Intelligent aromatherapy margin calculation method based on dynamic perception of vehicle body data

Abstract

The invention discloses an aromatherapy intelligent margin calculation method based on vehicle body data dynamic perception. The method comprises the following steps: S1, generating an initial data set; s2, inputting the initial data set into a multi-channel variational encoder, and outputting hierarchical feature representation; s3, inputting the hierarchical feature representation into a spatio-temporal context sensing module to generate spatio-temporal feature representation; s4, inputting the spatial-temporal feature representation into a heterogeneous feature mapping layer to generate a unified data representation; s5, inputting the unified data representation into a dynamic prediction regulation and control module, and generating an optimized feature representation; s6, inputting the optimized feature representation into an aromatherapy residual amount prediction model, and performing residual amount prediction by using a Gaussian mixture process; s7, inputting an aromatherapy residue prediction result into a feedback integrator, and carrying out self-supervised learning; and S8, integrating to a vehicle-mounted system, displaying an aromatherapy allowance state and sending a personalized notice. According to the method, accurate prediction and calculation of the vehicle-mounted aromatherapy allowance are realized by utilizing multi-channel variational coding, spatio-temporal context perception, a Gaussian process and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent vehicle-mounted systems, and in particular to an intelligent aromatherapy remainder calculation method based on dynamic perception of vehicle body data. Background Art

[0002] In the existing vehicle environment control system, in-vehicle air quality management and comfort adjustment have become important components in the development of intelligent vehicles. In the vehicle environment, in order to improve the comfort of drivers and passengers, aromatherapy systems are widely used. However, traditional aromatherapy systems have many limitations in management and control. They mainly rely on preset usage modes and simple time or quantitative control strategies, and cannot be dynamically adjusted and accurately predicted based on real-time environmental data and user behavior. This traditional method is difficult to cope with complex changes in the in-vehicle and in-vehicle environment and differences in user preferences, resulting in problems such as insufficient adjustment accuracy, inflexible response, and inaccurate management of aromatherapy residue during use.

[0003] In current research and technological development, some systems have introduced sensors to detect in-vehicle environmental data, such as parameters such as temperature, humidity, and air quality. However, these systems often lack data processing and multi-source information fusion capabilities, resulting in limited correlation analysis and feature extraction capabilities between data, and cannot achieve in-depth real-time intelligent analysis. This makes it difficult for existing systems to accurately model user usage behavior and external environmental changes in complex environments when predicting and optimizing aromatherapy residue. In addition, traditional data processing models usually rely on fixed algorithms and simple models when dealing with dynamically changing complex features, and are unable to adjust parameters and structures according to real-time environments or user feedback, making it difficult for the system to maintain long-term prediction performance and adaptability.

[0004] When regulating the aromatherapy system, existing methods usually lack the ability to comprehensively analyze multi-dimensional features, especially the simultaneous processing of different features such as high frequency, low frequency and emergencies. Existing technologies often use a simple analysis method of a single channel, which cannot effectively cope with complex and changing environmental inputs, especially when the external environment changes rapidly or the user behavior in the car is abnormal. The prediction accuracy and system stability will drop significantly. In addition, due to the inability to dynamically adjust the loss function and model structure to optimize system performance, the adaptability and robustness of traditional methods are weak, making it impossible for the aromatherapy system to achieve efficient margin management in different scenarios.

[0005] Although existing technologies have attempted to introduce machine learning and simple neural networks to improve the intelligence of the system, the complexity of the model and the computing power of these methods limit their performance when processing real-time data and multimodal inputs. For example, some systems may try to use variational autoencoders to handle nonlinear features, but they usually fail to combine multi-channel structures and adaptive loss adjustment strategies to deal with the problems of uncertainty and feature coupling in high-dimensional data. In addition, existing time series data analysis methods, such as traditional sliding windows and unidirectional attention mechanisms, have difficulty in fully capturing the spatiotemporal characteristics of the data and cannot fully reflect the relationship between historical and real-time contexts, thereby limiting the model's predictive ability and contextual adaptability.

[0006] In terms of feedback and self-learning, most current aromatherapy systems lack effective self-supervised learning and dynamic reconstruction mechanisms. Even if a few systems try to use feedback loops for simple parameter updates, they still lack the ability to effectively detect and respond to long-term deviations and new features, making it difficult to improve system performance through adaptive adjustments. Most existing learning mechanisms are static structures and fail to achieve automated model reconstruction and optimization, thus limiting the scalability and durability of the model in long-term use.

[0007] Therefore, how to provide an intelligent aromatherapy remainder calculation method based on dynamic perception of vehicle body data is an urgent problem that technicians in this field need to solve. Summary of the invention

[0008] One purpose of the present invention is to propose an intelligent aromatherapy remainder calculation method based on dynamic perception of vehicle body data. The present invention combines a multi-channel variational encoder, a spatiotemporal context perception module, and a mixed Gaussian process technology to achieve efficient collection and analysis of in-vehicle and out-of-vehicle environmental data and user behavior data. Feature fusion is performed through a bidirectional spatiotemporal attention mechanism and a heterogeneous feature mapping layer to ensure that the system has excellent feature recognition and dynamic response capabilities in different environments. With the help of real-time feedback and self-supervised learning mechanisms, the present invention can continuously optimize the model structure and parameters, improve the prediction accuracy and adaptive performance of the system in complex and changeable environments, and has the significant advantages of high precision, robustness, and rapid response.

[0009] According to an embodiment of the present invention, a method for calculating the intelligent fragrance remainder based on dynamic perception of vehicle body data includes the following steps:

[0010] S1. Collect environmental data and user behavior data through sensors inside and outside the vehicle, and perform noise filtering, outlier removal and standardization on the data to generate an initial data set;

[0011] S2, input the initial data set into the multi-channel variational encoder, classify and encode it according to high frequency, low frequency and sudden events, output hierarchical feature representation, and realize feature information exchange through the interactive network between channels, and dynamically adjust the variational loss function of the encoder;

[0012] S3, input the hierarchical feature representation into the spatiotemporal context perception module, capture the time series data features through the spatiotemporal sliding window, generate the spatiotemporal feature representation in combination with the bidirectional spatiotemporal attention mechanism, and inject the historical and real-time context weights;

[0013] S4, input the spatiotemporal feature representation into the heterogeneous feature mapping layer, perform spatiotemporal feature mapping and conversion, and map to generate a unified data representation;

[0014] S5. Input the unified data representation into the dynamic prediction and control module, dynamically adjust the input feature weights according to real-time feedback, and generate an optimized feature representation;

[0015] S6, inputting the optimized feature representation into the aromatherapy remainder prediction model, using the mixed Gaussian process to perform remainder prediction, and combining variational inference to update the parameters of the aromatherapy remainder prediction model;

[0016] S7, input the fragrance remaining amount prediction result into the feedback integrator, collect user feedback and in-vehicle and out-of-vehicle data for self-supervised learning, and when long-term deviations or new data features are detected, trigger the dynamic intelligent reconstruction and optimization learning mechanism to perform structural update and parameter optimization of the fragrance remaining amount prediction model;

[0017] S8. Integrate the updated aromatherapy remaining quantity prediction model into the vehicle-mounted system to display the aromatherapy remaining quantity status in real time, and issue personalized notifications and usage suggestions based on user behavior and aromatherapy remaining quantity prediction results.

