, the multi-model heterogeneous feature fusion branch projects feature vectors from different modalities onto a unified dimensional space through a feature alignment module, effectively addressing the semantic gap between multi-source data. Furthermore, an attention-weighted module dynamically learns the importance weights of each modality's features, achieving adaptive feature fusion. This hierarchical fusion mechanism preserves the unique information of each modality's features while also uncovering their inherent correlations, significantly improving the representational ability of multi-source pollution features. This provides more discriminative feature inputs for subsequent pollution level identification, thereby enhancing the system's perception accuracy and decision reliability in complex in-vehicle environments.

[0013] In some embodiments of this application, the smoke particle size multi-scale feature extraction branch includes an image preprocessing module, a hierarchical feature extraction module, and a global feature aggregation module connected in sequence. The image preprocessing module performs image preprocessing on the image data and divides it into a sequence of image blocks. The hierarchical feature extraction module extracts smoke particle size features at different scales from the image patch sequence through a multi-level transformation layer containing sliding window attention and cross-window connections, thereby obtaining a multi-scale smoke feature map. The global feature aggregation module performs global average pooling on the multi-scale smoke feature map to generate a fixed-dimensional smoke particle size feature vector.

[0014] In this technical solution, a smoke particle size feature extraction process is constructed to achieve deep feature learning of image data. The image preprocessing module standardizes the original image and segments it into a sequence of image patches, laying the foundation for subsequent feature extraction. The hierarchical feature extraction module employs a multi-level transformation layer structure with sliding window attention and cross-window connections to effectively capture smoke particle size features at different scales, forming a multi-scale smoke feature map. Finally, the global feature aggregation module compresses the multi-scale features into a fixed-dimensional feature vector. This hierarchical processing architecture preserves local detail features while integrating global contextual information, significantly improving the expressive power and discriminative power of smoke particle size features, and providing high-quality visual feature input for subsequent multimodal fusion and pollution level identification.

[0015] In some embodiments of this application, the carbon dioxide concentration time-series feature compression branch includes a data preprocessing module and a stacked autoencoder module connected in sequence. The data preprocessing module normalizes the carbon dioxide concentration time series data to construct a standardized concentration time series. The stacked autoencoder module nonlinearly maps the standardized concentration time series to a low-dimensional latent space through its encoder, and outputs a fixed-dimensional carbon dioxide concentration time series feature vector.

[0016] In this technical solution, a time-series feature extraction process for carbon dioxide concentration is constructed, enabling the effective extraction of dynamic changes in gas concentration. The data preprocessing module normalizes the raw concentration data, eliminating dimensional differences and constructing a standardized time-series sequence. The stacked autoencoder module nonlinearly maps the standardized sequence to a low-dimensional latent space through its encoder, extracting the essential features characterizing the concentration change trend. This not only effectively reduces data dimensionality and removes redundant information but also preserves the time-series patterns and dynamic features of concentration changes, providing a high-quality gas concentration feature representation for subsequent multimodal feature fusion and enhancing the system's ability to perceive pollution conditions.

[0017] In some embodiments of this application, the encoder of the stacked autoencoder module includes: Multiple fully connected layers are used to perform a layer-by-layer dimensionality reduction nonlinear transformation on the standardized concentration time series. Each fully connected layer is followed by a batch normalization layer, a nonlinear activation function layer, and a random deactivation layer. The low-dimensional vector output by the last fully connected layer of the encoder is used as the temporal feature vector of carbon dioxide concentration.

[0018] In the technical solution, the encoder of the stacked autoencoder module adopts a composite structure of multiple fully connected layers combined with batch normalization, nonlinear activation, and random deactivation layers to achieve deep feature learning of time-series concentration data. Through layer-by-layer nonlinear transformation, the essential features in the concentration sequence are effectively extracted. The batch normalization layer ensures training stability, the nonlinear activation enhances the model's expressive power, and the random deactivation layer effectively prevents overfitting. This hierarchical encoding structure significantly improves the model's generalization performance and robustness while maintaining feature expressive power, providing high-quality, low-dimensional temporal feature representations for subsequent multimodal fusion.

[0019] In some embodiments of this application, the air pollution level map branch includes a map construction module and a graph convolution inference module; The graph construction module uses the smoke particle size feature vector, the carbon dioxide concentration time series feature vector, and the environmental auxiliary feature vector converted from the environmental auxiliary data as nodes, and constructs weighted edges based on predefined physical associations and statistical correlations between nodes to form a dynamic heterogeneous graph. The graph convolutional inference module performs neighborhood information aggregation and feature updates on the dynamic heterogeneous graph through a multi-layer graph convolutional network, and outputs the pollution level prediction result through a classifier.

[0020] In this technical solution, the air pollution level mapping branch constructs a dynamic heterogeneous graph and performs graph convolutional inference, achieving structured modeling of multi-source pollution characteristics. The graph construction module uses feature vectors from different modalities as nodes, establishing weighted edges based on physical associations and statistical correlations to form a graph structure that represents the spatial relationships of pollution states. The graph convolutional inference module aggregates and iteratively updates the node features using a multi-layer graph convolutional network, effectively capturing the complex interactions between pollution factors. This graph-based modeling method not only fully utilizes the inherent correlations between multi-source features but also enhances the system's inference ability for complex pollution scenarios, significantly improving the accuracy and robustness of pollution level identification and providing a reliable decision-making basis for precise purification control.

[0021] In some embodiments of this application, determining the air purification control command corresponding to the pollution level prediction result and controlling the air purification actuator to execute the air purification control command includes: Based on the pollution level prediction results, the preset strategy mapping table is queried to obtain the query results; Based on the query results, an air purification control command is generated. The air purification control command includes at least a command to switch the vehicle air conditioning circulation mode, a command to adjust the power of the air purifier, and a prompt command for the user interface. The control commands are sent to the corresponding air purification actuators via the vehicle bus to coordinate the air purification operation.

[0022] The technical solution constructs a complete air purification control execution mechanism, achieving intelligent closed-loop control from pollution identification to purification execution. Based on pollution level prediction results and real-time vehicle operating parameters, a preset strategy mapping table is queried to ensure the accuracy of control decisions and scenario adaptability. This generates multi-dimensional control commands, including air conditioning mode switching, purifier power adjustment, and user prompts, achieving synergistic effects of purification measures. Finally, the commands are distributed to each actuator via the vehicle bus, ensuring efficient transmission and reliable execution of control signals. This structured control mechanism considers the differentiated response requirements under different pollution levels and fully integrates the actual vehicle operating status, achieving efficient and precise treatment of in-vehicle air quality, significantly improving the system's practicality and user experience.

