Oil fume purification path optimization method and system based on multi-source data fusion

Abstract

This invention discloses a method and system for optimizing oil fume purification paths based on multi-source data fusion, belonging to the field of data processing technology. It includes: collecting environmental perception data within a target space using a multi-source sensor array; inferring the attenuation basis function parameters of each sensor through a parameter decision model, and constructing an anisotropic concentration attenuation field for the sensors by combining a priori concentration attenuation field and the attenuation basis function parameters; constructing a concentration prediction field for each sensor based on discrete concentration measurements and the concentration attenuation field, performing weighted fusion to obtain an inverted concentration field for the target space; inputting the inverted concentration field and wind speed and direction data into a concentration trend prediction model to generate a prediction result of dynamic concentration change trends; and performing global path optimization with the goal of maximizing the total purification efficiency of the mobile purification equipment's path coverage area to generate the optimal purification path. This invention effectively improves the operational efficiency and adaptive governance capabilities of oil fume purification.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for optimizing oil fume purification paths based on multi-source data fusion. Background Technology

[0002] With the large-scale development of catering, industrial kitchens and other places, autonomous mobile platforms equipped with oil fume purification modules have become an emerging technology direction for kitchen oil fume treatment. At present, oil fume concentration field modeling mainly adopts three methods: CFD simulation, direct sensor measurement, and empirical models. Mobile purification path planning is mostly aimed at regional coverage or shortest path, while real-time obstacle avoidance relies on conventional methods such as artificial potential fields and dynamic windows. Together, they constitute the existing technical system of mobile oil fume purification.

[0003] However, while CFD simulation offers high accuracy, it comes at a huge computational cost and cannot meet real-time requirements. Sensor measurements offer good real-time performance, but are limited by quantity and suffer from spatial undersampling. Empirical models offer fast calculations, but differ significantly from the actual flow field and lack accuracy. They cannot simultaneously achieve both high accuracy and real-time performance in concentration field modeling, making it difficult to realize efficient and adaptive mobile fume purification. Summary of the Invention

[0004] This invention provides a method and system for optimizing the fume purification path based on multi-source data fusion, aiming to solve the technical problems of insufficient operation efficiency and adaptive treatment capability of fume purification in the prior art.

[0005] In view of the above problems, the present invention provides a method and system for optimizing the fume purification path based on multi-source data fusion.

[0006] In a first aspect, the present invention provides a method for optimizing the fume purification path based on multi-source data fusion, including: Real-time environmental perception data within the target space is collected using a multi-source sensor array; Based on the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through the parameter decision model. Combined with the prior concentration attenuation field and the attenuation basis function parameters, an anisotropic concentration attenuation field is constructed for each sensor. Based on the discrete concentration measurements of each sensor in the environmental perception data and the concentration decay field, a concentration prediction field is constructed for each sensor, and multiple concentration prediction fields are weighted and fused to obtain the inversion concentration field of the target space. The inverted concentration field and the collected wind speed and direction data are input into the concentration trend prediction model to generate the dynamic change trend prediction results of the concentration at each spatial location in the target space in the future time period. Based on the predicted results of the dynamic change trend of the concentration, with the core optimization objective of maximizing the total purification efficiency of the area covered by the mobile purification equipment path, global path optimization is performed to generate the optimal purification path.

[0007] Secondly, this invention provides an oil fume purification path optimization system based on multi-source data fusion, comprising: The environmental perception data acquisition module is used to acquire environmental perception data in the target space in real time through a multi-source sensor array; The concentration decay field construction module is used to combine the environmental sensing data, infer the decay basis function parameters corresponding to each sensor through the parameter decision model, and combine the prior concentration decay field with the decay basis function parameters to construct an anisotropic concentration decay field for each sensor. The inversion concentration field fusion module is used to construct a concentration prediction field for each sensor based on the discrete concentration measurement values ​​of each sensor in the environmental perception data and the concentration decay field, and to perform weighted fusion of multiple concentration prediction fields to obtain the inversion concentration field of the target space. The concentration trend prediction module is used to input the inverted concentration field and the collected wind speed and direction data into the concentration trend prediction model to generate the dynamic change trend prediction results of the concentration at each spatial location in the target space in the future time period. The optimal path optimization module is used to perform global path optimization based on the predicted results of the dynamic change trend of the concentration, with the core optimization objective of maximizing the total purification efficiency of the area covered by the path of the mobile purification equipment, and generate the optimal purification path.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a method and system for optimizing fume purification paths based on multi-source data fusion. It utilizes a multi-source sensor array to collect environmental perception data in real time, providing a comprehensive and real-time multi-source data foundation for fume concentration field modeling and purification path optimization. An anisotropic concentration decay field is constructed based on environmental perception data and a parameter decision model, accurately conforming to kitchen space conditions and the physical laws of fume diffusion, thus improving the accuracy and adaptability of concentration field modeling. A weighted fusion of the concentration decay field and discrete concentration measurements yields an inverted concentration field, effectively compensating for undersampling defects in sensor space and achieving high-resolution, high-precision concentration field reconstruction across the entire space under sparse sensing conditions. Combining the inverted concentration field with wind speed and direction data, dynamic concentration trend prediction is completed, providing a forward-looking decision-making basis for path planning and adapting to dynamic changes in cooking conditions and wind fields. Finally, with the core objective of maximizing total purification efficiency, global path optimization is conducted. The generated optimal purification path can accurately target high-concentration areas and dynamically respond to concentration changes, comprehensively improving the operational efficiency and adaptive governance capabilities of fume purification. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the oil fume purification path optimization method based on multi-source data fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the oil fume purification path optimization system based on multi-source data fusion provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: The system includes an environmental perception data acquisition module 11, a concentration decay field construction module 12, an inversion concentration field fusion module 13, a concentration trend prediction module 14, and an optimal path optimization module 15. Detailed Implementation

[0010] This invention provides a method and system for optimizing the fume purification path based on multi-source data fusion, which is used to address the technical problems of insufficient operational efficiency and adaptive treatment capability in existing fume purification technologies.

[0011] Example 1, as Figure 1 As shown, this invention provides a method for optimizing the fume purification path based on multi-source data fusion, the method comprising: S100: Real-time acquisition of environmental perception data within the target space through a multi-source sensor array.

[0012] In this embodiment of the invention, a multi-source sensor array is used to collect environmental perception data in the target space in real time. The concentration field modeling and path planning of mobile fume purification systems rely on spatiotemporally synchronized multi-source heterogeneous data. Single-source fume concentration data cannot reflect key constraints such as spatial layout, equipment location, and wind field. If there are temporal misalignments or spatial mismatches between different source data, it will directly lead to distortion in the concentration decay field construction and deviations from actual operating conditions in path optimization. Therefore, it is necessary to synchronously collect ambient air data through a multi-source sensor array, combine it with spatial layout and equipment location data, and form unified environmental perception data after temporal alignment. This provides an accurate and complete input foundation for subsequent concentration field inversion, trend prediction, and path optimization.

[0013] Step S100 in the method provided in this embodiment of the invention includes: Ambient air data of the target space is collected in real time through a multi-source sensor array. The ambient air data includes discrete oil fume concentration, wind speed, and wind direction, and the sampling position of each sensor in the multi-source sensor array is obtained simultaneously. Acquire spatial layout and relative equipment position data of the target space as environmental spatial data; After time alignment, the ambient air data and the ambient space data are merged and output as the environmental perception data.

[0014] First, ambient air data of the target space is collected in real time using a multi-source sensor array. This ambient air data includes discrete oil fume concentration, wind speed, and wind direction, and the sampling position of each sensor in the multi-source sensor array is acquired simultaneously. The multi-source sensor array is a data acquisition unit composed of various types of sensors at multiple locations, including oil fume concentration sensors and wind speed and direction sensors, used to acquire multi-dimensional environmental parameters. Ambient air data characterizes the real-time air quality and airflow state of the target space, including discrete oil fume concentration, wind speed, and wind direction. The sampling position refers to the fixed installation coordinates of the sensor in the three-dimensional coordinate system of the target space, which is the basis for spatial positioning of the concentration field.

[0015] Specifically, a multi-source sensor array is deployed in the target space according to preset rules to collect discrete data on oil fume concentration, wind speed, and wind direction at a fixed sampling period; the three-dimensional sampling position of each sensor is recorded through pre-calibration or coordinate measurement, and stored in conjunction with the real-time collected air data.

[0016] For example, taking a 4m×3m×2.8m commercial kitchen as the target space, an array consisting of 6 laser scattering type oil fume concentration sensors and 2 ultrasonic wind speed and direction sensors is deployed; the sensor sampling period is set to 1 second, and data is collected in real time: the sensor above the stove collects an oil fume concentration of 1.8mg / m³. 3 The ventilation outlet wind speed was 0.6 m / s and the wind direction was north; the sampling position of the oil fume sensor was recorded simultaneously as three-dimensional coordinates (1.0 m, 1.5 m, 0.5 m).

[0017] Secondly, acquire spatial layout and relative equipment position data of the target space as environmental spatial data. Environmental spatial data is static data characterizing the physical structure of the target space and the relative positions of equipment, including spatial layout, wall positions, and the relative positions of stoves and range hoods. Spatial layout refers to the physical structural information of the target space, such as geometric dimensions and the distribution of walls or obstacles. Relative equipment positions refer to the distances and orientations between smoke-generating or smoke-exhausting equipment such as stoves and range hoods.

[0018] Specifically, by installing pre-configured data, using depth camera vision recognition, or importing CAD drawings, the geometric dimensions of the target space, wall boundaries, and the installation positions and relative distances of the stove and range hood are extracted and organized into standardized environmental space data.

[0019] For example, the spatial layout of the above-mentioned commercial kitchen is a rectangular enclosed space with wall boundary coordinates from (0,0) to (4,3); the center position of the stove is (1.0, 1.5, 0.0), and the position of the range hood intake is (1.0, 1.5, 1.8). The vertical distance between the two is 1.8m and the horizontal distance is 0m. The above data constitutes the environmental space data of the kitchen.