[0018] Optionally, the vehicle body status data specifically includes the vehicle start / stop status, whether the windows are open, and whether the air conditioner in the vehicle is turned on.

[0019] Optionally, the S2 specifically includes:

[0020] S21, inputting the environment data and user behavior data in the initial data set into the multi-channel variational encoder;

[0021] S22. In a multi-channel variational encoder, the input data is hierarchically classified into three categories: high frequency, low frequency, and emergency events, wherein the high frequency channel extracts rapidly changing environmental features in the input data, the low frequency channel extracts long-term stable features in the input data, and the emergency event channel extracts unpredictable abnormal changes in the input data;

[0022] S23, feature encoding is performed on the data in each channel, and the high-frequency feature encoder generates a high-frequency feature vector V high, the low-frequency feature encoder generates a low-frequency feature vector V low , the emergency feature encoder generates the emergency feature vector V event ;

[0023] S24, feature information is exchanged through the interactive network between channels, and the high-frequency feature vector V high , low-frequency feature vector V low and the emergency feature vector V event Perform interactive processing and output the comprehensive feature matrix M exchange :

[0024]

[0025] Among them, k represents the dimension of the feature vector, and α represents the normalization coefficient;

[0026] S25. During the feature information exchange process, the variational loss function of the encoder is dynamically adjusted using co-correlation analysis to optimize the encoding quality of each feature channel:

[0027]

[0028] Among them, L var represents the loss function, n represents the total number of features in the input dataset, KL represents the divergence value, γ i represents the weight coefficient, M represents the number of additional features, q(z|x i ) represents the encoded posterior distribution, p(z represents the prior distribution, represents the standard deviation of the input feature, ∈ represents the smoothing term, θ represents the adjustment factor, μ represents the offset, β and δ represent the adjustment parameters, T represents the time step, and λ represents the scaling parameter;

[0029] S26. Output the adjusted hierarchical feature representation.

[0030] Optionally, the S3 specifically includes:

[0031] S31, inputting the hierarchical feature representation into a spatiotemporal context perception module, wherein the spatiotemporal context perception module performs a time series analysis on the input multi-channel features to capture the dynamic changes of environmental data and user behavior in the time and space dimensions;

[0032] S32, applying a spatiotemporal sliding window strategy to divide the input feature data into different time windows for processing, wherein the spatiotemporal sliding window dynamically adjusts the size and step length according to vehicle speed, environmental conditions and data changes to capture short-term and long-term time series features;

[0033] S33. In each spatiotemporal sliding window, a bidirectional spatiotemporal attention mechanism is used to perform feature analysis. The bidirectional spatiotemporal attention mechanism captures the association between the current time step and past features through a forward attention network, and captures the association between the current time step and future features through a backward attention network, generating an attention weight covering global temporal information;

[0034] S34, based on the attention weights covering global temporal information generated by the bidirectional spatiotemporal attention mechanism, weighted processing is performed on the features within the window, highlighting the attention to important features related to the current task, suppressing noise and irrelevant features, and forming a feature representation with temporal correlation and context association;

[0035] S35, injecting historical context information and real-time context information into the spatiotemporal feature representation, and dynamically combining the weight of historical data and the weight of real-time data through context weight injection technology;

[0036] S36. Output the spatiotemporal feature representation combining historical and real-time context information as input for unified data representation.

[0037] Optionally, the S4 specifically includes:

[0038] S41, inputting the spatiotemporal feature representation output by the spatiotemporal context perception module into a heterogeneous feature mapping layer, wherein the heterogeneous feature mapping layer is composed of a plurality of parallel processing units, and the units respectively perform preliminary analysis and block division on features of different sources and time scales;

[0039] S42. Through the double normalization strategy and adaptive filtering algorithm, the weight distribution of high-frequency features and low-frequency features is gradually balanced, and dynamic range adjustment is adopted to keep the features of each source in a stable distribution, thereby reducing the conversion deviation caused by inconsistent data scales;

[0040] S43. In the heterogeneous feature mapping conversion network, a multi-layer convolution combined with a deformable convolution is used to map and reconstruct the features layer by layer, and the original spatiotemporal features are gradually converted into a comprehensive representation with a unified feature space. The deformable convolution is used to automatically adjust the shape of the convolution kernel to adapt to the changes of the features in different time periods.

[0041] S44. In the feature mapping process, context features are embedded in combination with the output of the context perception module. The historical features are organically integrated with the current real-time features through the context attention mechanism to form a real-time context weight mapping. The input features are assigned fine-grained weights through dynamic adjustment of weight priorities.

[0042] S45. Adopting a multi-channel feature aggregation strategy, the mapped high-frequency, low-frequency and emergency event features are hierarchically aggregated, and an interaction weight matrix between channels is introduced to achieve interactive enhancement of features and complementarity of multi-dimensional information;

[0043] S46. Perform stability check on the final unified data representation before output, and perform output correction in combination with a feature smoothing mechanism.

[0044] Optionally, the S5 specifically includes:

[0045] S51, inputting the generated unified data representation into a dynamic prediction and control module, wherein the dynamic prediction and control module includes a feature weighting unit and an adaptive feedback unit, which are used to adjust the weights of input features in real time to adapt to environmental changes;

[0046] S52, in the feature weighting unit, a multi-layer adaptive weight allocation algorithm is used to perform weighted processing on the input unified data representation, the multi-layer adaptive weight allocation algorithm automatically allocates weights according to the change rate and correlation of the features at different time points, generates a preliminary optimized feature matrix, and updates the weights through an iterative process to adapt to the dynamic changes of the environment and user behavior;

[0047] S53, receiving the change information of the external environment data and the user behavior data in real time through the adaptive feedback unit, comparing the feedback result with the current weight distribution, and adjusting the weight distribution strategy, gradually learning and adapting to the input feature changes under different environmental conditions based on the incremental learning mechanism;

[0048] S54, screening the initially optimized feature matrix, using a feature selection algorithm to remove redundant and low-correlation features, retaining high-correlation features, and ensuring high information density, wherein the feature selection algorithm uses a weight pruning technology;

[0049] S55, passing the filtered feature matrix to the dynamic control strategy module, and finally optimizing the feature matrix in combination with the context association model, wherein the context association model adjusts the time and environment priority of the features according to historical data and real-time feedback to generate a comprehensive optimized feature representation;

[0050] S56. Output the optimized feature representation as the input of the aromatherapy remainder prediction module.