[0023] In a second aspect, the present invention provides a vehicle for implementing the in-vehicle air quality identification method as described in the first aspect, comprising: The data acquisition unit is used to acquire image data, carbon dioxide concentration time-series data, and environmental auxiliary data. The controller is communicatively connected to the data acquisition unit and is used to store at least an in-vehicle air quality identification model, obtain pollution level prediction results based on the in-vehicle air quality identification model, and determine the air purification control command corresponding to the pollution level prediction results. An air purification actuator, communicatively connected to the controller, is used to receive and execute the air purification control commands.

[0024] In this technical solution, multi-source environmental data is acquired through a data acquisition unit. The controller's built-in air quality identification model determines the pollution level and generates control commands, which are then used to adjust the environment via an air purification actuator. This ensures closed-loop control throughout the entire process, from data acquisition and intelligent analysis to purification execution. This not only improves the system's reliability and response speed but also achieves good adaptability to different vehicle platforms, providing a complete solution for intelligent management of in-vehicle air quality.

[0025] As can be seen from the above technical solutions, additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the in-vehicle air quality identification method according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an in-vehicle air quality identification model according to an embodiment of this application; Figure 3 This is a schematic diagram of the branch structure for extracting multi-scale features of smoke particle size according to an embodiment of this application; Figure 4 This is a schematic diagram of the compression branch structure based on the time-series characteristics of carbon dioxide concentration according to an embodiment of this application; Figure 5 A schematic diagram of the multi-model heterogeneous feature fusion branch structure according to an embodiment of this application; Figure 6 This is a schematic diagram of the branch structure of the air pollution level map according to the embodiments of this application; Figure 7 This is a schematic diagram of the vehicle air quality model training process according to the embodiments of this application; Figure 8 This is a structural schematic diagram of a vehicle according to an embodiment of this application; Figure 9 This is a schematic diagram of a computer device according to an embodiment of this application.

[0027] In the above figures: 1. Data acquisition unit; 2. Controller; 3. Air purification actuator; 40. Bus; 41. Processor; 42. Memory; 43. Communication interface; 100. Multi-scale feature extraction branch for smoke particle size; 200. Temporal feature compression branch for carbon dioxide concentration; 300. Multi-model heterogeneous feature fusion branch; 400. Air pollution level map branch; 110. Image preprocessing module; 120. Hierarchical feature extraction module; 130. Global feature aggregation module; 210. Data preprocessing module; 220. Stacked autoencoder module; 310. Feature alignment module; 320. Attention weighting module; 410. Graph construction module; 420. Graph convolution inference module. Detailed Implementation

[0028] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicating the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between components; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature. In this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0029] The present application will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments. In the automotive industry, the level of intelligence in in-vehicle air quality monitoring and purification systems directly impacts the comfort and health of the driving environment. With the continuous improvement of vehicle intelligence and increased consumer health awareness, in-vehicle air quality monitoring has evolved from simple temperature and humidity detection to comprehensive monitoring of multiple pollutants.

[0030] The characterization data of in-vehicle air quality mainly includes, but is not limited to, the following categories: Particulate matter concentration indicators include data on the concentration of particulate matter of different sizes, such as PM1.0, PM2.5, and PM10. These parameters can typically be monitored in real time using laser scattering sensors and can reflect the pollution level of solid pollutants such as dust and smoke inside the vehicle. Among them, PM2.5 has become a key monitoring indicator due to its significant impact on human health.

[0031] The concentration indicators for gaseous pollutants cover carbon dioxide (CO2), carbon monoxide (CO), nitrogen oxides (NOx), ozone (O3), and volatile organic compounds (TVOC). Among them, CO2 concentration is a key indicator for evaluating the ventilation efficiency of the vehicle interior; when the concentration exceeds 1000 ppm, it can reduce the attention of occupants. TVOC includes harmful substances such as formaldehyde and benzene compounds, which mainly come from interior trim materials.

[0032] Bioaerosol indicators include the concentration of microbial particles such as bacteria, fungi, and viruses. These data are typically acquired using specific biosensors and are of great significance for preventing the spread of respiratory diseases, especially during peak epidemic periods or in scenarios involving multiple people sharing a ride.

[0033] Environmental parameters, including basic physical parameters such as temperature, humidity, and air pressure, not only directly affect passenger comfort but also have synergistic effects with other pollutants. For example, high temperature and high humidity environments accelerate the release of TVOCs, while appropriate humidity levels can effectively inhibit virus transmission.

[0034] Currently, mainstream vehicle air quality systems primarily use a network of gas sensors deployed throughout the vehicle to monitor in real time the concentration of particulate matter, including PM2.5 and PM10, as well as the content of gaseous pollutants such as carbon dioxide, nitrogen oxides, formaldehyde, and total volatile organic compounds.

[0035] In the course of technological development, early systems mainly relied on single-type sensors for independent monitoring, but with technological advancements, they gradually evolved into multi-sensor fusion monitoring.

[0036] However, existing technologies still have significant limitations. First, most current vehicle-mounted air quality monitoring systems rely primarily on point-based gas concentration monitoring, lacking the ability to acquire and analyze visual information such as the morphology and distribution of pollution sources. For example, in smog pollution scenarios, particulate matter concentration alone cannot distinguish the source and nature of pollution, while image information can provide key features such as the smog's diffusion range and concentration distribution. Second, the control strategies of existing vehicle-mounted air quality monitoring systems are relatively simple, often employing fixed threshold triggering mechanisms, and cannot perform predictive control based on pollution trends. Furthermore, existing vehicle-mounted air quality monitoring systems generally lack comprehensive consideration of contextual information such as vehicle operating status and occupant behavior, making it difficult to achieve truly intelligent and personalized air quality management.

[0037] In complex and ever-changing real-world driving environments, systems relying solely on traditional sensor data often experience false alarms and missed alarms, and their control responses lack accuracy and timeliness. Especially in typical scenarios such as new car odors, external air pollution, and smoking inside the vehicle, existing systems struggle to provide satisfactory solutions. With the rapid development of intelligent cockpit technology in automobiles, users are placing higher demands on the quality of the in-vehicle environment, creating an urgent need to develop more intelligent and accurate in-vehicle air quality identification and purification technologies.

[0038] Based on this, this application proposes an in-vehicle air quality identification method and vehicle. By constructing a multimodal identification model that integrates images, gas concentrations and environmental parameters, and generating accurate air purification control commands based on the pollution level prediction results output by the model, it achieves high-precision identification and adaptive purification control of in-vehicle air pollution status. This solves the problems of low identification accuracy, lagging control strategies and poor system adaptability caused by the single perception information and lack of multimodal fusion capability in the existing technology.

[0039] In the following, embodiments of this application will be described in detail with reference to the accompanying drawings.