[0020] Finally, after time alignment, the ambient air data and the ambient spatial data are merged and output as the environmental perception data. Time alignment refers to unifying multi-source data with different sampling frequencies and collection times to the same timestamp, eliminating data time misalignment. The environmental perception data is a collection of multi-source data after time alignment and fusion, and it is the core input for subsequent model inference.

[0021] Specifically, a unified standard timestamp is assigned to ambient air data and ambient spatial data. The timestamps of high-frequency collected air data and static spatial data are matched, abnormal time-series data are removed, and finally the two types of data are merged and output as standardized environmental sensing data.

[0022] For example, the collected data is uniformly timestamped as 2026-04-13 10:00:00. The data on oil fume concentration, wind speed and direction, sensor sampling location, spatial layout, and relative equipment position are integrated to form complete environmental perception data containing spatiotemporal information, environmental parameters, and physical structure, which is then output to subsequent modules.

[0023] In this embodiment of the invention, environmental perception data that is time-synchronized, spatially matched, and dimensionally complete is constructed by collecting data from multiple sources and fusing spatiotemporal data. This solves the problems of insufficient single data dimension and spatiotemporal misalignment of heterogeneous data. It provides accurate input support for subsequent attenuation basis function parameter inference and concentration attenuation field construction, ensuring the authenticity and reliability of the entire process modeling and path optimization.

[0024] S200: Combining the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through the parameter decision model, and an anisotropic concentration attenuation field for each sensor is constructed by combining the prior concentration attenuation field with the attenuation basis function parameters.

[0025] In this embodiment of the invention, by combining the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through a parameter decision model. Combined with the prior concentration attenuation field and the attenuation basis function parameters, an anisotropic concentration attenuation field is constructed for each sensor. The diffusion of cooking fumes in a kitchen is influenced by multiple factors such as stove layout, range hood airflow, wall obstruction, and cooking conditions, exhibiting strong anisotropic characteristics. Traditional isotropic concentration attenuation models cannot match the actual diffusion patterns of cooking fumes. Furthermore, the concentration attenuation benchmarks corresponding to different sensor installation locations differ, and direct modeling can easily lead to distorted concentration field predictions. Therefore, it is necessary to first complete the construction of the working space, simulation calibration, and model training offline, and then combine the environmental perception data online. The parameter decision model quickly infers the specific attenuation basis function parameters for each sensor, ultimately constructing an anisotropic concentration attenuation field for each sensor, providing an accurate prior model that conforms to physical laws for subsequent concentration field inversion.

[0026] Step S200 in the method provided in this embodiment of the invention includes: Based on the environmental perception data, a scene condition feature vector for the target space is constructed; The scene condition feature vectors corresponding to each sensor are input into the trained parameter decision model for inference, and the decay basis function parameters corresponding to each sensor are output. Based on the attenuation basis function parameters corresponding to each sensor, and combined with the sampling position information of each sensor, the attenuation basis function of the prior concentration attenuation field is substituted into the attenuation basis function to construct an independent anisotropic concentration attenuation field for each sensor. The concentration attenuation field is used to output the concentration attenuation prediction value of any position in the target space relative to the sampling position of the corresponding sensor.

[0027] Specifically, by combining the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through a parameter decision model. Then, by combining the prior concentration attenuation field with the attenuation basis function parameters, an anisotropic concentration attenuation field is constructed for each sensor. Prior to this, the process includes: Construct a working condition parameter space, wherein the working conditions include at least: stove layout type, distance between stove and nearest wall, range hood air volume setting, number of stoves working at the same time, and cooking method type; Based on the aforementioned working condition parameter space, a simulation model covering multiple combinations of working conditions is established. The simulation model is used to solve the convection and diffusion equations of oil fume aerosol in the target space and obtain the simulated concentration field under each combination of working conditions. Define the mathematical form of the attenuation basis function based on the kernel function, wherein the attenuation basis function is a product of multiple spatial modulation factors, wherein the spatial modulation factors include at least: Source factor, used to describe the concentration distribution benchmark centered on the stove location and decreasing with spatial distance; Spatial attenuation factor, used to describe the anisotropic distribution characteristics of concentration attenuation with distance in each direction; The flow modulation factor is used to describe the enhanced modulation of the concentration distribution along the direction of the smoke machine by the smoke machine's suction effect. The wall constraint factor describes the attenuation modulation of the concentration distribution on the wall side due to the wall's blocking effect on the concentration diffusion path. For each combination of working conditions in the working condition parameter space, with the corresponding simulated concentration field as the target distribution, the optimal set of parameters that minimizes the error metric between the decay basis function and the simulated concentration field is obtained by parameter inversion optimization. The feature descriptions of each combination of working conditions are paired with the corresponding set of optimal parameters to form a calibration dataset.

[0028] First, a parameter space for operating conditions is constructed. These operating conditions include at least: cooktop layout type, distance between the cooktop and the nearest wall, range hood fan speed setting, number of cooktops operating simultaneously, and cooking method type. The parameter space is a set of parameters composed of variables such as cooktop layout, range hood fan speed, and number of cooktops, covering all typical kitchen fume diffusion conditions. Operating condition variables are selected, and their value ranges and types are defined to form a complete parameter space covering typical operating states in both commercial and residential kitchens.

[0029] For example, taking a 4m×3m×2.8m commercial kitchen as the target space, construct the working parameter space: the stove layout type is a single stove in the center, the distance between the stove and the nearest wall is 1.5m, the range hood air volume is medium at 1200m3 / h, the number of stoves working at the same time is 1, and the cooking method is stir-frying.

[0030] Secondly, based on the aforementioned operating condition parameter space, a simulation model covering various combinations of operating conditions is established. This simulation model is used to solve the convection-diffusion equations for oil fume aerosols in the target space, obtaining the simulated concentration field under each combination of operating conditions. The simulation model is a three-dimensional flow field model of a kitchen built based on computational fluid dynamics (CFD), used to simulate the oil fume diffusion process. The convection-diffusion equations are the physical governing equations describing the flow and diffusion of oil fume aerosols in space, and serve as the basis for calculating the simulated concentration field.

[0031] Specifically, for each combination of operating conditions in the operating condition parameter space, a three-dimensional geometric model of the kitchen is established, a computational mesh is divided, boundary conditions are set, the convection-diffusion equation is solved, and the three-dimensional simulation concentration field data under that operating condition is output after convergence.

[0032] For example, for the above-mentioned stir-frying conditions, a kitchen CFD simulation model is established, divided into 1 million level unstructured meshes, and the oil fume convection and diffusion equation is solved to obtain a steady-state simulation concentration field with the highest concentration above the stove, converging towards the range hood, and the concentration rapidly decreasing on the wall side.

[0033] Furthermore, the mathematical form of the attenuation basis function based on the kernel function is defined, wherein the attenuation basis function is a product of multiple spatial modulation factors, the spatial modulation factors including at least: Source factor, used to describe the concentration distribution benchmark centered on the stove location and decreasing with spatial distance; Spatial attenuation factor, used to describe the anisotropic distribution characteristics of concentration attenuation with distance in each direction; The flow modulation factor is used to describe the enhanced modulation of the concentration distribution along the direction of the smoke machine by the smoke machine's suction effect. The wall constraint factor describes the attenuation modulation of the concentration distribution on the wall side due to the wall's blocking effect on the concentration diffusion path.

[0034] The attenuation basis function is a mathematical model based on kernel functions that describes the spatial attenuation of oil fume concentration. It consists of the product of multiple spatial modulation factors. The spatial modulation factors characterize the influence of different physical factors on concentration attenuation, including the source term factor, spatial attenuation factor, flow modulation factor, and wall constraint factor.

[0035] Specifically, the attenuation basis function is defined as the product of four spatial modulation factors, and the physical meaning of each factor is clarified: source term factor: the concentration distribution benchmark centered on the stove; spatial attenuation factor: the anisotropic characteristics of concentration attenuation with distance in each direction; flow modulation factor: the directional enhancement effect of the range hood's suction on the concentration; wall constraint factor: the wall-side concentration attenuation effect caused by the wall obstruction.

[0036] For example, for the kitchen mentioned above, the attenuation basis function is defined as: Attenuation basis function = source term factor × spatial attenuation factor × flow modulation factor × wall constraint factor, which correspond to the four major physical effects of stove smoke generation, spatial diffusion, range hood suction, and wall obstruction, respectively.

[0037] Subsequently, for each combination of operating conditions in the parameter space, using the corresponding simulated concentration field as the target distribution, the optimal set of parameters that minimizes the error metric between the attenuation basis function and the simulated concentration field is obtained through parameter inversion optimization. The optimal set of parameters refers to the combination of parameters that minimizes the error between the attenuation basis function and the simulated concentration field, including diffusion parameters, drainage parameters, wall constraint parameters, and concentration distribution center offset parameters. Parameter inversion optimization refers to the process of finding the optimal parameters of the attenuation basis function using an optimization algorithm with the simulated concentration field as the target. Using the simulated concentration field as the target distribution, a nonlinear least squares algorithm is employed to iteratively find the optimal set of parameters that minimizes the error between the attenuation basis function output and the simulated concentration field.

[0038] For example, for the simulated concentration field under the stir-frying condition, the optimal parameter set is obtained by inversion: diffusion parameters: horizontal 0.8m, vertical 1.2m; diversion parameters: convergence coefficient in the direction of the smoke machine 0.9; wall constraint parameters: wall side attenuation coefficient 0.7; concentration distribution center offset parameters: offset in the direction of the smoke machine 0.3m.