[0051] Optionally, the S6 specifically includes:

[0052] S61, inputting the optimized feature representation into an aromatherapy residue prediction model, wherein the aromatherapy residue prediction model adopts a mixed Gaussian process structure and combines multi-dimensional feature input to achieve modeling of complex nonlinear relationships between features;

[0053] S62, the aromatherapy remaining quantity prediction model consists of a two-layer feature processing unit. The first layer is the input feature mapping layer, which embeds the optimized feature representation into the high-dimensional feature space; the second layer adopts a multi-core Gaussian process structure, integrating linear kernel, radial basis kernel and polynomial kernel functions to capture the complex interactive relationship between high-frequency, low-frequency and emergency event features;

[0054] S63. In the Gaussian process core layer, define the margin prediction formula, and realize adaptive adjustment of margin prediction by dynamically weighing different feature combinations:

[0055]

[0056] Among them, y 余量 (t) represents the prediction result of aromatherapy residue, μ 高频 (t) represents the predicted mean of high-frequency features, μ 低频 (t) represents the predicted mean of low-frequency features, μ 突发 (t) represents the predicted mean of the emergency event characteristics, σ 低频 (t represents the prediction standard deviation of low-frequency features, σ 高频 (t) represents the prediction standard deviation of high-frequency features, σ 突发 (t) represents the predicted standard deviation of the sudden event characteristics, α 1,j , α 2,l and λ r represents the weight adjustment coefficient, β 1,i and β 2,p represents the integral adjustment coefficient, γ 1,i and γ 2,p represents the scaling parameter, δ3 represents the global adjustment coefficient, and f j (t) represents the input feature gradient function at time t, h l (t) represents the input feature correction function at time t, g r (t) represents the correction function of the sudden event, k represents the number of weight coefficients in the high-frequency feature, m represents the number of adjustment parameters of the high-frequency feature, n represents the number of adjustment coefficients of the low-frequency feature, q represents the number of adjustment parameters used for time series integration in the low-frequency feature, s represents the number of correction functions in the sudden event, u represents the number of time steps used for cumulative standard deviation calculation in the sudden event part, and t v represents the low-frequency features at time step v, θ1 and θ2 represent threshold parameters;

[0057] S64, using variational inference technology to update model parameters in real time, optimize the posterior distribution of the Gaussian process, and ensure that the model can be adaptively adjusted according to changes in input;

[0058] S65. In the residual forecasting process, the uncertainty quantitative analysis of the model is combined to estimate the confidence interval of the forecast results to identify and deal with potential forecast deviations;

[0059] S66, output aromatherapy remaining amount prediction result y 余量 (t), combining multidimensional features and uncertainty information to generate aromatherapy remainder prediction results with high confidence.

[0060] Optionally, the S7 specifically includes:

[0061] S71, inputting the prediction result of the fragrance remaining amount prediction model into a feedback integrator, wherein the feedback integrator continuously collects data from the environment inside and outside the vehicle, including temperature, humidity, vehicle speed and user operation behavior information, and combines the user feedback to form a multi-dimensional data set;

[0062] S72. Based on the multi-dimensional data set, the actual environmental data is compared with the predicted results, a deviation analysis report is generated, the deviation between the predicted value and the actual value is identified, and whether the deviation exceeds a preset long-term threshold is determined;

[0063] S73. When the deviation analysis result shows that the long-term threshold is exceeded or a new feature appears, the self-supervised learning module is started to generate pseudo labels through unlabeled data, expand the training data set, automatically learn new environmental features, and improve the adaptive ability of the aromatherapy residue prediction model to cope with the environment;

[0064] S74, inputting the expanded training data set into the incremental learning unit of the aromatherapy residue prediction model, wherein the incremental learning unit performs progressive parameter updates on the aromatherapy residue prediction model to maintain efficient operation of the aromatherapy residue prediction model in a real-time environment and optimize prediction performance;

[0065] S75. When long-term deviations are detected frequently or new data features continue to appear, a dynamic intelligent reconstruction mechanism is triggered to automatically adjust the structure of the aromatherapy remainder prediction model, including optimizing the model hierarchy, reconfiguring the weight distribution, and adjusting the feature path;

[0066] S76. After completing the structural adjustment and parameter optimization of the aromatherapy remaining quantity prediction model, the improved aromatherapy remaining quantity prediction model is applied to the vehicle-mounted system, and the updated version information and adjustment details of the aromatherapy remaining quantity prediction model are recorded.

[0067] The beneficial effects of the present invention are:

[0068] First, by introducing a multi-channel variational encoder, the present invention can efficiently and hierarchically encode and analyze the in-vehicle and out-of-vehicle environmental data and user behavior data, fully capturing multi-dimensional features such as high frequency, low frequency, and emergencies. The application of this multi-channel structure improves the processing capability of complex data, and can more accurately identify and respond to different feature changes in a real-time environment, effectively making up for the limitations of traditional single-channel analysis methods when processing multi-source data. Combined with the interactive network between channels, the present invention realizes dynamic information exchange and collaborative analysis between features, improving the quality and robustness of the overall feature encoding.

[0069] Secondly, by introducing the spatiotemporal context perception module and the bidirectional spatiotemporal attention mechanism, the bidirectional time series characteristics of historical and real-time data are captured, achieving a comprehensive perception of the dynamic changes in the vehicle's internal and external environments. The technology of injecting historical and real-time context weights ensures that the system can accurately reflect user behavior and environmental changes at different time spans, thereby improving the accuracy of predictions and the system's ability to adapt to the environment. Compared with the existing unidirectional time series analysis method, this bidirectional spatiotemporal analysis method significantly enhances the system's spatiotemporal perception ability and solves the problems of prediction lag and missing context information.

[0070] In addition, through the heterogeneous feature mapping layer, the present invention has achieved new progress in multimodal data processing. This layer realizes the seamless mapping and conversion of features from different sources and time scales, allowing multi-source data to be comprehensively analyzed in a unified feature space. This not only improves the efficiency of feature fusion, but also ensures the stability and consistency of the system in processing complex inputs in a dynamic environment. At the same time, the dynamic prediction and control module adjusts the input feature weights through real-time feedback, allowing the system to quickly adapt when the environment changes or user behavior changes, ensuring the flexibility and continuous optimization capabilities of the prediction model.