[0040] As attached Figures 1 to 6 As shown in an illustrative embodiment of a vehicle air quality identification method according to this application, the vehicle air quality identification method includes: S1: Acquire real-time vehicle-mounted multi-source data, input the real-time vehicle-mounted multi-source data into the trained vehicle-mounted air quality recognition model, and obtain pollution level prediction results; real-time vehicle-mounted multi-source data includes image data, carbon dioxide concentration time series data, and environmental auxiliary data. The vehicle-mounted air quality identification model is used to identify features from real-time multi-source vehicle data, extract smoke particle size feature vectors and carbon dioxide concentration time-series feature vectors, and predict pollution levels based on smoke particle size feature vectors, carbon dioxide concentration time-series feature vectors and environmental auxiliary data, and output pollution level prediction results.

[0041] In some embodiments, real-time vehicle-mounted multi-source data mainly includes three core types of data: image data, carbon dioxide concentration time-series data, and environmental auxiliary data.

[0042] The image data is collected by high-definition cameras deployed in areas such as the car's interior dome lights and A-pillars, with a resolution of no less than 720P and a frame rate of no less than one frame per second.

[0043] Carbon dioxide concentration time-series data are collected by sensors deployed in the central control area. The sampling frequency can be set to 0.5-1Hz. The carbon dioxide concentration time-series data are carbon dioxide concentration values ​​with timestamps, and the unit is ppm.

[0044] Environmental auxiliary data includes pollution parameters such as PM2.5 concentration, temperature and humidity, and TVOC, as well as vehicle operating parameters such as air conditioning status, window status, and number of occupants. Environmental auxiliary data is obtained through the vehicle gateway or CAN bus.

[0045] Furthermore, to ensure data quality, all collected data must be time-aligned by a time synchronization module, and data consistency is guaranteed based on the system RTC (Real-Time Clock) timestamp and a custom software marking mechanism.

[0046] The tagging mechanism compensates for insufficient hardware clock precision by embedding logical timestamps and event markers into the data stream, ensuring logical timing consistency of multi-source data. Simultaneously, a real-time data quality detection mechanism is introduced to identify and handle issues such as image blurring, sensor failure, and numerical anomalies.

[0047] The following is combined Figures 2-6 This paper details the structure of the vehicle-mounted air quality identification model, its specific structure, and the process for obtaining pollution level prediction results.

[0048] In some embodiments, combined with Figure 2 As shown, the vehicle-mounted air quality identification model includes a smoke particle size multi-scale feature extraction branch 100, a carbon dioxide concentration temporal feature compression branch 200, a multi-model heterogeneous feature fusion branch 300, and an air pollution level map branch 400.

[0049] Based on the multi-scale feature extraction branch 100 for smoke particle size, multi-scale feature extraction is performed on the image data to obtain the smoke particle size feature vector. Based on the carbon dioxide concentration time series feature compression branch 200, time series pattern compression and feature extraction are performed on carbon dioxide concentration time series data to obtain carbon dioxide concentration time series feature vector; Based on the multi-model heterogeneous feature fusion branch 300, feature alignment and weighted fusion of smoke particle size feature vector and carbon dioxide concentration time series feature vector are performed to generate multi-modal fusion feature vector; Based on the 400-branch multimodal fusion feature vector and environmental auxiliary data of the air pollution level map, the pollution level prediction results are obtained by map inference.

[0050] In some embodiments, combined with Figure 3 As shown, the smoke particle size multi-scale feature extraction branch 100 includes an image preprocessing module 110, a hierarchical feature extraction module 120, and a global feature aggregation module 130 connected in sequence.

[0051] In some embodiments, the image preprocessing module 110 performs image preprocessing on the image data and divides it into a sequence of image blocks.

[0052] Preferably, the image preprocessing module 110 performs color normalization and illumination enhancement operations on the input RGB three-channel image data frame to eliminate visual deviations caused by changes in ambient light inside the vehicle and improve the visibility of smoke areas. The preprocessed image data is divided into several image blocks through a fixed window segmentation operation, forming an image block sequence that is input to subsequent modules.

[0053] In some embodiments, the hierarchical feature extraction module 120 extracts smoke particle size features at different scales from an image patch sequence through a multi-level transformation layer that includes sliding window attention and cross-window connections, thereby obtaining a multi-scale smoke feature map.

[0054] Preferably, the hierarchical feature extraction module 120 is constructed using a multi-level Swing Transformer layer. Swing Transformer (Shifted Window-based Transformer) is a hierarchical visual Transformer architecture based on a window shifting mechanism. By introducing local window self-attention and cross-window connections, it significantly reduces computational complexity while maintaining the powerful modeling capabilities of Transformer.

[0055] Furthermore, each layer of the hierarchical feature extraction module 120 includes an in-window self-attention mechanism, cross-window connections, layer normalization, and a feedforward neural network.

[0056] Specifically, each processing level of the hierarchical feature extraction module 120 implements multi-scale feature learning through the following components.

[0057] The in-window self-attention mechanism first recombines the input image patch sequence into multiple non-overlapping local windows, and independently calculates the self-attention weights within each window. While significantly reducing computational complexity, it can effectively capture the morphological features, density distribution, and spatial relationships of smoke particles within a local region.

[0058] Cross-window connectivity introduces a window shifting mechanism between adjacent layers, causing the window partitioning of the current layer to shift regularly relative to the previous layer. This breaks the boundary constraints of fixed windows, establishes a cross-window feature propagation path, and effectively integrates global contextual information such as smoke diffusion trends and asymmetric edge distribution.

[0059] The layer normalization component is deployed before each self-attention layer and feedforward network. By re-centering and scaling the activation values, it stabilizes the feature distribution and ensures the training stability and convergence efficiency of deep networks when processing complex smoke images.

[0060] The feedforward neural network employs a two-layer fully connected structure, with a nonlinear transformation introduced through the GELU activation function in between, further enhancing the model's ability to represent complex morphological features such as changes in smoke particle size and concentration gradients. Residual connections are used between all components, effectively mitigating the gradient vanishing problem in deep networks and ensuring the smooth transmission of feature information.

[0061] Furthermore, the hierarchical feature extraction module 120 introduces a hierarchical downsampling strategy into the deep network, gradually reducing the spatial resolution of the feature map through four stages of feature transformation, thereby constructing a multi-scale feature pyramid with different receptive fields. This enables the model to simultaneously capture local fine structures and global distribution patterns, ultimately outputting a smoke feature map that integrates multi-scale information, effectively improving the recognition ability of complex scenes such as sparse smoke distribution and asymmetric diffusion edges.

[0062] In some embodiments, the global feature aggregation module 130 performs global average pooling on the multi-scale smoke feature map to generate a fixed-dimensional smoke particle size feature vector.