[0039] Finally, the feature descriptions of each combination of operating conditions are paired with the corresponding optimal parameter sets to form a calibration dataset. The calibration dataset, composed of paired operating condition features and corresponding optimal parameter sets, is used to train the parameter decision model. The feature descriptions of each set of operating conditions are paired one-to-one with the corresponding inverted optimal parameter sets, and then organized into a standardized calibration dataset.

[0040] For example, the above operating conditions characteristics—single stove, 1.5m wall distance, medium air volume, 1 stove, stir-frying—are paired with the optimal parameter set and used as a sample in the calibration dataset.

[0041] This includes creating a calibration dataset, followed by: Construct a scene condition feature vector, which includes at least the following encoded features: stove layout type, distance between the stove and the nearest wall, range hood fan speed setting, number of stoves working simultaneously, and cooking method type. Based on the calibration dataset and the data format of the scenario condition feature vector, the feature descriptions of each working condition combination are combined into an input feature vector, and the corresponding optimal parameter set is defined as a label to construct the training sample set of the parameter decision model. Construct the parameter decision model based on a multi-output neural network, define a loss function based on physical constraints and parameter prediction error, and train the parameter decision model using the training sample set until the loss function converges to below a preset threshold to obtain the trained parameter decision model.

[0042] First, a scene condition feature vector is constructed. This scene condition feature vector includes at least the following encoded features: stove layout type, distance between the stove and the nearest wall, range hood fan speed setting, number of stoves operating simultaneously, and cooking method type. The scene condition feature vector refers to encoding the operating condition parameters into a numerical vector, which serves as the input format for the parameter decision model.

[0043] Specifically, the scene condition feature vector adopts a hybrid coding rule of fixed numerical mapping encoding for categorical variables and direct assignment of continuous / discrete numerical variables. The vector has 5 dimensions, corresponding to the stove layout type, the distance between the stove and the nearest wall, the range hood fan speed setting, the number of stoves working at the same time, and the cooking method type. Among them, the stove layout type, the range hood fan speed setting, and the cooking method type are categorical variables, which are assigned unique integer codes according to their types. The distance between the stove and the nearest wall is a continuous variable that directly uses the measured value, and the number of stoves working at the same time is a discrete numerical variable that directly uses the actual quantity value.

[0044] For example, according to the above coding rules, the following conditions are encoded as 1, 1.5, 2, 1, 3 respectively: single stove center layout, stove distance from the nearest wall is 1.5m, range hood medium airflow, one stove working at the same time, and stir-fry cooking mode. These conditions are combined to form a standardized scene condition feature vector [1, 1.5, 2, 1, 3].

[0045] Secondly, based on the calibration dataset and the data format of the scenario condition feature vector, the feature descriptions of each working condition combination are combined into an input feature vector, and the corresponding optimal parameter set is defined as a label to construct the training sample set of the parameter decision model.

[0046] Specifically, the training sample set is constructed using the calibration dataset as the sole source and strictly matches the 5-dimensional standardized format of the scene condition feature vector. Each combination of working conditions is converted into an input feature vector according to the hybrid encoding rule, and the optimal parameter set corresponding to the working condition is used as the label to complete one-to-one sample pairing. All samples are divided into training set, validation set and test set in a ratio of 8:1:1. The dimension of the input feature vector is fixed at 5 dimensions, and the dimension of the label is consistent with the number of parameters in the optimal parameter set.

[0047] For example, based on the previously calibrated dataset, the encoded 5-dimensional input feature vector [1, 1.5, 2, 1, 3] is used as the model input, and the corresponding paired optimal parameter set [diffusion parameter (0.8, 1.2), flow guidance parameter 0.9, wall constraint parameter 0.7, offset parameter 0.3] is used as the label to form a single training sample. After pairing the feature vectors of all working conditions with the optimal parameters in this way, the dataset is divided into 8:1:1 to complete the construction of the model training sample set.

[0048] Finally, a parameter decision model based on a multi-output neural network is constructed, and a loss function is defined based on physical constraints and parameter prediction error terms. The parameter decision model is then trained using the training sample set until the loss function converges to below a preset threshold, resulting in a trained parameter decision model. The parameter decision model is an inference model based on a multi-output neural network, comprising a shared feature extraction layer and multiple parameter prediction branches. The shared feature extraction layer extracts shared features related to the scene and location. The parameter prediction branches output diffusion parameters, flow parameters, and wall constraint parameters, respectively. The loss function, composed of physical constraints and parameter prediction error terms, is used for model training and optimization.

[0049] Specifically, the parameter decision model adopts a multi-output neural network structure, including one shared feature extraction layer and three independent parameter prediction branches. The shared feature extraction layer has two hidden layers: the first layer has 128 neurons and the second layer has 64 neurons. The activation function for both layers is ReLU, used to extract shared features of the scene and location. The three prediction branches correspond to the outputs of diffusion parameters, flow parameters, and wall constraint parameters, respectively, and have no activation functions. The loss function is a weighted combination of the parameter prediction error term (L2 mean squared error) and the physical constraint term. Training uses the Adam optimizer with an initial learning rate of 0.001. The validation set loss is used as the convergence criterion. The model is considered convergent when the loss does not decrease for 50 consecutive iterations or when the loss value drops to a preset threshold of 10. -4Training will stop at the following point, indicating that the model has converged.

[0050] For example, a multi-output neural network with two shared hidden layers and three parameter prediction branches is built using a 5-dimensional scene feature vector as input. The training sample set constructed above is input into the model, and the Adam optimizer is used for iterative training. The loss is calculated by combining the parameter prediction L2 error and the concentration diffusion physical constraint. The model is continuously iterated until the validation set loss does not decrease for 50 consecutive rounds and is lower than 10-4. The model converges, and the trained parameter decision model is obtained.

[0051] Based on this, and combined with the environmental perception data, a scene condition feature vector for the target space is constructed. The environmental perception data refers to the time-aligned multi-source fusion data output from step S100, which includes information such as spatial layout, equipment location, range hood operating status, sensor location, and cooking status.

[0052] For example, taking a 4m×3m×2.8m commercial kitchen as an example, the following are extracted from real-time environmental perception data: single stove in the center, distance between the stove and the nearest wall is 1.5m, range hood at medium airflow, 1 stove working, cooking method is stir-frying; and scene condition feature vector is generated according to the coding rules: [1, 1.5, 2, 1, 3].

[0053] Then, the scene condition feature vectors corresponding to each sensor are input into the trained parameter decision model for inference, and the attenuation basis function parameters corresponding to each sensor are output. The scene condition feature vectors corresponding to each sensor are input into the trained and converged parameter decision model; the parameter decision model extracts scene and location features through a shared feature extraction layer, and then each parameter prediction branch infers in parallel to output the diffusion parameters, flow parameters, and wall constraint parameters specific to that sensor.

[0054] For example, the scene feature vector [1, 1.5, 2, 1, 3] corresponding to the sensor above the stove is input into the parameter decision model, and the parameter decision model infers and outputs the attenuation basis function parameters of the sensor: diffusion parameter (horizontal 0.8m, vertical 1.2m), drainage parameter 0.9, and wall constraint parameter 0.7.

[0055] Finally, based on the attenuation basis function parameters corresponding to each sensor and combined with the sampling position information of each sensor, the attenuation basis function of the prior concentration attenuation field is substituted into the attenuation basis function to construct an independent anisotropic concentration attenuation field for each sensor. This concentration attenuation field is used to output the predicted concentration attenuation value at any position in the target space relative to the corresponding sensor sampling position. The prior concentration attenuation field refers to a predefined, parameter-free basic concentration attenuation field based on the physical laws of oil fume diffusion. The attenuation basis function is a function composed of the product of the source term, spatial attenuation, diversion modulation, and wall constraint factor. The anisotropic concentration attenuation field is a dedicated attenuation field with different concentration attenuation rates in different spatial directions, based on the sensor sampling position, and can calculate the predicted attenuation value at any position. The sampling position information refers to the fixed coordinates of the sensor in the three-dimensional coordinate system of the kitchen, providing a spatial reference for the attenuation field.

[0056] Specifically, the attenuation basis function parameters of each sensor and the sampling position coordinates are substituted into the attenuation basis function of the prior concentration attenuation field; with the sensor sampling position as the center, the concentration attenuation coefficient of all points in the space is calculated according to the anisotropic attenuation rule, and an independent anisotropic concentration attenuation field is generated for each sensor.

[0057] For example, the attenuation basis function parameters and sampling position of the sensor above the stove are substituted into the attenuation basis function of the prior concentration attenuation field; an anisotropic concentration attenuation field with the sensor as the center, slow attenuation in the direction of the range hood and fast attenuation in the direction of the wall is generated, and the predicted concentration attenuation value of any point in the kitchen relative to the sensor can be output.

[0058] In this embodiment of the invention, a two-layer architecture of offline simulation calibration and online model inference is adopted. This architecture ensures the physical authenticity of the decay basis function through CFD simulation and achieves millisecond-level fast parameter inference through neural networks. An independent anisotropic concentration decay field is constructed for each sensor, which accurately matches the sensor position, spatial layout and real-time operating conditions. This completely solves the problems of isotropy, poor adaptability to operating conditions and mismatch of sensor position in traditional decay models, and provides high-precision and high-real-time prior model support for the subsequent construction of the concentration prediction field.

[0059] S300: Based on the discrete concentration measurements of each sensor in the environmental perception data and the concentration decay field, a concentration prediction field is constructed for each sensor, and multiple concentration prediction fields are weighted and fused to obtain the inversion concentration field of the target space.

[0060] In this embodiment of the invention, based on the discrete concentration measurements of each sensor in the environmental perception data and the concentration decay field, a concentration prediction field is constructed for each sensor, and multiple concentration prediction fields are weighted and fused to obtain the inverted concentration field of the target space. A single sensor can only obtain local discrete concentration values, and its corresponding concentration decay field only represents the decay ratio relative to the sampling position. Using it alone cannot restore the true concentration distribution of the entire space. Moreover, the prediction reliability of different sensors varies due to spatial distance, model fit, and their own working state. Direct fusion will lead to distortion of the concentration field. Therefore, it is necessary to first generate a single-sensor concentration prediction field based on the sensor's measured values ​​and the dedicated decay field, and then fuse them through three-dimensional confidence weighted fusion, while marking low-confidence areas, to finally obtain a high-precision and high-reliability full-space inverted concentration field.