[0071] The aromatherapy residue prediction model uses a combination of mixed Gaussian processes and variational inference technology, providing the ability to model complex nonlinear feature relationships and dynamically update parameters. Through the introduction of real-time uncertainty quantification and variational inference, the system can make adaptive adjustments when long-term deviations or new features are detected, improving the accuracy of predictions and the robustness of the model in a changing environment. This capability enables the system to not only effectively predict current data, but also maintain its long-term prediction performance through self-learning.

[0072] Finally, through the feedback integrator and dynamic intelligent reconstruction mechanism, the present invention achieves continuous self-supervised learning and model optimization. The system can automatically update and reconstruct the model when it detects that the deviation exceeds the threshold or new data features appear, thereby ensuring the stability and scalability of the system in long-term use. Compared with traditional methods, the present invention provides a higher level of adaptability and intelligence, improving user experience and system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0074] Figure 1 This is a flow chart of the aromatherapy intelligent remainder calculation method based on dynamic perception of vehicle body data proposed by the present invention;

[0075] Figure 2 This is a structural schematic diagram of the aromatherapy intelligent remainder calculation method based on dynamic perception of vehicle body data proposed in the present invention. DETAILED DESCRIPTION

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0077] refer to Figure 1 and Figure 2 , the aromatherapy intelligent remainder calculation method based on dynamic perception of vehicle body data includes the following steps:

[0078] S1. Collect environmental data and user behavior data through sensors inside and outside the vehicle, and perform noise filtering, outlier removal and standardization on the data to generate an initial data set;

[0079] S2, input the initial data set into the multi-channel variational encoder, classify and encode it according to high frequency, low frequency and sudden events, output hierarchical feature representation, and realize feature information exchange through the interactive network between channels, and dynamically adjust the variational loss function of the encoder;

[0080] S3, input the hierarchical feature representation into the spatiotemporal context perception module, capture the time series data features through the spatiotemporal sliding window, generate the spatiotemporal feature representation in combination with the bidirectional spatiotemporal attention mechanism, and inject the historical and real-time context weights;

[0081] S4, input the spatiotemporal feature representation into the heterogeneous feature mapping layer, perform spatiotemporal feature mapping and conversion, and map to generate a unified data representation;

[0082] S5. Input the unified data representation into the dynamic prediction and control module, dynamically adjust the input feature weights according to real-time feedback, and generate an optimized feature representation;

[0083] S6, inputting the optimized feature representation into the aromatherapy remainder prediction model, using a mixed Gaussian process to perform remainder prediction, and combining variational inference to update the parameters of the aromatherapy remainder prediction model;

[0084] S7, input the fragrance remaining amount prediction result into the feedback integrator, collect user feedback and in-vehicle and out-of-vehicle data for self-supervised learning, and when long-term deviations or new data features are detected, trigger the dynamic intelligent reconstruction and optimization learning mechanism to perform structural update and parameter optimization of the fragrance remaining amount prediction model;

[0085] S8. Integrate the updated aromatherapy remaining quantity prediction model into the vehicle-mounted system to display the aromatherapy remaining quantity status in real time, and issue personalized notifications and usage suggestions based on user behavior and aromatherapy remaining quantity prediction results.

[0086] In this implementation, the vehicle body status data specifically includes the vehicle start / stop status, whether the vehicle windows are open, and whether the vehicle air conditioner is turned on.

[0087] In this implementation, S2 specifically includes:

[0088] S21, inputting the environment data and user behavior data in the initial data set into the multi-channel variational encoder;

[0089] S22. In a multi-channel variational encoder, the input data is hierarchically classified into three categories: high frequency, low frequency, and emergency events, wherein the high frequency channel extracts rapidly changing environmental features in the input data, the low frequency channel extracts long-term stable features in the input data, and the emergency event channel extracts unpredictable abnormal changes in the input data;

[0090] S23, feature encoding is performed on the data in each channel, and the high-frequency feature encoder generates a high-frequency feature vector V high , the low-frequency feature encoder generates a low-frequency feature vector V low , the emergency feature encoder generates the emergency feature vector V event ;

[0091] S24, feature information is exchanged through the interactive network between channels, and the high-frequency feature vector V high , low-frequency feature vector V low and the emergency feature vector V event Perform interactive processing and output the comprehensive feature matrix M exchange :

[0092]

[0093] Among them, k represents the dimension of the feature vector, and α represents the normalization coefficient;

[0094] S25. During the feature information exchange process, the variational loss function of the encoder is dynamically adjusted using co-correlation analysis to optimize the encoding quality of each feature channel:

[0095]

[0096] Among them, L var represents the loss function, n represents the total number of features in the input dataset, KL represents the divergence value, γ i represents the weight coefficient, M represents the number of additional features, q(z|x i ) represents the encoded posterior distribution, p(z represents the prior distribution, represents the standard deviation of the input feature, ∈ represents the smoothing term, θ represents the adjustment factor, μ represents the offset, β and δ represent the adjustment parameters, T represents the time step, and λ represents the scaling parameter;

[0097] S26. Output the adjusted hierarchical feature representation.

[0098] In this implementation, S3 specifically includes:

[0099] S31, inputting the hierarchical feature representation into a spatiotemporal context perception module, wherein the spatiotemporal context perception module performs a time series analysis on the input multi-channel features to capture the dynamic changes of environmental data and user behavior in the time and space dimensions;

[0100] S32, applying a spatiotemporal sliding window strategy to divide the input feature data into different time windows for processing, wherein the spatiotemporal sliding window dynamically adjusts the size and step length according to vehicle speed, environmental conditions and data changes to capture short-term and long-term time series features;

[0101] S33. In each spatiotemporal sliding window, a bidirectional spatiotemporal attention mechanism is used to perform feature analysis. The bidirectional spatiotemporal attention mechanism captures the association between the current time step and past features through a forward attention network, and captures the association between the current time step and future features through a backward attention network, generating an attention weight covering global temporal information;

[0102] S34, based on the attention weights covering global temporal information generated by the bidirectional spatiotemporal attention mechanism, weighted processing is performed on the features within the window, highlighting the attention to important features related to the current task, suppressing noise and irrelevant features, and forming a feature representation with temporal correlation and context association;

[0103] S35, injecting historical context information and real-time context information into the spatiotemporal feature representation, and dynamically combining the weight of historical data and the weight of real-time data through context weight injection technology;

[0104] S36. Output the spatiotemporal feature representation combining historical and real-time context information as input for unified data representation.