[0063] Preferably, the global feature aggregation module 130 performs a global average pooling operation on the multi-scale smoke feature map output by the hierarchical feature extraction module 120, compressing the multi-dimensional feature tensor into a fixed-dimensional feature vector. This vector is further mapped through a fully connected layer, ultimately generating a fixed-dimensional smoke particle size feature vector characterizing the smoke particle size distribution.

[0064] Furthermore, during the model training phase, a labeling mechanism based on simulated experimental cabin environment data was adopted. A high-precision particle size analyzer was used to label images under different particle size conditions, establishing a precise correspondence between image data and particle size distribution. During training, a multi-scale smoke sample set was constructed, and data augmentation was performed using typical in-vehicle environmental disturbances such as occlusion, blurring, and backlighting to improve the model's robustness in real-world in-vehicle environments.

[0065] The training process employs an iterative optimization strategy. First, pre-training is performed on a general image dataset. Then, transfer learning is conducted using image data collected from actual in-vehicle scenarios, enabling the model to better adapt to complex lighting conditions, material properties, and reflection interference within the vehicle. Through multiple rounds of fine-tuning, the model gradually masters the unique visual characteristics of smoke inside the vehicle, improving its generalization ability in real-world environments.

[0066] The multi-scale feature extraction branch 100 for smoke particle size ultimately outputs a fixed-dimensional visual feature vector, fully representing the particle size distribution of smoke in the image. This feature vector will serve as input to key nodes in subsequent graph neural networks, participating in multimodal feature fusion and pollution level determination. By introducing the Swing Transformer structure, this module overcomes the limitations of traditional convolutional neural networks in terms of receptive field, demonstrating excellent feature extraction capabilities when faced with complex and varied smoke diffusion patterns inside vehicles.

[0067] Furthermore, the smoke particle size multi-scale feature extraction branch 100 is deployed on the vehicle's main control chip using an edge computing mode, possessing real-time processing capabilities and meeting the image processing requirement of one frame per second. It supports flexible adaptation to different camera models and resolutions, demonstrating excellent hardware and software decoupling characteristics, thus facilitating the deployment of the system on different vehicle platforms.

[0068] In some embodiments, combined with Figure 4 As shown, the carbon dioxide concentration time-series feature compression branch 200 includes a data preprocessing module 210 and a stacked autoencoder module 220 connected in sequence.

[0069] Preferably, the carbon dioxide concentration temporal feature compression branch 200 is constructed based on a stacked autoencoder (SAE). A SAE is a deep neural network structure composed of multiple stacked autoencoder layers. It achieves efficient feature compression and abstract representation of input data through a combination of layer-by-layer unsupervised pre-training and supervised fine-tuning.

[0070] In some embodiments, the data preprocessing module 210 normalizes the carbon dioxide concentration time series data to construct a standardized concentration time series.

[0071] Preferably, the data preprocessing module 210 performs systematic preprocessing on the input carbon dioxide concentration time series data.

[0072] The data preprocessing module 210 receives carbon dioxide concentration sequences divided into fixed time windows. Each sequence contains continuous concentration values ​​within 10 to 15 seconds, and the sampling frequency is not less than 1 Hz, forming a one-dimensional floating-point array.

[0073] Furthermore, the data preprocessing module 210 performs multi-stage purification processing on the raw carbon dioxide concentration data. First, it identifies and corrects abnormal abrupt changes using a statistical outlier detection algorithm, and combines this with the sliding window mean method to achieve local smoothing, effectively suppressing random noise interference. Then, it performs zero-mean normalization processing, which eliminates the concentration baseline differences caused by vehicle model differences or environmental conditions by subtracting the sequence mean and dividing by the standard deviation, highlighting the relative trend characteristics of concentration changes.

[0074] After preprocessing by the data preprocessing module 210, a standardized carbon dioxide concentration time series that meets the model input requirements is constructed, laying the data foundation for subsequent feature extraction. This preprocessing process effectively improves data quality and ensures the stability and reliability of time series feature learning.

[0075] In some embodiments, the stacked autoencoder module 220 nonlinearly maps the standardized concentration time series to a low-dimensional latent space through its encoder, and outputs a fixed-dimensional carbon dioxide concentration time series feature vector.

[0076] In some embodiments, the encoder of the stacked autoencoder module 220 includes multiple layers of fully connected layers connected in sequence for performing a nonlinear transformation of the normalized concentration time series by layer-by-layer dimensionality reduction. Each fully connected layer is followed by a batch normalization layer, a nonlinear activation function layer, and a random deactivation layer. The low-dimensional vector output by the last fully connected layer of the encoder is used as the temporal feature vector of carbon dioxide concentration.

[0077] Preferably, the encoder of the stacked autoencoder module 220 includes three fully connected layers connected in sequence, forming a nonlinear transformation path of dimensionality reduction layer by layer.

[0078] Specifically, the first layer maps the input sequence to a 128-dimensional feature space, the second layer further compresses it to 64 dimensions, and the third layer finally encodes a 32-dimensional latent representation. The low-dimensional vector output by the fully connected third layer of the encoder serves as the temporal feature vector of carbon dioxide concentration.

[0079] Each fully connected layer is followed by a batch normalization layer, a ReLU activation function layer, and a Dropout layer in sequence. Batch normalization ensures training stability, ReLU activation introduces non-linear modeling capabilities, and the Dropout layer prevents overfitting through random deactivation.

[0080] It should be noted that the above settings for the number of layers and dimensions of the fully connected layer are illustrative and can be adjusted according to the computing resources, sensor sampling characteristics, and accuracy requirements of the specific vehicle model in actual applications.

[0081] For example, in entry-level vehicles with limited computing resources, a simplified two-layer coding structure can be used to reduce computational load while ensuring basic feature extraction capabilities. In enhanced configurations of high-end vehicles, this can be expanded to a four-layer coding structure to obtain more refined temporal feature representations. Furthermore, for sensor systems with high sampling frequencies, the dimensions of the first-layer network can be appropriately increased to 256 or 512 dimensions to better capture high-frequency fluctuation features.

[0082] Furthermore, the decoder portion of the stacked autoencoder module 220 adopts a network structure symmetrical to the encoder, progressively reconstructing the 32-dimensional latent features into the original sequence dimensions through three fully connected layers. During training, the goal is to minimize the reconstruction error, forcing the encoder to learn a compressed representation with high information fidelity.

[0083] Furthermore, during the model training phase, historical real-vehicle data containing different vehicle models, climate conditions, and ventilation settings were used to optimize the network parameters through unsupervised learning. The training objective was to minimize the mean absolute error between the original input and the reconstructed output, while auxiliary supervision signals were used to ensure that the compressed features had sufficient discriminative power for the pollution level classification task.