[0061] Step S300 in the method provided in this embodiment of the invention includes: Using the discrete concentration measurements from each sensor as the baseline intensity, and substituting them into the concentration decay field of each sensor, the concentration prediction field is obtained. The confidence scores of the concentration prediction fields of each sensor at each location in the target space are calculated. The confidence scores are composed of the product of three components: spatial distance confidence score, model prediction confidence score, and sensor state confidence score. The concentration prediction fields of each sensor are weighted and fused using the confidence scores of each sensor at each location in the target space to obtain the inverted concentration field. Calculate the fusion confidence score at each location in the target space, where the fusion confidence score is the sum of the confidence scores of each sensor at that location. When the fusion confidence score is lower than a preset threshold, the concentration estimate at the corresponding location is marked as unreliable.

[0062] First, the discrete concentration measurements from each sensor are used as the baseline intensity, and substituted into the concentration decay field of each sensor to obtain the concentration prediction field. The discrete concentration measurements are the real-time measured values ​​of oil fume concentration collected by each sensor in the environmental sensing data, providing the baseline intensity for concentration prediction. The concentration decay field is an anisotropic concentration decay field specific to each sensor, representing only the concentration decay ratio at any location in the target space relative to the sampling location of that sensor. The concentration prediction field refers to the full-space concentration distribution field from the sensor's perspective, calculated based on the sensor's measured concentration and the decay ratio. The real-time discrete concentration measurements from each sensor are used as the baseline intensity, multiplied by the decay ratio at each location in the corresponding concentration decay field of that sensor, to calculate the sensor's specific concentration prediction field point by point.

[0063] For example, in a 4m×3m×2.8m commercial kitchen, the discrete concentration measurement value of the sensor above the stove (sampling position (1.0, 1.5, 0.5)) is 1.8 mg / m³.3 The attenuation ratio at the center point of the kitchen (2.0, 1.5, 1.0) in the concentration attenuation field of the sensor is 0.6. Therefore, the predicted concentration at this point is 1.8 × 0.6 = 1.08 mg / m³. 3 The concentration prediction field of the sensor is generated by calculating point by point.

[0064] Secondly, the confidence scores for the concentration prediction fields of each sensor at each location in the target space are calculated. These confidence scores are the product of three components: spatial distance confidence score, model prediction confidence score, and sensor state confidence score. The confidence score characterizes the reliability of the sensor concentration prediction field at the target location and is obtained by multiplying the spatial distance confidence score, model prediction confidence score, and sensor state confidence score.

[0065] The spatial distance confidence score reflects the spatial distance relationship between the target location and the corresponding sensor sampling location. The closer the target location is to the corresponding sensor sampling location, the higher the spatial distance confidence score. The model prediction confidence score reflects the reliability of the parameter decision model's prediction of the corresponding sensor attenuation basis function parameters under the current scene conditions. The higher the density of the current scene conditional feature vector in the conditional feature space of the training data, the higher the model prediction confidence score. The sensor state confidence score reflects the current operational health status of the corresponding sensor. The sensor state confidence score is composed of the product of at least two of the following factors: sensor self-diagnostic health status factor, sensor calibration time freshness factor, and sensor most recent self-calibration deviation factor.

[0066] Specifically, the spatial distance confidence score is calculated using a Gaussian decay formula. First, the Euclidean distance between the target location and the sensor sampling location is calculated. The closer the distance, the higher the calculated spatial distance confidence score. A preset distance confidence coefficient is used in the calculation. The model prediction confidence score is calculated using a kernel density estimation method. The current scene condition feature vector is substituted into the feature space of the training data, the corresponding kernel density value is calculated, and normalization is performed. The higher the density in the feature space, the higher the model prediction confidence score. The sensor state confidence score is obtained by multiplying the sensor self-diagnostic health status factor and the sensor calibration time freshness factor. The self-diagnostic health status factor is a 0-1 range value from the sensor's real-time self-diagnostic output. The calibration time freshness factor is calculated based on the current time, the most recent calibration time, and the maximum effective calibration period, and is also a 0-1 range value. The total confidence score is obtained by directly multiplying the spatial distance confidence score, the model prediction confidence score, and the sensor state confidence score.

[0067] For example, the confidence level of the sensor above the stove at the center point (2.0, 1.5, 1.0) in the kitchen is calculated as follows: First, the Euclidean distance between this point and the sensor sampling location is calculated to be 1.1m. Using the Gaussian decay formula and substituting it into the preset distance confidence coefficient, the spatial distance confidence level is 0.82. The current scene condition feature vector is substituted into the training data feature space for kernel density estimation and normalization, resulting in a model prediction confidence level of 0.90. The sensor's real-time self-diagnostic health status factor is 0.95. Combining the current time and the most recent calibration time, the calibration time freshness factor is calculated to be 0.92. Multiplying the two together, the sensor status confidence level is 0.87. The three confidence components are multiplied sequentially to obtain the total confidence level of the sensor at this location, which is 0.64.

[0068] Next, using the confidence levels of each sensor at each location in the target space as weights, the concentration prediction fields of each sensor are weighted and fused to obtain the inverted concentration field. Weighted fusion refers to summing the concentration prediction values ​​of multiple sensors using the confidence levels of each sensor at the same location as weights, thus eliminating prediction biases from a single sensor. The inverted concentration field is a high-precision, high-reliability concentration distribution field across all points in the target space obtained after weighted fusion of multi-sensor confidence levels.

[0069] Specifically, the process iterates through all points in the target space. For each point, the predicted concentration value of each sensor is multiplied by the confidence level of that sensor at that point. The sum of these values ​​is then divided by the sum of the confidence levels of all sensors at that point to obtain the inverted concentration value for that point. The inverted concentration field for the entire space is then calculated point by point.

[0070] For example, at the kitchen center point (2.0, 1.5, 1.0), two sensors are involved in the fusion: Sensor 1: predicted value 1.08 mg / m³ 3 Confidence level 0.64; Sensor 2: Predicted value 1.12 mg / m³ 3 Confidence level 0.71; Inverted concentration value = (1.08 × 0.64 + 1.12 × 0.71) / (0.64 + 0.71) = 1.10 mg / m³ 3 After point-by-point calculation, a complete inversion concentration field is generated.

[0071] Finally, the fusion confidence score for each location in the target space is calculated. The fusion confidence score is the sum of the confidence scores of all sensors at that location. When the fusion confidence score is lower than a preset threshold, the concentration estimate at the corresponding location is marked as unreliable. The fusion confidence score refers to the cumulative sum of the confidence scores of all sensors at a single point in the target space, representing the overall reliability of the concentration estimate at that point. The preset threshold is a pre-defined critical value for fusion confidence; below this value, the concentration estimate is considered unreliable.

[0072] Specifically, the confidence scores of all sensors at each location are calculated to obtain a fused confidence score. This fused confidence score is then compared with a preset threshold. Locations with concentration estimates below the threshold are marked as unreliable. For example, if the preset threshold for fused confidence score is set to 0.3, and the fused confidence score at a point in a kitchen corner is only 0.25, which is below the threshold, the concentration estimate at that point is marked as unreliable and will not be included in concentration trend prediction and path planning.

[0073] In this embodiment of the invention, the problems of undersampling in sparse sensor space and insufficient reliability of single prediction field are solved by benchmark intensity calibration and three-dimensional confidence weighted fusion. The fusion process takes into account spatial distance, model adaptability and sensor health status, improving the accuracy and reliability of the inverted concentration field. At the same time, low confidence areas are marked to avoid invalid / distorted data from affecting subsequent processes, providing high-quality full-space concentration data support for concentration trend prediction and path optimization.

[0074] S400: Input the inverted concentration field and the collected wind speed and direction data into the concentration trend prediction model to generate the prediction results of the dynamic change trend of concentration at each spatial location in the target space in the future time period.

[0075] In this embodiment of the invention, the retrieved concentration field and the collected wind speed and direction data are input into a concentration trend prediction model to generate a prediction result of the dynamic change trend of concentration at various spatial locations in the target space over a future period. The retrieved concentration field only represents the spatial distribution of oil fume concentration at the current moment and cannot reflect the dynamic evolution of concentration caused by changes in wind speed and direction, temperature and humidity, and cooking conditions. Directly using it for path planning would cause the purification path to lag behind concentration changes. Furthermore, oil fume diffusion follows the physical laws of mass diffusion, and purely data-driven prediction is prone to results that violate physical logic. Therefore, it is necessary to integrate multi-dimensional environmental and state information and use a pre-trained spatiotemporal neural network model embedded with physical constraints to predict the concentration field in the future, providing a dynamic basis for path optimization.

[0076] Step S400 in the method provided in this embodiment of the invention includes: Based on the environmental sensing data, obtain environmental temperature and humidity data and cooking status inference results; Using a pre-trained concentration trend prediction model, with the inverted concentration field, the environmental temperature and humidity data, and the cooking state inference results as inputs, the dynamic change trend prediction results of the concentration are obtained. In the pre-training of the concentration trend prediction model, the training loss function includes at least the following: The data fitting error term measures the difference between the concentration field predicted by the model and the actual observed concentration field. The physical constraint term is used to penalize predictions that violate the fundamental laws of the matter diffusion equation, which include at least the conservation of concentration time variability and spatial diffusion flux.