[0105] In this implementation, S4 specifically includes:

[0106] S41, inputting the spatiotemporal feature representation output by the spatiotemporal context perception module into a heterogeneous feature mapping layer, wherein the heterogeneous feature mapping layer is composed of a plurality of parallel processing units, and the units respectively perform preliminary analysis and block division on features of different sources and time scales;

[0107] S42. Through the double normalization strategy and adaptive filtering algorithm, the weight distribution of high-frequency features and low-frequency features is gradually balanced, and dynamic range adjustment is adopted to keep the features of each source in a stable distribution, thereby reducing the conversion deviation caused by inconsistent data scales;

[0108] S43. In the heterogeneous feature mapping conversion network, a multi-layer convolution combined with a deformable convolution is used to map and reconstruct the features layer by layer, and the original spatiotemporal features are gradually converted into a comprehensive representation with a unified feature space. The deformable convolution is used to automatically adjust the shape of the convolution kernel to adapt to the changes of the features in different time periods.

[0109] S44. In the feature mapping process, context features are embedded in combination with the output of the context perception module. The historical features are organically integrated with the current real-time features through the context attention mechanism to form a real-time context weight mapping. The input features are assigned fine-grained weights through dynamic adjustment of weight priorities.

[0110] S45. Adopting a multi-channel feature aggregation strategy, the mapped high-frequency, low-frequency and emergency event features are hierarchically aggregated, and an interaction weight matrix between channels is introduced to achieve interactive enhancement of features and complementarity of multi-dimensional information;

[0111] S46. Perform stability check on the final unified data representation before output, and perform output correction in combination with a feature smoothing mechanism.

[0112] In this implementation manner, S5 specifically includes:

[0113] S51, inputting the generated unified data representation into a dynamic prediction and control module, wherein the dynamic prediction and control module includes a feature weighting unit and an adaptive feedback unit, which are used to adjust the weights of input features in real time to adapt to environmental changes;

[0114] S52, in the feature weighting unit, a multi-layer adaptive weight allocation algorithm is used to perform weighted processing on the input unified data representation, the multi-layer adaptive weight allocation algorithm automatically allocates weights according to the change rate and correlation of the features at different time points, generates a preliminary optimized feature matrix, and updates the weights through an iterative process to adapt to the dynamic changes of the environment and user behavior;

[0115] S53, receiving the change information of the external environment data and the user behavior data in real time through the adaptive feedback unit, comparing the feedback result with the current weight distribution, and adjusting the weight distribution strategy, gradually learning and adapting to the input feature changes under different environmental conditions based on the incremental learning mechanism;

[0116] S54, screening the initially optimized feature matrix, using a feature selection algorithm to remove redundant and low-correlation features, retaining high-correlation features, and ensuring high information density, wherein the feature selection algorithm uses a weight pruning technology;

[0117] S55, passing the filtered feature matrix to the dynamic control strategy module, and finally optimizing the feature matrix in combination with the context association model, wherein the context association model adjusts the time and environment priority of the features according to historical data and real-time feedback to generate a comprehensive optimized feature representation;

[0118] S56. Output the optimized feature representation as the input of the aromatherapy remainder prediction module.

[0119] In this implementation manner, S6 specifically includes:

[0120] S61, inputting the optimized feature representation into an aromatherapy residue prediction model, wherein the aromatherapy residue prediction model adopts a mixed Gaussian process structure and combines multi-dimensional feature input to achieve modeling of complex nonlinear relationships between features;

[0121] S62, the aromatherapy remaining quantity prediction model consists of a two-layer feature processing unit. The first layer is the input feature mapping layer, which embeds the optimized feature representation into the high-dimensional feature space; the second layer adopts a multi-core Gaussian process structure, integrating linear kernel, radial basis kernel and polynomial kernel functions to capture the complex interactive relationship between high-frequency, low-frequency and emergency event features;

[0122] S63. In the Gaussian process core layer, define the margin prediction formula, and realize adaptive adjustment of margin prediction by dynamically weighing different feature combinations:

[0123]

[0124] Among them, y 余量 (t) represents the prediction result of aromatherapy residue, μ 高频 (t) represents the predicted mean of high-frequency features, μ 低频 (t) represents the predicted mean of low-frequency features, μ 突发 (t) represents the predicted mean of the emergency event characteristics, σ 低频 (t represents the prediction standard deviation of low-frequency features, σ 高频 (t) represents the prediction standard deviation of high-frequency features, σ 突发(t) represents the predicted standard deviation of the sudden event characteristics, α 1,j , α 2,l and λ r represents the weight adjustment coefficient, β 1,i and β 2,p represents the integral adjustment coefficient, γ 1,i and γ 2,p represents the scaling parameter, δ3 represents the global adjustment coefficient, and f j (t) represents the input feature gradient function at time t, h l (t) represents the input feature correction function at time t, g r (t) represents the correction function of the sudden event, k represents the number of weight coefficients in the high-frequency feature, m represents the number of adjustment parameters of the high-frequency feature, n represents the number of adjustment coefficients of the low-frequency feature, q represents the number of adjustment parameters used for time series integration in the low-frequency feature, s represents the number of correction functions in the sudden event, u represents the number of time steps used for cumulative standard deviation calculation in the sudden event part, and t v represents the low-frequency features at time step v, θ1 and θ2 represent threshold parameters;

[0125] S64, using variational inference technology to update model parameters in real time, optimize the posterior distribution of the Gaussian process, and ensure that the model can be adaptively adjusted according to changes in input;

[0126] S65. In the residual forecasting process, the uncertainty quantitative analysis of the model is combined to estimate the confidence interval of the forecast results to identify and deal with potential forecast deviations;

[0127] S66, output aromatherapy remaining amount prediction result y 余量 (t), combining multidimensional features and uncertainty information to generate aromatherapy remainder prediction results with high confidence.