[0084] The carbon dioxide concentration temporal feature compression branch 200 ultimately outputs a fixed-dimensional carbon dioxide concentration temporal feature vector. This feature vector effectively captures the trend patterns, fluctuation characteristics, and dynamic laws of concentration changes, serving as the key node attributes input to subsequent graph neural networks. The carbon dioxide concentration temporal feature compression branch 200 optimizes computational complexity for in-vehicle environments and can be deployed on low-load cores of in-vehicle main control platforms, enabling low-power real-time inference at the edge.

[0085] In some embodiments, combined with Figure 5 As shown, the multi-model heterogeneous feature fusion branch 300 includes a feature alignment module 310 and an attention weighting module 320.

[0086] Preferably, the multi-model heterogeneous feature fusion branch 300 adopts a multi-stage hierarchical structure, including a feature alignment module 310, an attention weighting module 320, a cross-modal residual enhancement module, and a feature consistency monitoring module connected in sequence.

[0087] In some embodiments, the feature alignment module 310 projects the smoke particle size feature vector and the carbon dioxide concentration time series feature vector onto the same feature dimension through a fully connected layer to obtain the aligned feature vector.

[0088] Preferably, the feature alignment module 310 projects the smoke particle size feature vector and the carbon dioxide concentration time-series feature vector to the same feature space dimension through two independent trainable fully connected layers, effectively solving the problem of differences in the dimensions and distribution of different modal features, and establishing a unified representation basis for subsequent feature fusion.

[0089] In some embodiments, the attention weighting module 320 receives the aligned feature vector, calculates the corresponding weight coefficients through a neural network, and performs a weighted summation on the aligned feature vector based on the weight coefficients to obtain a multimodal fusion feature vector.

[0090] Preferably, the attention weighting module 320 uses an attention network based on a multilayer perceptron to achieve adaptive weighting of features of different modalities.

[0091] The attention weighting module 320 first concatenates the aligned smoke particle size feature vector and the carbon dioxide concentration time-series feature vector, and then calculates the attention weights through a neural network containing two fully connected layers. The first layer uses the ReLU activation function (non-linear activation function) for non-linear transformation, and the second layer uses the Softmax function (normalized exponential function) to generate normalized weight coefficients.

[0092] The attention weighting module 320 can dynamically adjust the contribution of different modal features according to the current in-vehicle environment. For example, in a static state with low smoke concentration and minimal image changes, the carbon dioxide concentration time-series feature vector should have a higher decision weight; while in high-temperature and high-humidity images where smoke visibility is high, the participation of the smoke particle size feature vector in the decision-making process should be increased.

[0093] The attention weighting module 320 sums each feature vector with its corresponding weight coefficient to generate a multimodal fusion feature vector that can adaptively highlight key pollution information, providing a more discriminative feature representation for subsequent pollution level identification.

[0094] Preferably, the multi-model heterogeneous feature fusion branch 300 also includes a cross-modal residual enhancement module, which retains a direct path for the original feature vector outside the main fusion path to prevent information from being diluted due to projection transformation or weighting mechanism during the fusion process, ensuring that nodes in the final graph retain their key semantic features and preventing the loss of effective features during the fusion process.

[0095] Preferably, the multi-model heterogeneous feature fusion branch 300 also includes a feature consistency monitoring module, which periodically evaluates the consistency between the smoke particle size feature vector and the carbon dioxide concentration time-series feature vector in judging pollution trends. When a significant discrepancy is detected, the feature consistency monitoring module dynamically adjusts the fusion strategy, temporarily increasing the fusion weight of environmental auxiliary features to ensure the reliability of the system's decision-making under complex operating conditions.

[0096] Furthermore, the multi-model heterogeneous feature fusion branch 300 supports parameter fine-tuning based on vehicle characteristics. For example, for vehicles with smaller cabin volume and faster gas accumulation, the attention weight base value of the carbon dioxide concentration time-series feature vector can be increased during system initialization; while for vehicles with efficient air circulation systems, the priority of smoke particle size feature vector in the initial pollution judgment can be enhanced.

[0097] The multi-model heterogeneous feature fusion branch 300 is designed with real-time performance in mind. All vector alignment, weighting calculation, and graph embedding operations can be completed within a second-level processing cycle, ensuring the system has the ability to assess pollution levels at high frequencies and adapt to dynamic vehicle operation scenarios. The fusion process is highly decoupled, facilitating independent upgrades or replacements of subsequent modules and providing an interface foundation for the future introduction of other modal sensors, such as odor recognition and acoustic detection.

[0098] The multi-model fusion achieved through the multi-model heterogeneous feature fusion branch 300 enables the system to simultaneously understand image, time series, and structured pollution signals, effectively enhancing the accuracy and robustness of pollution level determination. This is significantly superior to traditional single-modal perception models, providing a reliable and continuous decision-making basis for subsequent control strategies.

[0099] In some embodiments, combined with Figure 6 As shown, the air pollution level map branch 400 includes a map construction module 410 and a graph convolution inference module 420.

[0100] Preferably, the air pollution level map branch 400 can be constructed based on a graph convolutional network (GCN). GCN is a deep learning model specifically designed for processing graph-structured data, which can effectively capture complex topological relationships and dependency structures in a graph through information transfer and aggregation between neighboring nodes.

[0101] In some embodiments, the graph construction module 410 uses the smoke particle size feature vector, the carbon dioxide concentration time series feature vector, and the environmental auxiliary feature vector converted from environmental auxiliary data as nodes, and constructs weighted edges based on predefined physical associations and statistical correlations between nodes to form a dynamic heterogeneous graph.

[0102] Preferably, the map construction module 410 uses the smoke particle size feature vector, the carbon dioxide concentration time series feature vector from the multimodal fusion branch, and the environmental auxiliary feature vector obtained by converting PM2.5 concentration, TVOC concentration, temperature and humidity, air conditioning status, window status and number of occupants as nodes of the heterogeneous map.

[0103] Furthermore, in the construction of edges, edge weights are set based on the physical association and statistical correlation between nodes.

[0104] For example, a strong weighted edge is established between the smoke node and the PM2.5 node to reflect the correlation between visual perception and particulate matter concentration; a dynamic weighted edge is constructed between the carbon dioxide node and the occupant number node to characterize the impact of human activities on gas concentration; and a moderately weighted edge is established between the temperature and humidity node and all pollution nodes to reflect the modulating effect of environmental parameters on pollution diffusion.

[0105] Furthermore, the graph construction module 410 also introduces a dynamic edge weight adjustment mechanism, which updates the edge weights based on real-time data, enabling the graph to adapt to different working conditions.

[0106] In some embodiments, the graph convolutional inference module 420 performs neighborhood information aggregation and feature update on dynamic heterogeneous graphs through a multi-layer graph convolutional network, and outputs pollution level prediction results through a classifier.