[0077] First, based on the environmental sensing data, environmental temperature and humidity data and cooking state inference results are obtained. Environmental temperature and humidity data refer to the real-time spatial temperature and relative humidity data included in the environmental sensing data, which affect the diffusion rate and settling characteristics of cooking fumes. The cooking state inference results are obtained through comprehensive analysis of at least two of the following characteristics: the rate of change of discrete concentration measurements from each sensor, the surface temperature of the cookware, and the characteristics of the ambient sound spectrum. The cooking state includes at least: idle state, preheating state, stir-frying state, deep-frying state, steaming / boiling state, and cooling state after heat is turned off. It may also include time characteristic information, characterizing the temporal relationship between the current moment and the preset dining period to reflect the periodic pattern of concentration changes.

[0078] Specifically, real-time ambient temperature and humidity data of the target space are directly extracted from the environmental perception data; at least two features are selected from the discrete concentration measurement change rate of each sensor, the surface temperature of the cookware, and the ambient sound spectrum characteristics, combined with time feature information, and the current cooking state is comprehensively inferred through threshold judgment and feature fusion rules; standardized cooking state inference results are output, covering six states: idle, preheating, stir-frying, frying, steaming, and cooling after turning off the heat.

[0079] For example, taking a 4m×3m×2.8m commercial kitchen as an example, the real-time temperature and humidity are extracted from environmental sensing data as 26℃ and 60%RH; the concentration change rate is selected as 1.2mg / m³. 3 Based on the two characteristics of ·s and the surface temperature of the cookware being 280℃, combined with the time characteristic of being in the peak dinner period, it can be inferred that the current cooking state is stir-frying.

[0080] Secondly, using a pre-trained concentration trend prediction model, with the inverted concentration field, the environmental temperature and humidity data, and the cooking state inference results as inputs, the dynamic change trend prediction results of the concentration are obtained.

[0081] In the pre-training of the concentration trend prediction model, the training loss function includes at least the following: The data fitting error term measures the difference between the concentration field predicted by the model and the actual observed concentration field. The physical constraint term is used to penalize predictions that violate the fundamental laws of the matter diffusion equation, which include at least the conservation of concentration time variability and spatial diffusion flux.

[0082] First, the concentration trend prediction model adopts the standard architecture of a spatiotemporal convolutional recurrent neural network (ConvLSTM) encoder-decoder, consisting of three parts: a spatiotemporal encoding layer, a conditional fusion layer, and a spatiotemporal decoding layer. The spatiotemporal encoding layer extracts spatial distribution and temporal evolution features from the input concentration field sequence; the conditional fusion layer fuses the wind speed and direction data, the cooking state inference results, and the temporal feature information with the output of the spatiotemporal encoding layer; and the spatiotemporal decoding layer outputs the predicted results of the dynamic concentration change trend at each spatial location within a preset future time period.

[0083] Specifically, the spatiotemporal coding layer contains two ConvLSTM hidden layers, with 64 neurons in the first hidden layer and 128 neurons in the second hidden layer. The activation function is tanh, and the recurrent activation function is hard_sigmoid. This is used to extract high-precision spatial distribution features and temporal evolution features from the input inverted concentration field sequence. The conditional fusion layer is a fully connected feature splicing layer with 256 neurons and ReLU activation function. This is used to dimensionally align and weightedly fuse wind speed and direction data, environmental temperature and humidity data, cooking state inference results, temporal feature information, and the deep features output by the spatiotemporal coding layer. The spatiotemporal decoding layer contains two deconvolutional ConvLSTM hidden layers, with 128 neurons in the first hidden layer and 64 neurons in the second hidden layer. The activation function is tanh. This is used to decode and output the temporal prediction results of the concentration of all points in the target space within a preset future time period.

[0084] Secondly, the training dataset is constructed by fusing offline simulation data with online measured data. The data sources include three parts: first, time series data of oil fume concentration field generated by CFD simulation under multiple working conditions, covering all cooking states and environmental parameter combinations; second, time series data of multi-source sensors collected in real kitchen scenarios, including measured values ​​of inverted concentration field, temperature and humidity, and wind speed and direction; and third, manually labeled cooking state labels and time feature data. After all data are uniformly time-aligned and spatially normalized, a standardized dataset for model training is formed.

[0085] Furthermore, the concentration trend prediction model was trained using the Adam adaptive optimizer with an initial learning rate of 0.001 and a batch size of 16. The standardized dataset was divided into training, validation, and test sets in an 8:1:1 ratio. Iterative training was performed using mini-batch gradient descent. After each round of full training, the validation set data was used to evaluate the model's generalization ability and dynamically adjust the network parameters to avoid overfitting.

[0086] The concentration trend prediction model employs a two-component composite loss function. The first component is the data fitting error term, calculated using the mean square error method, which is used to accurately measure the numerical difference between the concentration field predicted by the model and the actual observed concentration field. The second component is the physical constraint term, which is used to identify and penalize prediction results that violate the laws of mass diffusion. The constraint rules include two core physical laws: the continuity of concentration time-varying rate and the conservation of spatial diffusion flux. During training, the two components are weighted and summed to serve as the basis for the total loss of model parameter updates.

[0087] Finally, the concentration trend prediction model training was implemented with dual convergence criteria, stopping training when either condition was met: first, the validation set loss function value showed no decrease for 50 consecutive iterations, indicating the model had reached optimal generalization ability; second, the total loss function value decreased to 10. -4 If the model's prediction accuracy is below the preset threshold, it is determined that the model's prediction accuracy meets the standard; after convergence, the model weights are saved as a pre-trained model for online inference.

[0088] For example, the inverted concentration field at the current moment, wind speed of 0.6 m / s (northbound), ambient temperature and humidity of 26℃ / 60%RH, inference results of stir-frying cooking, and time characteristics of the dinner peak period are organized into the standard input format of the model. After inputting into the trained concentration trend prediction model, through three layers of inference of spatiotemporal encoding, conditional fusion, and spatiotemporal decoding, the final output is the second-by-second change sequence of oil fume concentration in all spatial locations of the kitchen within the next 5 minutes, which is the dynamic change trend prediction result of the concentration.

[0089] In this embodiment of the invention, by integrating spatiotemporal concentration fields, meteorological parameters, cooking states, and temporal characteristics, and combining them with a spatiotemporal neural network model embedded with physical constraints, accurate forward prediction of future oil fume concentration fields is achieved, solving the problem that static concentration fields cannot adapt to dynamic cooking conditions. The prediction results strictly follow the physical laws of material diffusion, without logical distortion, providing a dynamic, forward-looking, and reliable basis for concentration changes for subsequent global path optimization, and improving the timeliness and targeting of the purification path.

[0090] S500: Based on the predicted results of the dynamic change trend of the concentration, with the core optimization objective of maximizing the total purification efficiency of the area covered by the mobile purification equipment path, global path optimization is performed to generate the optimal purification path.

[0091] In this embodiment of the invention, based on the predicted results of the dynamic change trend of the concentration, the core optimization objective is to maximize the total purification efficiency of the area covered by the mobile purification device's path. Global path optimization is then performed to generate the optimal purification path. The concentration of kitchen fumes changes dynamically in real time with cooking conditions and airflow. Traditional path planning only focuses on area coverage or the shortest path, failing to maximize purification efficiency. Simultaneously, the operation of the mobile purification device must meet hard constraints such as obstacle avoidance safety, smooth movement, and controllable energy consumption. Unconstrained, purely efficiency-oriented paths are not practically feasible. Therefore, path optimization needs to be modeled as a coverage optimization problem with maximizing the total purification efficiency as the core and multiple physical constraints as limitations. The initial path is guided by high-concentration anchor points, and global optimization is completed using swarm intelligence algorithms, ultimately generating the optimal purification path that balances purification efficiency, operational safety, smooth movement, and energy consumption control.

[0092] Step S500 in the method provided in this embodiment of the invention includes: With the goal of maximizing the total purification efficiency of the path coverage area, and with multiple constraints including path safety, motion smoothness, and movement energy consumption, the global path optimization is modeled as a constrained path coverage optimization problem, where: The path safety constraint is used to ensure that the distance between any location on the path and dynamic obstacles is not less than a preset safety distance; The motion smoothness constraint is used to ensure that the acceleration of the path does not exceed a preset acceleration limit; The mobile energy consumption constraint is used to ensure that the total energy consumption during path execution does not exceed the preset energy consumption budget; A predetermined number of high-concentration region centers are extracted from the inverted concentration field and used as path anchors. An initial path sequence connecting each anchor point is generated along the concentration gradient direction, and random perturbation is superimposed to obtain an initial path sequence set. Based on the initial path sequence set, a global path optimization algorithm combining swarm intelligence is used to solve the path coverage optimization problem, and the waypoint sequence output by the global path optimization is taken as the optimal purification path.

[0093] First, with the goal of maximizing the total purification efficiency of the path coverage area, and with multiple constraints including path safety, motion smoothness, and movement energy consumption, the global path optimization is modeled as a constrained path coverage optimization problem, where: The path safety constraint is used to ensure that the distance between any location on the path and dynamic obstacles is not less than a preset safety distance; The motion smoothness constraint is used to ensure that the acceleration of the path does not exceed a preset acceleration limit; The mobile energy consumption constraint is used to ensure that the total energy consumption during path execution does not exceed the preset energy consumption budget.

[0094] Among these, the total purification efficiency refers to the weighted cumulative total of oil fume concentration, purification efficiency, and coverage time within the covered area when the mobile purification equipment runs along the path, and is the core objective of path optimization. Path safety constraints ensure that the distance between any point on the path and dynamic obstacles is not less than a preset safe distance to avoid collisions. Motion smoothness constraints limit the path's running acceleration to the equipment's preset upper limit to ensure smooth operation without severe vibration. Mobile energy consumption constraints ensure that the total energy consumption throughout the path does not exceed the equipment's preset energy consumption budget, matching its endurance.