[0128] In this implementation manner, the S7 specifically includes:

[0129] S71, inputting the prediction result of the fragrance remaining amount prediction model into a feedback integrator, wherein the feedback integrator continuously collects data from the environment inside and outside the vehicle, including temperature, humidity, vehicle speed and user operation behavior information, and combines the user feedback to form a multi-dimensional data set;

[0130] S72. Based on the multi-dimensional data set, the actual environmental data is compared with the predicted results, a deviation analysis report is generated, the deviation between the predicted value and the actual value is identified, and whether the deviation exceeds a preset long-term threshold is determined;

[0131] S73. When the deviation analysis result shows that the long-term threshold is exceeded or a new feature appears, the self-supervised learning module is started to generate pseudo labels through unlabeled data, expand the training data set, automatically learn new environmental features, and improve the adaptive ability of the aromatherapy residue prediction model to cope with the environment;

[0132] S74, inputting the expanded training data set into the incremental learning unit of the aromatherapy residue prediction model, wherein the incremental learning unit performs progressive parameter updates on the aromatherapy residue prediction model to maintain efficient operation of the aromatherapy residue prediction model in a real-time environment and optimize prediction performance;

[0133] S75. When long-term deviations are detected frequently or new data features continue to appear, a dynamic intelligent reconstruction mechanism is triggered to automatically adjust the structure of the aromatherapy remainder prediction model, including optimizing the model hierarchy, reconfiguring the weight distribution, and adjusting the feature path;

[0134] S76. After completing the structural adjustment and parameter optimization of the aromatherapy remaining quantity prediction model, the improved aromatherapy remaining quantity prediction model is applied to the vehicle-mounted system, and the updated version information and adjustment details of the aromatherapy remaining quantity prediction model are recorded.

[0135] Embodiment 1:

[0136] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a high-end autonomous driving vehicle. During long-term use, the owner found that the traditional aromatherapy system showed inaccurate management and delayed response when responding to changes in the environment inside and outside the car. For example, when the vehicle opens the windows or uses the air conditioner for a long time, the airflow and temperature in the car change significantly. It is difficult for the traditional system to accurately adjust the aromatherapy remainder display according to these changes, resulting in aromatherapy waste or poor effect. In order to deal with this problem, the owner used the aromatherapy intelligent remainder calculation method based on dynamic perception of vehicle body data of the present invention, and tested its performance for three months.

[0137] In this embodiment, environmental data and user behavior data inside and outside the vehicle are collected in real time by a sensor network, including in-vehicle temperature, humidity, vehicle speed, air conditioning usage, and window opening status. After noise filtering and outlier removal, these data are input into a multi-channel variational encoder for hierarchical encoding of high frequency, low frequency, and emergencies. This step ensures comprehensive analysis and response to rapidly changing in-vehicle temperature and humidity, long-term stable air conditioning usage status, and sudden window opening operations. The spatiotemporal context perception module performs dynamic time series analysis on these encoded data through a spatiotemporal sliding window and a bidirectional spatiotemporal attention mechanism, enabling the system to identify and combine historical and real-time data for adaptive adjustment.

[0138] When the vehicle stops running for a long time, the temperature and humidity inside the vehicle gradually increase, especially in direct sunlight in summer, the temperature will quickly rise to above 35°C. At this time, the system of the present invention can detect the temperature change trend and adjust the fragrance remaining display data in real time through the dynamic prediction and control module and the fragrance remaining prediction model.

[0139] When driving with the windows open for a long time, the temperature and humidity in the car fluctuate significantly, and the frequent air flow affects the aromatherapy diffusion effect. Test data shows that when the car windows are open for more than 20 minutes and the vehicle speed is maintained at 40 kilometers per hour, the system of the present invention can capture these dynamic changes through the spatiotemporal context perception module, and adjust the aromatherapy remaining display data in real time in combination with the context weight injection mechanism. Compared with the traditional system, the present invention reduces 20% of the aromatherapy waste and improves the accuracy of the remaining management to more than 85%.

[0140] When the air conditioner is turned on for a long time, the humidity in the car will drop significantly. When the humidity in the car gradually drops from 50% to 30%, the system analyzes environmental data and user behavior through a multi-channel variational encoder and an aromatherapy remaining prediction model to ensure that the aromatherapy remaining can be accurately displayed even when the humidity changes greatly.

[0141] Table 1 Performance data of aromatherapy intelligent balance management in different scenarios

[0142]

[0143] Table 2 Analysis data of feedback mechanism and self-learning effect of the system of the present invention

[0144]

[0145] The data analysis results of Table 1 and Table 2 show that the accuracy of the margin management of the present invention in different in-car environments and usage scenarios is significantly better than that of the traditional system. Specifically, in the long-term parking scenario, when the temperature in the car rises to 35°C and the air conditioner is set to medium-speed cooling, the system of the present invention reduces the margin error rate from 20% of the traditional system to 8%. This shows that the system can flexibly adjust according to the changes in real-time temperature and humidity to ensure efficient margin management under static high-temperature conditions.

[0146] In the scenario of driving with the windows open for more than 20 minutes, when the temperature in the car is 30°C, the humidity is 70%, and the air conditioner is not used, the system of the present invention controls the margin error rate to 7%, which is 18% compared with the traditional system, showing excellent adaptability and stability. This data shows that the spatiotemporal context perception module of the present invention can achieve more accurate real-time response when detecting changes in air flow and user behavior in the car.

[0147] In the test scenario where the air conditioner was turned on for a long time, when the air conditioner was set to high wind speed cooling, the temperature in the car was maintained at 25°C and the humidity was reduced to 30%, the system's margin error rate was significantly reduced to 6.5%, while the error rate of the traditional system was 15%. This performance reflects the advantages of the present invention in adaptive optimization through feedback integration and self-supervised learning when the humidity fluctuates greatly.

[0148] The data of the night-time driving scene with open windows further verified the effectiveness of the present invention. When the temperature inside the car was 28°C, the humidity was 50%, and the air conditioner was set to low-speed heating, the system maintained the margin error rate at 7.5%, while the error rate of the traditional system was 17%. This shows that the present invention can achieve more sophisticated management and optimization in complex and changing environments through context weight injection and real-time feature adjustment mechanism.

[0149] The user feedback results in Table 2 further support these conclusions. In the scenarios of long-term parking and long-term operation of the air conditioner, the user satisfaction reached 8 and 9 points respectively, showing that users highly recognized the system's precise control in different situations. The system performed well in self-learning adjustment, and the self-learning frequency was maintained at 1 to 2 times / hour, indicating that the system can adapt to changes in the in-vehicle and out-of-vehicle environment and user behavior in a timely manner. For example, in the "long-term parking" scenario, the system adjusted the margin error from 10% to 7% through self-supervised learning, improving the prediction accuracy by 30%.

[0150] Comprehensive analysis of these data shows that the present invention demonstrates excellent margin management performance in different in-vehicle scenarios, especially under complex conditions such as long-term window opening, air conditioning use, and temperature and humidity changes. It has excellent adaptability and optimization capabilities, which significantly improves the user experience and system efficiency.