[0107] Preferably, the graph convolutional inference module 420 adopts a two-layer graph convolutional network architecture.

[0108] The first-layer graph convolutional network aggregates features from directly adjacent nodes to capture local pollution association patterns; the second-layer graph convolutional network extends to the second-order neighborhood, enhancing the system's ability to perceive complex pollution propagation structures.

[0109] Each layer of the graph convolutional network architecture includes operations such as linear transformation, weighted summation of neighborhood features, and non-linear activation, while residual connections preserve the integrity of the node's own features. Finally, global pooling and a softmax classifier are used to output the pollution level prediction result.

[0110] It should be noted that the number of layers in the multi-layer graph convolutional network is set as an example, and can be adjusted according to specific needs in actual applications.

[0111] Furthermore, the output of branch 400 of the air pollution level map is the predicted pollution level, in the form of multi-class labels, corresponding to the air quality grading standards set internally by the car manufacturer, such as five levels: excellent, good, lightly polluted, moderately polluted, and severely polluted. The final output of the model can be a single-value label or a pollution level score vector with a probability distribution, which is used for fine-grained mapping of subsequent control strategies and optimization of boundary strategies.

[0112] Furthermore, during model training, real-vehicle data samples labeled with pollution levels were used, covering typical scenarios such as different vehicle interior configurations, occupant states, and external environments. Pollution level labels were derived from simultaneous detection results by laser particle size analyzers and gas analyzers, and a quality assessment system was established by combining subjective driver experience feedback. During the training phase, a cross-entropy loss function was used for multi-class supervised learning, while a graph structure consistency regularization term was introduced to ensure that the model learns stable and transferable pollution pattern maps.

[0113] As the core of multimodal sensing information fusion, branch 400 of the air pollution level map directly determines the response logic and control level of the subsequent air purification system. Through graph structure modeling, not only is the system's ability to model the interactions between pollution sources improved, but it also has the ability to perform structural reasoning when nodes are incomplete or data is missing, significantly enhancing the model's stability and robustness in complex in-vehicle scenarios.

[0114] The Air Pollution Level Map Branch 400 is deployed in the vehicle's main control platform. It can support periodic updates to the map structure and perform real-time inference based on a sliding time window. It has high-frequency pollution identification capabilities and meets the response requirements of dynamic vehicle-mounted air treatment systems.

[0115] S2: Determine the air purification control command corresponding to the pollution level prediction result, and control the air purification actuator 3 to execute the air purification control command.

[0116] In some embodiments, determining the air purification control command corresponding to the pollution level prediction result and controlling the air purification actuator 3 to execute the air purification control command includes: Based on the pollution level prediction results, query the preset strategy mapping table to obtain the query results; Based on the query results, air purification control commands are generated. The air purification control commands include at least commands to switch the vehicle air conditioning circulation mode, commands to adjust the power of the air purifier, and prompts for the user interface. Control commands are sent to the corresponding air purification actuator 3 via the vehicle bus to coordinate the air purification operation.

[0117] Preferably, the input to the control system is the pollution level prediction result output by the air pollution level map branch 400, specifically the pollution level label or score result.

[0118] Furthermore, the pollution levels in the preset strategy mapping table are typically divided into five levels: excellent, good, lightly polluted, moderately polluted, and severely polluted. After receiving the pollution level prediction result, the control system first performs a decision conversion on the current level through the strategy mapping module, and then matches the corresponding control response according to the preset strategy table.

[0119] It should be noted that the strategy mapping table is defined by the car manufacturer's engineering team based on factors such as the actual vehicle's cabin ventilation efficiency, purification module performance, and user comfort feedback, and can be dynamically adjusted in subsequent OTA (Over-The-Air) updates.

[0120] Furthermore, based on the query results, control commands with at least three dimensions are generated, specifically commands to switch the vehicle air conditioning circulation mode, commands to adjust the power of the air purifier, and prompts for the user interface.

[0121] Specifically, in terms of environmental control, the heating, ventilation and air conditioning system (HVACS) is automatically switched according to the pollution level, such as switching from external circulation to internal circulation, increasing the wind speed level, and closing the front compartment air outlet.

[0122] In terms of purification control, the working power and operating mode of the air purifier can be adjusted, including the multi-stage filter start / stop combination and the intensity of the ion generator. For example, the PM2.5 filter, ionization purifier, activated carbon adsorption system and other devices can be activated, and the working time and frequency can be set according to the pollution level.

[0123] In terms of user interface, tiered alerts are provided through a multimodal human-machine interface, including dashboard pollution level indicator lights, voice broadcasts of air quality status, and vehicle infotainment system push notifications for window opening or air purification.

[0124] Furthermore, the control mechanism uses a state mapping diagram to determine the combination of actions, ensuring that control commands do not conflict.

[0125] For example, when the pollution level is severe, the system will prioritize shutting off the external air circulation, forcibly switching to internal air circulation mode, activating the purification module at maximum power, and displaying a message on the instrument panel stating "The air pollution inside the vehicle is severe; please do not open the windows." However, when the pollution level is only light, the system may only increase the airflow rate and prompt the user to manually operate the purification function.

[0126] It should be noted that the state mapping diagram can be refined according to vehicle model and driving scenario, such as dynamic weighted adjustment of strategy under conditions such as long-term idling, closed parking, and high occupant density.

[0127] Furthermore, to achieve precise control response, the system needs to communicate with multiple onboard ECUs (Electronic Control Units) via the CAN bus. Specific control signals must be encapsulated according to vehicle communication protocol specifications and interface with modules such as the HVAC (Heating, Ventilation and Air Conditioning) control unit, BCM (Body Control Module), and gateway controller. All commands support status readback functionality to ensure successful execution or record of failure states. To reduce system energy consumption, the purification module has maximum continuous operating time and periodic start-up limits to avoid prolonged high-power operation that could lead to a degraded user experience or energy waste.

[0128] In some embodiments, the following is combined with Figure 7 The training process of the vehicle air quality recognition model is described in detail.

[0129] The training process for the vehicle air quality recognition model includes: S100: Acquire image data, carbon dioxide concentration time series data, and environmental auxiliary data under different vehicle operating parameters, and label them with pollution level labels. Use the labeled image data, carbon dioxide concentration time series data, and environmental auxiliary data as a multi-source training dataset.

[0130] In some embodiments, step S100 is for constructing a multi-source training dataset, which aims to build a high-quality, multimodal dataset for model training.

[0131] The training data comes from two parts: real vehicle data and simulated test environment data. The real vehicle data covers different models, configurations, number of passengers and ventilation status under multiple regions and climate conditions. The simulated environment data is obtained by using industrial-grade smoke generators, laser particle size analyzers and infrared carbon dioxide sensors in the experimental chamber to precisely control the pollution release and accumulation process in order to obtain samples with precise labels.