[0095] Specifically, the core optimization objective is set as follows: maximize the total purification efficiency of the area covered by the mobile purification device's path; define three hard constraints: path safety constraint: the minimum distance between all points on the path and dynamic obstacles is greater than or equal to the preset safety distance; motion smoothness constraint: the acceleration of the device along the path is less than or equal to the preset acceleration limit; and mobile energy consumption constraint: the total energy consumption throughout the path execution is less than or equal to the preset energy consumption budget; integrate the optimization objective and constraints to complete the modeling of the constrained path coverage optimization problem.

[0096] For example, taking a 4m×3m commercial kitchen mobile purification device as an example, the core objective is to maximize the total purification efficiency within a 5-minute operation cycle; the following constraints are set: safety distance 0.3m, maximum acceleration 1.0m / s². 2 With an energy consumption budget of 100Wh, a complete constrained path optimization model is constructed.

[0097] Secondly, a predetermined number of high-concentration region centers are extracted from the inverted concentration field as path anchor points. An initial path sequence connecting these anchor points is generated along the concentration gradient direction, and random perturbations are superimposed to obtain an initial path sequence set. High-concentration region centers refer to the predetermined number of local maxima points with the highest concentration values ​​in the inverted concentration field, representing priority coverage areas for fume purification. Path anchor points are key nodes guiding the path direction, forcing the path to prioritize coverage of the high-concentration core areas. The initial path sequence refers to the basic path formed by connecting the anchor points along the concentration gradient direction. The initial path sequence set refers to the set of multiple initial paths generated after superimposing random perturbations on the basic paths, providing an initial population for the optimization algorithm.

[0098] Specifically, a predetermined number of high-concentration region centers are extracted from the inverted concentration field as path anchor points; the anchor points are connected sequentially according to the concentration gradient from high to low to generate a basic initial path sequence; small random perturbations are superimposed on the waypoints of the basic path to generate multiple sets of differentiated initial paths, forming an initial path sequence set.

[0099] For example, six high-concentration region centers are extracted from the kitchen inversion concentration field as path anchor points; the anchor points are connected sequentially from high to low concentration to generate a basic path; and a random perturbation of ±0.1m is superimposed on the waypoints of the basic path to generate 20 different initial paths, forming an initial path sequence set.

[0100] Finally, based on the initial path sequence set, a global path optimization algorithm combining swarm intelligence is applied to the path coverage optimization problem. The waypoint sequence output by the global path optimization is then used as the optimal purification path. Swarm intelligence algorithms are global optimization algorithms that simulate the behavior of biological groups, such as particle swarm optimization, ant colony optimization, and genetic algorithms, and are suitable for complex, multi-constraint path optimization problems. Global path optimization refers to searching for the optimal path that satisfies the constraints and has the best core objective in all feasible path spaces. The optimal purification path is the waypoint sequence output after the algorithm's iterative convergence, taking into account purification efficiency, safety, smoothness, and energy consumption. The waypoint sequence is an ordered set of spatial coordinate points that constitute the optimal path and is directly used for equipment motion control.

[0101] Specifically, using the initial path sequence set as the initial set of individuals for path optimization, firstly, constraint feasibility verification is performed on each initial path, eliminating invalid paths that do not meet the constraints of path safety, motion smoothness, and movement energy consumption; then, using the total purification efficiency as the core indicator, the fitness value of each feasible path is calculated, with a higher total purification efficiency corresponding to a larger fitness value; during the iterative optimization process, the waypoints of each path are gradually adjusted and updated based on the individual optimal path and the group optimal path, and constraint feasibility verification and fitness calculation are performed again after the update; the above iterative update and verification process is repeated until the fitness value no longer changes significantly and the preset convergence condition is met, and finally, the waypoint sequence corresponding to the feasible path with the highest fitness value is determined as the optimal purification path.

[0102] For example, 20 initial paths are used as the initial set of individuals for path optimization. Each path is then checked to ensure that it meets the requirements of a minimum distance of 0.3m from dynamic obstacles and an acceleration not exceeding 1.0m / s². 2 The system employs several constraints, including limiting total energy consumption to no more than 100Wh, to filter out non-compliant paths. For the remaining feasible paths, it calculates the total purification efficiency over their coverage area and uses this efficiency as the criterion for evaluating path quality. The system iterates through the waypoints of each path generation, re-verifying constraints and recalculating purification efficiency after each iteration, completing 50 rounds of continuous optimization. Finally, it selects the path that satisfies all constraints and has the highest total purification efficiency, and uses the ordered waypoint sequence of this path as the final optimal purification path.

[0103] In this embodiment of the invention, dynamic concentration trends are combined with multi-constraint optimization to maximize purification efficiency as the core objective. High-concentration anchor points are used to precisely guide the path direction, and a global optimal search is achieved by combining swarm intelligence algorithms. The generated optimal purification path can prioritize the coverage of high-concentration areas and dynamically adapt to concentration changes. At the same time, it strictly meets the engineering requirements of safe obstacle avoidance, smooth movement, and controllable energy consumption. This solves the problems of low path planning efficiency, poor adaptability, and impracticality in dynamic oil fume scenarios, and maximizes the oil fume treatment efficiency of mobile purification equipment.

[0104] The method provided in this embodiment of the invention further includes: Using a preset first cycle, the waypoint sequence of the optimal purification path is read as a reference path; By combining real-time collected dynamic obstacle location information, local trajectory planning is performed to generate local speed control commands that meet path safety constraints; The local speed control command is received at a preset second cycle and converted into a motor control command to drive the mobile purification device to run along the local trajectory, wherein the second cycle is shorter than the first cycle; Real-time monitoring of changes in the inverted concentration field, and calculation of the concentration field abrupt change measure between the current inverted concentration field and the inverted concentration field on which the optimal purification path was based; When the concentration field mutation metric exceeds a preset first threshold, global path replanning is triggered to generate an updated optimal purification path. When the path segment between the current waypoint and the next waypoint in the optimal purification path is continuously blocked by dynamic obstacles, and the blocking time exceeds a preset second threshold, global path replanning is triggered.

[0105] First, using a preset first cycle, the waypoint sequence of the optimal purification path is read as a reference path. The preset first cycle refers to a long period for global path reading and updating, on the order of seconds (e.g., 10 seconds), used to synchronize the latest global optimal path and avoid frequent replanning. The waypoint sequence is an ordered set of spatial coordinates of the optimal purification path, serving as a reference for the movement of the mobile purification equipment. The reference path is the baseline path traveled by the equipment within the current cycle, used as the guiding basis for local trajectory planning.

[0106] For example, the first cycle is set to 10 seconds. The waypoint sequence of the optimal purification path is read every 10 seconds. The coordinates of the stove, the stir-frying area, the ventilation vent, etc. are combined into a reference path for local trajectory planning.

[0107] Secondly, local trajectory planning is performed by combining real-time acquired dynamic obstacle location information to generate local speed control commands that meet path safety constraints. Dynamic obstacle location information refers to the real-time acquisition of the three-dimensional coordinates and movement trajectories of obstacles such as people and moving kitchen utensils within the kitchen. Local trajectory planning involves making minor path corrections for local dynamic obstacles based on a reference path, generating a short-distance safe driving trajectory. Local speed control commands include control signals for movement speed and steering angle, used for local obstacle avoidance and trajectory tracking.

[0108] For example, if people are detected walking on the reference path in real time, the local correction trajectory is planned by shifting 0.3m to the right from the original reference path, and a local speed control command with a speed of 0.5m / s and a right turn angle of 15° is generated.

[0109] Next, the local speed control command is received at a preset second cycle and converted into a motor control command to drive the mobile purification device to run along the local trajectory. The second cycle is shorter than the first cycle. The preset second cycle refers to the short cycle of local command execution, in the millisecond range (e.g., 50ms), which is shorter than the first cycle to ensure real-time obstacle avoidance response. The motor control command refers to the underlying control signals that can directly drive the mobile purification device's walking motor and steering motor.

[0110] For example, the second cycle is set to 50ms. Every 50ms, a local speed control command is received and converted into a motor drive pulse signal to drive the mobile purification device to move smoothly along the local obstacle avoidance trajectory.

[0111] Furthermore, the changes in the retrieved concentration field are monitored in real time, and the concentration field abrupt change metric between the current retrieved concentration field and the retrieved concentration field used in the optimal purification path planning is calculated. The change in the retrieved concentration field refers to the spatial distribution difference between the current retrieved concentration field and the concentration field at the time of path planning. The concentration field abrupt change metric is a numerical indicator used to quantify the degree of abrupt change in the concentration field distribution, reflecting the degree of drastic change in the distribution of oil fumes.

[0112] For example, by comparing the current concentration field with the concentration field during path planning in real time, the concentration difference and distribution offset between the stove area and the dining area are calculated, and the weighted concentration field mutation metric is 0.8.

[0113] When the concentration field abrupt change exceeds a preset first threshold, global path replanning is triggered to generate an updated optimal purification path. The preset first threshold is a critical value for determining a drastic change in the concentration field; exceeding this value indicates that the original path no longer matches the current concentration distribution. Global path replanning refers to re-executing the S500 process to generate a new optimal purification path based on the latest concentration trend prediction results.

[0114] For example, the preset first threshold is 0.5, and the current concentration field mutation metric is 0.8. If the preset first threshold is exceeded, global path replanning is triggered to generate the optimal purification path that adapts to the latest concentration distribution.

[0115] When a path segment between the current waypoint and the next waypoint in the optimal cleanup path is continuously blocked by dynamic obstacles, and the blocking time exceeds a preset second threshold, global path replanning is triggered. Path segment blocking means that the path from the current waypoint to the next waypoint is completely occupied by dynamic obstacles, preventing the equipment from passing. The preset second threshold is a timeout threshold for determining continuous path blocking; exceeding this value indicates that local obstacle avoidance cannot solve the passage problem.

[0116] For example, the second threshold is preset to 3 seconds. If the current waypoint to the next waypoint is continuously blocked by kitchen utensils for 4 seconds, exceeding the threshold, global path replanning is triggered to generate the optimal purification path that bypasses the blocked area.