[0151] It can be seen from the above embodiments that the present invention can achieve continuous optimization and dynamic adjustment through self-supervised learning mechanism and feedback integration in complex and changeable driving environments. The system can maintain a high level of user satisfaction in different scenarios and significantly reduce the error rate of aromatherapy remaining prediction. Compared with traditional systems, it has higher adaptability and prediction accuracy.

[0152] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. The intelligent aromatherapy remainder calculation method based on dynamic perception of vehicle body data is characterized by: The steps include: S1. Collect environmental data, user behavior data and vehicle status data through sensors inside and outside the vehicle, and perform noise filtering, outlier removal and standardization on the data to generate an initial data set; S2, input the initial data set into the multi-channel variational encoder, classify and encode it according to high frequency, low frequency and sudden events, output hierarchical feature representation, and realize feature information exchange through the interactive network between channels, and dynamically adjust the variational loss function of the encoder; S3, input the hierarchical feature representation into the spatiotemporal context perception module, capture the time series data features through the spatiotemporal sliding window, generate the spatiotemporal feature representation in combination with the bidirectional spatiotemporal attention mechanism, and inject the historical and real-time context weights; S4, input the spatiotemporal feature representation into the heterogeneous feature mapping layer, perform spatiotemporal feature mapping and conversion, and map to generate a unified data representation; S5. Input the unified data representation into the dynamic prediction and control module, dynamically adjust the input feature weights according to real-time feedback, and generate an optimized feature representation; S6, inputting the optimized feature representation into the aromatherapy remainder prediction model, using a mixed Gaussian process to perform remainder prediction, and combining variational inference to update the parameters of the aromatherapy remainder prediction model; S7, input the fragrance remaining amount prediction result into the feedback integrator, collect user feedback and in-vehicle and out-of-vehicle data for self-supervised learning, and when long-term deviations or new data features are detected, trigger the dynamic intelligent reconstruction and optimization learning mechanism to perform structural update and parameter optimization of the fragrance remaining amount prediction model; S8. Integrate the updated aromatherapy remaining quantity prediction model into the vehicle-mounted system to display the aromatherapy remaining quantity status in real time, and issue personalized notifications and usage suggestions based on user behavior and aromatherapy remaining quantity prediction results.

2. The method for calculating the intelligent aromatherapy remainder based on dynamic perception of vehicle body data according to claim 1 is characterized in that: The vehicle body status data specifically includes the vehicle start / stop status, whether the windows are open, and whether the air conditioner in the vehicle is turned on.

3. The method for calculating the intelligent aromatherapy remainder based on dynamic perception of vehicle body data according to claim 1 is characterized in that: The S2 specifically includes: S21, inputting the environment data and user behavior data in the initial data set into the multi-channel variational encoder; S22. In a multi-channel variational encoder, the input data is hierarchically classified into three categories: high frequency, low frequency, and emergency events, wherein the high frequency channel extracts rapidly changing environmental features in the input data, the low frequency channel extracts long-term stable features in the input data, and the emergency event channel extracts unpredictable abnormal changes in the input data; S23, feature encoding is performed on the data in each channel, and the high-frequency feature encoder generates a high-frequency feature vector V high , the low-frequency feature encoder generates a low-frequency feature vector V low , the emergency feature encoder generates the emergency feature vector V event ; S24, feature information is exchanged through the interactive network between channels, and the high-frequency feature vector V high , low-frequency feature vector V low and the emergency feature vector V event Perform interactive processing and output the comprehensive feature matrix M exchange : Among them, k represents the dimension of the feature vector, and α represents the normalization coefficient; S25. During the feature information exchange process, the variational loss function of the encoder is dynamically adjusted using co-correlation analysis to optimize the encoding quality of each feature channel: Among them, L var represents the loss function, n represents the total number of features in the input dataset, KL represents the divergence value, γ i represents the weight coefficient, M represents the number of additional features, q(z|x i ) represents the encoded posterior distribution, p(z represents the prior distribution, represents the standard deviation of the input feature, ∈ represents the smoothing term, θ represents the adjustment factor, μ represents the offset, β and δ represent the adjustment parameters, T represents the time step, and λ represents the scaling parameter; S26. Output the adjusted hierarchical feature representation.

4. The method for calculating the intelligent aromatherapy remainder based on dynamic perception of vehicle body data according to claim 1 is characterized in that: The S3 specifically includes: S31, inputting the hierarchical feature representation into a spatiotemporal context perception module, wherein the spatiotemporal context perception module performs a time series analysis on the input multi-channel features to capture the dynamic changes of environmental data and user behavior in the time and space dimensions; S32, applying a spatiotemporal sliding window strategy to divide the input feature data into different time windows for processing, wherein the spatiotemporal sliding window dynamically adjusts the size and step length according to vehicle speed, environmental conditions and data changes to capture short-term and long-term time series features; S33. In each spatiotemporal sliding window, a bidirectional spatiotemporal attention mechanism is used to perform feature analysis. The bidirectional spatiotemporal attention mechanism captures the association between the current time step and past features through a forward attention network, and captures the association between the current time step and future features through a backward attention network, generating an attention weight covering global temporal information; S34, based on the attention weights covering global temporal information generated by the bidirectional spatiotemporal attention mechanism, weighted processing is performed on the features within the window, highlighting the attention to important features related to the current task, suppressing noise and irrelevant features, and forming a feature representation with temporal correlation and context association; S35, injecting historical context information and real-time context information into the spatiotemporal feature representation, and dynamically combining the weight of historical data and the weight of real-time data through context weight injection technology; S36. Output the spatiotemporal feature representation combining historical and real-time context information as input for unified data representation.

5. The method for calculating the intelligent aromatherapy remainder based on dynamic perception of vehicle body data according to claim 1 is characterized in that: The S4 specifically includes: S41, inputting the spatiotemporal feature representation output by the spatiotemporal context perception module into a heterogeneous feature mapping layer, wherein the heterogeneous feature mapping layer is composed of a plurality of parallel processing units, and the units respectively perform preliminary analysis and block division on features of different sources and time scales; S42. Through the double normalization strategy and adaptive filtering algorithm, the weight distribution of high-frequency features and low-frequency features is gradually balanced, and dynamic range adjustment is adopted to keep the features of each source in a stable distribution, thereby reducing the conversion deviation caused by inconsistent data scales; S43. In the heterogeneous feature mapping conversion network, a multi-layer convolution combined with a deformable convolution is used to map and reconstruct the features layer by layer, and the original spatiotemporal features are gradually converted into a comprehensive representation with a unified feature space. The deformable convolution is used to automatically adjust the shape of the convolution kernel to adapt to the changes of the features in different time periods. S44. In the feature mapping process, context features are embedded in combination with the output of the context perception module. The historical features are organically integrated with the current real-time features through the context attention mechanism to form a real-time context weight mapping. The input features are assigned fine-grained weights through dynamic adjustment of weight priorities. S45. Adopting a multi-channel feature aggregation strategy, the mapped high-frequency, low-frequency and emergency event features are hierarchically aggregated, and an interaction weight matrix between channels is introduced to achieve interactive enhancement of features and complementarity of multi-dimensional information; S46. Perform stability check on the final unified data representation before output, and perform output correction in combination with a feature smoothing mechanism.