[0132] Furthermore, the pollution level labels are generated through a process where, in real-vehicle testing and simulated environments, professional air quality monitoring equipment, including laser particle size analyzers and multi-gas analyzers, is used to simultaneously collect precise pollutant concentration data. This data is then combined with air quality grading standards to generate corresponding pollution level labels. Simultaneously, subjective evaluations from multiple test personnel are used to further validate the labels, ensuring their accuracy and usability.

[0133] In addition, all raw data must undergo standardized format conversion, outlier removal, missing value imputation, time alignment, and label confirmation before entering the training process, ultimately forming a multi-source training dataset containing image data, carbon dioxide concentration time-series data, environmental auxiliary data, and corresponding pollution level labels. This multi-source training dataset is cached in a standardized format in the intermediate data layer of the vehicle controller computing platform, providing a standardized input interface for each algorithm model, which can then be used by the model training and inference modules.

[0134] S200: Construct an in-vehicle air quality recognition model, train the in-vehicle air quality recognition model based on a multi-source training dataset, and obtain the trained in-vehicle air quality recognition model.

[0135] Preferably, each core module adopts a differentiated training strategy.

[0136] The multi-scale feature extraction branch for smoke particle size can be based on the Swin Transformer architecture. In the initial stage, it uses publicly available image recognition datasets for transfer learning pre-training, and then uses in-vehicle smoke image data for fine-tuning training. During training, a distributed data parallel strategy is adopted, combined with a multi-scale loss function to control the feature extraction accuracy of the model at different scales, thereby improving the robustness of smoke particle size recognition under blurred, low-light, and occluded conditions.

[0137] The carbon dioxide concentration time-series feature compression branch 200 adopts a stacked autoencoder structure. It uses unsupervised training with reconstruction error as the objective function and performs multiple rounds of comparative training with data of different time window lengths to optimize its ability to express concentration change patterns. The air pollution level map branch 400 employs a graph neural network, with pollution level classification as the supervised learning objective, using labeled samples for supervised learning. During graph structure generation, a graph structure optimization module is introduced to automatically search and optimize node weights, edge connection methods, and edge weight distribution. Graph sampling techniques are used during the training phase to reduce computational complexity, ensuring real-time inference under embedded resource constraints. Simultaneously, an early stopping strategy and graph structure regularization term are introduced to prevent overfitting and improve generalization ability.

[0138] Furthermore, the multi-model fusion branch employs an end-to-end joint optimization strategy. After the smoke particle size multi-scale feature extraction branch 100 and the carbon dioxide concentration temporal feature compression branch 200 are trained, their parameters are frozen. The multi-model heterogeneous feature fusion branch 300 and the air pollution level map branch 400 are then fine-tuned together, enabling the attention weights to adaptively adjust the importance of each modality feature according to the pollution scenario. Multi-task learning is also introduced during training to predict auxiliary indicators such as purification delay, thereby improving the overall system performance.

[0139] Furthermore, the parameter tuning process employs a phased optimization strategy.

[0140] In the initial stage, grid search is used to perform coarse screening of key hyperparameters such as learning rate and batch size.

[0141] After the model stabilizes, Bayesian optimization is used for fine-tuning.

[0142] All trained models have undergone quantization compression and structural pruning to adapt to the storage and computing resource constraints of automotive-grade SOC (System on a Chip) platforms.

[0143] Furthermore, the system has established a complete performance evaluation system, with performance optimization indicators covering multiple dimensions such as recognition accuracy, response latency, system power consumption, false alarm rate, and fault tolerance.

[0144] For the multi-scale feature extraction branch 100 for smoke particle size, its recognition robustness under different lighting conditions was evaluated.

[0145] We evaluated the early warning capability of the compression branch 200 based on the time-series characteristics of carbon dioxide concentration in short-term pollution abrupt change scenarios.

[0146] For the fusion and control module, the overall control strategy trigger latency and execution success rate were evaluated. All model versions have a version control system and support incremental model updates and online learning via OTA (Over-The-Air) to ensure long-term performance stability of the system in new scenarios and new vehicle models.

[0147] Furthermore, the system is equipped with online learning and model self-adaptation capabilities. During vehicle operation, it continuously records pollution status and control responses, forming a feedback loop. Through remote data transmission and a cloud-based model optimization framework, it regularly performs incremental model updates, enabling automatic adaptation to new scenarios, new vehicle models, and new user habits. This improves long-term performance stability and provides automakers with a sustainably evolving intelligent air quality identification platform.

[0148] Combination Figure 8 As shown, the present invention provides a vehicle for implementing the in-vehicle air quality identification method as described in the first aspect, comprising: Data acquisition unit 1 is used to acquire image data, carbon dioxide concentration time series data, and environmental auxiliary data; Controller 2 is connected to data acquisition unit 1 and is used to store at least the vehicle air quality identification model, obtain pollution level prediction results based on the vehicle air quality identification model, and determine the air purification control command corresponding to the pollution level prediction results. Air purification actuator 3 is connected to controller 2 for receiving and executing air purification control commands.

[0149] Combination Figure 9 As shown, this embodiment discloses a specific implementation of a computer device. The computer device may include a processor 41 and a memory 42 storing computer program instructions.

[0150] Specifically, the processor 41 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0151] The memory 42 may include a large-capacity storage device for data or instructions. For example, and not limitingly, the memory 42 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory 42 may include removable or non-removable (or fixed) media. Where appropriate, the memory 42 may be internal or external to the data processing device. In a particular embodiment, the memory 42 is non-volatile. Volatile memory. In a particular embodiment, memory 42 includes read-only memory. ROM (ROM-only memory) and RAM (Random Access Memory). Where appropriate, the ROM can be a mask-programmed ROM or a programmable ROM. Only Memory (PROM) and Erasable Programmable Read Memory (PROM) The RAM can be a type of RAM, such as EPROM (Electrically Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), EAROM (Electrically Alterable Read-Only Memory), or FLASH (Flash Memory), or a combination of two or more of these. Where appropriate, the RAM can be a Static Random Access Memory (SRAM). Access Memory (SRAM) or Dynamic Random Access Memory (DRAM) can be Fast Page Mode Dynamic Random Access Memory (FPMDRAM), Extended Data Out Dynamic Random Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0152] The memory 42 can be used to store or cache various data files that need to be processed and / or used for communication, as well as possible computer program instructions executed by the processor 41.

[0153] The processor 41 implements the vehicle air quality identification method in the above embodiments by reading and executing computer program instructions stored in the memory 42.

[0154] In some embodiments, the computer device may further include a communication interface 43 and a bus 40. For example, Figure 9 As shown, the processor 41, memory 42, and communication interface 43 are connected through bus 40 and complete communication with each other.