[0117] In this embodiment of the invention, a dual-cycle control of long-cycle global synchronization and short-cycle local execution is adopted to ensure the optimality of the global path and to achieve rapid local avoidance of dynamic obstacles. Through a dual planning triggering mechanism of concentration field mutation and path continuous blockage, it adapts in real time to drastic changes in oil fume concentration distribution and permanent path blockage scenarios, completely solving the problem that fixed global paths cannot cope with dynamic kitchen environments, and ensuring that the mobile purification equipment always performs purification operations safely, efficiently and continuously.

[0118] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides a method and system for optimizing fume purification paths based on multi-source data fusion. It utilizes a multi-source sensor array to collect environmental perception data in real time, laying a complete and accurate data foundation for fume concentration field modeling and path optimization. An anisotropic concentration decay field is constructed based on a parameter decision model, ensuring that the concentration distribution model closely matches the working conditions of the kitchen space and the physical laws of fume diffusion. A high-resolution inversion concentration field is obtained through confidence-weighted fusion of multi-sensor concentration prediction fields, effectively compensating for the spatial undersampling defects of sparse sensors and achieving accurate reconstruction of the entire space concentration field. Combining the inversion concentration field with multi-source environmental parameters, dynamic trend prediction of concentration is completed, providing a forward-looking prediction capability for the spatiotemporal evolution of fume concentration. Finally, with the core objective of maximizing total purification efficiency, a constrained global path optimization is performed. The generated optimal purification path can accurately target high-concentration areas and dynamically adapt to concentration changes, while simultaneously meeting safety, smoothness, and energy consumption constraints, comprehensively improving the real-time performance, accuracy, and adaptive governance efficiency of mobile fume purification.

[0119] Example 2, as Figure 2 As shown, this invention provides an oil fume purification path optimization system based on multi-source data fusion, the system comprising: The environmental perception data acquisition module 11 is used to acquire environmental perception data in the target space in real time through a multi-source sensor array; The concentration decay field construction module 12 is used to combine the environmental perception data, infer the decay basis function parameters corresponding to each sensor through the parameter decision model, and combine the prior concentration decay field with the decay basis function parameters to construct an anisotropic concentration decay field for each sensor. The inversion concentration field fusion module 13 is used to construct a concentration prediction field for each sensor based on the discrete concentration measurement values ​​of each sensor in the environmental perception data and the concentration decay field, and to perform weighted fusion of multiple concentration prediction fields to obtain the inversion concentration field of the target space. The concentration trend prediction module 14 is used to input the inverted concentration field and the collected wind speed and direction data into the concentration trend prediction model to generate the dynamic change trend prediction results of the concentration at each spatial location in the target space in the future period. The optimal path optimization module 15 is used to perform global path optimization based on the predicted results of the dynamic change trend of the concentration, with the core optimization objective of maximizing the total purification efficiency of the area covered by the path of the mobile purification device, and generate the optimal purification path.

[0120] In one embodiment, the environmental perception data acquisition module 11 is further configured to: Ambient air data of the target space is collected in real time through a multi-source sensor array. The ambient air data includes discrete oil fume concentration, wind speed, and wind direction, and the sampling position of each sensor in the multi-source sensor array is obtained simultaneously. Acquire spatial layout and relative equipment position data of the target space as environmental spatial data; After time alignment, the ambient air data and the ambient space data are merged and output as the environmental perception data.

[0121] In one embodiment, the concentration decay field construction module 12 is further configured to: Based on the environmental perception data, a scene condition feature vector for the target space is constructed; The scene condition feature vectors corresponding to each sensor are input into the trained parameter decision model for inference, and the decay basis function parameters corresponding to each sensor are output. Based on the attenuation basis function parameters corresponding to each sensor, and combined with the sampling position information of each sensor, the attenuation basis function of the prior concentration attenuation field is substituted into the attenuation basis function to construct an independent anisotropic concentration attenuation field for each sensor. The concentration attenuation field is used to output the concentration attenuation prediction value of any position in the target space relative to the sampling position of the corresponding sensor.

[0122] Specifically, by combining the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through a parameter decision model. Then, by combining the prior concentration attenuation field with the attenuation basis function parameters, an anisotropic concentration attenuation field is constructed for each sensor. Prior to this, the process includes: Construct a working condition parameter space, wherein the working conditions include at least: stove layout type, distance between stove and nearest wall, range hood air volume setting, number of stoves working at the same time, and cooking method type; Based on the aforementioned working condition parameter space, a simulation model covering multiple combinations of working conditions is established. The simulation model is used to solve the convection and diffusion equations of oil fume aerosol in the target space and obtain the simulated concentration field under each combination of working conditions. Define the mathematical form of the attenuation basis function based on the kernel function, wherein the attenuation basis function is a product of multiple spatial modulation factors, wherein the spatial modulation factors include at least: Source factor, used to describe the concentration distribution benchmark centered on the stove location and decreasing with spatial distance; Spatial attenuation factor, used to describe the anisotropic distribution characteristics of concentration attenuation with distance in each direction; The flow modulation factor is used to describe the enhanced modulation of the concentration distribution along the direction of the smoke machine by the smoke machine's suction effect. The wall constraint factor describes the attenuation modulation of the concentration distribution on the wall side due to the wall's blocking effect on the concentration diffusion path. For each combination of working conditions in the working condition parameter space, with the corresponding simulated concentration field as the target distribution, the optimal set of parameters that minimizes the error metric between the decay basis function and the simulated concentration field is obtained by parameter inversion optimization. The feature descriptions of each combination of working conditions are paired with the corresponding set of optimal parameters to form a calibration dataset.

[0123] This includes creating a calibration dataset, followed by: Construct a scene condition feature vector, which includes at least the following encoded features: stove layout type, distance between the stove and the nearest wall, range hood fan speed setting, number of stoves working simultaneously, and cooking method type. Based on the calibration dataset and the data format of the scenario condition feature vector, the feature descriptions of each working condition combination are combined into an input feature vector, and the corresponding optimal parameter set is defined as a label to construct the training sample set of the parameter decision model. Construct the parameter decision model based on a multi-output neural network, define a loss function based on physical constraints and parameter prediction error, and train the parameter decision model using the training sample set until the loss function converges to below a preset threshold to obtain the trained parameter decision model.

[0124] In one embodiment, the inversion concentration field fusion module 13 is further configured to: Using the discrete concentration measurements from each sensor as the baseline intensity, and substituting them into the concentration decay field of each sensor, the concentration prediction field is obtained. The confidence scores of the concentration prediction fields of each sensor at each location in the target space are calculated. The confidence scores are composed of the product of three components: spatial distance confidence score, model prediction confidence score, and sensor state confidence score. The concentration prediction fields of each sensor are weighted and fused using the confidence scores of each sensor at each location in the target space to obtain the inverted concentration field. Calculate the fusion confidence score at each location in the target space, where the fusion confidence score is the sum of the confidence scores of each sensor at that location. When the fusion confidence score is lower than a preset threshold, the concentration estimate at the corresponding location is marked as unreliable.

[0125] In one embodiment, the concentration trend prediction module 14 is further configured to: Based on the environmental sensing data, obtain environmental temperature and humidity data and cooking status inference results; Using a pre-trained concentration trend prediction model, with the inverted concentration field, the environmental temperature and humidity data, and the cooking state inference results as inputs, the dynamic change trend prediction results of the concentration are obtained. In the pre-training of the concentration trend prediction model, the training loss function includes at least the following: The data fitting error term measures the difference between the concentration field predicted by the model and the actual observed concentration field. The physical constraint term is used to penalize predictions that violate the fundamental laws of the matter diffusion equation, which include at least the conservation of concentration time variability and spatial diffusion flux.

[0126] In one embodiment, the optimal path optimization module 15 is further configured to: With the goal of maximizing the total purification efficiency of the path coverage area, and with multiple constraints including path safety, motion smoothness, and movement energy consumption, the global path optimization is modeled as a constrained path coverage optimization problem, where: The path safety constraint is used to ensure that the distance between any location on the path and dynamic obstacles is not less than a preset safety distance; The motion smoothness constraint is used to ensure that the acceleration of the path does not exceed a preset acceleration limit; The mobile energy consumption constraint is used to ensure that the total energy consumption during path execution does not exceed the preset energy consumption budget; A predetermined number of high-concentration region centers are extracted from the inverted concentration field and used as path anchors. An initial path sequence connecting each anchor point is generated along the concentration gradient direction, and random perturbation is superimposed to obtain an initial path sequence set. Based on the initial path sequence set, a global path optimization algorithm combining swarm intelligence is used to solve the path coverage optimization problem, and the waypoint sequence output by the global path optimization is taken as the optimal purification path.

[0127] The system provided in this embodiment of the invention is also used for: Using a preset first cycle, the waypoint sequence of the optimal purification path is read as a reference path; By combining real-time collected dynamic obstacle location information, local trajectory planning is performed to generate local speed control commands that meet path safety constraints; The local speed control command is received at a preset second cycle and converted into a motor control command to drive the mobile purification device to run along the local trajectory, wherein the second cycle is shorter than the first cycle; Real-time monitoring of changes in the inverted concentration field, and calculation of the concentration field abrupt change measure between the current inverted concentration field and the inverted concentration field on which the optimal purification path was based; When the concentration field mutation metric exceeds a preset first threshold, global path replanning is triggered to generate an updated optimal purification path. When the path segment between the current waypoint and the next waypoint in the optimal purification path is continuously blocked by dynamic obstacles, and the blocking time exceeds a preset second threshold, global path replanning is triggered.