6. The method for calculating the intelligent aromatherapy remainder based on dynamic perception of vehicle body data according to claim 1 is characterized in that: The S5 specifically includes: S51, inputting the generated unified data representation into a dynamic prediction and control module, wherein the dynamic prediction and control module includes a feature weighting unit and an adaptive feedback unit, which are used to adjust the weights of input features in real time to adapt to environmental changes; S52, in the feature weighting unit, a multi-layer adaptive weight allocation algorithm is used to perform weighted processing on the input unified data representation, the multi-layer adaptive weight allocation algorithm automatically allocates weights according to the change rate and correlation of the features at different time points, generates a preliminary optimized feature matrix, and updates the weights through an iterative process to adapt to the dynamic changes of the environment and user behavior; S53, receiving the change information of the external environment data and the user behavior data in real time through the adaptive feedback unit, comparing the feedback result with the current weight distribution, and adjusting the weight distribution strategy, gradually learning and adapting to the input feature changes under different environmental conditions based on the incremental learning mechanism; S54, screening the initially optimized feature matrix, using a feature selection algorithm to remove redundant and low-correlation features, retaining high-correlation features, and ensuring high information density, wherein the feature selection algorithm uses a weight pruning technology; S55, passing the filtered feature matrix to the dynamic control strategy module, and finally optimizing the feature matrix in combination with the context association model, wherein the context association model adjusts the time and environment priority of the features according to historical data and real-time feedback to generate a comprehensive optimized feature representation; S56. Output the optimized feature representation as the input of the aromatherapy remainder prediction module.

7. The method for calculating the intelligent aromatherapy remainder based on dynamic perception of vehicle body data according to claim 1 is characterized in that: The S6 specifically includes: S61, inputting the optimized feature representation into an aromatherapy residue prediction model, wherein the aromatherapy residue prediction model adopts a mixed Gaussian process structure and combines multi-dimensional feature input to achieve modeling of complex nonlinear relationships between features; S62, the aromatherapy remaining quantity prediction model consists of a two-layer feature processing unit. The first layer is the input feature mapping layer, which embeds the optimized feature representation into the high-dimensional feature space; the second layer adopts a multi-core Gaussian process structure, integrating linear kernel, radial basis kernel and polynomial kernel functions to capture the complex interactive relationship between high-frequency, low-frequency and emergency event features; S63. In the Gaussian process core layer, define the margin prediction formula, and realize adaptive adjustment of margin prediction by dynamically weighing different feature combinations: Among them, y 余量 (t) represents the prediction result of aromatherapy residue, μ 高频 ( t ) Represents the predicted mean of high-frequency features, μ 低频 ( t ) Represents the predicted mean of low-frequency features, μ 突发 ( t ) represents the predicted mean of the emergency characteristics, σ 低频 ( t represents the prediction standard deviation of low-frequency features, σ 高频 ( t ) Represents the prediction standard deviation of high-frequency features, σ 突发 ( t ) represents the predicted standard deviation of the emergency event characteristics, α 1,j , α 2,l and λ r represents the weight adjustment coefficient, β 1,i and β 2,p represents the integral adjustment coefficient, γ 1,i and γ 2,p represents the scaling parameter, δ3 represents the global adjustment coefficient, and f j ( t ) represents the input feature gradient function at time t, h l (t) represents the input feature correction function at time t, g r (t) represents the correction function of the sudden event, k represents the number of weight coefficients in the high-frequency feature, m represents the number of adjustment parameters of the high-frequency feature, n represents the number of adjustment coefficients of the low-frequency feature, q represents the number of adjustment parameters used for time series integration in the low-frequency feature, s represents the number of correction functions in the sudden event, u represents the number of time steps used for cumulative standard deviation calculation in the sudden event part, and t v represents the low-frequency features at time step v, θ1 and θ2 represent threshold parameters; S64, using variational inference technology to update model parameters in real time, optimize the posterior distribution of the Gaussian process, and ensure that the model can be adaptively adjusted according to changes in input; S65. In the residual forecasting process, the uncertainty quantitative analysis of the model is combined to estimate the confidence interval of the forecast results to identify and deal with potential forecast deviations; S66, output aromatherapy remaining amount prediction result y 余量 (t), combining multidimensional features and uncertainty information to generate aromatherapy remainder prediction results with high confidence.

8. The method for calculating the intelligent aromatherapy remainder based on dynamic perception of vehicle body data according to claim 1 is characterized in that: The S7 specifically includes: S71, inputting the prediction result of the fragrance remaining amount prediction model into a feedback integrator, wherein the feedback integrator continuously collects data from the environment inside and outside the vehicle, including temperature, humidity, vehicle speed and user operation behavior information, and combines the user feedback to form a multi-dimensional data set; S72. Based on the multi-dimensional data set, the actual environmental data is compared with the predicted results, a deviation analysis report is generated, the deviation between the predicted value and the actual value is identified, and whether the deviation exceeds a preset long-term threshold is determined; S73. When the deviation analysis result shows that the long-term threshold is exceeded or a new feature appears, the self-supervised learning module is started to generate pseudo labels through unlabeled data, expand the training data set, automatically learn new environmental features, and improve the adaptive ability of the aromatherapy residue prediction model to cope with the environment; S74, inputting the expanded training data set into the incremental learning unit of the aromatherapy residue prediction model, wherein the incremental learning unit performs progressive parameter updates on the aromatherapy residue prediction model to maintain efficient operation of the aromatherapy residue prediction model in a real-time environment and optimize prediction performance; S75. When long-term deviations are detected frequently or new data features continue to appear, a dynamic intelligent reconstruction mechanism is triggered to automatically adjust the structure of the aromatherapy remainder prediction model, including optimizing the model hierarchy, reconfiguring the weight distribution, and adjusting the feature path; S76. After completing the structural adjustment and parameter optimization of the aromatherapy remaining quantity prediction model, the improved aromatherapy remaining quantity prediction model is applied to the vehicle-mounted system, and the updated version information and adjustment details of the aromatherapy remaining quantity prediction model are recorded.

Images