[0155] Communication interface 43 is used to enable communication between modules, devices, units and / or equipment in the embodiments of this application.

[0156] Communication port 43 can also enable data communication with other components such as external devices, image / data acquisition devices, databases, external storage, and image / data processing workstations.

[0157] Bus 40 includes hardware, software, or both, that couples components of a computer device together. Bus 40 includes, but is not limited to, at least one of the following: a data bus, an address bus, a control bus, an expansion bus, and a local bus. For example, and not as a limitation, bus 40 may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, and a PCI bus. Express (PCI X) bus, Serial Advanced Technology Accessory (Seria l Advanced) The bus may be a Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 40 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0158] Furthermore, in conjunction with the vehicle air quality identification method in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the vehicle air quality identification methods in the above embodiments.

[0159] The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned vehicle air quality identification method.

[0160] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described vehicle air quality identification method.

[0161] Although embodiments of the invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the invention.

Claims

1. A method for identifying in-vehicle air quality, characterized in that, It includes: Acquire real-time vehicle-mounted multi-source data, input the real-time vehicle-mounted multi-source data into the trained vehicle-mounted air quality recognition model, and obtain pollution level prediction results; The real-time vehicle-mounted multi-source data includes image data, carbon dioxide concentration time-series data, and environmental auxiliary data. Determine the air purification control command corresponding to the pollution level prediction result, and control the air purification actuator to execute the air purification control command; The vehicle-mounted air quality identification model is used to identify features of the real-time vehicle-mounted multi-source data, extract smoke particle size feature vectors and carbon dioxide concentration time-series feature vectors, predict pollution levels based on the smoke particle size feature vectors, the carbon dioxide concentration time-series feature vectors and the environmental auxiliary data, and output pollution level prediction results.

2. The vehicle-mounted air quality identification method according to claim 1, characterized in that, Also includes: Image data, carbon dioxide concentration time series data, and environmental auxiliary data under different vehicle operating parameters were acquired and labeled with pollution level. The labeled image data, carbon dioxide concentration time series data, and environmental auxiliary data were used as a multi-source training dataset. A vehicle air quality recognition model is constructed, and the vehicle air quality recognition model is trained based on the multi-source training dataset to obtain the trained vehicle air quality recognition model.

3. The vehicle-mounted air quality identification method according to claim 1, characterized in that, The vehicle-mounted air quality recognition model includes: The image data is subjected to multi-scale feature extraction based on the smoke particle size multi-scale feature extraction branch to obtain the smoke particle size feature vector. The carbon dioxide concentration time-series data is compressed and features are extracted based on the carbon dioxide concentration time-series feature compression branch to obtain the carbon dioxide concentration time-series feature vector. Based on the multi-model heterogeneous feature fusion branch, feature alignment and weighted fusion are performed on the smoke particle size feature vector and the carbon dioxide concentration time series feature vector to generate a multi-modal fusion feature vector; Based on the air pollution level map branch, the multimodal fusion feature vector and the environmental auxiliary data are used for map inference to obtain the pollution level prediction result.

4. The vehicle-mounted air quality identification method according to claim 2, characterized in that, The multi-model heterogeneous feature fusion branch includes: Based on the feature alignment module, the smoke particle size feature vector and the carbon dioxide concentration time series feature vector are projected to the same feature dimension through a fully connected layer to obtain the aligned feature vector. The attention weighting module receives the aligned feature vector, calculates the corresponding weight coefficients through a neural network, and performs a weighted summation on the aligned feature vector based on the weight coefficients to obtain a multimodal fusion feature vector.

5. The vehicle-mounted air quality identification method according to claim 3, characterized in that, The smoke particle size multi-scale feature extraction branch includes an image preprocessing module, a hierarchical feature extraction module, and a global feature aggregation module connected in sequence. The image preprocessing module performs image preprocessing on the image data and divides it into a sequence of image blocks. The hierarchical feature extraction module extracts smoke particle size features at different scales from the image patch sequence through a multi-level transformation layer containing sliding window attention and cross-window connections, thereby obtaining a multi-scale smoke feature map. The global feature aggregation module performs global average pooling on the multi-scale smoke feature map to generate a fixed-dimensional smoke particle size feature vector.

6. The vehicle-mounted air quality identification method according to claim 3, characterized in that, The carbon dioxide concentration temporal feature compression branch includes a data preprocessing module and a stacked autoencoder module connected in sequence. The data preprocessing module normalizes the carbon dioxide concentration time series data to construct a standardized concentration time series. The stacked autoencoder module nonlinearly maps the standardized concentration time series to a low-dimensional latent space through its encoder, and outputs a fixed-dimensional carbon dioxide concentration time series feature vector.

7. The vehicle-mounted air quality identification method according to claim 6, characterized in that, The encoder of the stacked autoencoder module includes: Multiple fully connected layers are used to perform a layer-by-layer dimensionality reduction nonlinear transformation on the standardized concentration time series. Each fully connected layer is followed by a batch normalization layer, a nonlinear activation function layer, and a random deactivation layer. The low-dimensional vector output by the last fully connected layer of the encoder is used as the temporal feature vector of carbon dioxide concentration.

8. The vehicle-mounted air quality identification method according to claim 3, characterized in that, The air pollution level map branch includes a map construction module and a graph convolution inference module; The graph construction module uses the smoke particle size feature vector, the carbon dioxide concentration time series feature vector, and the environmental auxiliary feature vector converted from the environmental auxiliary data as nodes, and constructs weighted edges based on predefined physical associations and statistical correlations between nodes to form a dynamic heterogeneous graph. The graph convolutional inference module performs neighborhood information aggregation and feature updates on the dynamic heterogeneous graph through a multi-layer graph convolutional network, and outputs the pollution level prediction result through a classifier.

9. The vehicle-mounted air quality identification method according to claim 1, characterized in that, The process of determining the air purification control command corresponding to the pollution level prediction result and controlling the air purification actuator to execute the air purification control command includes: Based on the pollution level prediction results, the preset strategy mapping table is queried to obtain the query results; Based on the query results, an air purification control command is generated. The air purification control command includes at least a command to switch the vehicle air conditioning circulation mode, a command to adjust the power of the air purifier, and a prompt command for the user interface. The control commands are sent to the corresponding air purification actuators via the vehicle bus to coordinate the air purification operation.

10. A vehicle for implementing the vehicle-mounted air quality identification method as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to acquire real-time multi-source vehicle data; The controller is communicatively connected to the data acquisition unit and is used to store at least the vehicle air quality identification model, input the real-time vehicle multi-source data into the vehicle air quality identification model, obtain the pollution level prediction result, and determine the air purification control command corresponding to the pollution level prediction result. An air purification actuator, communicatively connected to the controller, is used to receive and execute the air purification control commands.

Images