[0128] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing the fume purification path based on multi-source data fusion, characterized in that, include: Real-time environmental perception data within the target space is collected using a multi-source sensor array; Based on the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through the parameter decision model. Combined with the prior concentration attenuation field and the attenuation basis function parameters, an anisotropic concentration attenuation field is constructed for each sensor. Based on the discrete concentration measurements of each sensor in the environmental perception data and the concentration decay field, a concentration prediction field is constructed for each sensor, and multiple concentration prediction fields are weighted and fused to obtain the inversion concentration field of the target space. The inverted concentration field and the collected wind speed and direction data are input into the concentration trend prediction model to generate the dynamic change trend prediction results of the concentration at each spatial location in the target space in the future time period. Based on the predicted results of the dynamic change trend of the concentration, with the core optimization objective of maximizing the total purification efficiency of the area covered by the mobile purification equipment path, global path optimization is performed to generate the optimal purification path.

2. The method for optimizing the fume purification path based on multi-source data fusion as described in claim 1, characterized in that, Through a multi-source sensor array, environmental perception data within the target space is collected in real time, including: Ambient air data of the target space is collected in real time through a multi-source sensor array. The ambient air data includes discrete oil fume concentration, wind speed, and wind direction, and the sampling position of each sensor in the multi-source sensor array is obtained simultaneously. Acquire spatial layout and relative equipment position data of the target space as environmental spatial data; After time alignment, the ambient air data and the ambient space data are merged and output as the environmental perception data.

3. The method for optimizing the fume purification path based on multi-source data fusion as described in claim 1, characterized in that, Based on the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through a parameter decision model. Then, combining the prior concentration attenuation field with the attenuation basis function parameters, an anisotropic concentration attenuation field is constructed for each sensor. Prior to this, the following steps are taken: Construct a working condition parameter space, wherein the working conditions include at least: stove layout type, distance between stove and nearest wall, range hood air volume setting, number of stoves working at the same time, and cooking method type; Based on the aforementioned working condition parameter space, a simulation model covering multiple combinations of working conditions is established. The simulation model is used to solve the convection and diffusion equations of oil fume aerosol in the target space and obtain the simulated concentration field under each combination of working conditions. Define the mathematical form of the attenuation basis function based on the kernel function, wherein the attenuation basis function is a product of multiple spatial modulation factors, wherein the spatial modulation factors include at least: Source factor, used to describe the concentration distribution benchmark centered on the stove location and decreasing with spatial distance; Spatial attenuation factor, used to describe the anisotropic distribution characteristics of concentration attenuation with distance in each direction; The flow modulation factor is used to describe the enhanced modulation of the concentration distribution along the direction of the smoke machine by the smoke machine's suction effect. The wall constraint factor describes the attenuation modulation of the concentration distribution on the wall side due to the wall's blocking effect on the concentration diffusion path. For each combination of working conditions in the working condition parameter space, with the corresponding simulated concentration field as the target distribution, the optimal set of parameters that minimizes the error metric between the decay basis function and the simulated concentration field is obtained by parameter inversion optimization. The feature descriptions of each combination of working conditions are paired with the corresponding set of optimal parameters to form a calibration dataset.

4. The method for optimizing the fume purification path based on multi-source data fusion as described in claim 3, characterized in that, After creating the calibration dataset, the following steps are also included: Construct a scene condition feature vector, which includes at least the following encoded features: stove layout type, distance between the stove and the nearest wall, range hood fan speed setting, number of stoves working simultaneously, and cooking method type. Based on the calibration dataset and the data format of the scenario condition feature vector, the feature descriptions of each working condition combination are combined into an input feature vector, and the corresponding optimal parameter set is defined as a label to construct the training sample set of the parameter decision model. Construct the parameter decision model based on a multi-output neural network, define a loss function based on physical constraints and parameter prediction error, and train the parameter decision model using the training sample set until the loss function converges to below a preset threshold to obtain the trained parameter decision model.

5. The method for optimizing the fume purification path based on multi-source data fusion as described in claim 1, characterized in that, Based on the environmental perception data, the attenuation basis function parameters corresponding to each sensor are inferred through a parameter decision model. Then, by combining the prior concentration attenuation field with the attenuation basis function parameters, an anisotropic concentration attenuation field is constructed for each sensor, including: Based on the environmental perception data, a scene condition feature vector for the target space is constructed; The scene condition feature vectors corresponding to each sensor are input into the trained parameter decision model for inference, and the decay basis function parameters corresponding to each sensor are output. Based on the attenuation basis function parameters corresponding to each sensor, and combined with the sampling position information of each sensor, the attenuation basis function of the prior concentration attenuation field is substituted into the attenuation basis function to construct an independent anisotropic concentration attenuation field for each sensor. The concentration attenuation field is used to output the concentration attenuation prediction value of any position in the target space relative to the sampling position of the corresponding sensor.

6. The method for optimizing the fume purification path based on multi-source data fusion as described in claim 1, characterized in that, Based on the discrete concentration measurements from each sensor in the environmental perception data and the concentration decay field, a concentration prediction field is constructed for each sensor, and multiple concentration prediction fields are weighted and fused to obtain the inverted concentration field of the target space, including: Using the discrete concentration measurements from each sensor as the baseline intensity, and substituting them into the concentration decay field of each sensor, the concentration prediction field is obtained. The confidence scores of the concentration prediction fields of each sensor at each location in the target space are calculated. The confidence scores are composed of the product of three components: spatial distance confidence score, model prediction confidence score, and sensor state confidence score. The concentration prediction fields of each sensor are weighted and fused using the confidence scores of each sensor at each location in the target space to obtain the inverted concentration field. Calculate the fusion confidence score at each location in the target space, where the fusion confidence score is the sum of the confidence scores of each sensor at that location. When the fusion confidence score is lower than a preset threshold, the concentration estimate at the corresponding location is marked as unreliable.

7. The method for optimizing oil fume purification paths based on multi-source data fusion as described in claim 1, characterized in that, The retrieved concentration field and the collected wind speed and direction data are input into the concentration trend prediction model to generate a prediction result of the dynamic change trend of concentration at various spatial locations of the target kitchen in the future time period, including: Based on the environmental sensing data, obtain environmental temperature and humidity data and cooking status inference results; Using a pre-trained concentration trend prediction model, with the inverted concentration field, the environmental temperature and humidity data, and the cooking state inference results as inputs, the dynamic change trend prediction results of the concentration are obtained. In the pre-training of the concentration trend prediction model, the training loss function includes at least the following: The data fitting error term measures the difference between the concentration field predicted by the model and the actual observed concentration field. The physical constraint term is used to penalize predictions that violate the fundamental laws of the matter diffusion equation, which include at least the conservation of concentration time variability and spatial diffusion flux.

8. The method for optimizing the fume purification path based on multi-source data fusion as described in claim 1, characterized in that, Based on the predicted results of the dynamic concentration change trend, with the core optimization objective of maximizing the total purification efficiency of the area covered by the mobile purification device's path, global path optimization is performed to generate the optimal purification path, including: With the goal of maximizing the total purification efficiency of the path coverage area, and with multiple constraints including path safety, motion smoothness, and movement energy consumption, the global path optimization is modeled as a constrained path coverage optimization problem, where: The path safety constraint is used to ensure that the distance between any location on the path and dynamic obstacles is not less than a preset safety distance; The motion smoothness constraint is used to ensure that the acceleration of the path does not exceed a preset acceleration limit; The mobile energy consumption constraint is used to ensure that the total energy consumption during path execution does not exceed the preset energy consumption budget; A predetermined number of high-concentration region centers are extracted from the inverted concentration field and used as path anchors. An initial path sequence connecting each anchor point is generated along the concentration gradient direction, and random perturbation is superimposed to obtain an initial path sequence set. Based on the initial path sequence set, a global path optimization algorithm combining swarm intelligence is used to solve the path coverage optimization problem, and the waypoint sequence output by the global path optimization is taken as the optimal purification path.

9. The method for optimizing the fume purification path based on multi-source data fusion as described in claim 1, characterized in that, Also includes: Using a preset first cycle, the waypoint sequence of the optimal purification path is read as a reference path; By combining real-time collected dynamic obstacle location information, local trajectory planning is performed to generate local speed control commands that meet path safety constraints; The local speed control command is received at a preset second cycle and converted into a motor control command to drive the mobile purification device to run along the local trajectory, wherein the second cycle is shorter than the first cycle; Real-time monitoring of changes in the inverted concentration field, and calculation of the concentration field abrupt change measure between the current inverted concentration field and the inverted concentration field on which the optimal purification path was based; When the concentration field mutation metric exceeds a preset first threshold, global path replanning is triggered to generate an updated optimal purification path. When the path segment between the current waypoint and the next waypoint in the optimal purification path is continuously blocked by dynamic obstacles, and the blocking time exceeds a preset second threshold, global path replanning is triggered.

10. A fume purification path optimization system based on multi-source data fusion, characterized in that, The method for optimizing the fume purification path based on multi-source data fusion as described in any one of claims 1-9 includes: The environmental perception data acquisition module is used to acquire environmental perception data in the target space in real time through a multi-source sensor array; The concentration decay field construction module is used to combine the environmental sensing data, infer the decay basis function parameters corresponding to each sensor through the parameter decision model, and combine the prior concentration decay field with the decay basis function parameters to construct an anisotropic concentration decay field for each sensor. The inversion concentration field fusion module is used to construct a concentration prediction field for each sensor based on the discrete concentration measurement values ​​of each sensor in the environmental perception data and the concentration decay field, and to perform weighted fusion of multiple concentration prediction fields to obtain the inversion concentration field of the target space. The concentration trend prediction module is used to input the inverted concentration field and the collected wind speed and direction data into the concentration trend prediction model to generate the dynamic change trend prediction results of the concentration at each spatial location in the target space in the future time period. The optimal path optimization module is used to perform global path optimization based on the predicted results of the dynamic change trend of the concentration, with the core optimization objective of maximizing the total purification efficiency of the area covered by the path of the mobile purification equipment, and generate the optimal purification path.

Images